instance_id
stringclasses
7 values
text
stringlengths
11.4k
828k
repo
stringclasses
3 values
base_commit
stringclasses
7 values
problem_statement
stringclasses
6 values
hints_text
stringclasses
5 values
created_at
stringclasses
7 values
patch
stringclasses
7 values
test_patch
stringclasses
7 values
version
stringclasses
1 value
FAIL_TO_PASS
stringclasses
1 value
PASS_TO_PASS
stringclasses
1 value
environment_setup_commit
stringclasses
1 value
celery__celery-2840
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost. </issue> <code> [start of README.rst] 1 ================================= 2 celery - Distributed Task Queue 3 ================================= 4 5 .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png 6 7 |build-status| |coverage-status| 8 9 :Version: 3.2.0a1 (Cipater) 10 :Web: http://celeryproject.org/ 11 :Download: http://pypi.python.org/pypi/celery/ 12 :Source: http://github.com/celery/celery/ 13 :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, 14 python, webhooks, queue, distributed 15 16 -- 17 18 What is a Task Queue? 19 ===================== 20 21 Task queues are used as a mechanism to distribute work across threads or 22 machines. 23 24 A task queue's input is a unit of work, called a task, dedicated worker 25 processes then constantly monitor the queue for new work to perform. 26 27 Celery communicates via messages, usually using a broker 28 to mediate between clients and workers. To initiate a task a client puts a 29 message on the queue, the broker then delivers the message to a worker. 30 31 A Celery system can consist of multiple workers and brokers, giving way 32 to high availability and horizontal scaling. 33 34 Celery is a library written in Python, but the protocol can be implemented in 35 any language. So far there's RCelery_ for the Ruby programming language, and a 36 `PHP client`, but language interoperability can also be achieved 37 by using webhooks. 38 39 .. _RCelery: https://github.com/leapfrogonline/rcelery 40 .. _`PHP client`: https://github.com/gjedeer/celery-php 41 .. _`using webhooks`: 42 http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html 43 44 What do I need? 45 =============== 46 47 Celery version 3.0 runs on, 48 49 - Python (2.6, 2.7, 3.3, 3.4) 50 - PyPy (1.8, 1.9) 51 - Jython (2.5, 2.7). 52 53 This is the last version to support Python 2.5, 54 and from Celery 3.1, Python 2.6 or later is required. 55 The last version to support Python 2.4 was Celery series 2.2. 56 57 *Celery* is usually used with a message broker to send and receive messages. 58 The RabbitMQ, Redis transports are feature complete, 59 but there's also experimental support for a myriad of other solutions, including 60 using SQLite for local development. 61 62 *Celery* can run on a single machine, on multiple machines, or even 63 across datacenters. 64 65 Get Started 66 =========== 67 68 If this is the first time you're trying to use Celery, or you are 69 new to Celery 3.0 coming from previous versions then you should read our 70 getting started tutorials: 71 72 - `First steps with Celery`_ 73 74 Tutorial teaching you the bare minimum needed to get started with Celery. 75 76 - `Next steps`_ 77 78 A more complete overview, showing more features. 79 80 .. _`First steps with Celery`: 81 http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html 82 83 .. _`Next steps`: 84 http://docs.celeryproject.org/en/latest/getting-started/next-steps.html 85 86 Celery is... 87 ========== 88 89 - **Simple** 90 91 Celery is easy to use and maintain, and does *not need configuration files*. 92 93 It has an active, friendly community you can talk to for support, 94 including a `mailing-list`_ and and an IRC channel. 95 96 Here's one of the simplest applications you can make:: 97 98 from celery import Celery 99 100 app = Celery('hello', broker='amqp://guest@localhost//') 101 102 @app.task 103 def hello(): 104 return 'hello world' 105 106 - **Highly Available** 107 108 Workers and clients will automatically retry in the event 109 of connection loss or failure, and some brokers support 110 HA in way of *Master/Master* or *Master/Slave* replication. 111 112 - **Fast** 113 114 A single Celery process can process millions of tasks a minute, 115 with sub-millisecond round-trip latency (using RabbitMQ, 116 py-librabbitmq, and optimized settings). 117 118 - **Flexible** 119 120 Almost every part of *Celery* can be extended or used on its own, 121 Custom pool implementations, serializers, compression schemes, logging, 122 schedulers, consumers, producers, autoscalers, broker transports and much more. 123 124 It supports... 125 ============ 126 127 - **Message Transports** 128 129 - RabbitMQ_, Redis_, 130 - MongoDB_ (experimental), Amazon SQS (experimental), 131 - CouchDB_ (experimental), SQLAlchemy_ (experimental), 132 - Django ORM (experimental), `IronMQ`_ 133 - and more... 134 135 - **Concurrency** 136 137 - Prefork, Eventlet_, gevent_, threads/single threaded 138 139 - **Result Stores** 140 141 - AMQP, Redis 142 - memcached, MongoDB 143 - SQLAlchemy, Django ORM 144 - Apache Cassandra, IronCache 145 146 - **Serialization** 147 148 - *pickle*, *json*, *yaml*, *msgpack*. 149 - *zlib*, *bzip2* compression. 150 - Cryptographic message signing. 151 152 .. _`Eventlet`: http://eventlet.net/ 153 .. _`gevent`: http://gevent.org/ 154 155 .. _RabbitMQ: http://rabbitmq.com 156 .. _Redis: http://redis.io 157 .. _MongoDB: http://mongodb.org 158 .. _Beanstalk: http://kr.github.com/beanstalkd 159 .. _CouchDB: http://couchdb.apache.org 160 .. _SQLAlchemy: http://sqlalchemy.org 161 .. _`IronMQ`: http://iron.io 162 163 Framework Integration 164 ===================== 165 166 Celery is easy to integrate with web frameworks, some of which even have 167 integration packages: 168 169 +--------------------+----------------------------------------------------+ 170 | `Django`_ | not needed | 171 +--------------------+----------------------------------------------------+ 172 | `Pyramid`_ | `pyramid_celery`_ | 173 +--------------------+----------------------------------------------------+ 174 | `Pylons`_ | `celery-pylons`_ | 175 +--------------------+----------------------------------------------------+ 176 | `Flask`_ | not needed | 177 +--------------------+----------------------------------------------------+ 178 | `web2py`_ | `web2py-celery`_ | 179 +--------------------+----------------------------------------------------+ 180 | `Tornado`_ | `tornado-celery`_ | `another tornado-celery`_ | 181 +--------------------+----------------------------------------------------+ 182 183 The integration packages are not strictly necessary, but they can make 184 development easier, and sometimes they add important hooks like closing 185 database connections at ``fork``. 186 187 .. _`Django`: http://djangoproject.com/ 188 .. _`Pylons`: http://www.pylonsproject.org/ 189 .. _`Flask`: http://flask.pocoo.org/ 190 .. _`web2py`: http://web2py.com/ 191 .. _`Bottle`: http://bottlepy.org/ 192 .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html 193 .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ 194 .. _`django-celery`: http://pypi.python.org/pypi/django-celery 195 .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons 196 .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ 197 .. _`Tornado`: http://www.tornadoweb.org/ 198 .. _`tornado-celery`: http://github.com/mher/tornado-celery/ 199 .. _`another tornado-celery`: https://github.com/mayflaver/tornado-celery 200 201 .. _celery-documentation: 202 203 Documentation 204 ============= 205 206 The `latest documentation`_ with user guides, tutorials and API reference 207 is hosted at Read The Docs. 208 209 .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ 210 211 .. _celery-installation: 212 213 Installation 214 ============ 215 216 You can install Celery either via the Python Package Index (PyPI) 217 or from source. 218 219 To install using `pip`,:: 220 221 $ pip install -U Celery 222 223 To install using `easy_install`,:: 224 225 $ easy_install -U Celery 226 227 .. _bundles: 228 229 Bundles 230 ------- 231 232 Celery also defines a group of bundles that can be used 233 to install Celery and the dependencies for a given feature. 234 235 You can specify these in your requirements or on the ``pip`` comand-line 236 by using brackets. Multiple bundles can be specified by separating them by 237 commas. 238 :: 239 240 $ pip install "celery[librabbitmq]" 241 242 $ pip install "celery[librabbitmq,redis,auth,msgpack]" 243 244 The following bundles are available: 245 246 Serializers 247 ~~~~~~~~~~~ 248 249 :celery[auth]: 250 for using the auth serializer. 251 252 :celery[msgpack]: 253 for using the msgpack serializer. 254 255 :celery[yaml]: 256 for using the yaml serializer. 257 258 Concurrency 259 ~~~~~~~~~~~ 260 261 :celery[eventlet]: 262 for using the eventlet pool. 263 264 :celery[gevent]: 265 for using the gevent pool. 266 267 :celery[threads]: 268 for using the thread pool. 269 270 Transports and Backends 271 ~~~~~~~~~~~~~~~~~~~~~~~ 272 273 :celery[librabbitmq]: 274 for using the librabbitmq C library. 275 276 :celery[redis]: 277 for using Redis as a message transport or as a result backend. 278 279 :celery[mongodb]: 280 for using MongoDB as a message transport (*experimental*), 281 or as a result backend (*supported*). 282 283 :celery[sqs]: 284 for using Amazon SQS as a message transport (*experimental*). 285 286 :celery[memcache]: 287 for using memcached as a result backend. 288 289 :celery[cassandra]: 290 for using Apache Cassandra as a result backend. 291 292 :celery[couchdb]: 293 for using CouchDB as a message transport (*experimental*). 294 295 :celery[couchbase]: 296 for using CouchBase as a result backend. 297 298 :celery[beanstalk]: 299 for using Beanstalk as a message transport (*experimental*). 300 301 :celery[zookeeper]: 302 for using Zookeeper as a message transport. 303 304 :celery[zeromq]: 305 for using ZeroMQ as a message transport (*experimental*). 306 307 :celery[sqlalchemy]: 308 for using SQLAlchemy as a message transport (*experimental*), 309 or as a result backend (*supported*). 310 311 :celery[pyro]: 312 for using the Pyro4 message transport (*experimental*). 313 314 :celery[slmq]: 315 for using the SoftLayer Message Queue transport (*experimental*). 316 317 .. _celery-installing-from-source: 318 319 Downloading and installing from source 320 -------------------------------------- 321 322 Download the latest version of Celery from 323 http://pypi.python.org/pypi/celery/ 324 325 You can install it by doing the following,:: 326 327 $ tar xvfz celery-0.0.0.tar.gz 328 $ cd celery-0.0.0 329 $ python setup.py build 330 # python setup.py install 331 332 The last command must be executed as a privileged user if 333 you are not currently using a virtualenv. 334 335 .. _celery-installing-from-git: 336 337 Using the development version 338 ----------------------------- 339 340 With pip 341 ~~~~~~~~ 342 343 The Celery development version also requires the development 344 versions of ``kombu``, ``amqp`` and ``billiard``. 345 346 You can install the latest snapshot of these using the following 347 pip commands:: 348 349 $ pip install https://github.com/celery/celery/zipball/master#egg=celery 350 $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard 351 $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp 352 $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu 353 354 With git 355 ~~~~~~~~ 356 357 Please the Contributing section. 358 359 .. _getting-help: 360 361 Getting Help 362 ============ 363 364 .. _mailing-list: 365 366 Mailing list 367 ------------ 368 369 For discussions about the usage, development, and future of celery, 370 please join the `celery-users`_ mailing list. 371 372 .. _`celery-users`: http://groups.google.com/group/celery-users/ 373 374 .. _irc-channel: 375 376 IRC 377 --- 378 379 Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ 380 network. 381 382 .. _`Freenode`: http://freenode.net 383 384 .. _bug-tracker: 385 386 Bug tracker 387 =========== 388 389 If you have any suggestions, bug reports or annoyances please report them 390 to our issue tracker at http://github.com/celery/celery/issues/ 391 392 .. _wiki: 393 394 Wiki 395 ==== 396 397 http://wiki.github.com/celery/celery/ 398 399 400 .. _maintainers: 401 402 Maintainers 403 =========== 404 405 - `@ask`_ (primary maintainer) 406 - `@thedrow`_ 407 - `@chrisgogreen`_ 408 - `@PMickael`_ 409 - `@malinoff`_ 410 - And you? We really need more: https://github.com/celery/celery/issues/2534 411 412 .. _`@ask`: http://github.com/ask 413 .. _`@thedrow`: http://github.com/thedrow 414 .. _`@chrisgogreen`: http://github.com/chrisgogreen 415 .. _`@PMickael`: http://github.com/PMickael 416 .. _`@malinoff`: http://github.com/malinoff 417 418 419 .. _contributing-short: 420 421 Contributing 422 ============ 423 424 Development of `celery` happens at Github: http://github.com/celery/celery 425 426 You are highly encouraged to participate in the development 427 of `celery`. If you don't like Github (for some reason) you're welcome 428 to send regular patches. 429 430 Be sure to also read the `Contributing to Celery`_ section in the 431 documentation. 432 433 .. _`Contributing to Celery`: 434 http://docs.celeryproject.org/en/master/contributing.html 435 436 .. _license: 437 438 License 439 ======= 440 441 This software is licensed under the `New BSD License`. See the ``LICENSE`` 442 file in the top distribution directory for the full license text. 443 444 .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround 445 446 447 .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png 448 :alt: Bitdeli badge 449 :target: https://bitdeli.com/free 450 451 .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master 452 :target: https://travis-ci.org/celery/celery 453 .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg 454 :target: https://coveralls.io/r/celery/celery 455 [end of README.rst] [start of celery/app/defaults.py] ... 118 'EAGER_PROPAGATES_EXCEPTIONS': Option(False, type='bool'), 119 'ENABLE_UTC': Option(True, type='bool'), 120 'ENABLE_REMOTE_CONTROL': Option(True, type='bool'), 121 'EVENT_SERIALIZER': Option('json'), 122 'EVENT_QUEUE_EXPIRES': Option(60.0, type='float'), 123 'EVENT_QUEUE_TTL': Option(5.0, type='float'), 124 'IMPORTS': Option((), type='tuple'), 125 'INCLUDE': Option((), type='tuple'), 126 'IGNORE_RESULT': Option(False, type='bool'), 127 'MAX_CACHED_RESULTS': Option(100, type='int'), 128 'MESSAGE_COMPRESSION': Option(type='string'), 129 'MONGODB_BACKEND_SETTINGS': Option(type='dict'), 130 'REDIS_HOST': Option(type='string', **_REDIS_OLD), 131 'REDIS_PORT': Option(type='int', **_REDIS_OLD), 132 'REDIS_DB': Option(type='int', **_REDIS_OLD), 133 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 134 'REDIS_MAX_CONNECTIONS': Option(type='int'), 135 'RESULT_BACKEND': Option(type='string'), 136 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 137 'RESULT_DB_TABLENAMES': Option(type='dict'), 138 'RESULT_DBURI': Option(), ... [end of celery/app/defaults.py] [start of celery/app/task.py] ... 206 #: 207 #: The application default can be overridden using the 208 #: :setting:`CELERY_TRACK_STARTED` setting. 209 track_started = None 210 211 #: When enabled messages for this task will be acknowledged **after** 212 #: the task has been executed, and not *just before* which is the 213 #: default behavior. 214 #: 215 #: Please note that this means the task may be executed twice if the 216 #: worker crashes mid execution (which may be acceptable for some 217 #: applications). 218 #: 219 #: The application default can be overridden with the 220 #: :setting:`CELERY_ACKS_LATE` setting. 221 acks_late = None 222 223 #: Tuple of expected exceptions. 224 #: 225 #: These are errors that are expected in normal operation 226 #: and that should not be regarded as a real error by the worker. ... ... 234 #: Task request stack, the current request will be the topmost. 235 request_stack = None 236 237 #: Some may expect a request to exist even if the task has not been 238 #: called. This should probably be deprecated. 239 _default_request = None 240 241 _exec_options = None 242 243 __bound__ = False 244 245 from_config = ( 246 ('send_error_emails', 'CELERY_SEND_TASK_ERROR_EMAILS'), 247 ('serializer', 'CELERY_TASK_SERIALIZER'), 248 ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), 249 ('track_started', 'CELERY_TRACK_STARTED'), 250 ('acks_late', 'CELERY_ACKS_LATE'), 251 ('ignore_result', 'CELERY_IGNORE_RESULT'), 252 ('store_errors_even_if_ignored', 253 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), 254 ) ... [end of celery/app/task.py] [start of celery/worker/request.py] ... 312 if self.task.acks_late: 313 self.acknowledge() 314 315 self.send_event('task-succeeded', result=retval, runtime=runtime) 316 317 def on_retry(self, exc_info): 318 """Handler called if the task should be retried.""" 319 if self.task.acks_late: 320 self.acknowledge() 321 322 self.send_event('task-retried', 323 exception=safe_repr(exc_info.exception.exc), 324 traceback=safe_str(exc_info.traceback)) 325 326 def on_failure(self, exc_info, send_failed_event=True, return_ok=False): 327 """Handler called if the task raised an exception.""" 328 task_ready(self) 329 330 if isinstance(exc_info.exception, MemoryError): 331 raise MemoryError('Process got: %s' % (exc_info.exception,)) 332 elif isinstance(exc_info.exception, Reject): 333 return self.reject(requeue=exc_info.exception.requeue) ... ... 338 339 if isinstance(exc, Retry): 340 return self.on_retry(exc_info) 341 342 # These are special cases where the process would not have had 343 # time to write the result. 344 if self.store_errors: 345 if isinstance(exc, Terminated): 346 self._announce_revoked( 347 'terminated', True, string(exc), False) 348 send_failed_event = False # already sent revoked event 349 elif isinstance(exc, WorkerLostError) or not return_ok: 350 self.task.backend.mark_as_failure( 351 self.id, exc, request=self, 352 ) 353 # (acks_late) acknowledge after result stored. 354 if self.task.acks_late: 355 self.acknowledge() 356 357 if send_failed_event: 358 self.send_event( 359 'task-failed', ... [end of celery/worker/request.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
celery/celery
045b52f1450d6d5cc500e0057a4b498250dc5692
Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost.
This is deliberate as if a task is killed it may mean that the next invocation will also cause the same to happen. If the task is redelivered it may cause a loop where the same conditions occur again and again. Also, sadly you cannot distinguish processes killed by OOM from processes killed by other means, and if an administrator kills -9 a task going amok, you usually don't want that task to be called again. There could be a configuration option for not acking terminated tasks, but I'm not sure how useful that would be. A better solution could be to use `basic_reject(requeue=False)` instead of `basic_ack`, that way you can configure a dead letter queue so that the killed tasks will be sent to a queue for manual inspection. I must say, regardless of the status of this feature request, the documentation is misleading. Specifically, [this FAQ makes it seem that process failures would NOT acknowledge messages](http://celery.readthedocs.org/en/latest/faq.html#faq-acks-late-vs-retry). And [this FAQ boldface states](http://celery.readthedocs.org/en/latest/faq.html#id54) that in the event of a kill signal (9), that acks_late will allow the task to re-run (which again, is patently wrong based on this poorly documented behavior). Nowhere in the docs have I found that if the process _dies_, the message will be acknowledged, regardless of acks_late or not. (for instance, I have a set of 10k+ tasks, and some 1% of tasks wind up acknowledged but incomplete when a WorkerLostError is thrown in connection with the worker, although there are no other errors of any kind in any of my logs related to that task). TL;DR at the least, appropriately document the current state when describing the functionality and limitations of acks_late. A work-around would be helpful -- I'm not sure I understand the solution of using `basic_reject`, although I'll keep looking into it. The docs are referring to killing the worker process with KILL, not the child processes. The term worker will always refer to the worker instance, not the pool processes. The section within about acks_late is probably not very helpful and should be removed
2015-10-06T05:34:34Z
<patch> <patch> diff --git a/celery/app/defaults.py b/celery/app/defaults.py --- a/celery/app/defaults.py +++ b/celery/app/defaults.py @@ -132,6 +132,7 @@ def __repr__(self): 'REDIS_DB': Option(type='int', **_REDIS_OLD), 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 'REDIS_MAX_CONNECTIONS': Option(type='int'), + 'REJECT_ON_WORKER_LOST': Option(type='bool'), 'RESULT_BACKEND': Option(type='string'), 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 'RESULT_DB_TABLENAMES': Option(type='dict'), diff --git a/celery/app/task.py b/celery/app/task.py --- a/celery/app/task.py +++ b/celery/app/task.py @@ -220,6 +220,12 @@ class Task(object): #: :setting:`CELERY_ACKS_LATE` setting. acks_late = None + #: When CELERY_ACKS_LATE is set to True, the default behavior to + #: handle worker crash is to acknowledge the message. Setting + #: this to true allows the message to be rejected and requeued so + #: it will be executed again by another worker. + reject_on_worker_lost = None + #: Tuple of expected exceptions. #: #: These are errors that are expected in normal operation @@ -248,6 +254,7 @@ class Task(object): ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), ('track_started', 'CELERY_TRACK_STARTED'), ('acks_late', 'CELERY_ACKS_LATE'), + ('reject_on_worker_lost', 'CELERY_REJECT_ON_WORKER_LOST'), ('ignore_result', 'CELERY_IGNORE_RESULT'), ('store_errors_even_if_ignored', 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), diff --git a/celery/worker/request.py b/celery/worker/request.py --- a/celery/worker/request.py +++ b/celery/worker/request.py @@ -326,7 +326,6 @@ def on_retry(self, exc_info): def on_failure(self, exc_info, send_failed_event=True, return_ok=False): """Handler called if the task raised an exception.""" task_ready(self) - if isinstance(exc_info.exception, MemoryError): raise MemoryError('Process got: %s' % (exc_info.exception,)) elif isinstance(exc_info.exception, Reject): @@ -352,7 +351,13 @@ def on_failure(self, exc_info, send_failed_event=True, return_ok=False): ) # (acks_late) acknowledge after result stored. if self.task.acks_late: - self.acknowledge() + reject_and_requeue = (self.task.reject_on_worker_lost and + isinstance(exc, WorkerLostError) and + self.delivery_info.get('redelivered', False) is False) + if reject_and_requeue: + self.reject(requeue=True) + else: + self.acknowledge() if send_failed_event: self.send_event( </patch> </s> </patch>
diff --git a/celery/tests/worker/test_request.py b/celery/tests/worker/test_request.py --- a/celery/tests/worker/test_request.py +++ b/celery/tests/worker/test_request.py @@ -325,6 +325,20 @@ def test_on_failure_Reject_rejects_with_requeue(self): req_logger, req.connection_errors, True, ) + def test_on_failure_WrokerLostError_rejects_with_requeue(self): + einfo = None + try: + raise WorkerLostError() + except: + einfo = ExceptionInfo(internal=True) + req = self.get_request(self.add.s(2, 2)) + req.task.acks_late = True + req.task.reject_on_worker_lost = True + req.delivery_info['redelivered'] = False + req.on_failure(einfo) + req.on_reject.assert_called_with(req_logger, + req.connection_errors, True) + def test_tzlocal_is_cached(self): req = self.get_request(self.add.s(2, 2)) req._tzlocal = 'foo'
1.0
NVIDIA__NeMo-473
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ``` </issue> <code> [start of README.rst] 1 .. image:: http://www.repostatus.org/badges/latest/active.svg 2 :target: http://www.repostatus.org/#active 3 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 4 5 .. image:: https://img.shields.io/badge/documentation-github.io-blue.svg 6 :target: https://nvidia.github.io/NeMo/ 7 :alt: NeMo documentation on GitHub pages 8 9 .. image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 10 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 11 :alt: NeMo core license and license for collections in this repo 12 13 .. image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 14 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python 15 :alt: Language grade: Python 16 17 .. image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 18 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ 19 :alt: Total alerts 20 21 .. image:: https://img.shields.io/badge/code%20style-black-000000.svg 22 :target: https://github.com/psf/black 23 :alt: Code style: black 24 25 26 27 NVIDIA Neural Modules: NeMo 28 =========================== 29 30 NeMo is a toolkit for defining and building `Conversational AI <https://developer.nvidia.com/conversational-ai#started>`_ applications. 31 32 Goal of the NeMo toolkit is to make it possible for researchers to easily compose complex neural network architectures for conversational AI using reusable components. Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 33 34 **Neural Modules** are conceptual blocks of neural networks that take *typed* inputs and produce *typed* outputs. Such modules typically represent data layers, encoders, decoders, language models, loss functions, or methods of combining activations. 35 36 The toolkit comes with extendable collections of pre-built modules for automatic speech recognition (ASR), natural language processing (NLP) and text synthesis (TTS). 37 38 **Introduction** 39 40 * Watch `this video <https://nvidia.github.io/NeMo/>`_ for a quick walk-through. 41 42 * Documentation (latest released version): https://nvidia.github.io/NeMo/ 43 44 * Read NVIDIA `Developer Blog for example applications <https://devblogs.nvidia.com/how-to-build-domain-specific-automatic-speech-recognition-models-on-gpus/>`_ 45 46 * Read NVIDIA `Developer Blog for Quartznet ASR model <https://devblogs.nvidia.com/develop-smaller-speech-recognition-models-with-nvidias-nemo-framework/>`_ 47 48 * Recommended version to install is **0.9.0** via ``pip install nemo-toolkit`` 49 50 * Recommended NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ 51 52 * Pretrained models are available on NVIDIA `NGC Model repository <https://ngc.nvidia.com/catalog/models?orderBy=modifiedDESC&query=nemo&quickFilter=models&filters=>`_ 53 54 55 Getting started 56 ~~~~~~~~~~~~~~~ 57 58 THE LATEST STABLE VERSION OF NeMo is **0.9.0** (Available via PIP). 59 60 **Requirements** 61 62 1) Python 3.6 or 3.7 63 2) PyTorch 1.4.* with GPU support 64 3) (optional, for best performance) NVIDIA APEX. Install from here: https://github.com/NVIDIA/apex 65 66 **NeMo Docker Container** 67 NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ is now available. 68 69 * Pull the docker: ``docker pull nvcr.io/nvidia/nemo:v0.9`` 70 * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/nemo:v0.9`` 71 72 If you are using the NVIDIA `NGC PyTorch container <https://ngc.nvidia.com/catalog/containers/nvidia:pytorch>`_ follow these instructions 73 74 * Pull the docker: ``docker pull nvcr.io/nvidia/pytorch:20.01-py3`` 75 * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/pytorch:20.01-py3`` 76 * ``apt-get update && apt-get install -y libsndfile1`` 77 * ``pip install nemo_toolkit`` NeMo core 78 * ``pip install nemo_asr`` NeMo ASR (Speech Recognition) collection 79 * ``pip install nemo_nlp`` NeMo NLP (Natural Language Processing) collection 80 * ``pip install nemo_tts`` NeMo TTS (Speech Synthesis) collection 81 82 See `examples/start_here` to get started with the simplest example. The folder `examples` contains several examples to get you started with various tasks in NLP and ASR. 83 84 **Tutorials** 85 86 * `Speech recognition <https://nvidia.github.io/NeMo/asr/intro.html>`_ 87 * `Natural language processing <https://nvidia.github.io/NeMo/nlp/intro.html>`_ 88 * `Speech Synthesis <https://nvidia.github.io/NeMo/tts/intro.html>`_ 89 90 91 DEVELOPMENT 92 ~~~~~~~~~~~ 93 If you'd like to use master branch and/or develop NeMo you can run "reinstall.sh" script. 94 95 `Documentation (master branch) <http://nemo-master-docs.s3-website.us-east-2.amazonaws.com/>`_. 96 97 **Installing From Github** 98 99 If you prefer to use NeMo's latest development version (from GitHub) follow the steps below: 100 101 1) Clone the repository ``git clone https://github.com/NVIDIA/NeMo.git`` 102 2) Go to NeMo folder and re-install the toolkit with collections: 103 104 .. code-block:: bash 105 106 ./reinstall.sh 107 108 **Style tests** 109 110 .. code-block:: bash 111 112 python setup.py style # Checks overall project code style and output issues with diff. 113 python setup.py style --fix # Tries to fix error in-place. 114 python setup.py style --scope=tests # Operates within certain scope (dir of file). 115 116 **Unittests** 117 118 This command runs unittests: 119 120 .. code-block:: bash 121 122 ./reinstall.sh 123 python pytest tests 124 125 126 Citation 127 ~~~~~~~~ 128 129 If you are using NeMo please cite the following publication 130 131 .. code-block:: tex 132 133 @misc{nemo2019, 134 title={NeMo: a toolkit for building AI applications using Neural Modules}, 135 author={Oleksii Kuchaiev and Jason Li and Huyen Nguyen and Oleksii Hrinchuk and Ryan Leary and Boris Ginsburg and Samuel Kriman and Stanislav Beliaev and Vitaly Lavrukhin and Jack Cook and Patrice Castonguay and Mariya Popova and Jocelyn Huang and Jonathan M. Cohen}, 136 year={2019}, 137 eprint={1909.09577}, 138 archivePrefix={arXiv}, 139 primaryClass={cs.LG} 140 } 141 142 [end of README.rst] [start of nemo/core/neural_modules.py] ... 379 def input_ports(self) -> Optional[Dict[str, NeuralType]]: 380 """Returns definitions of module input ports 381 382 Returns: 383 A (dict) of module's input ports names to NeuralTypes mapping 384 """ 385 386 @property 387 @abstractmethod 388 def output_ports(self) -> Optional[Dict[str, NeuralType]]: 389 """Returns definitions of module output ports 390 391 Returns: 392 A (dict) of module's output ports names to NeuralTypes mapping 393 """ 394 395 @property 396 def disabled_deployment_input_ports(self) -> Optional[Set[str]]: 397 """Returns names of input ports that will not be included in an export 398 399 Returns: 400 A (set) of module's input port names that are not exportable ... ... 388 def output_ports(self) -> Optional[Dict[str, NeuralType]]: 389 """Returns definitions of module output ports 390 391 Returns: 392 A (dict) of module's output ports names to NeuralTypes mapping 393 """ 394 395 @property 396 def disabled_deployment_input_ports(self) -> Optional[Set[str]]: 397 """Returns names of input ports that will not be included in an export 398 399 Returns: 400 A (set) of module's input port names that are not exportable 401 """ 402 return set([]) 403 404 @property 405 def disabled_deployment_output_ports(self) -> Optional[Set[str]]: 406 """Returns names of output ports that will not be included in an export 407 408 Returns: 409 A (set) of module's output port names that are not exportable ... ... 396 def disabled_deployment_input_ports(self) -> Optional[Set[str]]: 397 """Returns names of input ports that will not be included in an export 398 399 Returns: 400 A (set) of module's input port names that are not exportable 401 """ 402 return set([]) 403 404 @property 405 def disabled_deployment_output_ports(self) -> Optional[Set[str]]: 406 """Returns names of output ports that will not be included in an export 407 408 Returns: 409 A (set) of module's output port names that are not exportable 410 """ 411 return set([]) 412 413 def prepare_for_deployment(self) -> None: 414 """Patch the module if required to prepare for deployment 415 416 """ 417 return ... [end of nemo/core/neural_modules.py] [start of nemo/backends/pytorch/actions.py] ... 923 @staticmethod 924 def __module_export(module, output, d_format: DeploymentFormat, input_example=None, output_example=None): 925 # Check if output already exists 926 destination = Path(output) 927 if destination.exists(): 928 raise FileExistsError(f"Destination {output} already exists. " f"Aborting export.") 929 930 input_names = list(module.input_ports.keys()) 931 output_names = list(module.output_ports.keys()) 932 dynamic_axes = defaultdict(list) 933 934 def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defaultdict): 935 if ntype.axes: 936 for ind, axis in enumerate(ntype.axes): 937 if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: 938 dynamic_axes[port_name].append(ind) 939 940 # This is a hack for Jasper to Jarvis export -- need re-design for this 941 inputs_to_drop = set() 942 outputs_to_drop = set() 943 if type(module).__name__ == "JasperEncoder": 944 logging.info( 945 "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " 946 "deployment" 947 ) 948 inputs_to_drop.add("length") 949 outputs_to_drop.add("encoded_lengths") 950 951 # for input_ports 952 for port_name, ntype in module.input_ports.items(): 953 if port_name in inputs_to_drop: 954 input_names.remove(port_name) 955 continue 956 __extract_dynamic_axes(port_name, ntype, dynamic_axes) 957 # for output_ports 958 for port_name, ntype in module.output_ports.items(): 959 if port_name in outputs_to_drop: 960 output_names.remove(port_name) 961 continue 962 __extract_dynamic_axes(port_name, ntype, dynamic_axes) 963 ... [end of nemo/backends/pytorch/actions.py] [start of nemo/collections/asr/jasper.py] ... 104 } 105 106 @property 107 @add_port_docs() 108 def output_ports(self): 109 """Returns definitions of module output ports. 110 """ 111 return { 112 # "outputs": NeuralType( 113 # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag),} 114 # ), 115 # "encoded_lengths": NeuralType({0: AxisType(BatchTag)}), 116 "outputs": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), 117 "encoded_lengths": NeuralType(tuple('B'), LengthsType()), 118 } 119 120 @property 121 def disabled_deployment_input_ports(self): 122 return set(["length"]) 123 124 @property 125 def disabled_deployment_output_ports(self): 126 return set(["encoded_lengths"]) 127 128 def prepare_for_deployment(self): 129 m_count = 0 130 for m in self.modules(): 131 if type(m).__name__ == "MaskedConv1d": 132 m.use_mask = False ... [end of nemo/collections/asr/jasper.py] [start of nemo/core/neural_factory.py] ... 596 raise TypeError(f"All callbacks passed to the eval action must" f"be inherited from EvaluatorCallback") 597 self.train( 598 tensors_to_optimize=None, optimizer='sgd', callbacks=callbacks, optimization_params={'num_epochs': 1}, 599 ) 600 601 def deployment_export( 602 self, module, output: str, d_format: DeploymentFormat, input_example=None, output_example=None 603 ): 604 """Exports Neural Module instance for deployment. 605 606 Args: 607 module: neural module to export 608 output (str): where export results should be saved 609 d_format (DeploymentFormat): which deployment format to use 610 input_example: sometimes tracing will require input examples 611 output_example: Should match inference on input_example 612 """ 613 module.prepare_for_deployment() 614 615 return self._trainer.deployment_export( 616 module=module, 617 output=output, ... [end of nemo/core/neural_factory.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
ba4616f1f011d599de87f0cb3315605e715d402a
Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ```
2020-03-10T03:03:23Z
<patch> <patch> diff --git a/nemo/backends/pytorch/actions.py b/nemo/backends/pytorch/actions.py --- a/nemo/backends/pytorch/actions.py +++ b/nemo/backends/pytorch/actions.py @@ -937,26 +937,16 @@ def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defa if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: dynamic_axes[port_name].append(ind) - # This is a hack for Jasper to Jarvis export -- need re-design for this - inputs_to_drop = set() - outputs_to_drop = set() - if type(module).__name__ == "JasperEncoder": - logging.info( - "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " - "deployment" - ) - inputs_to_drop.add("length") - outputs_to_drop.add("encoded_lengths") - + # extract dynamic axes and remove unnecessary inputs/outputs # for input_ports for port_name, ntype in module.input_ports.items(): - if port_name in inputs_to_drop: + if port_name in module._disabled_deployment_input_ports: input_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) # for output_ports for port_name, ntype in module.output_ports.items(): - if port_name in outputs_to_drop: + if port_name in module._disabled_deployment_output_ports: output_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) diff --git a/nemo/collections/asr/jasper.py b/nemo/collections/asr/jasper.py --- a/nemo/collections/asr/jasper.py +++ b/nemo/collections/asr/jasper.py @@ -118,14 +118,14 @@ def output_ports(self): } @property - def disabled_deployment_input_ports(self): + def _disabled_deployment_input_ports(self): return set(["length"]) @property - def disabled_deployment_output_ports(self): + def _disabled_deployment_output_ports(self): return set(["encoded_lengths"]) - def prepare_for_deployment(self): + def _prepare_for_deployment(self): m_count = 0 for m in self.modules(): if type(m).__name__ == "MaskedConv1d": diff --git a/nemo/core/neural_factory.py b/nemo/core/neural_factory.py --- a/nemo/core/neural_factory.py +++ b/nemo/core/neural_factory.py @@ -610,7 +610,7 @@ def deployment_export( input_example: sometimes tracing will require input examples output_example: Should match inference on input_example """ - module.prepare_for_deployment() + module._prepare_for_deployment() return self._trainer.deployment_export( module=module, diff --git a/nemo/core/neural_modules.py b/nemo/core/neural_modules.py --- a/nemo/core/neural_modules.py +++ b/nemo/core/neural_modules.py @@ -393,7 +393,7 @@ def output_ports(self) -> Optional[Dict[str, NeuralType]]: """ @property - def disabled_deployment_input_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_input_ports(self) -> Optional[Set[str]]: """Returns names of input ports that will not be included in an export Returns: @@ -402,7 +402,7 @@ def disabled_deployment_input_ports(self) -> Optional[Set[str]]: return set([]) @property - def disabled_deployment_output_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_output_ports(self) -> Optional[Set[str]]: """Returns names of output ports that will not be included in an export Returns: @@ -410,7 +410,7 @@ def disabled_deployment_output_ports(self) -> Optional[Set[str]]: """ return set([]) - def prepare_for_deployment(self) -> None: + def _prepare_for_deployment(self) -> None: """Patch the module if required to prepare for deployment """ </patch> </s> </patch>
diff --git a/tests/unit/core/test_deploy_export.py b/tests/unit/core/test_deploy_export.py --- a/tests/unit/core/test_deploy_export.py +++ b/tests/unit/core/test_deploy_export.py @@ -46,9 +46,11 @@ import nemo.collections.nlp.nm.trainables.common.token_classification_nm from nemo import logging +TRT_ONNX_DISABLED = False + # Check if the required libraries and runtimes are installed. +# Only initialize GPU after this runner is activated. try: - # Only initialize GPU after this runner is activated. import pycuda.autoinit # This import causes pycuda to automatically manage CUDA context creation and cleanup. @@ -63,16 +65,17 @@ ) from .tensorrt_runner import TensorRTRunnerV2 except: - # Skip tests. - pytestmark = pytest.mark.skip + TRT_ONNX_DISABLED = True @pytest.mark.usefixtures("neural_factory") class TestDeployExport(TestCase): - def setUp(self): - logging.setLevel(logging.WARNING) - device = nemo.core.DeviceType.GPU - self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) + # def setUp(self): + # super().setUp() + + # logging.setLevel(logging.WARNING) + # device = nemo.core.DeviceType.GPU + # self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) def __test_export_route(self, module, out_name, mode, input_example=None): out = Path(out_name) @@ -112,7 +115,13 @@ def __test_export_route(self, module, out_name, mode, input_example=None): loader_cache = DataLoaderCache(data_loader) profile_shapes = OrderedDict() names = list(module.input_ports) + list(module.output_ports) - + names = list( + filter( + lambda x: x + not in (module._disabled_deployment_input_ports | module._disabled_deployment_output_ports), + names, + ) + ) if isinstance(input_example, tuple): si = [tuple(input_example[i].shape) for i in range(len(input_example))] elif isinstance(input_example, OrderedDict): @@ -152,7 +161,7 @@ def __test_export_route(self, module, out_name, mode, input_example=None): input_names = list(input_metadata.keys()) for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: + if input_name in module._disabled_deployment_input_ports: continue inputs[input_name] = ( input_example[input_name].cpu().numpy() @@ -209,8 +218,8 @@ def __test_export_route(self, module, out_name, mode, input_example=None): ort_inputs = ort_session.get_inputs() for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: - input_name = ort_inputs[i].name + if input_name in module._disabled_deployment_input_ports: + continue inputs[input_name] = ( input_example[input_name].cpu().numpy() if isinstance(input_example, OrderedDict) @@ -263,9 +272,10 @@ def __test_export_route(self, module, out_name, mode, input_example=None): def __test_export_route_all(self, module, out_name, input_example=None): if input_example is not None: - self.__test_export_route( - module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example - ) + if not TRT_ONNX_DISABLED: + self.__test_export_route( + module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example + ) self.__test_export_route(module, out_name + '.onnx', nemo.core.DeploymentFormat.ONNX, input_example) self.__test_export_route(module, out_name + '.pt', nemo.core.DeploymentFormat.PYTORCH, input_example) self.__test_export_route(module, out_name + '.ts', nemo.core.DeploymentFormat.TORCHSCRIPT, input_example) @@ -323,9 +333,7 @@ def test_jasper_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="jasper_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randn(256).cuda()), + module=jasper_encoder, out_name="jasper_encoder", input_example=torch.randn(16, 64, 256).cuda(), ) @pytest.mark.unit @@ -343,7 +351,5 @@ def test_quartz_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="quartz_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randint(20, (16,)).cuda()), + module=jasper_encoder, out_name="quartz_encoder", input_example=torch.randn(16, 64, 256).cuda(), )
1.0
NVIDIA__NeMo-3632
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> ./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`. </issue> <code> [start of README.rst] 1 2 |status| |documentation| |license| |lgtm_grade| |lgtm_alerts| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |lgtm_grade| image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 17 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python 18 :alt: Language grade: Python 19 20 .. |lgtm_alerts| image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 21 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ 22 :alt: Total alerts 23 24 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 25 :target: https://github.com/psf/black 26 :alt: Code style: black 27 28 .. _main-readme: 29 30 **NVIDIA NeMo** 31 =============== 32 33 Introduction 34 ------------ 35 36 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), natural language processing (NLP), and text-to-speech synthesis (TTS). 37 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 38 39 `Pre-trained NeMo models. <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_ 40 41 `Introductory video. <https://www.youtube.com/embed/wBgpMf_KQVw>`_ 42 43 Key Features 44 ------------ 45 46 * Speech processing 47 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 48 * Supported models: Jasper, QuartzNet, CitriNet, Conformer-CTC, Conformer-Transducer, ContextNet, ... 49 * Supports CTC and Transducer/RNNT losses/decoders 50 * Beam Search decoding 51 * `Language Modelling for ASR <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 52 * Streaming and Buffered ASR (CTC/Transdcer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/main/examples/asr/asr_chunked_inference>`_ 53 * `Speech Classification and Speech Command Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition) 54 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 55 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 56 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 57 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 58 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 59 * Natural Language Processing 60 * `Compatible with Hugging Face Transformers and NVIDIA Megatron <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html>`_ 61 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation.html>`_ 62 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 63 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 64 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 65 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 66 * `BERT pre-training <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/bert_pretraining.html>`_ 67 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 68 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 69 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 70 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 71 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 72 * `Neural Duplex Text Normalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization.html>`_ 73 * `Prompt Tuning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html#prompt-tuning>`_ 74 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 75 * `Speech synthesis (TTS) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 76 * Spectrogram generation: Tacotron2, GlowTTS, TalkNet, FastPitch, FastSpeech2, Mixer-TTS, Mixer-TTS-X 77 * Vocoders: WaveGlow, SqueezeWave, UniGlow, MelGAN, HiFiGAN, UnivNet 78 * End-to-end speech generation: FastPitch_HifiGan_E2E, FastSpeech2_HifiGan_E2E 79 * `NGC collection of pre-trained TTS models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 80 * `Tools <https://github.com/NVIDIA/NeMo/tree/main/tools>`_ 81 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/text_processing_deployment.html>`_ 82 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 83 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 84 85 86 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 87 88 Requirements 89 ------------ 90 91 1) Python 3.6, 3.7 or 3.8 92 2) Pytorch 1.10.0 or above 93 3) NVIDIA GPU for training 94 95 Documentation 96 ------------- 97 98 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 99 :alt: Documentation Status 100 :scale: 100% 101 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 102 103 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 104 :alt: Documentation Status 105 :scale: 100% 106 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 107 108 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 109 | Version | Status | Description | 110 +=========+=============+==========================================================================================================================================+ 111 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 112 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 113 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 114 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 115 116 Tutorials 117 --------- 118 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 119 120 Getting help with NeMo 121 ---------------------- 122 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 123 124 125 Installation 126 ------------ 127 128 Pip 129 ~~~ 130 Use this installation mode if you want the latest released version. 131 132 .. code-block:: bash 133 134 apt-get update && apt-get install -y libsndfile1 ffmpeg 135 pip install Cython 136 pip install nemo_toolkit['all'] 137 138 .. note:: 139 140 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 141 142 Pip from source 143 ~~~~~~~~~~~~~~~ 144 Use this installation mode if you want the a version from particular GitHub branch (e.g main). 145 146 .. code-block:: bash 147 148 apt-get update && apt-get install -y libsndfile1 ffmpeg 149 pip install Cython 150 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 151 152 153 From source 154 ~~~~~~~~~~~ 155 Use this installation mode if you are contributing to NeMo. 156 157 .. code-block:: bash 158 159 apt-get update && apt-get install -y libsndfile1 ffmpeg 160 git clone https://github.com/NVIDIA/NeMo 161 cd NeMo 162 ./reinstall.sh 163 164 .. note:: 165 166 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 167 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 168 169 RNNT 170 ~~~~ 171 Note that RNNT requires numba to be installed from conda. 172 173 .. code-block:: bash 174 175 conda remove numba 176 pip uninstall numba 177 conda install -c conda-forge numba 178 179 Megatron GPT 180 ~~~~~~~~~~~~ 181 Megatron GPT training requires NVIDIA Apex to be installed. 182 183 .. code-block:: bash 184 185 git clone https://github.com/NVIDIA/apex 186 cd apex 187 git checkout c8bcc98176ad8c3a0717082600c70c907891f9cb 188 pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" ./ 189 190 Docker containers: 191 ~~~~~~~~~~~~~~~~~~ 192 To build a nemo container with Dockerfile from a branch, please run 193 194 .. code-block:: bash 195 196 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 197 198 199 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 22.01-py3 and then installing from GitHub. 200 201 .. code-block:: bash 202 203 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 204 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 205 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:22.01-py3 206 207 Examples 208 -------- 209 210 Many examples can be found under `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 211 212 213 Contributing 214 ------------ 215 216 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 217 218 Publications 219 ------------ 220 221 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/blob/main/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 222 223 Citation 224 -------- 225 226 .. code-block:: bash 227 228 @article{kuchaiev2019nemo, 229 title={Nemo: a toolkit for building ai applications using neural modules}, 230 author={Kuchaiev, Oleksii and Li, Jason and Nguyen, Huyen and Hrinchuk, Oleksii and Leary, Ryan and Ginsburg, Boris and Kriman, Samuel and Beliaev, Stanislav and Lavrukhin, Vitaly and Cook, Jack and others}, 231 journal={arXiv preprint arXiv:1909.09577}, 232 year={2019} 233 } 234 235 License 236 ------- 237 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 238 [end of README.rst] [start of nemo_text_processing/text_normalization/normalize.py] ... 32 except (ModuleNotFoundError, ImportError): 33 PYNINI_AVAILABLE = False 34 35 try: 36 from nemo.collections.common.tokenizers.moses_tokenizers import MosesProcessor 37 from nemo.collections.nlp.data.text_normalization.utils import post_process_punct 38 39 NLP_AVAILABLE = True 40 except (ModuleNotFoundError, ImportError): 41 NLP_AVAILABLE = False 42 43 44 SPACE_DUP = re.compile(' {2,}') 45 46 47 class Normalizer: 48 """ 49 Normalizer class that converts text from written to spoken form. 50 Useful for TTS preprocessing. 51 52 Args: 53 input_case: expected input capitalization 54 lang: language specifying the TN rules, by default: English ... ... 69 assert input_case in ["lower_cased", "cased"] 70 71 if not PYNINI_AVAILABLE: 72 raise ImportError(get_installation_msg()) 73 74 if lang == 'en' and deterministic: 75 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify import ClassifyFst 76 from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst 77 elif lang == 'en' and not deterministic: 78 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify_with_audio import ClassifyFst 79 from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst 80 elif lang == 'ru': 81 # Ru TN only support non-deterministic cases and produces multiple normalization options 82 # use normalize_with_audio.py 83 from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst 84 from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst 85 elif lang == 'de': 86 # Ru TN only support non-deterministic cases and produces multiple normalization options 87 # use normalize_with_audio.py 88 from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst 89 from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst 90 self.tagger = ClassifyFst( 91 input_case=input_case, 92 deterministic=deterministic, 93 cache_dir=cache_dir, ... ... 92 deterministic=deterministic, 93 cache_dir=cache_dir, 94 overwrite_cache=overwrite_cache, 95 whitelist=whitelist, 96 ) 97 self.verbalizer = VerbalizeFinalFst(deterministic=deterministic) 98 self.parser = TokenParser() 99 self.lang = lang 100 101 if NLP_AVAILABLE: 102 self.processor = MosesProcessor(lang_id=lang) 103 else: 104 self.processor = None 105 print("NeMo NLP is not available. Moses de-tokenization will be skipped.") 106 107 def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: 108 """ 109 NeMo text normalizer 110 111 Args: 112 texts: list of input strings 113 verbose: whether to print intermediate meta information ... ... 343 344 def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: 345 """ 346 Given verbalized lattice return shortest path 347 348 Args: 349 lattice: verbalization lattice 350 351 Returns: shortest path 352 """ 353 output = pynini.shortestpath(lattice, nshortest=1, unique=True).string() 354 return output 355 356 357 def parse_args(): 358 parser = ArgumentParser() 359 parser.add_argument("input_string", help="input string", type=str) 360 parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) 361 parser.add_argument( 362 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 363 ) 364 parser.add_argument("--verbose", help="print info for debugging", action='store_true') ... [end of nemo_text_processing/text_normalization/normalize.py] [start of nemo_text_processing/text_normalization/__init__.py] ... 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from nemo.utils import logging 16 17 try: 18 import pynini 19 20 PYNINI_AVAILABLE = True 21 except (ModuleNotFoundError, ImportError): 22 logging.warning( 23 "`pynini` is not installed ! \n" 24 "Please run the `nemo_text_processing/setup.sh` script" 25 "prior to usage of this toolkit." 26 ) 27 28 PYNINI_AVAILABLE = False ... [end of nemo_text_processing/text_normalization/__init__.py] [start of nemo_text_processing/text_normalization/en/graph_utils.py] ... 145 146 def get_singulars(fst): 147 """ 148 Given plural returns singulars 149 150 Args: 151 fst: Fst 152 153 Returns singulars to given plural forms 154 """ 155 return PLURAL_TO_SINGULAR @ fst 156 157 158 def convert_space(fst) -> 'pynini.FstLike': 159 """ 160 Converts space to nonbreaking space. 161 Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" 162 This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. 163 164 Args: 165 fst: input fst 166 ... ... 194 """ 195 Returns true if FAR can be loaded 196 """ 197 return self.far_path.exists() 198 199 @property 200 def fst(self) -> 'pynini.FstLike': 201 return self._fst 202 203 @fst.setter 204 def fst(self, fst): 205 self._fst = fst 206 207 def add_tokens(self, fst) -> 'pynini.FstLike': 208 """ 209 Wraps class name around to given fst 210 211 Args: 212 fst: input fst 213 214 Returns: 215 Fst: fst 216 """ 217 return pynutil.insert(f"{self.name} {{ ") + fst + pynutil.insert(" }") ... [end of nemo_text_processing/text_normalization/en/graph_utils.py] [start of nemo_text_processing/text_normalization/normalize_with_audio.py] ... 41 42 try: 43 from nemo.collections.nlp.data.text_normalization.utils import post_process_punct 44 from nemo_text_processing.text_normalization.data_loader_utils import pre_process 45 46 NLP_AVAILABLE = True 47 except (ModuleNotFoundError, ImportError): 48 NLP_AVAILABLE = False 49 50 """ 51 The script provides multiple normalization options and chooses the best one that minimizes CER of the ASR output 52 (most of the semiotic classes use deterministic=False flag). 53 54 To run this script with a .json manifest file, the manifest file should contain the following fields: 55 "audio_data" - path to the audio file 56 "text" - raw text 57 "pred_text" - ASR model prediction 58 59 See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions 60 61 When the manifest is ready, run: 62 python normalize_with_audio.py \ 63 --audio_data PATH/TO/MANIFEST.JSON \ 64 --language en 65 66 67 To run with a single audio file, specify path to audio and text with: 68 python normalize_with_audio.py \ 69 --audio_data PATH/TO/AUDIO.WAV \ 70 --language en \ ... ... 57 "pred_text" - ASR model prediction 58 59 See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions 60 61 When the manifest is ready, run: 62 python normalize_with_audio.py \ 63 --audio_data PATH/TO/MANIFEST.JSON \ 64 --language en 65 66 67 To run with a single audio file, specify path to audio and text with: 68 python normalize_with_audio.py \ 69 --audio_data PATH/TO/AUDIO.WAV \ 70 --language en \ 71 --text raw text OR PATH/TO/.TXT/FILE 72 --model QuartzNet15x5Base-En \ 73 --verbose 74 75 To see possible normalization options for a text input without an audio file (could be used for debugging), run: 76 python python normalize_with_audio.py --text "RAW TEXT" 77 78 Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference 79 """ 80 81 82 class NormalizerWithAudio(Normalizer): 83 """ 84 Normalizer class that converts text from written to spoken form. 85 Useful for TTS preprocessing. 86 87 Args: 88 input_case: expected input capitalization 89 lang: language ... ... 268 asr_model = ASRModel.restore_from(asr_model) 269 elif args.model in ASRModel.get_available_model_names(): 270 asr_model = ASRModel.from_pretrained(asr_model) 271 else: 272 raise ValueError( 273 f'Provide path to the pretrained checkpoint or choose from {ASRModel.get_available_model_names()}' 274 ) 275 return asr_model 276 277 278 def parse_args(): 279 parser = ArgumentParser() 280 parser.add_argument("--text", help="input string or path to a .txt file", default=None, type=str) 281 parser.add_argument( 282 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 283 ) 284 parser.add_argument( 285 "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str 286 ) 287 parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") 288 parser.add_argument( 289 '--model', type=str, default='QuartzNet15x5Base-En', help='Pre-trained model name or path to model checkpoint' ... [end of nemo_text_processing/text_normalization/normalize_with_audio.py] [start of /dev/null] ... [end of /dev/null] [start of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] ... ... 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 17 from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space 18 19 try: 20 import pynini 21 from pynini.lib import pynutil 22 23 PYNINI_AVAILABLE = True 24 except (ModuleNotFoundError, ImportError): 25 PYNINI_AVAILABLE = False 26 27 ... [end of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] [start of tools/text_processing_deployment/pynini_export.py] ... 53 54 def tn_grammars(**kwargs): 55 d = {} 56 d['classify'] = { 57 'TOKENIZE_AND_CLASSIFY': TNClassifyFst( 58 input_case=kwargs["input_case"], 59 deterministic=True, 60 cache_dir=kwargs["cache_dir"], 61 overwrite_cache=kwargs["overwrite_cache"], 62 ).fst 63 } 64 d['verbalize'] = {'ALL': TNVerbalizeFst(deterministic=True).fst, 'REDUP': pynini.accep("REDUP")} 65 return d 66 67 68 def export_grammars(output_dir, grammars): 69 """ 70 Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. 71 72 Args: 73 output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. 74 grammars: grammars to be exported ... ... 95 ) 96 parser.add_argument( 97 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 98 ) 99 parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") 100 parser.add_argument( 101 "--cache_dir", 102 help="path to a dir with .far grammar file. Set to None to avoid using cache", 103 default=None, 104 type=str, 105 ) 106 return parser.parse_args() 107 108 109 if __name__ == '__main__': 110 args = parse_args() 111 112 if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': 113 raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') 114 115 if args.language == 'en': 116 from nemo_text_processing.inverse_text_normalization.en.taggers.tokenize_and_classify import ( ... ... 134 ClassifyFst as TNClassifyFst, 135 ) 136 from nemo_text_processing.text_normalization.de.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst 137 elif args.language == 'ru': 138 from nemo_text_processing.inverse_text_normalization.ru.taggers.tokenize_and_classify import ( 139 ClassifyFst as ITNClassifyFst, 140 ) 141 from nemo_text_processing.inverse_text_normalization.ru.verbalizers.verbalize import ( 142 VerbalizeFst as ITNVerbalizeFst, 143 ) 144 elif args.language == 'es': 145 from nemo_text_processing.inverse_text_normalization.es.taggers.tokenize_and_classify import ( 146 ClassifyFst as ITNClassifyFst, 147 ) 148 from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( 149 VerbalizeFst as ITNVerbalizeFst, 150 ) 151 elif args.language == 'fr': 152 from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( 153 ClassifyFst as ITNClassifyFst, 154 ) ... [end of tools/text_processing_deployment/pynini_export.py] [start of nemo_text_processing/text_normalization/en/verbalizers/word.py] ... ... 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space 17 18 try: 19 import pynini 20 from pynini.lib import pynutil 21 22 PYNINI_AVAILABLE = True 23 except (ModuleNotFoundError, ImportError): 24 PYNINI_AVAILABLE = False 25 26 ... [end of nemo_text_processing/text_normalization/en/verbalizers/word.py] [start of nemo_text_processing/text_normalization/en/taggers/punctuation.py] ... 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 import sys 17 from unicodedata import category 18 19 from nemo_text_processing.text_normalization.en.graph_utils import GraphFst 20 21 try: 22 import pynini 23 from pynini.lib import pynutil 24 25 PYNINI_AVAILABLE = False 26 except (ModuleNotFoundError, ImportError): 27 PYNINI_AVAILABLE = False 28 29 ... [end of nemo_text_processing/text_normalization/en/taggers/punctuation.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
022f0292aecbc98d591d49423d5045235394f793
./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`.
2022-02-09T05:12:31Z
<patch> <patch> diff --git a/nemo_text_processing/text_normalization/__init__.py b/nemo_text_processing/text_normalization/__init__.py --- a/nemo_text_processing/text_normalization/__init__.py +++ b/nemo_text_processing/text_normalization/__init__.py @@ -21,7 +21,7 @@ except (ModuleNotFoundError, ImportError): logging.warning( "`pynini` is not installed ! \n" - "Please run the `nemo_text_processing/setup.sh` script" + "Please run the `nemo_text_processing/setup.sh` script " "prior to usage of this toolkit." ) diff --git a/nemo_text_processing/text_normalization/en/graph_utils.py b/nemo_text_processing/text_normalization/en/graph_utils.py --- a/nemo_text_processing/text_normalization/en/graph_utils.py +++ b/nemo_text_processing/text_normalization/en/graph_utils.py @@ -159,7 +159,7 @@ def convert_space(fst) -> 'pynini.FstLike': """ Converts space to nonbreaking space. Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" - This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. + This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. Args: fst: input fst @@ -208,9 +208,9 @@ def add_tokens(self, fst) -> 'pynini.FstLike': """ Wraps class name around to given fst - Args: + Args: fst: input fst - + Returns: Fst: fst """ diff --git a/nemo_text_processing/text_normalization/en/taggers/punctuation.py b/nemo_text_processing/text_normalization/en/taggers/punctuation.py --- a/nemo_text_processing/text_normalization/en/taggers/punctuation.py +++ b/nemo_text_processing/text_normalization/en/taggers/punctuation.py @@ -22,7 +22,7 @@ import pynini from pynini.lib import pynutil - PYNINI_AVAILABLE = False + PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py --- a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py @@ -12,8 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -21,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/word.py b/nemo_text_processing/text_normalization/en/verbalizers/word.py --- a/nemo_text_processing/text_normalization/en/verbalizers/word.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/word.py @@ -12,7 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -20,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/es/__init__.py b/nemo_text_processing/text_normalization/es/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/__init__.py @@ -0,0 +1,15 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +LOCALIZATION = "eu" # Set to am for alternate formatting diff --git a/nemo_text_processing/text_normalization/es/data/__init__.py b/nemo_text_processing/text_normalization/es/data/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/dates/__init__.py b/nemo_text_processing/text_normalization/es/data/dates/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/dates/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/electronic/__init__.py b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/fractions/__init__.py b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/measures/__init__.py b/nemo_text_processing/text_normalization/es/data/measures/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/measures/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/money/__init__.py b/nemo_text_processing/text_normalization/es/data/money/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/money/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/numbers/__init__.py b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/roman/__init__.py b/nemo_text_processing/text_normalization/es/data/roman/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/roman/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/time/__init__.py b/nemo_text_processing/text_normalization/es/data/time/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/time/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/graph_utils.py b/nemo_text_processing/text_normalization/es/graph_utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/graph_utils.py @@ -0,0 +1,179 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, NEMO_SPACE +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + digits = pynini.project(pynini.string_file(get_abs_path("data/numbers/digit.tsv")), "input") + tens = pynini.project(pynini.string_file(get_abs_path("data/numbers/ties.tsv")), "input") + teens = pynini.project(pynini.string_file(get_abs_path("data/numbers/teen.tsv")), "input") + twenties = pynini.project(pynini.string_file(get_abs_path("data/numbers/twenties.tsv")), "input") + hundreds = pynini.project(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv")), "input") + + accents = pynini.string_map([("รก", "a"), ("รฉ", "e"), ("รญ", "i"), ("รณ", "o"), ("รบ", "u")]) + + if LOCALIZATION == "am": # Setting localization for central and northern america formatting + cardinal_separator = pynini.string_map([",", NEMO_SPACE]) + decimal_separator = pynini.accep(".") + else: + cardinal_separator = pynini.string_map([".", NEMO_SPACE]) + decimal_separator = pynini.accep(",") + + ones = pynini.union("un", "รบn") + fem_ones = pynini.union(pynini.cross("un", "una"), pynini.cross("รบn", "una"), pynini.cross("uno", "una")) + one_to_one_hundred = pynini.union(digits, tens, teens, twenties, tens + pynini.accep(" y ") + digits) + fem_hundreds = hundreds @ pynini.cdrewrite(pynini.cross("ientos", "ientas"), "", "", NEMO_SIGMA) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digits = None + tens = None + teens = None + twenties = None + hundreds = None + + accents = None + + cardinal_separator = None + decimal_separator = None + + ones = None + fem_ones = None + one_to_one_hundred = None + fem_hundreds = None + + PYNINI_AVAILABLE = False + + +def strip_accent(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Converts all accented vowels to non-accented equivalents + + Args: + fst: Any fst. Composes vowel conversion onto fst's output strings + """ + return fst @ pynini.cdrewrite(accents, "", "", NEMO_SIGMA) + + +def shift_cardinal_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Applies gender conversion rules to a cardinal string. These include: rendering all masculine forms of "uno" (including apocopated forms) as "una" and + Converting all gendered numbers in the hundreds series (200,300,400...) to feminine equivalent (e.g. "doscientos" -> "doscientas"). Converssion only applies + to value place for <1000 and multiple of 1000. (e.g. "doscientos mil doscientos" -> "doscientas mil doscientas".) For place values greater than the thousands, there + is no gender shift as the higher powers of ten ("millones", "billones") are masculine nouns and any conversion would be formally + ungrammatical. + e.g. + "doscientos" -> "doscientas" + "doscientos mil" -> "doscientas mil" + "doscientos millones" -> "doscientos millones" + "doscientos mil millones" -> "doscientos mil millones" + "doscientos millones doscientos mil doscientos" -> "doscientos millones doscientas mil doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + before_mil = ( + NEMO_SPACE + + (pynini.accep("mil") | pynini.accep("milรฉsimo")) + + pynini.closure(NEMO_SPACE + hundreds, 0, 1) + + pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + + pynini.union(pynini.accep("[EOS]"), pynini.accep("\""), decimal_separator) + ) + before_double_digits = pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + pynini.union( + pynini.accep("[EOS]"), pynini.accep("\"") + ) + + fem_allign = pynini.cdrewrite(fem_hundreds, "", before_mil, NEMO_SIGMA) # doscientas mil dosciento + fem_allign @= pynini.cdrewrite(fem_hundreds, "", before_double_digits, NEMO_SIGMA) # doscientas mil doscienta + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union("[EOS]", "\"", decimal_separator), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def shift_number_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Performs gender conversion on all verbalized numbers in output. All values in the hundreds series (200,300,400) are changed to + feminine gender (e.g. "doscientos" -> "doscientas") and all forms of "uno" (including apocopated forms) are converted to "una". + This has no boundary restriction and will perform shift across all values in output string. + e.g. + "doscientos" -> "doscientas" + "doscientos millones" -> "doscientas millones" + "doscientos millones doscientos" -> "doscientas millones doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + fem_allign = pynini.cdrewrite(fem_hundreds, "", "", NEMO_SIGMA) + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union(NEMO_SPACE, pynini.accep("[EOS]"), pynini.accep("\"")), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def strip_cardinal_apocope(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Reverts apocope on cardinal strings in line with formation rules. e.g. "un" -> "uno". Due to cardinal formation rules, this in effect only + affects strings where the final value is a variation of "un". + e.g. + "un" -> "uno" + "veintiรบn" -> "veintiuno" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + # Since cardinals use apocope by default for large values (e.g. "millรณn"), this only needs to act on the last instance of one + strip = pynini.cross("un", "uno") | pynini.cross("รบn", "uno") + strip = pynini.cdrewrite(strip, "", pynini.union("[EOS]", "\""), NEMO_SIGMA) + return fst @ strip + + +def roman_to_int(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Alters given fst to convert Roman integers (lower and upper cased) into Arabic numerals. Valid for values up to 1000. + e.g. + "V" -> "5" + "i" -> "1" + + Args: + fst: Any fst. Composes fst onto Roman conversion outputs. + """ + + def _load_roman(file: str): + roman = load_labels(get_abs_path(file)) + roman_numerals = [(x, y) for x, y in roman] + [(x.upper(), y) for x, y in roman] + return pynini.string_map(roman_numerals) + + digit = _load_roman("data/roman/digit.tsv") + ties = _load_roman("data/roman/ties.tsv") + hundreds = _load_roman("data/roman/hundreds.tsv") + + graph = ( + digit + | ties + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + | ( + hundreds + + (ties | pynutil.add_weight(pynutil.insert("0"), 0.01)) + + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + ) + ).optimize() + + return graph @ fst diff --git a/nemo_text_processing/text_normalization/es/taggers/__init__.py b/nemo_text_processing/text_normalization/es/taggers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/taggers/cardinal.py b/nemo_text_processing/text_normalization/es/taggers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/cardinal.py @@ -0,0 +1,190 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import cardinal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + teen = pynini.invert(pynini.string_file(get_abs_path("data/numbers/teen.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/ties.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/twenties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + zero = None + digit = None + teen = None + ties = None + twenties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def filter_punctuation(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Helper function for parsing number strings. Converts common cardinal strings (groups of three digits delineated by 'cardinal_separator' - see graph_utils) + and converts to a string of digits: + "1 000" -> "1000" + "1.000.000" -> "1000000" + Args: + fst: Any pynini.FstLike object. Function composes fst onto string parser fst + + Returns: + fst: A pynini.FstLike object + """ + exactly_three_digits = NEMO_DIGIT ** 3 # for blocks of three + up_to_three_digits = pynini.closure(NEMO_DIGIT, 1, 3) # for start of string + + cardinal_string = pynini.closure( + NEMO_DIGIT, 1 + ) # For string w/o punctuation (used for page numbers, thousand series) + + cardinal_string |= ( + up_to_three_digits + + pynutil.delete(cardinal_separator) + + pynini.closure(exactly_three_digits + pynutil.delete(cardinal_separator)) + + exactly_three_digits + ) + + return cardinal_string @ fst + + +class CardinalFst(GraphFst): + """ + Finite state transducer for classifying cardinals, e.g. + "1000" -> cardinal { integer: "mil" } + "2.000.000" -> cardinal { integer: "dos millones" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="classify", deterministic=deterministic) + + # Any single digit + graph_digit = digit + digits_no_one = (NEMO_DIGIT - "1") @ graph_digit + + # Any double digit + graph_tens = teen + graph_tens |= ties + (pynutil.delete('0') | (pynutil.insert(" y ") + graph_digit)) + graph_tens |= twenties + + self.tens = graph_tens.optimize() + + self.two_digit_non_zero = pynini.union( + graph_digit, graph_tens, (pynini.cross("0", NEMO_SPACE) + graph_digit) + ).optimize() + + # Three digit strings + graph_hundreds = hundreds + pynini.union( + pynutil.delete("00"), (insert_space + graph_tens), (pynini.cross("0", NEMO_SPACE) + graph_digit) + ) + graph_hundreds |= pynini.cross("100", "cien") + graph_hundreds |= ( + pynini.cross("1", "ciento") + insert_space + pynini.union(graph_tens, pynutil.delete("0") + graph_digit) + ) + + self.hundreds = graph_hundreds.optimize() + + # For all three digit strings with leading zeroes (graph appends '0's to manage place in string) + graph_hundreds_component = pynini.union(graph_hundreds, pynutil.delete("0") + graph_tens) + + graph_hundreds_component_at_least_one_none_zero_digit = graph_hundreds_component | ( + pynutil.delete("00") + graph_digit + ) + graph_hundreds_component_at_least_one_none_zero_digit_no_one = graph_hundreds_component | ( + pynutil.delete("00") + digits_no_one + ) + + graph_thousands_component_at_least_one_none_zero_digit = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_thousands_component_at_least_one_none_zero_digit_no_one = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit_no_one, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_million = pynutil.add_weight(pynini.cross("000001", "un millรณn"), -0.001) + graph_million |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" millones") + graph_million |= pynutil.delete("000000") + graph_million += insert_space + + graph_billion = pynutil.add_weight(pynini.cross("000001", "un billรณn"), -0.001) + graph_billion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" billones") + graph_billion |= pynutil.delete("000000") + graph_billion += insert_space + + graph_trillion = pynutil.add_weight(pynini.cross("000001", "un trillรณn"), -0.001) + graph_trillion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" trillones") + graph_trillion |= pynutil.delete("000000") + graph_trillion += insert_space + + graph = ( + graph_trillion + + graph_billion + + graph_million + + (graph_thousands_component_at_least_one_none_zero_digit | pynutil.delete("000000")) + ) + + self.graph = ( + ((NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 0)) + @ pynini.cdrewrite(pynini.closure(pynutil.insert("0")), "[BOS]", "", NEMO_SIGMA) + @ NEMO_DIGIT ** 24 + @ graph + @ pynini.cdrewrite(delete_space, "[BOS]", "", NEMO_SIGMA) + @ pynini.cdrewrite(delete_space, "", "[EOS]", NEMO_SIGMA) + @ pynini.cdrewrite( + pynini.cross(pynini.closure(NEMO_WHITE_SPACE, 2), NEMO_SPACE), NEMO_ALPHA, NEMO_ALPHA, NEMO_SIGMA + ) + ) + self.graph |= zero + + self.graph = filter_punctuation(self.graph).optimize() + + optional_minus_graph = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + final_graph = optional_minus_graph + pynutil.insert("integer: \"") + self.graph + pynutil.insert("\"") + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/date.py b/nemo_text_processing/text_normalization/es/taggers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/date.py @@ -0,0 +1,107 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_DIGIT, NEMO_SPACE, GraphFst, delete_extra_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + articles = pynini.union("de", "del", "el", "del aรฑo") + delete_leading_zero = (pynutil.delete("0") | (NEMO_DIGIT - "0")) + NEMO_DIGIT + month_numbers = pynini.string_file(get_abs_path("data/dates/months.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + articles = None + delete_leading_zero = None + month_numbers = None + + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for classifying date, e.g. + "01.04.2010" -> date { day: "un" month: "enero" year: "dos mil diez" preserve_order: true } + "marzo 4 2000" -> date { month: "marzo" day: "cuatro" year: "dos mil" } + "1990-20-01" -> date { year: "mil novecientos noventa" day: "veinte" month: "enero" } + + Args: + cardinal: cardinal GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool): + super().__init__(name="date", kind="classify", deterministic=deterministic) + + number_to_month = month_numbers.optimize() + month_graph = pynini.project(number_to_month, "output") + + numbers = cardinal.graph + optional_leading_zero = delete_leading_zero | NEMO_DIGIT + + # 01, 31, 1 + digit_day = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 32)]) @ numbers + day = (pynutil.insert("day: \"") + digit_day + pynutil.insert("\"")).optimize() + + digit_month = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 13)]) + number_to_month = digit_month @ number_to_month + + month_name = (pynutil.insert("month: \"") + month_graph + pynutil.insert("\"")).optimize() + month_number = (pynutil.insert("month: \"") + number_to_month + pynutil.insert("\"")).optimize() + + # prefer cardinal over year + year = (NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 1, 3) # 90, 990, 1990 + year @= numbers + self.year = year + + year_only = pynutil.insert("year: \"") + year + pynutil.insert("\"") + year_with_articles = ( + pynutil.insert("year: \"") + pynini.closure(articles + NEMO_SPACE, 0, 1) + year + pynutil.insert("\"") + ) + + graph_dmy = ( + day + + pynini.closure(pynutil.delete(" de")) + + NEMO_SPACE + + month_name + + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + graph_mdy = ( # English influences on language + month_name + delete_extra_space + day + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + separators = [".", "-", "/"] + for sep in separators: + year_optional = pynini.closure(pynini.cross(sep, NEMO_SPACE) + year_only, 0, 1) + new_graph = day + pynini.cross(sep, NEMO_SPACE) + month_number + year_optional + graph_dmy |= new_graph + if not deterministic: + new_graph = month_number + pynini.cross(sep, NEMO_SPACE) + day + year_optional + graph_mdy |= new_graph + + dash = "-" + day_optional = pynini.closure(pynini.cross(dash, NEMO_SPACE) + day, 0, 1) + graph_ymd = NEMO_DIGIT ** 4 @ year_only + pynini.cross(dash, NEMO_SPACE) + month_number + day_optional + + final_graph = graph_dmy + pynutil.insert(" preserve_order: true") + final_graph |= graph_ymd + final_graph |= graph_mdy + + self.final_graph = final_graph.optimize() + self.fst = self.add_tokens(self.final_graph).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/decimals.py b/nemo_text_processing/text_normalization/es/taggers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/decimals.py @@ -0,0 +1,138 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + cardinal_separator, + decimal_separator, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + quantities = pynini.string_file(get_abs_path("data/numbers/quantities.tsv")) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + quantities = None + digit = None + zero = None + + PYNINI_AVAILABLE = False + + +def get_quantity(decimal_graph: 'pynini.FstLike', cardinal_graph: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Returns FST that transforms either a cardinal or decimal followed by a quantity into a numeral, + e.g. 2 millones -> integer_part: "dos" quantity: "millones" + e.g. 2,4 millones -> integer_part: "dos" fractional_part: "quatro" quantity: "millones" + e.g. 2,400 millones -> integer_part: "dos mil cuatrocientos" fractional_part: "quatro" quantity: "millones" + + Args: + decimal_graph: DecimalFST + cardinal_graph: CardinalFST + """ + numbers = pynini.closure(NEMO_DIGIT, 1, 6) @ cardinal_graph + numbers = pynini.cdrewrite(pynutil.delete(cardinal_separator), "", "", NEMO_SIGMA) @ numbers + + res = ( + pynutil.insert("integer_part: \"") + + numbers # The cardinal we're passing only produces 'un' for one, so gender agreement is safe (all quantities are masculine). Limit to 10^6 power. + + pynutil.insert("\"") + + NEMO_SPACE + + pynutil.insert("quantity: \"") + + quantities + + pynutil.insert("\"") + ) + res |= decimal_graph + NEMO_SPACE + pynutil.insert("quantity: \"") + quantities + pynutil.insert("\"") + return res + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + -11,4006 billones -> decimal { negative: "true" integer_part: "once" fractional_part: "cuatro cero cero seis" quantity: "billones" preserve_order: true } + 1 billรณn -> decimal { integer_part: "un" quantity: "billรณn" preserve_order: true } + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + graph_digit = digit | zero + + if not deterministic: + graph = pynini.union(graph_digit, cardinal.hundreds, cardinal.tens) + graph += pynini.closure(insert_space + graph) + + else: + # General pattern seems to be 1-3 digits: map as cardinal, default to digits otherwise \ + graph = pynini.union( + graph_digit, + cardinal.tens, + cardinal.hundreds, + graph_digit + pynini.closure(insert_space + graph_digit, 3), + zero + + pynini.closure(insert_space + zero) + + pynini.closure(insert_space + graph_digit), # For cases such as "1,010" + ) + + # Need to strip apocope everywhere BUT end of string + reverse_apocope = pynini.string_map([("un", "uno"), ("รบn", "uno")]) + apply_reverse_apocope = pynini.cdrewrite(reverse_apocope, "", NEMO_SPACE, NEMO_SIGMA) + graph @= apply_reverse_apocope + + # Technically decimals should be space delineated groups of three, e.g. (1,333 333). This removes any possible spaces + strip_formatting = pynini.cdrewrite(delete_space, "", "", NEMO_SIGMA) + graph = strip_formatting @ graph + + self.graph = graph.optimize() + + graph_separator = pynutil.delete(decimal_separator) + optional_graph_negative = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + self.graph_fractional = pynutil.insert("fractional_part: \"") + self.graph + pynutil.insert("\"") + + # Integer graph maintains apocope except for ones place + graph_integer = ( + strip_cardinal_apocope(cardinal.graph) + if deterministic + else pynini.union(cardinal.graph, strip_cardinal_apocope(cardinal.graph)) + ) # Gives us forms w/ and w/o apocope + self.graph_integer = pynutil.insert("integer_part: \"") + graph_integer + pynutil.insert("\"") + final_graph_wo_sign = self.graph_integer + graph_separator + insert_space + self.graph_fractional + + self.final_graph_wo_negative = ( + final_graph_wo_sign | get_quantity(final_graph_wo_sign, cardinal.graph).optimize() + ) + final_graph = optional_graph_negative + self.final_graph_wo_negative + + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/electronic.py b/nemo_text_processing/text_normalization/es/taggers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/electronic.py @@ -0,0 +1,84 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_ALPHA, NEMO_DIGIT, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + common_domains = [x[0] for x in load_labels(get_abs_path("data/electronic/domain.tsv"))] + symbols = [x[0] for x in load_labels(get_abs_path("data/electronic/symbols.tsv"))] + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + common_domains = None + symbols = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for classifying electronic: email addresses + e.g. "abc@hotmail.com" -> electronic { username: "abc" domain: "hotmail.com" preserve_order: true } + e.g. "www.abc.com/123" -> electronic { protocol: "www." domain: "abc.com/123" preserve_order: true } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="classify", deterministic=deterministic) + + dot = pynini.accep(".") + accepted_common_domains = pynini.union(*common_domains) + accepted_symbols = pynini.union(*symbols) - dot + accepted_characters = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols) + acceepted_characters_with_dot = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols | dot) + + # email + username = ( + pynutil.insert("username: \"") + + acceepted_characters_with_dot + + pynutil.insert("\"") + + pynini.cross('@', ' ') + ) + domain_graph = accepted_characters + dot + accepted_characters + domain_graph = pynutil.insert("domain: \"") + domain_graph + pynutil.insert("\"") + domain_common_graph = ( + pynutil.insert("domain: \"") + + accepted_characters + + accepted_common_domains + + pynini.closure((accepted_symbols | dot) + pynini.closure(accepted_characters, 1), 0, 1) + + pynutil.insert("\"") + ) + graph = (username + domain_graph) | domain_common_graph + + # url + protocol_start = pynini.accep("https://") | pynini.accep("http://") + protocol_end = ( + pynini.accep("www.") + if deterministic + else pynini.accep("www.") | pynini.cross("www.", "doble ve doble ve doble ve.") + ) + protocol = protocol_start | protocol_end | (protocol_start + protocol_end) + protocol = pynutil.insert("protocol: \"") + protocol + pynutil.insert("\"") + graph |= protocol + insert_space + (domain_graph | domain_common_graph) + self.graph = graph + + final_graph = self.add_tokens(self.graph + pynutil.insert(" preserve_order: true")) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/fraction.py b/nemo_text_processing/text_normalization/es/taggers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/fraction.py @@ -0,0 +1,124 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + ordinal_exceptions = pynini.string_file(get_abs_path("data/fractions/ordinal_exceptions.tsv")) + higher_powers_of_ten = pynini.string_file(get_abs_path("data/fractions/powers_of_ten.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + ordinal_exceptions = None + higher_powers_of_ten = None + + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for classifying fraction + "23 4/5" -> + tokens { fraction { integer: "veintitrรฉs" numerator: "cuatro" denominator: "quinto" mophosyntactic_features: "ordinal" } } + + Args: + cardinal: CardinalFst + ordinal: OrdinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, ordinal: GraphFst, deterministic: bool = True): + super().__init__(name="fraction", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + ordinal_graph = ordinal.graph + + # 2-10 are all ordinals + three_to_ten = pynini.string_map(["2", "3", "4", "5", "6", "7", "8", "9", "10",]) + block_three_to_ten = pynutil.delete(three_to_ten) # To block cardinal productions + if not deterministic: # Multiples of tens are sometimes rendered as ordinals + three_to_ten |= pynini.string_map(["20", "30", "40", "50", "60", "70", "80", "90",]) + graph_three_to_ten = three_to_ten @ ordinal_graph + graph_three_to_ten @= pynini.cdrewrite(ordinal_exceptions, "", "", NEMO_SIGMA) + + # Higher powers of tens (and multiples) are converted to ordinals. + hundreds = pynini.string_map(["100", "200", "300", "400", "500", "600", "700", "800", "900",]) + graph_hundreds = hundreds @ ordinal_graph + + multiples_of_thousand = ordinal.multiples_of_thousand # So we can have X milรฉsimos + + graph_higher_powers_of_ten = ( + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + + pynini.closure("mil ", 0, 1) + + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + ) # x millones / x mil millones / x mil z millones + graph_higher_powers_of_ten += higher_powers_of_ten + graph_higher_powers_of_ten = cardinal_graph @ graph_higher_powers_of_ten + graph_higher_powers_of_ten @= pynini.cdrewrite( + pynutil.delete("un "), pynini.accep("[BOS]"), pynini.project(higher_powers_of_ten, "output"), NEMO_SIGMA + ) # we drop 'un' from these ordinals (millionths, not one-millionths) + + graph_higher_powers_of_ten = multiples_of_thousand | graph_hundreds | graph_higher_powers_of_ten + block_higher_powers_of_ten = pynutil.delete( + pynini.project(graph_higher_powers_of_ten, "input") + ) # For cardinal graph + + graph_fractions_ordinals = graph_higher_powers_of_ten | graph_three_to_ten + graph_fractions_ordinals += pynutil.insert( + "\" morphosyntactic_features: \"ordinal\"" + ) # We note the root for processing later + + # Blocking the digits and hundreds from Cardinal graph + graph_fractions_cardinals = pynini.cdrewrite( + block_three_to_ten | block_higher_powers_of_ten, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fractions_cardinals @= NEMO_CHAR.plus @ pynini.cdrewrite( + pynutil.delete("0"), pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) # Empty characters become '0' for NEMO_CHAR fst, so ned to block + graph_fractions_cardinals @= cardinal_graph + graph_fractions_cardinals += pynutil.insert( + "\" morphosyntactic_features: \"add_root\"" + ) # blocking these entries to reduce erroneous possibilities in debugging + + if deterministic: + graph_fractions_cardinals = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ graph_fractions_cardinals + ) # Past hundreds the conventional scheme can be hard to read. For determinism we stop here + + graph_denominator = pynini.union( + graph_fractions_ordinals, + graph_fractions_cardinals, + pynutil.add_weight(cardinal_graph + pynutil.insert("\""), 0.001), + ) # Last form is simply recording the cardinal. Weighting so last resort + + integer = pynutil.insert("integer_part: \"") + cardinal_graph + pynutil.insert("\"") + NEMO_SPACE + numerator = ( + pynutil.insert("numerator: \"") + cardinal_graph + (pynini.cross("/", "\" ") | pynini.cross(" / ", "\" ")) + ) + denominator = pynutil.insert("denominator: \"") + graph_denominator + + self.graph = pynini.closure(integer, 0, 1) + numerator + denominator + + final_graph = self.add_tokens(self.graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/measure.py b/nemo_text_processing/text_normalization/es/taggers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/measure.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_NON_BREAKING_SPACE, + NEMO_SPACE, + GraphFst, + convert_space, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit = pynini.string_file(get_abs_path("data/measures/measurements.tsv")) + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit = None + unit_plural_fem = None + unit_plural_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for classifying measure, e.g. + "2,4 g" -> measure { cardinal { integer_part: "dos" fractional_part: "cuatro" units: "gramos" preserve_order: true } } + "1 g" -> measure { cardinal { integer: "un" units: "gramo" preserve_order: true } } + "1 millรณn g" -> measure { cardinal { integer: "un quantity: "millรณn" units: "gramos" preserve_order: true } } + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + This class also converts words containing numbers and letters + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + + + Args: + cardinal: CardinalFst + decimal: DecimalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, fraction: GraphFst, deterministic: bool = True): + super().__init__(name="measure", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + + unit_singular = unit + unit_plural = unit_singular @ (unit_plural_fem | unit_plural_masc) + + graph_unit_singular = convert_space(unit_singular) + graph_unit_plural = convert_space(unit_plural) + + optional_graph_negative = pynini.closure("-", 0, 1) + + graph_unit_denominator = ( + pynini.cross("/", "por") + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_singular + ) + + optional_unit_denominator = pynini.closure( + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_denominator, 0, 1, + ) + + unit_plural = ( + pynutil.insert("units: \"") + + ((graph_unit_plural + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + unit_singular_graph = ( + pynutil.insert("units: \"") + + ((graph_unit_singular + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + subgraph_decimal = decimal.fst + insert_space + pynini.closure(NEMO_SPACE, 0, 1) + unit_plural + + subgraph_cardinal = ( + (optional_graph_negative + (pynini.closure(NEMO_DIGIT) - "1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_plural + ) + + subgraph_cardinal |= ( + (optional_graph_negative + pynini.accep("1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_singular_graph + ) + + subgraph_fraction = fraction.fst + insert_space + pynini.closure(delete_space, 0, 1) + unit_plural + + decimal_times = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_times = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.insert("\" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_dash_alpha = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.delete('-') + + pynutil.insert("\" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + decimal_dash_alpha = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.delete('-') + + pynutil.insert(" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + alpha_dash_cardinal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" cardinal { integer: \"") + + cardinal_graph + + pynutil.insert("\" } preserve_order: true") + ) + + alpha_dash_decimal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } preserve_order: true") + ) + + final_graph = ( + subgraph_decimal + | subgraph_cardinal + | subgraph_fraction + | cardinal_dash_alpha + | alpha_dash_cardinal + | decimal_dash_alpha + | decimal_times + | cardinal_times + | alpha_dash_decimal + ) + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/money.py b/nemo_text_processing/text_normalization/es/taggers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/money.py @@ -0,0 +1,194 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import decimal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + maj_singular_labels = load_labels(get_abs_path("data/money/currency_major.tsv")) + maj_singular = pynini.string_file((get_abs_path("data/money/currency_major.tsv"))) + min_singular = pynini.string_file(get_abs_path("data/money/currency_minor.tsv")) + fem_plural = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc_plural = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + maj_singular_labels = None + min_singular = None + maj_singular = None + fem_plural = None + masc_plural = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for classifying money, e.g. + "โ‚ฌ1" -> money { currency_maj: "euro" integer_part: "un"} + "โ‚ฌ1,000" -> money { currency_maj: "euro" integer_part: "un" } + "โ‚ฌ1,001" -> money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un" } + "ยฃ1,4" -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true } + -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "penique" preserve_order: true } + "0,01 ยฃ" -> money { fractional_part: "un" currency_min: "penique" preserve_order: true } + "0,02 ยฃ" -> money { fractional_part: "dos" currency_min: "peniques" preserve_order: true } + "ยฃ0,01 million" -> money { currency_maj: "libra" integer_part: "cero" fractional_part: "cero un" quantity: "million" } + + Args: + cardinal: CardinalFst + decimal: DecimalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + graph_decimal_final = decimal.final_graph_wo_negative + + maj_singular_graph = maj_singular + min_singular_graph = min_singular + maj_plural_graph = maj_singular @ (fem_plural | masc_plural) + min_plural_graph = min_singular @ (fem_plural | masc_plural) + + graph_maj_singular = pynutil.insert("currency_maj: \"") + maj_singular_graph + pynutil.insert("\"") + graph_maj_plural = pynutil.insert("currency_maj: \"") + maj_plural_graph + pynutil.insert("\"") + + graph_integer_one = pynutil.insert("integer_part: \"") + pynini.cross("1", "un") + pynutil.insert("\"") + + decimal_with_quantity = (NEMO_SIGMA + NEMO_ALPHA) @ graph_decimal_final + + graph_decimal_plural = pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, # 1,05 $ + ) + graph_decimal_plural = ( + (NEMO_SIGMA - "1") + decimal_separator + NEMO_SIGMA + ) @ graph_decimal_plural # Can't have "un euros" + + graph_decimal_singular = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, # 1,05 $ + ) + graph_decimal_singular = (pynini.accep("1") + decimal_separator + NEMO_SIGMA) @ graph_decimal_singular + + graph_decimal = pynini.union( + graph_decimal_singular, + graph_decimal_plural, + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + decimal_with_quantity, + ) + + graph_integer = ( + pynutil.insert("integer_part: \"") + ((NEMO_SIGMA - "1") @ cardinal_graph) + pynutil.insert("\"") + ) + + graph_integer_only = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer_one, + graph_integer_one + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, + ) + graph_integer_only |= pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer, + graph_integer + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, + ) + + graph = graph_integer_only | graph_decimal + + # remove trailing zeros of non zero number in the first 2 digits and fill up to 2 digits + # e.g. 2000 -> 20, 0200->02, 01 -> 01, 10 -> 10 + # not accepted: 002, 00, 0, + two_digits_fractional_part = ( + pynini.closure(NEMO_DIGIT) + (NEMO_DIGIT - "0") + pynini.closure(pynutil.delete("0")) + ) @ ( + (pynutil.delete("0") + (NEMO_DIGIT - "0")) + | ((NEMO_DIGIT - "0") + pynutil.insert("0")) + | ((NEMO_DIGIT - "0") + NEMO_DIGIT) + ) + + graph_min_singular = pynutil.insert("currency_min: \"") + min_singular_graph + pynutil.insert("\"") + graph_min_plural = pynutil.insert("currency_min: \"") + min_plural_graph + pynutil.insert("\"") + + # format ** euro ** cent + decimal_graph_with_minor = None + for curr_symbol, _ in maj_singular_labels: + preserve_order = pynutil.insert(" preserve_order: true") + + integer_plus_maj = pynini.union( + graph_integer + insert_space + pynutil.insert(curr_symbol) @ graph_maj_plural, + graph_integer_one + insert_space + pynutil.insert(curr_symbol) @ graph_maj_singular, + ) + # non zero integer part + integer_plus_maj = (pynini.closure(NEMO_DIGIT) - "0") @ integer_plus_maj + + graph_fractional_one = ( + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ pynini.cross("1", "un") + + pynutil.insert("\"") + ) + + graph_fractional = ( + two_digits_fractional_part @ (pynini.closure(NEMO_DIGIT, 1, 2) - "1") @ cardinal.two_digit_non_zero + ) + graph_fractional = pynutil.insert("fractional_part: \"") + graph_fractional + pynutil.insert("\"") + + fractional_plus_min = pynini.union( + graph_fractional + insert_space + pynutil.insert(curr_symbol) @ graph_min_plural, + graph_fractional_one + insert_space + pynutil.insert(curr_symbol) @ graph_min_singular, + ) + + decimal_graph_with_minor_curr = ( + integer_plus_maj + pynini.cross(decimal_separator, NEMO_SPACE) + fractional_plus_min + ) + decimal_graph_with_minor_curr |= pynutil.add_weight( + integer_plus_maj + + pynini.cross(decimal_separator, NEMO_SPACE) + + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ cardinal.two_digit_non_zero + + pynutil.insert("\""), + weight=0.0001, + ) + + decimal_graph_with_minor_curr |= pynutil.delete("0,") + fractional_plus_min + decimal_graph_with_minor_curr = pynini.union( + pynutil.delete(curr_symbol) + + pynini.closure(delete_space, 0, 1) + + decimal_graph_with_minor_curr + + preserve_order, + decimal_graph_with_minor_curr + + preserve_order + + pynini.closure(delete_space, 0, 1) + + pynutil.delete(curr_symbol), + ) + + decimal_graph_with_minor = ( + decimal_graph_with_minor_curr + if decimal_graph_with_minor is None + else pynini.union(decimal_graph_with_minor, decimal_graph_with_minor_curr) + ) + + final_graph = graph | pynutil.add_weight(decimal_graph_with_minor, -0.001) + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/ordinal.py b/nemo_text_processing/text_normalization/es/taggers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/ordinal.py @@ -0,0 +1,186 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import roman_to_int, strip_accent +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/digit.tsv"))) + teens = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/teen.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/twenties.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/ties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ImportError, ModuleNotFoundError): + digit = None + teens = None + twenties = None + ties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def get_one_to_one_thousand(cardinal: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Produces an acceptor for verbalizations of all numbers from 1 to 1000. Needed for ordinals and fractions. + + Args: + cardinal: CardinalFst + + Returns: + fst: A pynini.FstLike object + """ + numbers = pynini.string_map([str(_) for _ in range(1, 1000)]) @ cardinal + return pynini.project(numbers, "output").optimize() + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for classifying ordinal + "21.ยบ" -> ordinal { integer: "vigรฉsimo primero" morphosyntactic_features: "gender_masc" } + This class converts ordinal up to the millionth (millonรฉsimo) order (exclusive). + + This FST also records the ending of the ordinal (called "morphosyntactic_features"): + either as gender_masc, gender_fem, or apocope. Also introduces plural feature for non-deterministic graphs. + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="ordinal", kind="classify") + cardinal_graph = cardinal.graph + + graph_digit = digit.optimize() + graph_teens = teens.optimize() + graph_ties = ties.optimize() + graph_twenties = twenties.optimize() + graph_hundreds = hundreds.optimize() + + if not deterministic: + # Some alternative derivations + graph_ties = graph_ties | pynini.cross("sesenta", "setuagรฉsimo") + + graph_teens = graph_teens | pynini.cross("once", "decimoprimero") + graph_teens |= pynini.cross("doce", "decimosegundo") + + graph_digit = graph_digit | pynini.cross("nueve", "nono") + graph_digit |= pynini.cross("siete", "sรฉtimo") + + graph_tens_component = ( + graph_teens + | (graph_ties + pynini.closure(pynini.cross(" y ", NEMO_SPACE) + graph_digit, 0, 1)) + | graph_twenties + ) + + graph_hundred_component = pynini.union( + graph_hundreds + pynini.closure(NEMO_SPACE + pynini.union(graph_tens_component, graph_digit), 0, 1), + graph_tens_component, + graph_digit, + ) + + # Need to go up to thousands for fractions + self.one_to_one_thousand = get_one_to_one_thousand(cardinal_graph) + + thousands = pynini.cross("mil", "milรฉsimo") + + graph_thousands = ( + strip_accent(self.one_to_one_thousand) + NEMO_SPACE + thousands + ) # Cardinals become prefix for thousands series. Snce accent on the powers of ten we strip accent from leading words + graph_thousands @= pynini.cdrewrite(delete_space, "", "milรฉsimo", NEMO_SIGMA) # merge as a prefix + graph_thousands |= thousands + + self.multiples_of_thousand = (cardinal_graph @ graph_thousands).optimize() + + if ( + not deterministic + ): # Formally the words preceding the power of ten should be a prefix, but some maintain word boundaries. + graph_thousands |= (self.one_to_one_thousand @ graph_hundred_component) + NEMO_SPACE + thousands + + graph_thousands += pynini.closure(NEMO_SPACE + graph_hundred_component, 0, 1) + + ordinal_graph = graph_thousands | graph_hundred_component + ordinal_graph = cardinal_graph @ ordinal_graph + + if not deterministic: + # The 10's and 20's series can also be two words + split_words = pynini.cross("decimo", "dรฉcimo ") | pynini.cross("vigesimo", "vigรฉsimo ") + split_words = pynini.cdrewrite(split_words, "", NEMO_CHAR, NEMO_SIGMA) + ordinal_graph |= ordinal_graph @ split_words + + # If "octavo" is preceeded by the "o" within string, it needs deletion + ordinal_graph @= pynini.cdrewrite(pynutil.delete("o"), "", "octavo", NEMO_SIGMA) + + self.graph = ordinal_graph.optimize() + + masc = pynini.accep("gender_masc") + fem = pynini.accep("gender_fem") + apocope = pynini.accep("apocope") + + delete_period = pynini.closure(pynutil.delete("."), 0, 1) # Sometimes the period is omitted f + + accept_masc = delete_period + pynini.cross("ยบ", masc) + accep_fem = delete_period + pynini.cross("ยช", fem) + accep_apocope = delete_period + pynini.cross("แต‰สณ", apocope) + + # Managing Romanization + graph_roman = pynutil.insert("integer: \"") + roman_to_int(ordinal_graph) + pynutil.insert("\"") + if not deterministic: + # Introduce plural + plural = pynini.closure(pynutil.insert("/plural"), 0, 1) + accept_masc += plural + accep_fem += plural + + # Romanizations have no morphology marker, so in non-deterministic case we provide option for all + insert_morphology = pynutil.insert(pynini.union(masc, fem)) + plural + insert_morphology |= pynutil.insert(apocope) + insert_morphology = ( + pynutil.insert(" morphosyntactic_features: \"") + insert_morphology + pynutil.insert("\"") + ) + + graph_roman += insert_morphology + + else: + # We assume masculine gender as default + graph_roman += pynutil.insert(" morphosyntactic_features: \"gender_masc\"") + + # Rest of graph + convert_abbreviation = accept_masc | accep_fem | accep_apocope + + graph = ( + pynutil.insert("integer: \"") + + ordinal_graph + + pynutil.insert("\"") + + pynutil.insert(" morphosyntactic_features: \"") + + convert_abbreviation + + pynutil.insert("\"") + ) + graph = pynini.union(graph, graph_roman) + + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/telephone.py b/nemo_text_processing/text_normalization/es/taggers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/telephone.py @@ -0,0 +1,156 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.graph_utils import ones +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + graph_digit = pynini.string_file(get_abs_path("data/numbers/digit.tsv")) + graph_ties = pynini.string_file(get_abs_path("data/numbers/ties.tsv")) + graph_teen = pynini.string_file(get_abs_path("data/numbers/teen.tsv")) + graph_twenties = pynini.string_file(get_abs_path("data/numbers/twenties.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + graph_digit = None + graph_ties = None + graph_teen = None + graph_twenties = None + + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for classifying telephone numbers, e.g. + 123-123-5678 -> { number_part: "uno dos tres uno dos tres cinco seis siete ocho" }. + In Spanish, digits are generally read individually, or as 2-digit numbers, + eg. "123" = "uno dos tres", + "1234" = "doce treinta y cuatro". + This will verbalize sequences of 10 (3+3+4 e.g. 123-456-7890). + 9 (3+3+3 e.g. 123-456-789) and 8 (4+4 e.g. 1234-5678) digits. + + (we ignore more complicated cases such as "doscientos y dos" or "tres nueves"). + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="classify") + + # create `single_digits` and `double_digits` graphs as these will be + # the building blocks of possible telephone numbers + single_digits = pynini.invert(graph_digit).optimize() | pynini.cross("0", "cero") + + double_digits = pynini.union( + graph_twenties, + graph_teen, + (graph_ties + pynutil.delete("0")), + (graph_ties + insert_space + pynutil.insert("y") + insert_space + graph_digit), + ) + double_digits = pynini.invert(double_digits) + + # define `ten_digit_graph`, `nine_digit_graph`, `eight_digit_graph` + # which produces telephone numbers spoken (1) only with single digits, + # or (2) spoken with double digits (and sometimes single digits) + + # 10-digit option (1): all single digits + ten_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + # 9-digit option (1): all single digits + nine_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 2, 2) + + single_digits + ) + + # 8-digit option (1): all single digits + eight_digit_graph = ( + pynini.closure(single_digits + insert_space, 4, 4) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + if not deterministic: + # 10-digit option (2): (1+2) + (1+2) + (2+2) digits + ten_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + # 9-digit option (2): (1+2) + (1+2) + (1+2) digits + nine_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + ) + + # 8-digit option (2): (2+2) + (2+2) digits + eight_digit_graph |= ( + double_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + number_part = pynini.union(ten_digit_graph, nine_digit_graph, eight_digit_graph) + number_part @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", "", NEMO_SIGMA) + + number_part = pynutil.insert("number_part: \"") + number_part + pynutil.insert("\"") + + graph = number_part + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/time.py b/nemo_text_processing/text_normalization/es/taggers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/time.py @@ -0,0 +1,218 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + time_zone_graph = pynini.string_file(get_abs_path("data/time/time_zone.tsv")) + suffix = pynini.string_file(get_abs_path("data/time/time_suffix.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + time_zone_graph = None + suffix = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for classifying time, e.g. + "02:15 est" -> time { hours: "dos" minutes: "quince" zone: "e s t"} + "2 h" -> time { hours: "dos" } + "9 h" -> time { hours: "nueve" } + "02:15:10 h" -> time { hours: "dos" minutes: "quince" seconds: "diez"} + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="time", kind="classify", deterministic=deterministic) + + delete_time_delimiter = pynutil.delete(pynini.union(".", ":")) + + one = pynini.string_map([("un", "una"), ("รบn", "una")]) + change_one = pynini.cdrewrite(one, "", "", NEMO_SIGMA) + cardinal_graph = cardinal.graph @ change_one + + day_suffix = pynutil.insert("suffix: \"") + suffix + pynutil.insert("\"") + day_suffix = delete_space + insert_space + day_suffix + + delete_hora_suffix = delete_space + insert_space + pynutil.delete("h") + delete_minute_suffix = delete_space + insert_space + pynutil.delete("min") + delete_second_suffix = delete_space + insert_space + pynutil.delete("s") + + labels_hour_24 = [ + str(x) for x in range(0, 25) + ] # Can see both systems. Twelve hour requires am/pm for ambiguity resolution + labels_hour_12 = [str(x) for x in range(1, 13)] + labels_minute_single = [str(x) for x in range(1, 10)] + labels_minute_double = [str(x) for x in range(10, 60)] + + delete_leading_zero_to_double_digit = ( + pynini.closure(pynutil.delete("0") | (NEMO_DIGIT - "0"), 0, 1) + NEMO_DIGIT + ) + + graph_24 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_24) + ) + graph_12 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_12) + ) + + graph_hour_24 = graph_24 @ cardinal_graph + graph_hour_12 = graph_12 @ cardinal_graph + + graph_minute_single = delete_leading_zero_to_double_digit @ pynini.union(*labels_minute_single) + graph_minute_double = pynini.union(*labels_minute_double) + + graph_minute = pynini.union(graph_minute_single, graph_minute_double) @ cardinal_graph + + final_graph_hour_only_24 = ( + pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + delete_hora_suffix + ) + final_graph_hour_only_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + day_suffix + + final_graph_hour_24 = pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + final_graph_hour_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + + final_graph_minute = pynutil.insert("minutes: \"") + graph_minute + pynutil.insert("\"") + final_graph_second = pynutil.insert("seconds: \"") + graph_minute + pynutil.insert("\"") + final_time_zone_optional = pynini.closure( + delete_space + insert_space + pynutil.insert("zone: \"") + time_zone_graph + pynutil.insert("\""), 0, 1, + ) + + # 02.30 h + graph_hm = ( + final_graph_hour_24 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 h + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_24 + + delete_hora_suffix + + delete_space + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + delete_minute_suffix + + pynini.closure( + delete_space + + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)) + + delete_second_suffix, + 0, + 1, + ) # For seconds + + final_time_zone_optional + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_12 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 a. m. + + day_suffix + + final_time_zone_optional + ) + + graph_h = ( + pynini.union(final_graph_hour_only_24, final_graph_hour_only_12) + final_time_zone_optional + ) # Should always have a time indicator, else we'll pass to cardinals + + if not deterministic: + # This includes alternate vocalization (hour menos min, min para hour), here we shift the times and indicate a `style` tag + hour_shift_24 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_24.tsv"))) + hour_shift_12 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_12.tsv"))) + minute_shift = pynini.string_file(get_abs_path("data/time/minute_to.tsv")) + + graph_hour_to_24 = graph_24 @ hour_shift_24 @ cardinal_graph + graph_hour_to_12 = graph_12 @ hour_shift_12 @ cardinal_graph + + graph_minute_to = pynini.union(graph_minute_single, graph_minute_double) @ minute_shift @ cardinal_graph + + final_graph_hour_to_24 = pynutil.insert("hours: \"") + graph_hour_to_24 + pynutil.insert("\"") + final_graph_hour_to_12 = pynutil.insert("hours: \"") + graph_hour_to_12 + pynutil.insert("\"") + + final_graph_minute_to = pynutil.insert("minutes: \"") + graph_minute_to + pynutil.insert("\"") + + graph_menos = pynutil.insert(" style: \"1\"") + graph_para = pynutil.insert(" style: \"2\"") + + final_graph_style = graph_menos | graph_para + + # 02.30 h (omitting seconds since a bit awkward) + graph_hm |= ( + final_graph_hour_to_24 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + + final_graph_style + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_to_24 + + delete_hora_suffix + + delete_space + + insert_space + + final_graph_minute_to + + delete_minute_suffix + + final_time_zone_optional + + final_graph_style + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_to_12 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + day_suffix + + final_time_zone_optional + + final_graph_style + ) + + final_graph = graph_hm | graph_h + if deterministic: + final_graph = final_graph + pynutil.insert(" preserve_order: true") + final_graph = final_graph.optimize() + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py @@ -0,0 +1,157 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os + +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_space, + generator_main, +) +from nemo_text_processing.text_normalization.en.taggers.punctuation import PunctuationFst +from nemo_text_processing.text_normalization.es.taggers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.taggers.date import DateFst +from nemo_text_processing.text_normalization.es.taggers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.taggers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.taggers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.taggers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.taggers.money import MoneyFst +from nemo_text_processing.text_normalization.es.taggers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.taggers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.taggers.time import TimeFst +from nemo_text_processing.text_normalization.es.taggers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.taggers.word import WordFst + +from nemo.utils import logging + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class ClassifyFst(GraphFst): + """ + Final class that composes all other classification grammars. This class can process an entire sentence, that is lower cased. + For deployment, this grammar will be compiled and exported to OpenFst Finate State aRchive (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. + overwrite_cache: set to True to overwrite .far files + whitelist: path to a file with whitelist replacements + """ + + def __init__( + self, + input_case: str, + deterministic: bool = False, + cache_dir: str = None, + overwrite_cache: bool = False, + whitelist: str = None, + ): + super().__init__(name="tokenize_and_classify", kind="classify", deterministic=deterministic) + far_file = None + if cache_dir is not None and cache_dir != "None": + os.makedirs(cache_dir, exist_ok=True) + whitelist_file = os.path.basename(whitelist) if whitelist else "" + far_file = os.path.join( + cache_dir, f"_{input_case}_es_tn_{deterministic}_deterministic{whitelist_file}.far" + ) + if not overwrite_cache and far_file and os.path.exists(far_file): + self.fst = pynini.Far(far_file, mode="r")["tokenize_and_classify"] + logging.info(f"ClassifyFst.fst was restored from {far_file}.") + else: + logging.info(f"Creating ClassifyFst grammars. This might take some time...") + + self.cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = self.cardinal.fst + + self.ordinal = OrdinalFst(cardinal=self.cardinal, deterministic=deterministic) + ordinal_graph = self.ordinal.fst + + self.decimal = DecimalFst(cardinal=self.cardinal, deterministic=deterministic) + decimal_graph = self.decimal.fst + + self.fraction = FractionFst(cardinal=self.cardinal, ordinal=self.ordinal, deterministic=deterministic) + fraction_graph = self.fraction.fst + self.measure = MeasureFst( + cardinal=self.cardinal, decimal=self.decimal, fraction=self.fraction, deterministic=deterministic + ) + measure_graph = self.measure.fst + self.date = DateFst(cardinal=self.cardinal, deterministic=deterministic) + date_graph = self.date.fst + word_graph = WordFst(deterministic=deterministic).fst + self.time = TimeFst(self.cardinal, deterministic=deterministic) + time_graph = self.time.fst + self.telephone = TelephoneFst(deterministic=deterministic) + telephone_graph = self.telephone.fst + self.electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = self.electronic.fst + self.money = MoneyFst(cardinal=self.cardinal, decimal=self.decimal, deterministic=deterministic) + money_graph = self.money.fst + self.whitelist = WhiteListFst(input_case=input_case, deterministic=deterministic, input_file=whitelist) + whitelist_graph = self.whitelist.fst + punct_graph = PunctuationFst(deterministic=deterministic).fst + + classify = ( + pynutil.add_weight(whitelist_graph, 1.01) + | pynutil.add_weight(time_graph, 1.09) + | pynutil.add_weight(measure_graph, 1.08) + | pynutil.add_weight(cardinal_graph, 1.1) + | pynutil.add_weight(fraction_graph, 1.09) + | pynutil.add_weight(date_graph, 1.1) + | pynutil.add_weight(ordinal_graph, 1.1) + | pynutil.add_weight(decimal_graph, 1.1) + | pynutil.add_weight(money_graph, 1.1) + | pynutil.add_weight(telephone_graph, 1.1) + | pynutil.add_weight(electronic_graph, 1.1) + | pynutil.add_weight(word_graph, 200) + ) + punct = pynutil.insert("tokens { ") + pynutil.add_weight(punct_graph, weight=2.1) + pynutil.insert(" }") + punct = pynini.closure( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct), + 1, + ) + token = pynutil.insert("tokens { ") + classify + pynutil.insert(" }") + token_plus_punct = ( + pynini.closure(punct + pynutil.insert(" ")) + token + pynini.closure(pynutil.insert(" ") + punct) + ) + + graph = token_plus_punct + pynini.closure( + ( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct + pynutil.insert(" ")) + ) + + token_plus_punct + ) + + graph = delete_space + graph + delete_space + graph |= punct + + self.fst = graph.optimize() + + if far_file: + generator_main(far_file, {"tokenize_and_classify": self.fst}) + logging.info(f"ClassifyFst grammars are saved to {far_file}.") diff --git a/nemo_text_processing/text_normalization/es/taggers/whitelist.py b/nemo_text_processing/text_normalization/es/taggers/whitelist.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/whitelist.py @@ -0,0 +1,69 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, convert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WhiteListFst(GraphFst): + """ + Finite state transducer for classifying whitelist, e.g. + "sr." -> tokens { name: "seรฑor" } + This class has highest priority among all classifier grammars. Whitelisted tokens are defined and loaded from "data/whitelist.tsv". + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + input_file: path to a file with whitelist replacements + """ + + def __init__(self, input_case: str, deterministic: bool = True, input_file: str = None): + super().__init__(name="whitelist", kind="classify", deterministic=deterministic) + + def _get_whitelist_graph(input_case, file): + whitelist = load_labels(file) + if input_case == "lower_cased": + whitelist = [[x[0].lower()] + x[1:] for x in whitelist] + graph = pynini.string_map(whitelist) + return graph + + graph = _get_whitelist_graph(input_case, get_abs_path("data/whitelist.tsv")) + if not deterministic and input_case != "lower_cased": + graph |= pynutil.add_weight( + _get_whitelist_graph("lower_cased", get_abs_path("data/whitelist.tsv")), weight=0.0001 + ) + + if input_file: + whitelist_provided = _get_whitelist_graph(input_case, input_file) + if not deterministic: + graph |= whitelist_provided + else: + graph = whitelist_provided + + if not deterministic: + units_graph = _get_whitelist_graph(input_case, file=get_abs_path("data/measures/measurements.tsv")) + graph |= units_graph + + self.graph = graph + self.final_graph = convert_space(self.graph).optimize() + self.fst = (pynutil.insert("name: \"") + self.final_graph + pynutil.insert("\"")).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/word.py b/nemo_text_processing/text_normalization/es/taggers/word.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/word.py @@ -0,0 +1,39 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_SPACE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WordFst(GraphFst): + """ + Finite state transducer for classifying word. + e.g. dormir -> tokens { name: "dormir" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="word", kind="classify") + word = pynutil.insert("name: \"") + pynini.closure(NEMO_NOT_SPACE, 1) + pynutil.insert("\"") + self.fst = word.optimize() diff --git a/nemo_text_processing/text_normalization/es/utils.py b/nemo_text_processing/text_normalization/es/utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/utils.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import csv +import os + + +def get_abs_path(rel_path): + """ + Get absolute path + + Args: + rel_path: relative path to this file + + Returns absolute path + """ + return os.path.dirname(os.path.abspath(__file__)) + '/' + rel_path + + +def load_labels(abs_path): + """ + loads relative path file as dictionary + + Args: + abs_path: absolute path + + Returns dictionary of mappings + """ + label_tsv = open(abs_path) + labels = list(csv.reader(label_tsv, delimiter="\t")) + return labels diff --git a/nemo_text_processing/text_normalization/es/verbalizers/__init__.py b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py @@ -0,0 +1,57 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_cardinal_gender, strip_cardinal_apocope + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class CardinalFst(GraphFst): + """ + Finite state transducer for verbalizing cardinals + e.g. cardinal { integer: "dos" } -> "dos" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="verbalize", deterministic=deterministic) + optional_sign = pynini.closure(pynini.cross("negative: \"true\" ", "menos "), 0, 1) + self.optional_sign = optional_sign + + integer = pynini.closure(NEMO_NOT_QUOTE, 1) + self.integer = pynutil.delete(" \"") + integer + pynutil.delete("\"") + + integer = pynutil.delete("integer:") + self.integer + self.numbers = integer + graph = optional_sign + self.numbers + + if not deterministic: + # For alternate renderings + no_adjust = graph + fem_adjust = shift_cardinal_gender(graph) + apocope_adjust = strip_cardinal_apocope(graph) + graph = no_adjust | fem_adjust | apocope_adjust + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/date.py b/nemo_text_processing/text_normalization/es/verbalizers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/date.py @@ -0,0 +1,86 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.taggers.date import articles + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for verbalizing date, e.g. + date { day: "treinta y uno" month: "marzo" year: "dos mil" } -> "treinta y uno de marzo de dos mil" + date { day: "uno" month: "mayo" year: "del mil novecientos noventa" } -> "primero de mayo del mil novecientos noventa" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="date", kind="verbalize", deterministic=deterministic) + + day_cardinal = pynutil.delete("day: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + day = strip_cardinal_apocope(day_cardinal) + + primero = pynini.cdrewrite(pynini.cross("uno", "primero"), "[BOS]", "[EOS]", NEMO_SIGMA) + day = ( + (day @ primero) if deterministic else pynini.union(day, day @ primero) + ) # Primero for first day is traditional, but will vary depending on region + + month = pynutil.delete("month: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + year = ( + pynutil.delete("year: \"") + + articles + + NEMO_SPACE + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # Insert preposition if wasn't originally with the year. This would mean a space was present + year = pynutil.add_weight(year, -0.001) + year |= ( + pynutil.delete("year: \"") + + pynutil.insert("de ") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # day month year + graph_dmy = day + pynini.cross(NEMO_SPACE, " de ") + month + pynini.closure(pynini.accep(" ") + year, 0, 1) + + graph_mdy = month + NEMO_SPACE + day + pynini.closure(NEMO_SPACE + year, 0, 1) + if deterministic: + graph_mdy += pynutil.delete(" preserve_order: true") # Only accepts this if was explicitly passed + + self.graph = graph_dmy | graph_mdy + final_graph = self.graph + delete_preserve_order + + delete_tokens = self.delete_tokens(final_graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/decimals.py b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py @@ -0,0 +1,87 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + decimal { negative: "true" integer_part: "dos" fractional_part: "cuatro cero" quantity: "billones" } -> menos dos coma quatro cero billones + decimal { integer_part: "un" quantity: "billรณn" } -> un billรณn + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + + self.optional_sign = pynini.closure(pynini.cross("negative: \"true\"", "menos ") + delete_space, 0, 1) + self.integer = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + self.fractional_default = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + conjunction = pynutil.insert(" punto ") if LOCALIZATION == "am" else pynutil.insert(" coma ") + if not deterministic: + conjunction |= pynutil.insert(pynini.union(" con ", " y ")) + self.fractional_default |= strip_cardinal_apocope(self.fractional_default) + self.fractional = conjunction + self.fractional_default + + self.quantity = ( + delete_space + + insert_space + + pynutil.delete("quantity: \"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + self.optional_quantity = pynini.closure(self.quantity, 0, 1) + + graph = self.optional_sign + pynini.union( + (self.integer + self.quantity), (self.integer + delete_space + self.fractional + self.optional_quantity) + ) + + self.numbers = graph.optimize() + self.numbers_no_quantity = self.integer + delete_space + self.fractional + self.optional_quantity + + if not deterministic: + graph |= self.optional_sign + ( + shift_cardinal_gender(self.integer + delete_space) + shift_number_gender(self.fractional) + ) + + graph += delete_preserve_order + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/electronic.py b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py @@ -0,0 +1,91 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit_no_zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + graph_symbols = pynini.string_file(get_abs_path("data/electronic/symbols.tsv")) + server_common = pynini.string_file(get_abs_path("data/electronic/server_name.tsv")) + domain_common = pynini.string_file(get_abs_path("data/electronic/domain.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digit_no_zero = None + zero = None + + graph_symbols = None + server_common = None + domain_common = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for verbalizing electronic + e.g. electronic { username: "abc" domain: "hotmail.com" } -> "a b c arroba hotmail punto com" + -> "a b c arroba h o t m a i l punto c o m" + -> "a b c arroba hotmail punto c o m" + -> "a b c at h o t m a i l punto com" + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="verbalize", deterministic=deterministic) + + graph_digit_no_zero = ( + digit_no_zero @ pynini.cdrewrite(pynini.cross("un", "uno"), "", "", NEMO_SIGMA).optimize() + ) + graph_digit = graph_digit_no_zero | zero + + def add_space_after_char(): + return pynini.closure(NEMO_NOT_QUOTE - pynini.accep(" ") + insert_space) + ( + NEMO_NOT_QUOTE - pynini.accep(" ") + ) + + verbalize_characters = pynini.cdrewrite(graph_symbols | graph_digit, "", "", NEMO_SIGMA) + + user_name = pynutil.delete("username: \"") + add_space_after_char() + pynutil.delete("\"") + user_name @= verbalize_characters + + convert_defaults = pynutil.add_weight(NEMO_NOT_QUOTE, weight=0.0001) | domain_common | server_common + domain = convert_defaults + pynini.closure(insert_space + convert_defaults) + domain @= verbalize_characters + + domain = pynutil.delete("domain: \"") + domain + pynutil.delete("\"") + protocol = ( + pynutil.delete("protocol: \"") + + add_space_after_char() @ pynini.cdrewrite(graph_symbols, "", "", NEMO_SIGMA) + + pynutil.delete("\"") + ) + self.graph = (pynini.closure(protocol + pynini.accep(" "), 0, 1) + domain) | ( + user_name + pynini.accep(" ") + pynutil.insert("arroba ") + domain + ) + delete_tokens = self.delete_tokens(self.graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/fraction.py b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_NOT_QUOTE, + NEMO_NOT_SPACE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + accents, + shift_cardinal_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for verbalizing fraction + e.g. tokens { fraction { integer: "treinta y tres" numerator: "cuatro" denominator: "quinto" } } -> + treinta y tres y cuatro quintos + + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="fraction", kind="verbalize", deterministic=deterministic) + + # Derivational strings append 'avo' as a suffix. Adding space for processing aid + fraction_stem = pynutil.insert(" avo") + plural = pynutil.insert("s") + + integer = ( + pynutil.delete("integer_part: \"") + + strip_cardinal_apocope(pynini.closure(NEMO_NOT_QUOTE)) + + pynutil.delete("\"") + ) + + numerator_one = pynutil.delete("numerator: \"") + pynini.accep("un") + pynutil.delete("\" ") + numerator = ( + pynutil.delete("numerator: \"") + + pynini.difference(pynini.closure(NEMO_NOT_QUOTE), "un") + + pynutil.delete("\" ") + ) + + denominator_add_stem = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + + fraction_stem + + pynutil.delete("\" morphosyntactic_features: \"add_root\"") + ) + denominator_ordinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\" morphosyntactic_features: \"ordinal\"") + ) + denominator_cardinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + ) + + denominator_singular = pynini.union(denominator_add_stem, denominator_ordinal) + denominator_plural = denominator_singular + plural + + if not deterministic: + # Occasional exceptions + denominator_singular |= denominator_add_stem @ pynini.string_map( + [("once avo", "undรฉcimo"), ("doce avo", "duodรฉcimo")] + ) + + # Merging operations + merge = pynini.cdrewrite( + pynini.cross(" y ", "i"), "", "", NEMO_SIGMA + ) # The denominator must be a single word, with the conjunction "y" replaced by i + merge @= pynini.cdrewrite(delete_space, "", pynini.difference(NEMO_CHAR, "parte"), NEMO_SIGMA) + + # The merger can produce duplicate vowels. This is not allowed in orthography + delete_duplicates = pynini.string_map([("aa", "a"), ("oo", "o")]) # Removes vowels + delete_duplicates = pynini.cdrewrite(delete_duplicates, "", "", NEMO_SIGMA) + + remove_accents = pynini.cdrewrite( + accents, + pynini.union(NEMO_SPACE, pynini.accep("[BOS]")) + pynini.closure(NEMO_NOT_SPACE), + pynini.closure(NEMO_NOT_SPACE) + pynini.union("avo", "ava", "รฉsimo", "รฉsima"), + NEMO_SIGMA, + ) + merge_into_single_word = merge @ remove_accents @ delete_duplicates + + fraction_default = numerator + delete_space + insert_space + (denominator_plural @ merge_into_single_word) + fraction_with_one = ( + numerator_one + delete_space + insert_space + (denominator_singular @ merge_into_single_word) + ) + + fraction_with_cardinal = strip_cardinal_apocope(numerator | numerator_one) + fraction_with_cardinal += ( + delete_space + pynutil.insert(" sobre ") + strip_cardinal_apocope(denominator_cardinal) + ) + + conjunction = pynutil.insert(" y ") + + if not deterministic: + # There is an alternative rendering where ordinals act as adjectives for 'parte'. This requires use of the feminine + # Other rules will manage use of "un" at end, so just worry about endings + exceptions = pynini.string_map([("tercia", "tercera")]) + apply_exceptions = pynini.cdrewrite(exceptions, "", "", NEMO_SIGMA) + vowel_change = pynini.cdrewrite(pynini.cross("o", "a"), "", pynini.accep("[EOS]"), NEMO_SIGMA) + + denominator_singular_fem = shift_cardinal_gender(denominator_singular) @ vowel_change @ apply_exceptions + denominator_plural_fem = denominator_singular_fem + plural + + numerator_one_fem = shift_cardinal_gender(numerator_one) + numerator_fem = shift_cardinal_gender(numerator) + + fraction_with_cardinal |= ( + (numerator_one_fem | numerator_fem) + + delete_space + + pynutil.insert(" sobre ") + + shift_cardinal_gender(denominator_cardinal) + ) + + # Still need to manage stems + merge_stem = pynini.cdrewrite( + delete_space, "", pynini.union("avo", "ava", "avos", "avas"), NEMO_SIGMA + ) # For managing alternative spacing + merge_stem @= remove_accents @ delete_duplicates + + fraction_with_one_fem = numerator_one_fem + delete_space + insert_space + fraction_with_one_fem += pynini.union( + denominator_singular_fem @ merge_stem, denominator_singular_fem @ merge_into_single_word + ) # Both forms exists + fraction_with_one_fem @= pynini.cdrewrite(pynini.cross("una media", "media"), "", "", NEMO_SIGMA) + fraction_with_one_fem += pynutil.insert(" parte") + + fraction_default_fem = numerator_fem + delete_space + insert_space + fraction_default_fem += pynini.union( + denominator_plural_fem @ merge_stem, denominator_plural_fem @ merge_into_single_word + ) + fraction_default_fem += pynutil.insert(" partes") + + fraction_default |= ( + numerator + delete_space + insert_space + denominator_plural @ merge_stem + ) # Case of no merger + fraction_default |= fraction_default_fem + + fraction_with_one |= numerator_one + delete_space + insert_space + denominator_singular @ merge_stem + fraction_with_one |= fraction_with_one_fem + + # Integers are influenced by dominant noun, need to allow feminine forms as well + integer |= shift_cardinal_gender(integer) + + # Remove 'un medio' + fraction_with_one @= pynini.cdrewrite(pynini.cross("un medio", "medio"), "", "", NEMO_SIGMA) + + integer = pynini.closure(integer + delete_space + conjunction, 0, 1) + + fraction = fraction_with_one | fraction_default | fraction_with_cardinal + + graph = integer + fraction + + self.graph = graph + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/measure.py b/nemo_text_processing/text_normalization/es/verbalizers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/measure.py @@ -0,0 +1,110 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ones, shift_cardinal_gender +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + unit_singular_fem = pynini.project(unit_plural_fem, "input") + unit_singular_masc = pynini.project(unit_plural_masc, "input") + + unit_plural_fem = pynini.project(unit_plural_fem, "output") + unit_plural_masc = pynini.project(unit_plural_masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit_plural_fem = None + unit_plural_masc = None + + unit_singular_fem = None + unit_singular_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for verbalizing measure, e.g. + measure { cardinal { integer: "dos" units: "gramos" } } -> "dos gramos" + measure { cardinal { integer_part: "dos" quantity: "millones" units: "gramos" } } -> "dos millones de gramos" + + Args: + decimal: DecimalFst + cardinal: CardinalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, cardinal: GraphFst, fraction: GraphFst, deterministic: bool): + super().__init__(name="measure", kind="verbalize", deterministic=deterministic) + + graph_decimal = decimal.fst + graph_cardinal = cardinal.fst + graph_fraction = fraction.fst + + unit_masc = (unit_plural_masc | unit_singular_masc) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_masc |= "por" + pynini.closure(NEMO_NOT_QUOTE, 1) + unit_masc = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_masc) + pynutil.delete("\"") + + unit_fem = (unit_plural_fem | unit_singular_fem) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_fem = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_fem) + pynutil.delete("\"") + + graph_masc = (graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_masc + graph_fem = ( + shift_cardinal_gender(graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_fem + ) + graph = graph_masc | graph_fem + + graph = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph + ) # billones de xyz + + graph @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", NEMO_WHITE_SPACE + "por", NEMO_SIGMA) + + # To manage alphanumeric combonations ("a-8, 5x"), we let them use a weighted default path. + alpha_num_unit = pynutil.delete("units: \"") + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + graph_alpha_num = pynini.union( + (graph_cardinal | graph_decimal) + NEMO_SPACE + alpha_num_unit, + alpha_num_unit + delete_extra_space + (graph_cardinal | graph_decimal), + ) + + graph |= pynutil.add_weight(graph_alpha_num, 0.01) + + graph += delete_preserve_order + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/money.py b/nemo_text_processing/text_normalization/es/verbalizers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/money.py @@ -0,0 +1,195 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + fem = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + fem_singular = pynini.project(fem, "input") + masc_singular = pynini.project(masc, "input") + + fem_plural = pynini.project(fem, "output") + masc_plural = pynini.project(masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + fem_plural = None + masc_plural = None + + fem_singular = None + masc_singular = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for verbalizing money, e.g. + money { currency_maj: "euro" integer_part: "un"} -> "un euro" + money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un"} -> "uno coma cero cero uno euros" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true} -> "una libra cuarenta" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "peniques" preserve_order: true} -> "una libra con cuarenta peniques" + money { fractional_part: "un" currency_min: "penique" preserve_order: true} -> "un penique" + + Args: + decimal: GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="verbalize", deterministic=deterministic) + + maj_singular_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + maj_singular_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + maj_plural_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + maj_plural_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + maj_masc = maj_plural_masc | maj_singular_masc # Tagger kept quantity resolution stable + maj_fem = maj_plural_fem | maj_singular_fem + + min_singular_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + min_singular_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + min_plural_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + min_plural_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + min_masc = min_plural_masc | min_singular_masc + min_fem = min_plural_fem | min_singular_fem + + fractional_part = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + integer_part = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + optional_add_and = pynini.closure(pynutil.insert(pynini.union("con ", "y ")), 0, 1) + + # *** currency_maj + graph_integer_masc = integer_part + NEMO_SPACE + maj_masc + graph_integer_fem = shift_cardinal_gender(integer_part) + NEMO_SPACE + maj_fem + graph_integer = graph_integer_fem | graph_integer_masc + + # *** currency_maj + (***) | ((con) *** current_min) + graph_integer_with_minor_masc = ( + integer_part + + NEMO_SPACE + + maj_masc + + NEMO_SPACE + + pynini.union( + optional_add_and + strip_cardinal_apocope(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor_fem = ( + shift_cardinal_gender(integer_part) + + NEMO_SPACE + + maj_fem + + NEMO_SPACE + + pynini.union( + optional_add_and + shift_cardinal_gender(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor = graph_integer_with_minor_fem | graph_integer_with_minor_masc + + # *** coma *** currency_maj + graph_decimal_masc = decimal.numbers + NEMO_SPACE + maj_masc + + # Need to fix some of the inner parts, so don't use decimal here (note: quantities covered by masc) + graph_decimal_fem = ( + pynini.accep("integer_part: \"") + + shift_cardinal_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SPACE + + pynini.accep("fractional_part: \"") + + shift_number_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SIGMA + ) + graph_decimal_fem @= decimal.numbers_no_quantity + graph_decimal_fem += NEMO_SPACE + maj_fem + + graph_decimal = graph_decimal_fem | graph_decimal_masc + graph_decimal = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph_decimal + ) # formally it's millones/billones de *** + + # *** current_min + graph_minor_masc = fractional_part + NEMO_SPACE + min_masc + delete_preserve_order + graph_minor_fem = shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem + delete_preserve_order + graph_minor = graph_minor_fem | graph_minor_masc + + graph = graph_integer | graph_integer_with_minor | graph_decimal | graph_minor + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py @@ -0,0 +1,76 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, NEMO_SIGMA, NEMO_SPACE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_number_gender + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for verbalizing ordinals + e.g. ordinal { integer: "tercer" } } -> "tercero" + -> "tercera" + -> "tercer" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="ordinal", kind="verbalize", deterministic=deterministic) + + graph = pynutil.delete("integer: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + # masculne gender we leave as is + graph_masc = graph + pynutil.delete(" morphosyntactic_features: \"gender_masc") + + # shift gender + graph_fem_ending = graph @ pynini.cdrewrite( + pynini.cross("o", "a"), "", NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fem = shift_number_gender(graph_fem_ending) + pynutil.delete(" morphosyntactic_features: \"gender_fem") + + # Apocope just changes tercero and primero. May occur if someone wrote 11.er (uncommon) + graph_apocope = ( + pynini.cross("tercero", "tercer") + | pynini.cross("primero", "primer") + | pynini.cross("undรฉcimo", "decimoprimer") + ) # In case someone wrote 11.er with deterministic + graph_apocope = (graph @ pynini.cdrewrite(graph_apocope, "", "", NEMO_SIGMA)) + pynutil.delete( + " morphosyntactic_features: \"apocope" + ) + + graph = graph_apocope | graph_masc | graph_fem + + if not deterministic: + # Plural graph + graph_plural = pynini.cdrewrite( + pynutil.insert("s"), pynini.union("o", "a"), NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + + graph |= (graph @ graph_plural) + pynutil.delete("/plural") + + self.graph = graph + pynutil.delete("\"") + + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/telephone.py b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for verbalizing telephone, e.g. + telephone { number_part: "uno dos tres uno dos tres cinco seis siete ocho" } + -> uno dos tres uno dos tres cinco seis siete ocho + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="verbalize") + + number_part = pynutil.delete("number_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + delete_tokens = self.delete_tokens(number_part) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/time.py b/nemo_text_processing/text_normalization/es/verbalizers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/time.py @@ -0,0 +1,269 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + alt_minutes = pynini.string_file(get_abs_path("data/time/alt_minutes.tsv")) + + morning_times = pynini.string_file(get_abs_path("data/time/morning_times.tsv")) + afternoon_times = pynini.string_file(get_abs_path("data/time/afternoon_times.tsv")) + evening_times = pynini.string_file(get_abs_path("data/time/evening_times.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + alt_minutes = None + + morning_times = None + afternoon_times = None + evening_times = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for verbalizing time, e.g. + time { hours: "doce" minutes: "media" suffix: "a m" } -> doce y media de la noche + time { hours: "doce" } -> twelve o'clock + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="time", kind="verbalize", deterministic=deterministic) + + change_minutes = pynini.cdrewrite(alt_minutes, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA) + + morning_phrases = pynini.cross("am", "de la maรฑana") + afternoon_phrases = pynini.cross("pm", "de la tarde") + evening_phrases = pynini.cross("pm", "de la noche") + + # For the 12's + mid_times = pynini.accep("doce") + mid_phrases = ( + pynini.string_map([("pm", "del mediodรญa"), ("am", "de la noche")]) + if deterministic + else pynini.string_map( + [ + ("pm", "de la maรฑana"), + ("pm", "del dรญa"), + ("pm", "del mediodรญa"), + ("am", "de la noche"), + ("am", "de la medianoche"), + ] + ) + ) + + hour = ( + pynutil.delete("hours:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = ( + pynutil.delete("minutes:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = (minute @ change_minutes) if deterministic else pynini.union(minute, minute @ change_minutes) + + suffix = ( + pynutil.delete("suffix:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + zone = ( + pynutil.delete("zone:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + optional_zone = pynini.closure(delete_space + insert_space + zone, 0, 1) + second = ( + pynutil.delete("seconds:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + graph_hms = ( + hour + + pynutil.insert(" horas ") + + delete_space + + minute + + pynutil.insert(" minutos y ") + + delete_space + + second + + pynutil.insert(" segundos") + ) + + graph_hm = hour + delete_space + pynutil.insert(" y ") + minute + graph_hm |= pynini.union( + (hour @ morning_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases), + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases), + (hour @ evening_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases), + (hour @ mid_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases), + ) + + graph_h = pynini.union( + hour, + (hour @ morning_times) + delete_space + insert_space + (suffix @ morning_phrases), + (hour @ afternoon_times) + delete_space + insert_space + (suffix @ afternoon_phrases), + (hour @ evening_times) + delete_space + insert_space + (suffix @ evening_phrases), + (hour @ mid_times) + delete_space + insert_space + (suffix @ mid_phrases), + ) + + graph = (graph_hms | graph_hm | graph_h) + optional_zone + + if not deterministic: + graph_style_1 = pynutil.delete(" style: \"1\"") + graph_style_2 = pynutil.delete(" style: \"2\"") + + graph_menos = hour + delete_space + pynutil.insert(" menos ") + minute + graph_style_1 + graph_menos |= ( + (hour @ morning_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ evening_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ mid_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_1 + ) + graph_menos += optional_zone + + graph_para = minute + pynutil.insert(" para las ") + delete_space + hour + graph_style_2 + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ morning_times) + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ afternoon_times) + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ evening_times) + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ mid_times) + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_2 + ) + graph_para += optional_zone + graph_para @= pynini.cdrewrite( + pynini.cross(" las ", " la "), "para", "una", NEMO_SIGMA + ) # Need agreement with one + + graph |= graph_menos | graph_para + delete_tokens = self.delete_tokens(graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py @@ -0,0 +1,73 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst +from nemo_text_processing.text_normalization.en.verbalizers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.verbalizers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.verbalizers.date import DateFst +from nemo_text_processing.text_normalization.es.verbalizers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.verbalizers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.verbalizers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.verbalizers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.verbalizers.money import MoneyFst +from nemo_text_processing.text_normalization.es.verbalizers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.verbalizers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.verbalizers.time import TimeFst + + +class VerbalizeFst(GraphFst): + """ + Composes other verbalizer grammars. + For deployment, this grammar will be compiled and exported to OpenFst Finate State Archiv (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize", kind="verbalize", deterministic=deterministic) + cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = cardinal.fst + ordinal = OrdinalFst(deterministic=deterministic) + ordinal_graph = ordinal.fst + decimal = DecimalFst(deterministic=deterministic) + decimal_graph = decimal.fst + fraction = FractionFst(deterministic=deterministic) + fraction_graph = fraction.fst + date = DateFst(deterministic=deterministic) + date_graph = date.fst + measure = MeasureFst(cardinal=cardinal, decimal=decimal, fraction=fraction, deterministic=deterministic) + measure_graph = measure.fst + electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = electronic.fst + whitelist_graph = WhiteListFst(deterministic=deterministic).fst + money_graph = MoneyFst(decimal=decimal, deterministic=deterministic).fst + telephone_graph = TelephoneFst(deterministic=deterministic).fst + time_graph = TimeFst(deterministic=deterministic).fst + + graph = ( + cardinal_graph + | measure_graph + | decimal_graph + | ordinal_graph + | date_graph + | electronic_graph + | money_graph + | fraction_graph + | whitelist_graph + | telephone_graph + | time_graph + ) + self.fst = graph diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py @@ -0,0 +1,52 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, delete_extra_space, delete_space +from nemo_text_processing.text_normalization.en.verbalizers.word import WordFst +from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class VerbalizeFinalFst(GraphFst): + """ + Finite state transducer that verbalizes an entire sentence + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize_final", kind="verbalize", deterministic=deterministic) + verbalize = VerbalizeFst(deterministic=deterministic).fst + word = WordFst(deterministic=deterministic).fst + types = verbalize | word + graph = ( + pynutil.delete("tokens") + + delete_space + + pynutil.delete("{") + + delete_space + + types + + delete_space + + pynutil.delete("}") + ) + graph = delete_space + pynini.closure(graph + delete_extra_space) + graph + delete_space + self.fst = graph diff --git a/nemo_text_processing/text_normalization/normalize.py b/nemo_text_processing/text_normalization/normalize.py --- a/nemo_text_processing/text_normalization/normalize.py +++ b/nemo_text_processing/text_normalization/normalize.py @@ -46,8 +46,8 @@ class Normalizer: """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -83,10 +83,11 @@ def __init__( from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst elif lang == 'de': - # Ru TN only support non-deterministic cases and produces multiple normalization options - # use normalize_with_audio.py from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst + elif lang == 'es': + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ClassifyFst + from nemo_text_processing.text_normalization.es.verbalizers.verbalize_final import VerbalizeFinalFst self.tagger = ClassifyFst( input_case=input_case, deterministic=deterministic, @@ -106,7 +107,7 @@ def __init__( def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: """ - NeMo text normalizer + NeMo text normalizer Args: texts: list of input strings @@ -357,7 +358,7 @@ def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: def parse_args(): parser = ArgumentParser() parser.add_argument("input_string", help="input string", type=str) - parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) + parser.add_argument("--language", help="language", choices=["en", "de", "es"], default="en", type=str) parser.add_argument( "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) diff --git a/nemo_text_processing/text_normalization/normalize_with_audio.py b/nemo_text_processing/text_normalization/normalize_with_audio.py --- a/nemo_text_processing/text_normalization/normalize_with_audio.py +++ b/nemo_text_processing/text_normalization/normalize_with_audio.py @@ -55,15 +55,15 @@ "audio_data" - path to the audio file "text" - raw text "pred_text" - ASR model prediction - + See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions - + When the manifest is ready, run: python normalize_with_audio.py \ --audio_data PATH/TO/MANIFEST.JSON \ - --language en - - + --language en + + To run with a single audio file, specify path to audio and text with: python normalize_with_audio.py \ --audio_data PATH/TO/AUDIO.WAV \ @@ -71,18 +71,18 @@ --text raw text OR PATH/TO/.TXT/FILE --model QuartzNet15x5Base-En \ --verbose - + To see possible normalization options for a text input without an audio file (could be used for debugging), run: python python normalize_with_audio.py --text "RAW TEXT" - + Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference """ class NormalizerWithAudio(Normalizer): """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -282,7 +282,7 @@ def parse_args(): "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) parser.add_argument( - "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str + "--language", help="Select target language", choices=["en", "ru", "de", "es"], default="en", type=str ) parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") parser.add_argument( diff --git a/tools/text_processing_deployment/pynini_export.py b/tools/text_processing_deployment/pynini_export.py --- a/tools/text_processing_deployment/pynini_export.py +++ b/tools/text_processing_deployment/pynini_export.py @@ -67,7 +67,7 @@ def tn_grammars(**kwargs): def export_grammars(output_dir, grammars): """ - Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. + Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. Args: output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. @@ -109,7 +109,7 @@ def parse_args(): if __name__ == '__main__': args = parse_args() - if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': + if args.language in ['ru', 'fr', 'vi'] and args.grammars == 'tn_grammars': raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') if args.language == 'en': @@ -148,6 +148,10 @@ def parse_args(): from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ( + ClassifyFst as TNClassifyFst, + ) + from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst elif args.language == 'fr': from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, </patch> </s> </patch>
diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt @@ -0,0 +1,86 @@ +1~un +2~dos +3~tres +4~cuatro +5~cinco +6~seis +7~siete +8~ocho +9~nueve +10~diez +11~once +12~doce +13~trece +14~catorce +15~quince +16~diecisรฉis +17~diecisiete +18~dieciocho +19~diecinueve +20~veinte +21~veintiรบn +22~veintidรณs +23~veintitrรฉs +24~veinticuatro +25~veinticinco +26~veintisรฉis +27~veintisiete +28~veintiocho +29~veintinueve +30~treinta +31~treinta y un +40~cuarenta +41~cuarenta y un +50~cincuenta +51~cincuenta y un +60~sesenta +70~setenta +80~ochenta +90~noventa +100~cien +101~ciento un +120~ciento veinte +121~ciento veintiรบn +130~ciento treinta +131~ciento treinta y un +200~doscientos +201~doscientos un +300~trescientos +301~trescientos un +1000~mil +1 000~mil +1.000~mil +1001~mil un +1010~mil diez +1020~mil veinte +1021~mil veintiรบn +1100~mil cien +1101~mil ciento un +1110~mil ciento diez +1111~mil ciento once +1234~mil doscientos treinta y cuatro +2000~dos mil +2001~dos mil un +2010~dos mil diez +2020~dos mil veinte +2100~dos mil cien +2101~dos mil ciento un +2110~dos mil ciento diez +2111~dos mil ciento once +2222~dos mil doscientos veintidรณs +10000~diez mil +10 000~diez mil +10.000~diez mil +100000~cien mil +100 000~cien mil +100.000~cien mil +1 000 000~un millรณn +1.000.000~un millรณn +1 234 568~un millรณn doscientos treinta y cuatro mil quinientos sesenta y ocho +2.000.000~dos millones +1.000.000.000~mil millones +2.000.000.000~dos mil millones +3 000 000 000 000~tres billones +3.000.000.000.000~tres billones +100 000 000 000 000 000 000 000~cien mil trillones +100 000 000 000 000 000 000 001~cien mil trillones un diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt @@ -0,0 +1,13 @@ +1 enero~primero de enero +5 febrero~cinco de febrero +20 de marzo~veinte de marzo +abril 30~treinta de abril +31 marzo~treinta y uno de marzo +10 mayo 1990~diez de mayo de mil novecientos noventa +junio 11 2000~once de junio de dos mil +30 julio del 2020~treinta de julio del dos mil veinte +30-2-1990~treinta de febrero de mil novecientos noventa +30/2/1990~treinta de febrero de mil novecientos noventa +30.2.1990~treinta de febrero de mil novecientos noventa +1990-2-30~treinta de febrero de mil novecientos noventa +1990-02-30~treinta de febrero de mil novecientos noventa \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt @@ -0,0 +1,27 @@ +0,1~cero coma un +0,01~cero coma cero un +0,010~cero coma cero uno cero +1,0101~uno coma cero uno cero un +0,0~cero coma cero +1,0~uno coma cero +1,00~uno coma cero cero +1,1~uno coma un +233,32~doscientos treinta y tres coma treinta y dos +32,22 millones~treinta y dos coma veintidรณs millones +320 320,22 millones~trescientos veinte mil trescientos veinte coma veintidรณs millones +5.002,232~cinco mil dos coma doscientos treinta y dos +3,2 trillones~tres coma dos trillones +3 millones~tres millones +3 000 millones~tres mil millones +3000 millones~tres mil millones +3.000 millones~tres mil millones +3.001 millones~tres mil un millones +1 millรณn~un millรณn +1 000 millones~mil millones +1000 millones~mil millones +1.000 millones~mil millones +2,33302 millones~dos coma tres tres tres cero dos millones +1,5332 millรณn~uno coma cinco tres tres dos millรณn +1,53322 millรณn~uno coma cinco tres tres dos dos millรณn +1,53321 millรณn~uno coma cinco tres tres dos un millรณn +101,010101 millones~ciento uno coma cero uno cero uno cero un millones \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt @@ -0,0 +1,12 @@ +a.bc@gmail.com~a punto b c arroba gmail punto com +cdf@abc.edu~c d f arroba a b c punto e d u +abc@gmail.abc~a b c arroba gmail punto a b c +abc@abc.com~a b c arroba a b c punto com +asdf123@abc.com~a s d f uno dos tres arroba a b c punto com +a1b2@abc.com~a uno b dos arroba a b c punto com +ab3.sdd.3@gmail.com~a b tres punto s d d punto tres arroba gmail punto com +https://www.nvidia.com~h t t p s dos puntos barra barra w w w punto nvidia punto com +www.nvidia.com~w w w punto nvidia punto com +www.abc.es/efg~w w w punto a b c punto es barra e f g +www.abc.es~w w w punto a b c punto es +http://www.ourdailynews.com.sm~h t t p dos puntos barra barra w w w punto o u r d a i l y n e w s punto com punto s m \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt @@ -0,0 +1,76 @@ +1/2~medio +1 1/2~uno y medio +3/2~tres medios +1 3/2~uno y tres medios +1/3~un tercio +2/3~dos tercios +1/4~un cuarto +2/4~dos cuartos +1/5~un quinto +2/5~dos quintos +1/6~un sexto +2/6~dos sextos +1/7~un sรฉptimo +2/7~dos sรฉptimos +1/8~un octavo +2/8~dos octavos +1/9~un noveno +2/9~dos novenos +1/10~un dรฉcimo +2/10~dos dรฉcimos +1/11~un onceavo +1/12~un doceavo +1/13~un treceavo +1/14~un catorceavo +1/15~un quinceavo +1/16~un dieciseisavo +1/17~un diecisieteavo +1/18~un dieciochoavo +1/19~un diecinueveavo +1/20~un veinteavo +1/21~un veintiunavo +1/22~un veintidosavo +1/30~un treintavo +1/31~un treintaiunavo +1/40~un cuarentavo +1/41~un cuarentaiunavo +1/50~un cincuentavo +1/60~un sesentavo +1/70~un setentavo +1/80~un ochentavo +1/90~un noventavo +1/100~un centรฉsimo +2/100~dos centรฉsimos +1 2/100~uno y dos centรฉsimos +1/101~uno sobre ciento uno +1/110~uno sobre ciento diez +1/111~uno sobre ciento once +1/112~uno sobre ciento doce +1/123~uno sobre ciento veintitrรฉs +1/134~uno sobre ciento treinta y cuatro +1/200~un ducentรฉsimo +1/201~uno sobre doscientos uno +1/234~uno sobre doscientos treinta y cuatro +1/300~un tricentรฉsimo +1/345~uno sobre trescientos cuarenta y cinco +1/400~un cuadringentรฉsimo +1/456~uno sobre cuatrocientos cincuenta y seis +1/500~un quingentรฉsimo +1/600~un sexcentรฉsimo +1/700~un septingentรฉsimo +1/800~un octingentรฉsimo +1/900~un noningentรฉsimo +1/1000~un milรฉsimo +2/1000~dos milรฉsimos +1 2/1000~uno y dos milรฉsimos +1/1001~uno sobre mil uno +1/1100~uno sobre mil cien +1/1200~uno sobre mil doscientos +1/1234~uno sobre mil doscientos treinta y cuatro +1/2000~un dosmilรฉsimo +1/5000~un cincomilรฉsimo +1/10000~un diezmilรฉsimo +1/100.000~un cienmilรฉsimo +1/1.000.000~un millonรฉsimo +1/100.000.000~un cienmillonรฉsimo +1/1.200.000.000~un mildoscientosmillonรฉsimo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt @@ -0,0 +1,17 @@ +1,2-a~uno coma dos a +a-5~a cinco +200 m~doscientos metros +3 h~tres horas +1 h~una hora +245 mph~doscientas cuarenta y cinco millas por hora +2 kg~dos kilogramos +60,2400 kg~sesenta coma dos cuatro cero cero kilogramos +-60,2400 kg~menos sesenta coma dos cuatro cero cero kilogramos +8,52 %~ocho coma cincuenta y dos por ciento +-8,52 %~menos ocho coma cincuenta y dos por ciento +1 %~uno por ciento +3 cm~tres centรญmetros +4 s~cuatro segundos +5 l~cinco litros +4,51/s~cuatro coma cincuenta y uno por segundo +0,0101 s~cero coma cero uno cero un segundos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt @@ -0,0 +1,24 @@ +$1~un dรณlar +1 $~un dรณlar +$1,50~un dรณlar cincuenta centavos +1,50 $~un dรณlar cincuenta centavos +ยฃ200.000.001~doscientos millones una libras +200.000.001 ยฃ~doscientos millones una libras +2 billones de euros~dos billones de euros +โ‚ฌ2 billones~dos billones de euros +โ‚ฌ 2 billones~dos billones de euros +โ‚ฌ 2,3 billones~dos coma tres billones de euros +2,3 billones de euros~dos coma tres billones de euros +โ‚ฌ5,50~cinco euros cincuenta cรฉntimos +5,50 โ‚ฌ~cinco euros cincuenta cรฉntimos +5,01 โ‚ฌ~cinco euros un cรฉntimo +5,01 ยฃ~cinco libras un penique +21 czk~veintiuna coronas checas +czk21~veintiuna coronas checas +czk21,1 millones~veintiuna coma una millones de coronas checas +czk 5,50 billones~cinco coma cincuenta billones de coronas checas +rs 5,50 billones~cinco coma cincuenta billones de rupias +czk5,50 billones~cinco coma cincuenta billones de coronas checas +0,55 $~cincuenta y cinco centavos +1,01 $~un dรณlar un centavo +ยฅ12,05~doce yenes cinco centavos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt @@ -0,0 +1,120 @@ +~121 +ciento veintiรบn +ciento veintiuno +ciento veintiuna +121 +~200 +doscientos +doscientas +200 +~201 +doscientos un +doscientos uno +doscientas una +201 +~1 +un +uno +una +1 +~550.000.001 +quinientos cincuenta millones un +quinientos cincuenta millones una +quinientos cincuenta millones uno +550.000.001 +~500.501 +quinientos mil quinientos un +quinientos mil quinientos uno +quinientas mil quinientas una +500.501 +~500.001.ยบ +quinientosmilรฉsimo primero +quingentรฉsimo milรฉsimo primero +quinientosmilรฉsimos primeros +quingentรฉsimos milรฉsimos primeros +500.001.ยบ +~500.001.ยช +quinientasmilรฉsima primera +quingentรฉsima milรฉsima primera +quinientasmilรฉsimas primeras +quingentรฉsimas milรฉsimas primeras +500.001.ยช +~11.ยช +dรฉcima primera +decimoprimera +dรฉcimas primeras +decimoprimeras +undรฉcima +undรฉcimas +11.ยช +~11.ยบ +dรฉcimo primero +decimoprimero +dรฉcimos primeros +decimoprimeros +undรฉcimo +undรฉcimos +11.ยบ +~12.ยบ +dรฉcimo segundo +decimosegundo +dรฉcimos segundos +decimosegundos +duodรฉcimo +duodรฉcimos +12.ยบ +~200,0101 +doscientos coma cero uno cero un +doscientos coma cero uno cero uno +doscientas coma cero una cero una +200,0101 +~1.000.200,21 +un millรณn doscientos coma veintiรบn +un millรณn doscientos coma veintiuno +un millรณn doscientas coma veintiuna +un millรณn doscientos coma dos un +un millรณn doscientos coma dos uno +un millรณn doscientas coma dos una +1.000.200,21 +~1/12 +un doceavo +una doceava parte +un duodรฉcimo +una duodรฉcima parte +uno sobre doce +1/12 +~5/200 +cinco ducentรฉsimos +cinco ducentรฉsimas partes +cinco sobre doscientos +5/200 +~1 5/3 +uno y cinco tercios +una y cinco terceras partes +uno y cinco sobre tres +una y cinco sobre tres +~1/5/2020 +primero de mayo de dos mil veinte +uno de mayo de dos mil veinte +cinco de enero de dos mil veinte +~$5,50 +cinco dรณlares con cincuenta +cinco dรณlares y cincuenta +cinco dรณlares cincuenta +cinco dรณlares con cincuenta centavos +cinco dรณlares y cincuenta centavos +cinco dรณlares cincuenta centavos +~2.30 h +dos y treinta +dos y media +tres menos treinta +tres menos media +treinta para las tres +~12.30 a.m. +doce y treinta de la medianoche +doce y treinta de la noche +doce y media de la medianoche +doce y media de la noche +una menos treinta de la maรฑana +una menos media de la maรฑana +treinta para la una de la maรฑana \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt @@ -0,0 +1,137 @@ +1.แต‰สณ~primer +1.ยบ~primero +1.ยช~primera +2.ยบ~segundo +2.ยช~segunda +ii~segundo +II~segundo +3.แต‰สณ~tercer +3.ยบ~tercero +3.ยช~tercera +4.ยบ~cuarto +4.ยช~cuarta +5.ยบ~quinto +5.ยช~quinta +6.ยบ~sexto +6.ยช~sexta +7.ยบ~sรฉptimo +7.ยช~sรฉptima +8.ยบ~octavo +8.ยช~octava +9.ยบ~noveno +9.ยช~novena +10.ยบ~dรฉcimo +10.ยช~dรฉcima +11.แต‰สณ~decimoprimer +11.ยบ~undรฉcimo +11.ยช~undรฉcima +12.ยบ~duodรฉcimo +12.ยช~duodรฉcima +13.แต‰สณ~decimotercer +13.ยบ~decimotercero +13.ยช~decimotercera +14.ยบ~decimocuarto +14.ยช~decimocuarta +15.ยบ~decimoquinto +15.ยช~decimoquinta +16.ยบ~decimosexto +16.ยช~decimosexta +17.ยบ~decimosรฉptimo +17.ยช~decimosรฉptima +18.ยบ~decimoctavo +18.ยช~decimoctava +19.ยบ~decimonoveno +19.ยช~decimonovena +20.ยบ~vigรฉsimo +20.ยช~vigรฉsima +21.แต‰สณ~vigesimoprimer +21.ยบ~vigesimoprimero +21.ยช~vigesimoprimera +30.ยบ~trigรฉsimo +30.ยช~trigรฉsima +31.แต‰สณ~trigรฉsimo primer +31.ยบ~trigรฉsimo primero +31.ยช~trigรฉsima primera +40.ยบ~cuadragรฉsimo +40.ยช~cuadragรฉsima +41.แต‰สณ~cuadragรฉsimo primer +41.ยบ~cuadragรฉsimo primero +41.ยช~cuadragรฉsima primera +50.ยบ~quincuagรฉsimo +50.ยช~quincuagรฉsima +51.แต‰สณ~quincuagรฉsimo primer +51.ยบ~quincuagรฉsimo primero +51.ยช~quincuagรฉsima primera +60.ยบ~sexagรฉsimo +60.ยช~sexagรฉsima +70.ยบ~septuagรฉsimo +70.ยช~septuagรฉsima +80.ยบ~octogรฉsimo +80.ยช~octogรฉsima +90.ยบ~nonagรฉsimo +90.ยช~nonagรฉsima +100.ยบ~centรฉsimo +100.ยช~centรฉsima +101.แต‰สณ~centรฉsimo primer +101.ยบ~centรฉsimo primero +101.ยช~centรฉsima primera +134.ยบ~centรฉsimo trigรฉsimo cuarto +134.ยช~centรฉsima trigรฉsima cuarta +200.ยบ~ducentรฉsimo +200.ยช~ducentรฉsima +300.ยบ~tricentรฉsimo +300.ยช~tricentรฉsima +400.ยบ~cuadringentรฉsimo +400.ยช~cuadringentรฉsima +500.ยบ~quingentรฉsimo +500.ยช~quingentรฉsima +600.ยบ~sexcentรฉsimo +600.ยช~sexcentรฉsima +700.ยบ~septingentรฉsimo +700.ยช~septingentรฉsima +800.ยบ~octingentรฉsimo +800.ยช~octingentรฉsima +900.ยบ~noningentรฉsimo +900.ยช~noningentรฉsima +1000.ยบ~milรฉsimo +1000.ยช~milรฉsima +1001.แต‰สณ~milรฉsimo primer +1 000.ยบ~milรฉsimo +1 000.ยช~milรฉsima +1 001.แต‰สณ~milรฉsimo primer +1.000.ยบ~milรฉsimo +1.000.ยช~milรฉsima +1.001.แต‰สณ~milรฉsimo primer +1248.ยบ~milรฉsimo ducentรฉsimo cuadragรฉsimo octavo +1248.ยช~milรฉsima ducentรฉsima cuadragรฉsima octava +2000.ยบ~dosmilรฉsimo +100 000.ยบ~cienmilรฉsimo +i~primero +I~primero +ii~segundo +II~segundo +iii~tercero +III~tercero +iv~cuarto +IV~cuarto +V~quinto +VI~sexto +VII~sรฉptimo +VIII~octavo +IX~noveno +X~dรฉcimo +XI~undรฉcimo +XII~duodรฉcimo +XIII~decimotercero +XX~vigรฉsimo +XXI~vigesimoprimero +XXX~trigรฉsimo +XL~cuadragรฉsimo +L~quincuagรฉsimo +XC~nonagรฉsimo +C~centรฉsimo +CD~cuadringentรฉsimo +D~quingentรฉsimo +CM~noningentรฉsimo +999.ยบ~noningentรฉsimo nonagรฉsimo noveno +cmxcix~noningentรฉsimo nonagรฉsimo noveno \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt @@ -0,0 +1,3 @@ +123-123-5678~uno dos tres uno dos tres cinco seis siete ocho +123-456-789~uno dos tres cuatro cinco seis siete ocho nueve +1234-5678~uno dos tres cuatro cinco seis siete ocho \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt @@ -0,0 +1,26 @@ +1.00~una +1:00~una +01:00~una +01 h~una +3 h~tres horas +1 h~una hora +1.05 h~una y cinco +01.05 h~una y cinco +1.00 h~una +1.00 a.m.~una de la maรฑana +1.00 a.m~una de la maรฑana +1.00 p.m.~una de la tarde +1.00 p.m est~una de la tarde e s t +1.00 est~una e s t +5:02 est~cinco y dos e s t +5:02 p.m pst~cinco y dos de la noche p s t +5:02 p.m.~cinco y dos de la noche +12.15~doce y cuarto +12.15 a.m.~doce y cuarto de la noche +12.15 p.m.~doce y cuarto del mediodรญa +13.30~trece y media +14.05~catorce y cinco +24:50~veinticuatro y cincuenta +3:02:32 pst~tres horas dos minutos y treinta y dos segundos p s t +00:52~cero y cincuenta y dos +0:52~cero y cincuenta y dos diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt @@ -0,0 +1,3 @@ +el dr.~el doctor +sr. rodriguez~seรฑor rodriguez +182 esq. toledo~ciento ochenta y dos esquina toledo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt @@ -0,0 +1,48 @@ +~ +yahoo!~yahoo! +veinte!~veinte! +โ€”~โ€” +aaa~aaa +aabach~aabach +aabenraa~aabenraa +aabye~aabye +aaccessed~aaccessed +aach~aach +aachen's~aachen's +aadri~aadri +aafia~aafia +aagaard~aagaard +aagadu~aagadu +aagard~aagard +aagathadi~aagathadi +aaghart's~aaghart's +aagnes~aagnes +aagomoni~aagomoni +aagon~aagon +aagoo~aagoo +aagot~aagot +aahar~aahar +aahh~aahh +aahperd~aahperd +aaibinterstate~aaibinterstate +aajab~aajab +aakasa~aakasa +aakervik~aakervik +aakirkeby~aakirkeby +aalam~aalam +aalbaek~aalbaek +aaldiu~aaldiu +aalem~aalem +a'ali~a'ali +aalilaassamthey~aalilaassamthey +aalin~aalin +aaliyan~aaliyan +aaliyan's~aaliyan's +aamadu~aamadu +aamara~aamara +aambala~aambala +aamera~aamera +aamer's~aamer's +aamina~aamina +aaminah~aaminah +aamjiwnaang~aamjiwnaang diff --git a/tests/nemo_text_processing/es/test_cardinal.py b/tests/nemo_text_processing/es/test_cardinal.py --- a/tests/nemo_text_processing/es/test_cardinal.py +++ b/tests/nemo_text_processing/es/test_cardinal.py @@ -22,7 +22,8 @@ class TestCardinal: - inverse_normalizer_es = ( + + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +33,34 @@ class TestCardinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_cardinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_date.py b/tests/nemo_text_processing/es/test_date.py --- a/tests/nemo_text_processing/es/test_date.py +++ b/tests/nemo_text_processing/es/test_date.py @@ -22,7 +22,7 @@ class TestDate: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDate: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_date.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_decimal.py b/tests/nemo_text_processing/es/test_decimal.py --- a/tests/nemo_text_processing/es/test_decimal.py +++ b/tests/nemo_text_processing/es/test_decimal.py @@ -22,7 +22,7 @@ class TestDecimal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDecimal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_decimal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_electronic.py b/tests/nemo_text_processing/es/test_electronic.py --- a/tests/nemo_text_processing/es/test_electronic.py +++ b/tests/nemo_text_processing/es/test_electronic.py @@ -35,3 +35,31 @@ class TestElectronic: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_electronic.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_fraction.py b/tests/nemo_text_processing/es/test_fraction.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_fraction.py @@ -0,0 +1,51 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import pytest +from nemo_text_processing.text_normalization.normalize import Normalizer +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, parse_test_case_file + + +class TestFraction: + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_fraction.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_measure.py b/tests/nemo_text_processing/es/test_measure.py --- a/tests/nemo_text_processing/es/test_measure.py +++ b/tests/nemo_text_processing/es/test_measure.py @@ -36,3 +36,31 @@ class TestMeasure: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_measure.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_money.py b/tests/nemo_text_processing/es/test_money.py --- a/tests/nemo_text_processing/es/test_money.py +++ b/tests/nemo_text_processing/es/test_money.py @@ -23,7 +23,7 @@ class TestMoney: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,34 @@ class TestMoney: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_money.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_normalization_with_audio.py b/tests/nemo_text_processing/es/test_normalization_with_audio.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_normalization_with_audio.py @@ -0,0 +1,43 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import pytest +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, get_test_cases_multiple + + +class TestNormalizeWithAudio: + + normalizer_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + @parameterized.expand(get_test_cases_multiple('es/data_text_normalization/test_cases_normalize_with_audio.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, n_tagged=1000, punct_post_process=False) + print(expected) + print("pred") + print(pred) + assert len(set(pred).intersection(set(expected))) == len( + expected + ), f'missing: {set(expected).difference(set(pred))}' diff --git a/tests/nemo_text_processing/es/test_ordinal.py b/tests/nemo_text_processing/es/test_ordinal.py --- a/tests/nemo_text_processing/es/test_ordinal.py +++ b/tests/nemo_text_processing/es/test_ordinal.py @@ -23,7 +23,7 @@ class TestOrdinal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,33 @@ class TestOrdinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_ordinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=30, punct_post_process=False, + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh @@ -0,0 +1,84 @@ +#! /bin/sh + +PROJECT_DIR=/workspace/tests + +runtest () { + input=$1 + cd /workspace/sparrowhawk/documentation/grammars + + # read test file + while read testcase; do + IFS='~' read written spoken <<< $testcase + denorm_pred=$(echo $written | normalizer_main --config=sparrowhawk_configuration.ascii_proto 2>&1 | tail -n 1) + + # trim white space + spoken="$(echo -e "${spoken}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + denorm_pred="$(echo -e "${denorm_pred}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + + # input expected actual + assertEquals "$written" "$spoken" "$denorm_pred" + done < "$input" +} + +testTNCardinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_cardinal.txt + runtest $input +} + +testTNDate() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_date.txt + runtest $input +} + +testTNDecimal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_decimal.txt + runtest $input +} + +testTNElectronic() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_electronic.txt + runtest $input +} + +testTNFraction() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_fraction.txt + runtest $input +} + +testTNMoney() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_money.txt + runtest $input +} + +testTNOrdinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTelephone() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTime() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_time.txt + runtest $input +} + +testTNMeasure() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_measure.txt + runtest $input +} + +testTNWhitelist() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_whitelist.txt + runtest $input +} + +testTNWord() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_word.txt + runtest $input +} + +# Load shUnit2 +. $PROJECT_DIR/../shunit2/shunit2 diff --git a/tests/nemo_text_processing/es/test_telephone.py b/tests/nemo_text_processing/es/test_telephone.py --- a/tests/nemo_text_processing/es/test_telephone.py +++ b/tests/nemo_text_processing/es/test_telephone.py @@ -36,3 +36,31 @@ class TestTelephone: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_telephone.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_time.py b/tests/nemo_text_processing/es/test_time.py --- a/tests/nemo_text_processing/es/test_time.py +++ b/tests/nemo_text_processing/es/test_time.py @@ -35,3 +35,31 @@ class TestTime: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_time.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_whitelist.py b/tests/nemo_text_processing/es/test_whitelist.py --- a/tests/nemo_text_processing/es/test_whitelist.py +++ b/tests/nemo_text_processing/es/test_whitelist.py @@ -35,3 +35,30 @@ class TestWhitelist: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_whitelist.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=10, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_word.py b/tests/nemo_text_processing/es/test_word.py --- a/tests/nemo_text_processing/es/test_word.py +++ b/tests/nemo_text_processing/es/test_word.py @@ -35,3 +35,30 @@ class TestWord: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer_es = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_word.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, verbose=False) + assert pred == expected, f"input: {test_input}" + + if self.normalizer_with_audio_es: + pred_non_deterministic = self.normalizer_with_audio_es.normalize( + test_input, n_tagged=150, punct_post_process=False + ) + assert expected in pred_non_deterministic, f"input: {test_input}"
1.0
NVIDIA__NeMo-7582
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] 1 2 |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg 17 :target: https://badge.fury.io/py/nemo-toolkit 18 :alt: Release version 19 20 .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg 21 :target: https://badge.fury.io/py/nemo-toolkit 22 :alt: Python version 23 24 .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads 25 :target: https://pepy.tech/project/nemo-toolkit 26 :alt: PyPi total downloads 27 28 .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push 29 :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml 30 :alt: CodeQL 31 32 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 33 :target: https://github.com/psf/black 34 :alt: Code style: black 35 36 .. _main-readme: 37 38 **NVIDIA NeMo** 39 =============== 40 41 Introduction 42 ------------ 43 44 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), 45 text-to-speech synthesis (TTS), large language models (LLMs), and 46 natural language processing (NLP). 47 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) 48 and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 49 50 All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and 51 training is automatically scalable to 1000s of GPUs. 52 Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. 53 NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. 54 55 Getting started with NeMo is simple. 56 State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and 57 `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. 58 These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. 59 60 We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that 61 can all be run on `Google Colab <https://colab.research.google.com>`_. 62 63 For advanced users that want to train NeMo models from scratch or finetune existing NeMo models 64 we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. 65 66 For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. 67 The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ 68 which can be used to find the optimal model parallel configuration for training on a specific cluster. 69 70 Also see our `introductory video <https://www.youtube.com/embed/wBgpMf_KQVw>`_ for a high level overview of NeMo. 71 72 Key Features 73 ------------ 74 75 * Speech processing 76 * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ 77 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 78 * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ 79 * Jasper, QuartzNet, CitriNet, ContextNet 80 * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer 81 * Squeezeformer-CTC and Squeezeformer-Transducer 82 * LSTM-Transducer (RNNT) and LSTM-CTC 83 * Supports the following decoders/losses: 84 * CTC 85 * Transducer/RNNT 86 * Hybrid Transducer/CTC 87 * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ 88 * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ 89 * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ 90 * Beam Search decoding 91 * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 92 * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ 93 * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) 94 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 95 * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ 96 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 97 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ 98 * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet 99 * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) 100 * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer 101 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 102 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 103 * Natural Language Processing 104 * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ 105 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ 106 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 107 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 108 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 109 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 110 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 111 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 112 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 113 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 114 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 115 * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ 116 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 117 * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ 118 * Text-to-Speech Synthesis (TTS): 119 * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 120 * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 121 * Vocoders: HiFiGAN, UnivNet, WaveGlow 122 * End-to-End Models: VITS 123 * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 124 * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ 125 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ 126 * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ 127 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 128 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 129 * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ 130 131 132 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 133 134 Requirements 135 ------------ 136 137 1) Python 3.10 or above 138 2) Pytorch 1.13.1 or above 139 3) NVIDIA GPU for training 140 141 Documentation 142 ------------- 143 144 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 145 :alt: Documentation Status 146 :scale: 100% 147 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 148 149 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 150 :alt: Documentation Status 151 :scale: 100% 152 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 153 154 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 155 | Version | Status | Description | 156 +=========+=============+==========================================================================================================================================+ 157 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 158 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 159 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 160 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 161 162 Tutorials 163 --------- 164 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 165 166 Getting help with NeMo 167 ---------------------- 168 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 169 170 171 Installation 172 ------------ 173 174 Conda 175 ~~~~~ 176 177 We recommend installing NeMo in a fresh Conda environment. 178 179 .. code-block:: bash 180 181 conda create --name nemo python==3.10.12 182 conda activate nemo 183 184 Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. 185 186 .. code-block:: bash 187 188 conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia 189 190 The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. 191 192 Pip 193 ~~~ 194 Use this installation mode if you want the latest released version. 195 196 .. code-block:: bash 197 198 apt-get update && apt-get install -y libsndfile1 ffmpeg 199 pip install Cython 200 pip install nemo_toolkit['all'] 201 202 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 203 204 Pip from source 205 ~~~~~~~~~~~~~~~ 206 Use this installation mode if you want the version from a particular GitHub branch (e.g main). 207 208 .. code-block:: bash 209 210 apt-get update && apt-get install -y libsndfile1 ffmpeg 211 pip install Cython 212 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 213 214 215 From source 216 ~~~~~~~~~~~ 217 Use this installation mode if you are contributing to NeMo. 218 219 .. code-block:: bash 220 221 apt-get update && apt-get install -y libsndfile1 ffmpeg 222 git clone https://github.com/NVIDIA/NeMo 223 cd NeMo 224 ./reinstall.sh 225 226 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 227 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 228 229 RNNT 230 ~~~~ 231 Note that RNNT requires numba to be installed from conda. 232 233 .. code-block:: bash 234 235 conda remove numba 236 pip uninstall numba 237 conda install -c conda-forge numba 238 239 NeMo Megatron 240 ~~~~~~~~~~~~~ 241 NeMo Megatron training requires NVIDIA Apex to be installed. 242 Install it manually if not using the NVIDIA PyTorch container. 243 244 To install Apex, run 245 246 .. code-block:: bash 247 248 git clone https://github.com/NVIDIA/apex.git 249 cd apex 250 git checkout 52e18c894223800cb611682dce27d88050edf1de 251 pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ 252 253 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. 254 255 While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. 256 This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 257 258 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: 259 260 .. code-block:: bash 261 262 conda install -c nvidia cuda-nvprof=11.8 263 264 packaging is also needed: 265 266 .. code-block:: bash 267 268 pip install packaging 269 270 With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. 271 272 273 Transformer Engine 274 ~~~~~~~~~~~~~~~~~~ 275 NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ 276 Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. 277 `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. 278 279 .. code-block:: bash 280 281 pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable 282 283 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. 284 285 Transformer Engine requires PyTorch to be built with CUDA 11.8. 286 287 288 Flash Attention 289 ~~~~~~~~~~~~~~~~~~~~ 290 Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. 291 292 .. code-block:: bash 293 294 pip install flash-attn 295 pip install triton==2.0.0.dev20221202 296 297 NLP inference UI 298 ~~~~~~~~~~~~~~~~~~~~ 299 To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. 300 301 .. code-block:: bash 302 303 pip install gradio==3.34.0 304 305 NeMo Text Processing 306 ~~~~~~~~~~~~~~~~~~~~ 307 NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. 308 309 Docker containers: 310 ~~~~~~~~~~~~~~~~~~ 311 We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. 312 313 To use built container, please run 314 315 .. code-block:: bash 316 317 docker pull nvcr.io/nvidia/nemo:23.06 318 319 To build a nemo container with Dockerfile from a branch, please run 320 321 .. code-block:: bash 322 323 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 324 325 326 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. 327 328 .. code-block:: bash 329 330 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 331 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 332 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 333 334 Examples 335 -------- 336 337 Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 338 339 340 Contributing 341 ------------ 342 343 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 344 345 Publications 346 ------------ 347 348 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 349 350 License 351 ------- 352 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 353 [end of README.rst] [start of nemo/collections/asr/metrics/wer.py] ... ... 1283 1284 # preserve decoding alignments 1285 preserve_alignments: Optional[bool] = None 1286 1287 # compute ctc time stamps 1288 compute_timestamps: Optional[bool] = None 1289 1290 # token representing word seperator 1291 word_seperator: str = " " 1292 1293 # type of timestamps to calculate 1294 ctc_timestamp_type: str = "all" # can be char, word or all for both 1295 1296 # batch dimension 1297 batch_dim_index: int = 0 1298 1299 # greedy decoding config 1300 greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() 1301 1302 # beam decoding config 1303 beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) 1304 1305 # confidence config 1306 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1307 1308 # can be used to change temperature for decoding 1309 temperature: float = 1.0 1310 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/tts/models/fastpitch.py] ... ... 56 @dataclass 57 class G2PConfig: 58 _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" 59 phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" 60 heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" 61 phoneme_probability: float = 0.5 62 63 64 @dataclass 65 class TextTokenizer: 66 _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" 67 punct: bool = True 68 stresses: bool = True 69 chars: bool = True 70 apostrophe: bool = True 71 pad_with_space: bool = True 72 add_blank_at: bool = True 73 g2p: G2PConfig = G2PConfig() 74 75 76 @dataclass 77 class TextTokenizerConfig: 78 text_tokenizer: TextTokenizer = TextTokenizer() 79 80 81 class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): 82 """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" ... [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] ... ... 60 61 # bucketing params 62 bucketing_strategy: str = "synced_randomized" 63 bucketing_batch_size: Optional[Any] = None 64 bucketing_weights: Optional[List[int]] = None 65 66 67 @dataclass 68 class EncDecCTCConfig(model_cfg.ModelConfig): 69 # Model global arguments 70 sample_rate: int = 16000 71 repeat: int = 1 72 dropout: float = 0.0 73 separable: bool = False 74 labels: List[str] = MISSING 75 76 # Dataset configs 77 train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) 78 validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 79 test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 80 81 # Optimizer / Scheduler config 82 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 83 84 # Model component configs 85 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 86 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 87 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 88 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 89 decoding: CTCDecodingConfig = CTCDecodingConfig() 90 91 92 @dataclass 93 class EncDecCTCModelConfig(model_cfg.NemoConfig): 94 model: EncDecCTCConfig = EncDecCTCConfig() 95 96 97 @dataclass 98 class CacheAwareStreamingConfig: ... [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] ... ... 64 # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. 65 word_ts_anchor_offset: Optional[float] = None 66 # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. 67 word_ts_anchor_pos: str = "start" 68 # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. 69 fix_word_ts_with_VAD: bool = False 70 # If True, use colored text to distinguish speakers in the output transcript. 71 colored_text: bool = False 72 # If True, the start and end time of each speaker turn is printed in the output transcript. 73 print_time: bool = True 74 # If True, the output transcript breaks the line to fix the line width (default is 90 chars) 75 break_lines: bool = False 76 77 78 @dataclass 79 class ASRDiarizerConfig(DiarizerComponentConfig): 80 model_path: Optional[str] = "stt_en_conformer_ctc_large" 81 parameters: ASRDiarizerParams = ASRDiarizerParams() 82 ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() 83 realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() 84 85 86 @dataclass 87 class VADParams(DiarizerComponentConfig): ... ... 88 window_length_in_sec: float = 0.15 # Window length in sec for VAD context input 89 shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction 90 smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) 91 overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter 92 onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech 93 offset: float = 0.1 # Offset threshold for detecting the end of a speech 94 pad_onset: float = 0.1 # Adding durations before each speech segment 95 pad_offset: float = 0 # Adding durations after each speech segment 96 min_duration_on: float = 0 # Threshold for small non_speech deletion 97 min_duration_off: float = 0.2 # Threshold for short speech segment deletion 98 filter_speech_first: bool = True 99 100 101 @dataclass 102 class VADConfig(DiarizerComponentConfig): 103 model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name 104 external_vad_manifest: Optional[str] = None 105 parameters: VADParams = VADParams() 106 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): ... ... 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): 110 # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] 111 window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) 112 # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] 113 shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) 114 # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] 115 multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) 116 # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. 117 save_embeddings: bool = True 118 119 120 @dataclass 121 class SpeakerEmbeddingsConfig(DiarizerComponentConfig): 122 # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) 123 model_path: Optional[str] = None 124 parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() 125 126 127 @dataclass 128 class ClusteringParams(DiarizerComponentConfig): ... ... 128 class ClusteringParams(DiarizerComponentConfig): 129 # If True, use num of speakers value provided in manifest file. 130 oracle_num_speakers: bool = False 131 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. 132 max_num_speakers: int = 8 133 # If the number of segments is lower than this number, enhanced speaker counting is activated. 134 enhanced_count_thres: int = 80 135 # Determines the range of p-value search: 0 < p <= max_rp_threshold. 136 max_rp_threshold: float = 0.25 137 # The higher the number, the more values will be examined with more time. 138 sparse_search_volume: int = 30 139 # If True, take a majority vote on multiple p-values to estimate the number of speakers. 140 maj_vote_spk_count: bool = False 141 142 143 @dataclass 144 class ClusteringConfig(DiarizerComponentConfig): 145 parameters: ClusteringParams = ClusteringParams() 146 147 148 @dataclass 149 class MSDDParams(DiarizerComponentConfig): ... ... 152 # Batch size for MSDD inference. 153 infer_batch_size: int = 25 154 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. 155 sigmoid_threshold: Tuple[float] = (0.7,) 156 # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. 157 seq_eval_mode: bool = False 158 # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. 159 split_infer: bool = True 160 # The length of split short sequence when split_infer is True. 161 diar_window_length: int = 50 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): ... ... 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): 174 manifest_filepath: Optional[str] = None 175 out_dir: Optional[str] = None 176 oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps 177 collar: float = 0.25 # Collar value for scoring 178 ignore_overlap: bool = True # Consider or ignore overlap segments while scoring 179 vad: VADConfig = VADConfig() 180 speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() 181 clustering: ClusteringConfig = ClusteringConfig() 182 msdd_model: MSDDConfig = MSDDConfig() 183 asr: ASRDiarizerConfig = ASRDiarizerConfig() 184 185 186 @dataclass 187 class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): 188 diarizer: DiarizerConfig = DiarizerConfig() 189 device: str = "cpu" 190 verbose: bool = False 191 batch_size: int = 64 192 num_workers: int = 1 ... [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/parts/k2/classes.py] ... ... 29 30 search_beam: float = 20.0 31 output_beam: float = 10.0 32 min_active_states: int = 30 33 max_active_states: int = 10000 34 35 36 @dataclass 37 class GraphModuleConfig: 38 """Config for graph modules. 39 Typically used with graph losses and decoders. 40 """ 41 42 topo_type: str = "default" 43 topo_with_self_loops: bool = True 44 token_lm: Optional[Any] = None 45 intersect_pruned: bool = False 46 intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() 47 boost_coeff: float = 0.0 48 predictor_window_size: int = 0 49 predictor_step_size: int = 1 50 ... [end of nemo/collections/asr/parts/k2/classes.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] ... 101 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 102 num_workers: int = 1 # Number of workers for DataLoader 103 104 # The decoding scheme to be used for evaluation 105 decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] 106 107 # Beam Search hyperparameters 108 beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding 109 beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding 110 111 maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding 112 maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding 113 114 # HAT related parameters (only for internal lm subtraction) 115 hat_subtract_ilm: bool = False 116 hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) 117 118 decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) 119 120 121 # fmt: on 122 ... [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of nemo/core/config/modelPT.py] ... 44 # ... 45 train_ds: Optional[DatasetConfig] = None 46 validation_ds: Optional[DatasetConfig] = None 47 test_ds: Optional[DatasetConfig] = None 48 optim: Optional[OptimConfig] = None 49 50 51 @dataclass 52 class HydraConfig: 53 run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) 54 job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) 55 56 57 @dataclass 58 class NemoConfig: 59 name: str = MISSING 60 model: ModelConfig = MISSING 61 trainer: config.TrainerConfig = config.TrainerConfig( 62 strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' 63 ) 64 exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() 65 hydra: HydraConfig = HydraConfig() 66 67 68 class ModelConfigBuilder: 69 def __init__(self, model_cfg: ModelConfig): ... [end of nemo/core/config/modelPT.py] [start of nemo/collections/asr/models/configs/aligner_config.py] ... ... 21 class AlignerCTCConfig: 22 prob_suppress_index: int = -1 23 prob_suppress_value: float = 1.0 24 25 26 @dataclass 27 class AlignerRNNTConfig: 28 predictor_window_size: int = 0 29 predictor_step_size: int = 1 30 31 32 @dataclass 33 class AlignerWrapperModelConfig: 34 alignment_type: str = "forced" 35 word_output: bool = True 36 cpu_decoding: bool = False 37 decode_batch_size: int = 0 38 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() 39 rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() 40 41 42 @dataclass 43 class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): 44 decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() 45 [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] ... ... 239 240 if self.preserve_frame_confidence: 241 raise ValueError( 242 "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." 243 ) 244 245 return hypothesis 246 247 def __call__(self, *args, **kwargs): 248 return self.forward(*args, **kwargs) 249 250 251 @dataclass 252 class GreedyCTCInferConfig: 253 preserve_alignments: bool = False 254 compute_timestamps: bool = False 255 preserve_frame_confidence: bool = False 256 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 257 confidence_method_cfg: str = "DEPRECATED" 258 259 def __post_init__(self): 260 # OmegaConf.structured ensures that post_init check is always executed ... [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/common/parts/adapter_modules.py] ... ... 146 def forward(self, x): 147 x = self.module(x) 148 149 # Add dropout if available 150 if self.dropout is not None: 151 x = self.dropout(x) 152 153 return x 154 155 156 @dataclass 157 class LinearAdapterConfig: 158 in_features: int 159 dim: int 160 activation: str = 'swish' 161 norm_position: str = 'pre' 162 dropout: float = 0.0 163 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 164 _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) 165 [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] ... ... 169 return super().forward(query, key, value, mask, pos_emb, cache=cache) 170 171 def reset_parameters(self): 172 with torch.no_grad(): 173 nn.init.zeros_(self.linear_out.weight) 174 nn.init.zeros_(self.linear_out.bias) 175 176 def get_default_strategy_config(self) -> 'dataclass': 177 return MHAResidualAddAdapterStrategyConfig() 178 179 180 @dataclass 181 class MultiHeadAttentionAdapterConfig: 182 n_head: int 183 n_feat: int 184 dropout_rate: float = 0.0 185 proj_dim: Optional[int] = None 186 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 187 _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) 188 189 190 class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): ... ... 273 # However: 274 # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 275 # For some reason at init sometimes it will cause the value of the tensor to become NaN 276 # All operations to compute matrix_ac and matrix_bd will then fail. 277 nn.init.zeros_(self.pos_bias_u) 278 nn.init.zeros_(self.pos_bias_v) 279 280 def get_default_strategy_config(self) -> 'dataclass': 281 return MHAResidualAddAdapterStrategyConfig() 282 283 284 @dataclass 285 class RelPositionMultiHeadAttentionAdapterConfig: 286 n_head: int 287 n_feat: int 288 dropout_rate: float = 0.0 289 proj_dim: Optional[int] = None 290 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 291 _target_: str = "{0}.{1}".format( 292 RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ 293 ) 294 ... ... 322 323 super().__init__( 324 d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, 325 ) 326 327 # Setup adapter strategy 328 self.setup_adapter_strategy(adapter_strategy) 329 330 def get_default_strategy_config(self) -> 'dataclass': 331 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 332 333 334 @dataclass 335 class PositionalEncodingAdapterConfig: 336 d_model: int 337 max_len: int = 5000 338 xscale: float = 1.0 339 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 340 _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) 341 342 343 class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): ... ... 364 xscale=1.0, 365 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 366 ): 367 super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) 368 369 # Setup adapter strategy 370 self.setup_adapter_strategy(adapter_strategy) 371 372 def get_default_strategy_config(self) -> 'dataclass': 373 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 374 375 376 @dataclass 377 class RelPositionalEncodingAdapterConfig: 378 d_model: int 379 max_len: int = 5000 380 xscale: float = 1.0 381 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 382 _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) 383 [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] ... 98 # Parameters for inference 99 acoustic_batch_size: int = 16 # The batch size to calculate log probabilities 100 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 101 device: str = "cuda" # The device to load the model onto to calculate log probabilities 102 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 103 104 # Beam Search hyperparameters 105 106 # The decoding scheme to be used for evaluation. 107 # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] 108 decoding_mode: str = "beamsearch_ngram" 109 110 beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding 111 beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding 112 beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding 113 114 decoding_strategy: str = "beam" 115 decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) 116 117 text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( 118 punctuation_marks = ".,?", 119 separate_punctuation = False, 120 do_lowercase = False, 121 rm_punctuation = False, 122 ) 123 # fmt: on 124 125 126 def beam_search_eval( ... [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] ... ... 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """Transformer based language model.""" 16 from MeCab import Model 17 from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule 18 from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule 19 from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( 20 MegatronRetrievalTransformerEncoderModule, 21 ) 22 from nemo.collections.nlp.modules.common.megatron.utils import ( 23 ApexGuardDefaults, 24 init_method_normal, 25 scaled_init_method_normal, 26 ) 27 28 try: 29 from apex.transformer.enums import AttnMaskType, ModelType 30 31 HAVE_APEX = True ... [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] ... 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 # Copyright 2017 Johns Hopkins University (Shinji Watanabe) 16 # 17 # Licensed under the Apache License, Version 2.0 (the "License"); 18 # you may not use this file except in compliance with the License. 19 # You may obtain a copy of the License at 20 # 21 # http://www.apache.org/licenses/LICENSE-2.0 22 # 23 # Unless required by applicable law or agreed to in writing, software 24 # distributed under the License is distributed on an "AS IS" BASIS, 25 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 26 # See the License for the specific language governing permissions and 27 # limitations under the License. 28 29 from dataclasses import dataclass 30 from typing import List, Optional, Tuple, Union 31 32 import numpy as np 33 import torch ... ... 2171 if self.preserve_frame_confidence: 2172 for batch_idx in range(batchsize): 2173 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2174 del hypotheses[batch_idx].frame_confidence[-1] 2175 2176 # Preserve states 2177 for batch_idx in range(batchsize): 2178 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2179 2180 return hypotheses 2181 2182 2183 @dataclass 2184 class GreedyRNNTInferConfig: 2185 max_symbols_per_step: Optional[int] = 10 2186 preserve_alignments: bool = False 2187 preserve_frame_confidence: bool = False 2188 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 2189 confidence_method_cfg: str = "DEPRECATED" 2190 2191 def __post_init__(self): 2192 # OmegaConf.structured ensures that post_init check is always executed ... ... 2203 2204 # TODO (alaptev): delete the following two lines sometime in the future 2205 logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") 2206 # OmegaConf.structured ensures that post_init check is always executed 2207 self.confidence_measure_cfg = OmegaConf.structured( 2208 self.confidence_method_cfg 2209 if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) 2210 else ConfidenceMeasureConfig(**self.confidence_method_cfg) 2211 ) 2212 self.confidence_method_cfg = "DEPRECATED" 2213 2214 2215 @dataclass 2216 class GreedyBatchedRNNTInferConfig: 2217 max_symbols_per_step: Optional[int] = 10 2218 preserve_alignments: bool = False 2219 preserve_frame_confidence: bool = False 2220 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 2221 confidence_method_cfg: str = "DEPRECATED" 2222 2223 def __post_init__(self): 2224 # OmegaConf.structured ensures that post_init check is always executed ... [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/tts/models/tacotron2.py] ... ... 39 LengthsType, 40 LogitsType, 41 MelSpectrogramType, 42 SequenceToSequenceAlignmentType, 43 ) 44 from nemo.core.neural_types.neural_type import NeuralType 45 from nemo.utils import logging, model_utils 46 47 48 @dataclass 49 class Preprocessor: 50 _target_: str = MISSING 51 pad_value: float = MISSING 52 53 54 @dataclass 55 class Tacotron2Config: 56 preprocessor: Preprocessor = Preprocessor() 57 encoder: Dict[Any, Any] = MISSING 58 decoder: Dict[Any, Any] = MISSING 59 postnet: Dict[Any, Any] = MISSING 60 labels: List = MISSING ... [end of nemo/collections/tts/models/tacotron2.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] ... 188 num_shards: int = -1 189 shuffle: bool = False 190 max_duration: Optional[float] = None 191 min_duration: Optional[float] = None 192 shuffle_seed: Optional[int] = None 193 sort_in_shards: bool = True 194 shard_manifests: bool = True 195 keep_files_together: bool = False 196 197 198 @dataclass 199 class ASRTarredDatasetMetadata: 200 created_datetime: Optional[str] = None 201 version: int = 0 202 num_samples_per_shard: Optional[int] = None 203 is_concatenated_manifest: bool = False 204 205 dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() 206 history: Optional[List[Any]] = field(default_factory=lambda: []) 207 208 def __post_init__(self): 209 self.created_datetime = self.get_current_datetime() ... [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] ... ... 58 bucketing_strategy: str = "synced_randomized" 59 bucketing_batch_size: Optional[Any] = None 60 bucketing_weights: Optional[List[int]] = None 61 62 63 @dataclass 64 class EncDecClassificationConfig(model_cfg.ModelConfig): 65 # Model global arguments 66 sample_rate: int = 16000 67 repeat: int = 1 68 dropout: float = 0.0 69 separable: bool = True 70 kernel_size_factor: float = 1.0 71 labels: List[str] = MISSING 72 timesteps: int = MISSING 73 74 # Dataset configs 75 train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 76 manifest_filepath=None, shuffle=True, trim_silence=False 77 ) 78 validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 79 manifest_filepath=None, shuffle=False 80 ) 81 test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 82 manifest_filepath=None, shuffle=False 83 ) 84 85 # Optimizer / Scheduler config 86 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 87 88 # Model component configs 89 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() 90 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 91 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 92 audio_length=timesteps 93 ) 94 95 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 96 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 97 98 99 @dataclass 100 class EncDecClassificationModelConfig(model_cfg.NemoConfig): 101 model: EncDecClassificationConfig = EncDecClassificationConfig() 102 [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] ... ... 167 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 168 169 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 170 When the alpha equals one, scaling is not applied to 'max_prob', 171 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 172 173 entropy_norm: A mapping of the entropy value to the interval [0,1]. 174 Supported values: 175 - 'lin' for using the linear mapping. 176 - 'exp' for using exponential mapping with linear shift. 177 """ 178 179 preserve_frame_confidence: bool = False 180 preserve_token_confidence: bool = False 181 preserve_word_confidence: bool = False 182 exclude_blank: bool = True 183 aggregation: str = "min" 184 measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() 185 method_cfg: str = "DEPRECATED" 186 187 def __post_init__(self): 188 # OmegaConf.structured ensures that post_init check is always executed ... [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of scripts/confidence_ensembles/build_ensemble.py] ... 61 <all arguments as in the previous examples> 62 tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set 63 tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set 64 65 You can check the dataclasses in this file for the full list of supported 66 arguments and their default values. 67 """ 68 69 import atexit 70 71 # using default logging to be able to silence unnecessary messages from nemo 72 import logging 73 import os 74 import random 75 import sys 76 import tempfile 77 from copy import deepcopy 78 from dataclasses import dataclass 79 from pathlib import Path 80 from typing import Dict, List, Optional, Tuple 81 82 import joblib ... ... 195 class_weight: Tuple = (None, "balanced") 196 197 # increase if getting many warnings that algorithm didn't converge 198 max_iter: int = 1000 199 200 201 @dataclass 202 class BuildEnsembleConfig: 203 # where to save the resulting ensemble model 204 output_path: str = MISSING 205 206 # each model specification 207 ensemble: List[EnsembleConfig] = MISSING 208 209 random_seed: int = 0 # for reproducibility 210 211 # default confidence, can override 212 confidence: ConfidenceConfig = ConfidenceConfig( 213 # we keep frame confidences and apply aggregation manually to get full-utterance confidence 214 preserve_frame_confidence=True, 215 exclude_blank=True, 216 aggregation="mean", 217 measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False ... ... 215 exclude_blank=True, 216 aggregation="mean", 217 measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False 229 # used to specify what to tune over. By default runs tuning over some 230 # reasonalbe grid, so that it does not take forever. 231 # Can be changed as needed 232 tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() 233 234 # very fast to tune and can be important in case of imbalanced datasets 235 # will automatically set to False if dev data is not available 236 tune_logistic_regression: bool = True 237 tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() 238 239 def __post_init__(self): 240 """Checking that if any dev data is provided, all are provided. 241 ... [end of scripts/confidence_ensembles/build_ensemble.py] [start of examples/asr/experimental/k2/align_speech_parallel.py] ... 60 61 You may control the aligner's config by setting the aligner_args: 62 aligner_args.alignment_type=argmax \ 63 aligner_args.word_output=False \ 64 aligner_args.cpu_decoding=True \ 65 aligner_args.decode_batch_size=8 \ 66 aligner_args.ctc_cfg.prob_suppress_index=-1 \ 67 aligner_args.ctc_cfg.prob_suppress_value=0.5 \ 68 aligner_args.rnnt_cfg.predictor_window_size=10 \ 69 aligner_args.decoder_module_cfg.intersect_pruned=true \ 70 aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ 71 ... 72 73 """ 74 75 76 import os 77 from dataclasses import dataclass, is_dataclass 78 from typing import Optional 79 80 import pytorch_lightning as ptl 81 import torch ... ... 80 import pytorch_lightning as ptl 81 import torch 82 from omegaconf import MISSING, OmegaConf 83 84 from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter 85 from nemo.collections.asr.models import ASRModel 86 from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig 87 from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig 88 from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel 89 from nemo.core.config import TrainerConfig, hydra_runner 90 from nemo.utils import logging 91 from nemo.utils.get_rank import is_global_rank_zero 92 93 94 @dataclass 95 class ParallelAlignmentConfig: 96 model: Optional[str] = None # name 97 predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) 98 aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() 99 output_path: str = MISSING 100 model_stride: int = 8 101 102 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") 103 104 # there arguments will be ignored 105 return_predictions: bool = False 106 use_cer: bool = False ... [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] ... 93 94 95 # fmt: on 96 97 98 @dataclass 99 class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): 100 # Model global arguments 101 sample_rate: int = 16000 102 repeat: int = 1 103 dropout: float = 0.0 104 separable: bool = True 105 kernel_size_factor: float = 1.0 106 timesteps: int = 128 107 labels: List[str] = MISSING 108 109 # Dataset configs 110 train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 111 manifest_filepath=None, shuffle=True, trim_silence=False 112 ) 113 validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 114 manifest_filepath=None, shuffle=False 115 ) 116 test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 117 manifest_filepath=None, shuffle=False 118 ) 119 120 # Optimizer / Scheduler config 121 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 122 123 # Model general component configs 124 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) 125 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( 126 freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 127 ) 128 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 129 audio_length=128 130 ) 131 132 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 133 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 134 135 136 @dataclass 137 class MatchboxNetVADModelConfig(MatchboxNetModelConfig): ... [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] ... ... 201 of ``model.class_labels`` files.""" 202 203 204 @dataclass 205 class PunctuationCapitalizationModelConfig: 206 """ 207 A configuration of 208 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 209 model. 210 211 See an example of model config in 212 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 213 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 214 215 This config is a part of :class:`~PunctuationCapitalizationConfig`. 216 """ 217 218 class_labels: ClassLabelsConfig = ClassLabelsConfig() 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" ... ... 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" 229 230 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 231 """A configuration for creating validation datasets and data loaders.""" 232 233 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 234 """A configuration for creating test datasets and data loaders.""" 235 236 punct_head: HeadConfig = HeadConfig() 237 """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" 238 239 capit_head: HeadConfig = HeadConfig() 240 """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" 241 242 tokenizer: Any = TokenizerConfig() 243 """A configuration for source text tokenizer.""" 244 245 language_model: LanguageModelConfig = LanguageModelConfig() 246 """A configuration of a BERT-like language model which serves as a model body.""" 247 248 optim: Optional[Any] = None 249 """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For ... ... 297 """ 298 299 pretrained_model: Optional[str] = None 300 """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options 301 by calling method 302 :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. 303 """ 304 305 name: Optional[str] = 'Punctuation_and_Capitalization' 306 """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" 307 308 do_training: bool = True 309 """Whether to perform training of the model.""" 310 311 do_testing: bool = False 312 """Whether ot perform testing of the model.""" 313 314 model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() 315 """A configuration for the 316 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 317 model.""" 318 319 trainer: Optional[TrainerConfig] = TrainerConfig() 320 """Contains ``Trainer`` Lightning class constructor parameters.""" 321 322 exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) 323 """A configuration with various NeMo training options such as output directories, resuming from checkpoint, 324 tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" 325 326 327 @dataclass 328 class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): 329 model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() 330 331 332 def is_legacy_model_config(model_cfg: DictConfig) -> bool: 333 """ ... [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of tools/nemo_forced_aligner/align.py] ... 135 viterbi_device: Optional[str] = None 136 batch_size: int = 1 137 use_local_attention: bool = True 138 additional_segment_grouping_separator: Optional[str] = None 139 audio_filepath_parts_in_utt_id: int = 1 140 141 # Buffered chunked streaming configs 142 use_buffered_chunked_streaming: bool = False 143 chunk_len_in_secs: float = 1.6 144 total_buffer_in_secs: float = 4.0 145 chunk_batch_size: int = 32 146 147 # Cache aware streaming configs 148 simulate_cache_aware_streaming: Optional[bool] = False 149 150 # Output file configs 151 save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) 152 ctm_file_config: CTMFileConfig = CTMFileConfig() 153 ass_file_config: ASSFileConfig = ASSFileConfig() 154 155 156 @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) 157 def main(cfg: AlignmentConfig): ... [end of tools/nemo_forced_aligner/align.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] ... ... 160 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 161 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 162 ] 163 return config 164 # fmt: on 165 166 167 @dataclass 168 class JasperModelConfig(ctc_cfg.EncDecCTCConfig): 169 # Model global arguments 170 sample_rate: int = 16000 171 repeat: int = 1 172 dropout: float = 0.0 173 separable: bool = False 174 labels: List[str] = MISSING 175 176 # Dataset configs 177 train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( 178 manifest_filepath=None, shuffle=True, trim_silence=True 179 ) 180 validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 181 test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 182 183 # Optimizer / Scheduler config 184 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 185 186 # Model general component configs 187 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 188 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 189 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 190 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 191 192 193 @dataclass 194 class QuartzNetModelConfig(JasperModelConfig): ... [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] ... ... 53 return ' '.join(tokens) 54 55 def normalize(self, text) -> str: 56 # Normalization doesn't handle Japanese periods correctly; 57 # 'ใ€‚'becomes '.'. 58 if self.lang_id == 'en': 59 return self.normalizer.normalize(text) 60 else: 61 return text 62 63 64 class JaMecabProcessor: 65 """ 66 Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English 67 """ 68 69 def __init__(self): 70 self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") 71 72 def detokenize(self, text: List[str]) -> str: 73 RE_WS_IN_FW = re.compile( ... [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import re 17 from abc import abstractmethod 18 from dataclasses import dataclass, is_dataclass 19 from typing import Callable, Dict, List, Optional, Tuple, Union 20 21 import editdistance 22 import numpy as np ... ... 1285 1286 def compute(self): 1287 wer = self.scores.float() / self.words 1288 return wer, self.scores.detach(), self.words.detach() 1289 1290 1291 @dataclass 1292 class RNNTDecodingConfig: 1293 model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" 1294 strategy: str = "greedy_batch" 1295 1296 compute_hypothesis_token_set: bool = False 1297 1298 # preserve decoding alignments 1299 preserve_alignments: Optional[bool] = None 1300 1301 # confidence config 1302 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 ... ... 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 1307 # compute RNNT time stamps 1308 compute_timestamps: Optional[bool] = None 1309 1310 # compute language IDs 1311 compute_langs: bool = False 1312 1313 # token representing word seperator 1314 word_seperator: str = " " 1315 1316 # type of timestamps to calculate 1317 rnnt_timestamp_type: str = "all" # can be char, word or all for both 1318 1319 # greedy decoding config 1320 greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() 1321 1322 # beam decoding config 1323 beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) 1324 1325 # can be used to change temperature for decoding 1326 temperature: float = 1.0 1327 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] ... 620 return augmented_spec 621 622 623 class CropOrPadSpectrogramAugmentation(NeuralModule): 624 """ 625 Pad or Crop the incoming Spectrogram to a certain shape. 626 627 Args: 628 audio_length (int): the final number of timesteps that is required. 629 The signal will be either padded or cropped temporally to this 630 size. 631 """ 632 633 def __init__(self, audio_length): 634 super(CropOrPadSpectrogramAugmentation, self).__init__() 635 self.audio_length = audio_length 636 637 @typecheck() 638 @torch.no_grad() 639 def forward(self, input_signal, length): 640 image = input_signal ... [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] ... ... 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig 18 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig 19 from nemo.core.config.modelPT import NemoConfig 20 21 22 @dataclass 23 class GraphModuleConfig: 24 criterion_type: str = "ml" 25 loss_type: str = "ctc" 26 split_batch_size: int = 0 27 dec_type: str = "topo" 28 transcribe_training: bool = True 29 backend_cfg: BackendConfig = BackendConfig() 30 31 32 @dataclass 33 class EncDecK2SeqConfig(EncDecCTCConfig): 34 graph_module_cfg: GraphModuleConfig = GraphModuleConfig() 35 36 37 @dataclass 38 class EncDecK2SeqModelConfig(NemoConfig): 39 model: EncDecK2SeqConfig = EncDecK2SeqConfig() 40 [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] ... ... 110 pretrained_name: Optional[str] = None # Name of a pretrained model 111 dataset_manifest: str = MISSING 112 output_dir: str = MISSING 113 114 # General configs 115 batch_size: int = 32 116 num_workers: int = 4 117 118 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA 119 # device anyway, and do inference on CPU only if CUDA device is not found. 120 # If `cuda` is a negative number, inference will be on CPU only. 121 cuda: Optional[int] = None 122 amp: bool = False 123 audio_type: str = "wav" 124 125 # Confidence configs 126 target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) 127 confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) 128 grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on 129 130 131 @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) ... [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] ... ... 32 from nemo.core.config.modelPT import OptimConfig, SchedConfig 33 34 35 @dataclass 36 class MTSchedConfig(SchedConfig): 37 name: str = 'InverseSquareRootAnnealing' 38 warmup_ratio: Optional[float] = None 39 last_epoch: int = -1 40 41 42 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). 43 @dataclass 44 class MTOptimConfig(OptimConfig): 45 name: str = 'adam' 46 lr: float = 1e-3 47 betas: Tuple[float, float] = (0.9, 0.98) 48 weight_decay: float = 0.0 49 sched: Optional[MTSchedConfig] = MTSchedConfig() 50 51 52 @dataclass 53 class MTEncDecModelConfig(EncDecNLPModelConfig): ... ... 60 len_pen: float = 0.0 61 src_language: Any = 'en' # Any = str or List[str] 62 tgt_language: Any = 'en' # Any = str or List[str] 63 find_unused_parameters: Optional[bool] = True 64 shared_tokenizer: Optional[bool] = True 65 multilingual: Optional[bool] = False 66 preproc_out_dir: Optional[str] = None 67 validate_input_ids: Optional[bool] = True 68 shared_embeddings: bool = False 69 70 # network architecture configuration 71 encoder_tokenizer: Any = MISSING 72 encoder: Any = MISSING 73 74 decoder_tokenizer: Any = MISSING 75 decoder: Any = MISSING 76 77 head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) 78 79 # dataset configurations 80 train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 81 src_file_name=MISSING, 82 tgt_file_name=MISSING, 83 tokens_in_batch=512, 84 clean=True, 85 shuffle=True, 86 cache_ids=False, 87 use_cache=False, 88 ) 89 validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 90 src_file_name=MISSING, 91 tgt_file_name=MISSING, 92 tokens_in_batch=512, 93 clean=False, 94 shuffle=False, 95 cache_ids=False, 96 use_cache=False, 97 ) 98 test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 99 src_file_name=MISSING, 100 tgt_file_name=MISSING, 101 tokens_in_batch=512, 102 clean=False, 103 shuffle=False, 104 cache_ids=False, 105 use_cache=False, 106 ) 107 optim: Optional[OptimConfig] = MTOptimConfig() 108 109 110 @dataclass 111 class AAYNBaseConfig(MTEncDecModelConfig): 112 113 # Attention is All You Need Base Configuration 114 encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 115 decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 116 117 encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( 118 library='nemo', 119 model_name=None, 120 pretrained=False, 121 hidden_size=512, 122 inner_size=2048, 123 num_layers=6, 124 num_attention_heads=8, 125 ffn_dropout=0.1, 126 attn_score_dropout=0.1, 127 attn_layer_dropout=0.1, 128 ) 129 130 decoder: NeMoTransformerConfig = NeMoTransformerConfig( 131 library='nemo', 132 model_name=None, 133 pretrained=False, 134 hidden_size=512, 135 inner_size=2048, 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass ... ... 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass 145 class MTBottleneckModelConfig(AAYNBaseConfig): 146 model_type: str = 'nll' 147 min_logv: float = -6 148 latent_size: int = -1 # -1 will take value of encoder hidden 149 non_recon_warmup_batches: int = 200000 150 recon_per_token: bool = True 151 log_timing: bool = True 152 153 encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( 154 library='nemo', 155 model_name=None, 156 pretrained=False, 157 hidden_size=512, 158 inner_size=2048, 159 num_layers=6, 160 num_attention_heads=8, 161 ffn_dropout=0.1, 162 attn_score_dropout=0.1, 163 attn_layer_dropout=0.1, 164 arch='seq2seq', 165 hidden_steps=32, 166 hidden_blocks=1, 167 hidden_init_method='params', 168 ) 169 170 decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( 171 library='nemo', 172 model_name=None, 173 pretrained=False, 174 inner_size=2048, 175 num_layers=6, 176 num_attention_heads=8, 177 ffn_dropout=0.1, 178 attn_score_dropout=0.1, 179 attn_layer_dropout=0.1, 180 arch='seq2seq', 181 ) 182 [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/utils/exp_manager.py] ... 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import glob 16 import os 17 import subprocess 18 import sys 19 import time 20 import warnings 21 from dataclasses import dataclass 22 from datetime import timedelta 23 from pathlib import Path 24 from shutil import copy, move 25 from typing import Any, Dict, List, Optional, Tuple, Union ... ... 132 133 # Log dir creation parameters 134 explicit_log_dir: Optional[str] = None 135 exp_dir: Optional[str] = None 136 name: Optional[str] = None 137 version: Optional[str] = None 138 use_datetime_version: Optional[bool] = True 139 resume_if_exists: Optional[bool] = False 140 resume_past_end: Optional[bool] = False 141 resume_ignore_no_checkpoint: Optional[bool] = False 142 resume_from_checkpoint: Optional[str] = None 143 # Logging parameters 144 create_tensorboard_logger: Optional[bool] = True 145 summary_writer_kwargs: Optional[Dict[Any, Any]] = None 146 create_wandb_logger: Optional[bool] = False 147 wandb_logger_kwargs: Optional[Dict[Any, Any]] = None 148 create_mlflow_logger: Optional[bool] = False 149 mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() 150 create_dllogger_logger: Optional[bool] = False 151 dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() 152 create_clearml_logger: Optional[bool] = False 153 clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() 154 # Checkpointing parameters 155 create_checkpoint_callback: Optional[bool] = True 156 checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() 157 create_early_stopping_callback: Optional[bool] = False 158 early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() 159 create_preemption_callback: Optional[bool] = True 160 # Additional exp_manager arguments 161 files_to_copy: Optional[List[str]] = None 162 # logs timing of train/val/test steps 163 log_step_timing: Optional[bool] = True 164 step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() 165 # Configures creation of log files for different ranks 166 log_local_rank_0_only: Optional[bool] = False 167 log_global_rank_0_only: Optional[bool] = False 168 # disable initial validation when resuming from a checkpoint saved during validation 169 disable_validation_on_resume: Optional[bool] = True 170 ema: Optional[EMAParams] = EMAParams() 171 # Wall clock time limit 172 max_time_per_run: Optional[str] = None 173 # time to sleep non 0 ranks during initialization 174 seconds_to_sleep: float = 5 ... [end of nemo/utils/exp_manager.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] ... ... 588 beam_threshold: float = 20.0 589 unk_weight: float = -math.inf 590 sil_weight: float = 0.0 591 592 593 @dataclass 594 class BeamCTCInferConfig: 595 beam_size: int 596 search_type: str = 'default' 597 preserve_alignments: bool = False 598 compute_timestamps: bool = False 599 return_best_hypothesis: bool = True 600 601 beam_alpha: float = 1.0 602 beam_beta: float = 0.0 603 kenlm_path: Optional[str] = None 604 605 flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() 606 pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() 607 [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
8a892b86186dbdf61803d75570cb5c58471e9dda
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-09-30T01:26:50Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,9 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,9 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): @@ -2217,7 +2219,9 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -181,7 +181,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() + measure_cfg: ConfidenceMeasureConfig = field(default_factory=lambda: ConfidenceMeasureConfig()) method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -110,7 +110,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
NVIDIA__NeMo-7616
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] 1 2 |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg 17 :target: https://badge.fury.io/py/nemo-toolkit 18 :alt: Release version 19 20 .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg 21 :target: https://badge.fury.io/py/nemo-toolkit 22 :alt: Python version 23 24 .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads 25 :target: https://pepy.tech/project/nemo-toolkit 26 :alt: PyPi total downloads 27 28 .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push 29 :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml 30 :alt: CodeQL 31 32 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 33 :target: https://github.com/psf/black 34 :alt: Code style: black 35 36 .. _main-readme: 37 38 **NVIDIA NeMo** 39 =============== 40 41 Introduction 42 ------------ 43 44 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), 45 text-to-speech synthesis (TTS), large language models (LLMs), and 46 natural language processing (NLP). 47 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) 48 and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 49 50 All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and 51 training is automatically scalable to 1000s of GPUs. 52 Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. 53 NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. 54 55 Getting started with NeMo is simple. 56 State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and 57 `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. 58 These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. 59 60 We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that 61 can all be run on `Google Colab <https://colab.research.google.com>`_. 62 63 For advanced users that want to train NeMo models from scratch or finetune existing NeMo models 64 we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. 65 66 For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. 67 The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ 68 which can be used to find the optimal model parallel configuration for training on a specific cluster. 69 70 Also see the two introductory videos below for a high level overview of NeMo. 71 72 * Developing State-Of-The-Art Conversational AI Models in Three Lines of Code. 73 * NVIDIA NeMo: Toolkit for Conversational AI at PyData Yerevan 2022. 74 75 |three_lines| |pydata| 76 77 .. |pydata| image:: https://img.youtube.com/vi/J-P6Sczmas8/maxres3.jpg 78 :target: https://www.youtube.com/embed/J-P6Sczmas8?mute=0&start=14&autoplay=0 79 :width: 600 80 :alt: Develop Conversational AI Models in 3 Lines 81 82 .. |three_lines| image:: https://img.youtube.com/vi/wBgpMf_KQVw/maxresdefault.jpg 83 :target: https://www.youtube.com/embed/wBgpMf_KQVw?mute=0&start=0&autoplay=0 84 :width: 600 85 :alt: Introduction at PyData@Yerevan 2022 86 87 Key Features 88 ------------ 89 90 * Speech processing 91 * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ 92 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 93 * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ 94 * Jasper, QuartzNet, CitriNet, ContextNet 95 * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer 96 * Squeezeformer-CTC and Squeezeformer-Transducer 97 * LSTM-Transducer (RNNT) and LSTM-CTC 98 * Supports the following decoders/losses: 99 * CTC 100 * Transducer/RNNT 101 * Hybrid Transducer/CTC 102 * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ 103 * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ 104 * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ 105 * Beam Search decoding 106 * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 107 * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ 108 * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) 109 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 110 * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ 111 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 112 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ 113 * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet 114 * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) 115 * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer 116 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 117 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 118 * Natural Language Processing 119 * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ 120 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ 121 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 122 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 123 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 124 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 125 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 126 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 127 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 128 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 129 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 130 * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ 131 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 132 * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ 133 * Text-to-Speech Synthesis (TTS): 134 * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 135 * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 136 * Vocoders: HiFiGAN, UnivNet, WaveGlow 137 * End-to-End Models: VITS 138 * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 139 * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ 140 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ 141 * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ 142 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 143 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 144 * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ 145 146 147 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 148 149 Requirements 150 ------------ 151 152 1) Python 3.10 or above 153 2) Pytorch 1.13.1 or above 154 3) NVIDIA GPU for training 155 156 Documentation 157 ------------- 158 159 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 160 :alt: Documentation Status 161 :scale: 100% 162 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 163 164 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 165 :alt: Documentation Status 166 :scale: 100% 167 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 168 169 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 170 | Version | Status | Description | 171 +=========+=============+==========================================================================================================================================+ 172 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 173 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 174 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 175 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 176 177 Tutorials 178 --------- 179 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 180 181 Getting help with NeMo 182 ---------------------- 183 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 184 185 186 Installation 187 ------------ 188 189 Conda 190 ~~~~~ 191 192 We recommend installing NeMo in a fresh Conda environment. 193 194 .. code-block:: bash 195 196 conda create --name nemo python==3.10.12 197 conda activate nemo 198 199 Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. 200 201 .. code-block:: bash 202 203 conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia 204 205 The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. 206 207 Pip 208 ~~~ 209 Use this installation mode if you want the latest released version. 210 211 .. code-block:: bash 212 213 apt-get update && apt-get install -y libsndfile1 ffmpeg 214 pip install Cython 215 pip install nemo_toolkit['all'] 216 217 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 218 219 Pip from source 220 ~~~~~~~~~~~~~~~ 221 Use this installation mode if you want the version from a particular GitHub branch (e.g main). 222 223 .. code-block:: bash 224 225 apt-get update && apt-get install -y libsndfile1 ffmpeg 226 pip install Cython 227 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 228 229 230 From source 231 ~~~~~~~~~~~ 232 Use this installation mode if you are contributing to NeMo. 233 234 .. code-block:: bash 235 236 apt-get update && apt-get install -y libsndfile1 ffmpeg 237 git clone https://github.com/NVIDIA/NeMo 238 cd NeMo 239 ./reinstall.sh 240 241 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 242 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 243 244 RNNT 245 ~~~~ 246 Note that RNNT requires numba to be installed from conda. 247 248 .. code-block:: bash 249 250 conda remove numba 251 pip uninstall numba 252 conda install -c conda-forge numba 253 254 NeMo Megatron 255 ~~~~~~~~~~~~~ 256 NeMo Megatron training requires NVIDIA Apex to be installed. 257 Install it manually if not using the NVIDIA PyTorch container. 258 259 To install Apex, run 260 261 .. code-block:: bash 262 263 git clone https://github.com/NVIDIA/apex.git 264 cd apex 265 git checkout 52e18c894223800cb611682dce27d88050edf1de 266 pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ 267 268 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. 269 270 While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. 271 This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 272 273 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: 274 275 .. code-block:: bash 276 277 conda install -c nvidia cuda-nvprof=11.8 278 279 packaging is also needed: 280 281 .. code-block:: bash 282 283 pip install packaging 284 285 With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. 286 287 288 Transformer Engine 289 ~~~~~~~~~~~~~~~~~~ 290 NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ 291 Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. 292 `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. 293 294 .. code-block:: bash 295 296 pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable 297 298 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. 299 300 Transformer Engine requires PyTorch to be built with CUDA 11.8. 301 302 303 Flash Attention 304 ~~~~~~~~~~~~~~~~~~~~ 305 Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. 306 307 .. code-block:: bash 308 309 pip install flash-attn 310 pip install triton==2.0.0.dev20221202 311 312 NLP inference UI 313 ~~~~~~~~~~~~~~~~~~~~ 314 To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. 315 316 .. code-block:: bash 317 318 pip install gradio==3.34.0 319 320 NeMo Text Processing 321 ~~~~~~~~~~~~~~~~~~~~ 322 NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. 323 324 Docker containers: 325 ~~~~~~~~~~~~~~~~~~ 326 We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. 327 328 To use built container, please run 329 330 .. code-block:: bash 331 332 docker pull nvcr.io/nvidia/nemo:23.06 333 334 To build a nemo container with Dockerfile from a branch, please run 335 336 .. code-block:: bash 337 338 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 339 340 341 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. 342 343 .. code-block:: bash 344 345 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 346 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 347 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 348 349 Examples 350 -------- 351 352 Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 353 354 355 Contributing 356 ------------ 357 358 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 359 360 Publications 361 ------------ 362 363 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 364 365 License 366 ------- 367 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 368 [end of README.rst] [start of nemo/collections/asr/metrics/wer.py] ... ... 1283 1284 # preserve decoding alignments 1285 preserve_alignments: Optional[bool] = None 1286 1287 # compute ctc time stamps 1288 compute_timestamps: Optional[bool] = None 1289 1290 # token representing word seperator 1291 word_seperator: str = " " 1292 1293 # type of timestamps to calculate 1294 ctc_timestamp_type: str = "all" # can be char, word or all for both 1295 1296 # batch dimension 1297 batch_dim_index: int = 0 1298 1299 # greedy decoding config 1300 greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() 1301 1302 # beam decoding config 1303 beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) 1304 1305 # confidence config 1306 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1307 1308 # can be used to change temperature for decoding 1309 temperature: float = 1.0 1310 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/tts/models/fastpitch.py] ... ... 56 @dataclass 57 class G2PConfig: 58 _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" 59 phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" 60 heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" 61 phoneme_probability: float = 0.5 62 63 64 @dataclass 65 class TextTokenizer: 66 _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" 67 punct: bool = True 68 stresses: bool = True 69 chars: bool = True 70 apostrophe: bool = True 71 pad_with_space: bool = True 72 add_blank_at: bool = True 73 g2p: G2PConfig = G2PConfig() 74 75 76 @dataclass 77 class TextTokenizerConfig: 78 text_tokenizer: TextTokenizer = TextTokenizer() 79 80 81 class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): 82 """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" ... [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] ... ... 60 61 # bucketing params 62 bucketing_strategy: str = "synced_randomized" 63 bucketing_batch_size: Optional[Any] = None 64 bucketing_weights: Optional[List[int]] = None 65 66 67 @dataclass 68 class EncDecCTCConfig(model_cfg.ModelConfig): 69 # Model global arguments 70 sample_rate: int = 16000 71 repeat: int = 1 72 dropout: float = 0.0 73 separable: bool = False 74 labels: List[str] = MISSING 75 76 # Dataset configs 77 train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) 78 validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 79 test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 80 81 # Optimizer / Scheduler config 82 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 83 84 # Model component configs 85 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 86 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 87 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 88 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 89 decoding: CTCDecodingConfig = CTCDecodingConfig() 90 91 92 @dataclass 93 class EncDecCTCModelConfig(model_cfg.NemoConfig): 94 model: EncDecCTCConfig = EncDecCTCConfig() 95 96 97 @dataclass 98 class CacheAwareStreamingConfig: ... [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] ... ... 64 # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. 65 word_ts_anchor_offset: Optional[float] = None 66 # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. 67 word_ts_anchor_pos: str = "start" 68 # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. 69 fix_word_ts_with_VAD: bool = False 70 # If True, use colored text to distinguish speakers in the output transcript. 71 colored_text: bool = False 72 # If True, the start and end time of each speaker turn is printed in the output transcript. 73 print_time: bool = True 74 # If True, the output transcript breaks the line to fix the line width (default is 90 chars) 75 break_lines: bool = False 76 77 78 @dataclass 79 class ASRDiarizerConfig(DiarizerComponentConfig): 80 model_path: Optional[str] = "stt_en_conformer_ctc_large" 81 parameters: ASRDiarizerParams = ASRDiarizerParams() 82 ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() 83 realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() 84 85 86 @dataclass 87 class VADParams(DiarizerComponentConfig): ... ... 88 window_length_in_sec: float = 0.15 # Window length in sec for VAD context input 89 shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction 90 smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) 91 overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter 92 onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech 93 offset: float = 0.1 # Offset threshold for detecting the end of a speech 94 pad_onset: float = 0.1 # Adding durations before each speech segment 95 pad_offset: float = 0 # Adding durations after each speech segment 96 min_duration_on: float = 0 # Threshold for small non_speech deletion 97 min_duration_off: float = 0.2 # Threshold for short speech segment deletion 98 filter_speech_first: bool = True 99 100 101 @dataclass 102 class VADConfig(DiarizerComponentConfig): 103 model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name 104 external_vad_manifest: Optional[str] = None 105 parameters: VADParams = VADParams() 106 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): ... ... 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): 110 # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] 111 window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) 112 # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] 113 shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) 114 # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] 115 multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) 116 # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. 117 save_embeddings: bool = True 118 119 120 @dataclass 121 class SpeakerEmbeddingsConfig(DiarizerComponentConfig): 122 # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) 123 model_path: Optional[str] = None 124 parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() 125 126 127 @dataclass 128 class ClusteringParams(DiarizerComponentConfig): ... ... 128 class ClusteringParams(DiarizerComponentConfig): 129 # If True, use num of speakers value provided in manifest file. 130 oracle_num_speakers: bool = False 131 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. 132 max_num_speakers: int = 8 133 # If the number of segments is lower than this number, enhanced speaker counting is activated. 134 enhanced_count_thres: int = 80 135 # Determines the range of p-value search: 0 < p <= max_rp_threshold. 136 max_rp_threshold: float = 0.25 137 # The higher the number, the more values will be examined with more time. 138 sparse_search_volume: int = 30 139 # If True, take a majority vote on multiple p-values to estimate the number of speakers. 140 maj_vote_spk_count: bool = False 141 142 143 @dataclass 144 class ClusteringConfig(DiarizerComponentConfig): 145 parameters: ClusteringParams = ClusteringParams() 146 147 148 @dataclass 149 class MSDDParams(DiarizerComponentConfig): ... ... 152 # Batch size for MSDD inference. 153 infer_batch_size: int = 25 154 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. 155 sigmoid_threshold: Tuple[float] = (0.7,) 156 # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. 157 seq_eval_mode: bool = False 158 # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. 159 split_infer: bool = True 160 # The length of split short sequence when split_infer is True. 161 diar_window_length: int = 50 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): ... ... 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): 174 manifest_filepath: Optional[str] = None 175 out_dir: Optional[str] = None 176 oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps 177 collar: float = 0.25 # Collar value for scoring 178 ignore_overlap: bool = True # Consider or ignore overlap segments while scoring 179 vad: VADConfig = VADConfig() 180 speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() 181 clustering: ClusteringConfig = ClusteringConfig() 182 msdd_model: MSDDConfig = MSDDConfig() 183 asr: ASRDiarizerConfig = ASRDiarizerConfig() 184 185 186 @dataclass 187 class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): 188 diarizer: DiarizerConfig = DiarizerConfig() 189 device: str = "cpu" 190 verbose: bool = False 191 batch_size: int = 64 192 num_workers: int = 1 ... [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/parts/k2/classes.py] ... ... 29 30 search_beam: float = 20.0 31 output_beam: float = 10.0 32 min_active_states: int = 30 33 max_active_states: int = 10000 34 35 36 @dataclass 37 class GraphModuleConfig: 38 """Config for graph modules. 39 Typically used with graph losses and decoders. 40 """ 41 42 topo_type: str = "default" 43 topo_with_self_loops: bool = True 44 token_lm: Optional[Any] = None 45 intersect_pruned: bool = False 46 intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() 47 boost_coeff: float = 0.0 48 predictor_window_size: int = 0 49 predictor_step_size: int = 1 50 ... [end of nemo/collections/asr/parts/k2/classes.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] ... 101 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 102 num_workers: int = 1 # Number of workers for DataLoader 103 104 # The decoding scheme to be used for evaluation 105 decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] 106 107 # Beam Search hyperparameters 108 beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding 109 beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding 110 111 maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding 112 maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding 113 114 # HAT related parameters (only for internal lm subtraction) 115 hat_subtract_ilm: bool = False 116 hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) 117 118 decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) 119 120 121 # fmt: on 122 ... [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of nemo/core/config/modelPT.py] ... 44 # ... 45 train_ds: Optional[DatasetConfig] = None 46 validation_ds: Optional[DatasetConfig] = None 47 test_ds: Optional[DatasetConfig] = None 48 optim: Optional[OptimConfig] = None 49 50 51 @dataclass 52 class HydraConfig: 53 run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) 54 job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) 55 56 57 @dataclass 58 class NemoConfig: 59 name: str = MISSING 60 model: ModelConfig = MISSING 61 trainer: config.TrainerConfig = config.TrainerConfig( 62 strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' 63 ) 64 exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() 65 hydra: HydraConfig = HydraConfig() 66 67 68 class ModelConfigBuilder: 69 def __init__(self, model_cfg: ModelConfig): ... [end of nemo/core/config/modelPT.py] [start of nemo/collections/asr/models/configs/aligner_config.py] ... ... 21 class AlignerCTCConfig: 22 prob_suppress_index: int = -1 23 prob_suppress_value: float = 1.0 24 25 26 @dataclass 27 class AlignerRNNTConfig: 28 predictor_window_size: int = 0 29 predictor_step_size: int = 1 30 31 32 @dataclass 33 class AlignerWrapperModelConfig: 34 alignment_type: str = "forced" 35 word_output: bool = True 36 cpu_decoding: bool = False 37 decode_batch_size: int = 0 38 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() 39 rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() 40 41 42 @dataclass 43 class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): 44 decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() 45 [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] ... ... 239 240 if self.preserve_frame_confidence: 241 raise ValueError( 242 "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." 243 ) 244 245 return hypothesis 246 247 def __call__(self, *args, **kwargs): 248 return self.forward(*args, **kwargs) 249 250 251 @dataclass 252 class GreedyCTCInferConfig: 253 preserve_alignments: bool = False 254 compute_timestamps: bool = False 255 preserve_frame_confidence: bool = False 256 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 257 258 def __post_init__(self): 259 # OmegaConf.structured ensures that post_init check is always executed 260 self.confidence_method_cfg = OmegaConf.structured( ... [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/common/parts/adapter_modules.py] ... ... 146 def forward(self, x): 147 x = self.module(x) 148 149 # Add dropout if available 150 if self.dropout is not None: 151 x = self.dropout(x) 152 153 return x 154 155 156 @dataclass 157 class LinearAdapterConfig: 158 in_features: int 159 dim: int 160 activation: str = 'swish' 161 norm_position: str = 'pre' 162 dropout: float = 0.0 163 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 164 _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) 165 [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] ... ... 169 return super().forward(query, key, value, mask, pos_emb, cache=cache) 170 171 def reset_parameters(self): 172 with torch.no_grad(): 173 nn.init.zeros_(self.linear_out.weight) 174 nn.init.zeros_(self.linear_out.bias) 175 176 def get_default_strategy_config(self) -> 'dataclass': 177 return MHAResidualAddAdapterStrategyConfig() 178 179 180 @dataclass 181 class MultiHeadAttentionAdapterConfig: 182 n_head: int 183 n_feat: int 184 dropout_rate: float = 0.0 185 proj_dim: Optional[int] = None 186 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 187 _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) 188 189 190 class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): ... ... 273 # However: 274 # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 275 # For some reason at init sometimes it will cause the value of the tensor to become NaN 276 # All operations to compute matrix_ac and matrix_bd will then fail. 277 nn.init.zeros_(self.pos_bias_u) 278 nn.init.zeros_(self.pos_bias_v) 279 280 def get_default_strategy_config(self) -> 'dataclass': 281 return MHAResidualAddAdapterStrategyConfig() 282 283 284 @dataclass 285 class RelPositionMultiHeadAttentionAdapterConfig: 286 n_head: int 287 n_feat: int 288 dropout_rate: float = 0.0 289 proj_dim: Optional[int] = None 290 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 291 _target_: str = "{0}.{1}".format( 292 RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ 293 ) 294 ... ... 322 323 super().__init__( 324 d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, 325 ) 326 327 # Setup adapter strategy 328 self.setup_adapter_strategy(adapter_strategy) 329 330 def get_default_strategy_config(self) -> 'dataclass': 331 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 332 333 334 @dataclass 335 class PositionalEncodingAdapterConfig: 336 d_model: int 337 max_len: int = 5000 338 xscale: float = 1.0 339 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 340 _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) 341 342 343 class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): ... ... 364 xscale=1.0, 365 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 366 ): 367 super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) 368 369 # Setup adapter strategy 370 self.setup_adapter_strategy(adapter_strategy) 371 372 def get_default_strategy_config(self) -> 'dataclass': 373 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 374 375 376 @dataclass 377 class RelPositionalEncodingAdapterConfig: 378 d_model: int 379 max_len: int = 5000 380 xscale: float = 1.0 381 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 382 _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) 383 [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] ... 98 # Parameters for inference 99 acoustic_batch_size: int = 16 # The batch size to calculate log probabilities 100 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 101 device: str = "cuda" # The device to load the model onto to calculate log probabilities 102 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 103 104 # Beam Search hyperparameters 105 106 # The decoding scheme to be used for evaluation. 107 # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] 108 decoding_mode: str = "beamsearch_ngram" 109 110 beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding 111 beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding 112 beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding 113 114 decoding_strategy: str = "beam" 115 decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) 116 117 text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( 118 punctuation_marks = ".,?", 119 separate_punctuation = False, 120 do_lowercase = False, 121 rm_punctuation = False, 122 ) 123 # fmt: on 124 125 126 def beam_search_eval( ... [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] ... ... 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """Transformer based language model.""" 16 from MeCab import Model 17 from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule 18 from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule 19 from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( 20 MegatronRetrievalTransformerEncoderModule, 21 ) 22 from nemo.collections.nlp.modules.common.megatron.utils import ( 23 ApexGuardDefaults, 24 init_method_normal, 25 scaled_init_method_normal, 26 ) 27 28 try: 29 from apex.transformer.enums import AttnMaskType, ModelType 30 31 HAVE_APEX = True ... [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] ... 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 # Copyright 2017 Johns Hopkins University (Shinji Watanabe) 16 # 17 # Licensed under the Apache License, Version 2.0 (the "License"); 18 # you may not use this file except in compliance with the License. 19 # You may obtain a copy of the License at 20 # 21 # http://www.apache.org/licenses/LICENSE-2.0 22 # 23 # Unless required by applicable law or agreed to in writing, software 24 # distributed under the License is distributed on an "AS IS" BASIS, 25 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 26 # See the License for the specific language governing permissions and 27 # limitations under the License. 28 29 from dataclasses import dataclass 30 from typing import List, Optional, Tuple, Union 31 32 import numpy as np 33 import torch ... ... 2171 if self.preserve_frame_confidence: 2172 for batch_idx in range(batchsize): 2173 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2174 del hypotheses[batch_idx].frame_confidence[-1] 2175 2176 # Preserve states 2177 for batch_idx in range(batchsize): 2178 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2179 2180 return hypotheses 2181 2182 2183 @dataclass 2184 class GreedyRNNTInferConfig: 2185 max_symbols_per_step: Optional[int] = 10 2186 preserve_alignments: bool = False 2187 preserve_frame_confidence: bool = False 2188 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 2189 2190 def __post_init__(self): 2191 # OmegaConf.structured ensures that post_init check is always executed 2192 self.confidence_method_cfg = OmegaConf.structured( ... ... 2187 preserve_frame_confidence: bool = False 2188 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 2189 2190 def __post_init__(self): 2191 # OmegaConf.structured ensures that post_init check is always executed 2192 self.confidence_method_cfg = OmegaConf.structured( 2193 self.confidence_method_cfg 2194 if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) 2195 else ConfidenceMethodConfig(**self.confidence_method_cfg) 2196 ) 2197 2198 2199 @dataclass 2200 class GreedyBatchedRNNTInferConfig: 2201 max_symbols_per_step: Optional[int] = 10 2202 preserve_alignments: bool = False 2203 preserve_frame_confidence: bool = False 2204 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 2205 2206 def __post_init__(self): 2207 # OmegaConf.structured ensures that post_init check is always executed 2208 self.confidence_method_cfg = OmegaConf.structured( ... [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/tts/models/tacotron2.py] ... ... 39 LengthsType, 40 LogitsType, 41 MelSpectrogramType, 42 SequenceToSequenceAlignmentType, 43 ) 44 from nemo.core.neural_types.neural_type import NeuralType 45 from nemo.utils import logging, model_utils 46 47 48 @dataclass 49 class Preprocessor: 50 _target_: str = MISSING 51 pad_value: float = MISSING 52 53 54 @dataclass 55 class Tacotron2Config: 56 preprocessor: Preprocessor = Preprocessor() 57 encoder: Dict[Any, Any] = MISSING 58 decoder: Dict[Any, Any] = MISSING 59 postnet: Dict[Any, Any] = MISSING 60 labels: List = MISSING ... [end of nemo/collections/tts/models/tacotron2.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] ... 188 num_shards: int = -1 189 shuffle: bool = False 190 max_duration: Optional[float] = None 191 min_duration: Optional[float] = None 192 shuffle_seed: Optional[int] = None 193 sort_in_shards: bool = True 194 shard_manifests: bool = True 195 keep_files_together: bool = False 196 197 198 @dataclass 199 class ASRTarredDatasetMetadata: 200 created_datetime: Optional[str] = None 201 version: int = 0 202 num_samples_per_shard: Optional[int] = None 203 is_concatenated_manifest: bool = False 204 205 dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() 206 history: Optional[List[Any]] = field(default_factory=lambda: []) 207 208 def __post_init__(self): 209 self.created_datetime = self.get_current_datetime() ... [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] ... ... 58 bucketing_strategy: str = "synced_randomized" 59 bucketing_batch_size: Optional[Any] = None 60 bucketing_weights: Optional[List[int]] = None 61 62 63 @dataclass 64 class EncDecClassificationConfig(model_cfg.ModelConfig): 65 # Model global arguments 66 sample_rate: int = 16000 67 repeat: int = 1 68 dropout: float = 0.0 69 separable: bool = True 70 kernel_size_factor: float = 1.0 71 labels: List[str] = MISSING 72 timesteps: int = MISSING 73 74 # Dataset configs 75 train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 76 manifest_filepath=None, shuffle=True, trim_silence=False 77 ) 78 validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 79 manifest_filepath=None, shuffle=False 80 ) 81 test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 82 manifest_filepath=None, shuffle=False 83 ) 84 85 # Optimizer / Scheduler config 86 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 87 88 # Model component configs 89 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() 90 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 91 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 92 audio_length=timesteps 93 ) 94 95 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 96 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 97 98 99 @dataclass 100 class EncDecClassificationModelConfig(model_cfg.NemoConfig): 101 model: EncDecClassificationConfig = EncDecClassificationConfig() 102 [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] ... ... 161 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 162 163 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 164 When the alpha equals one, scaling is not applied to 'max_prob', 165 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 166 167 entropy_norm: A mapping of the entropy value to the interval [0,1]. 168 Supported values: 169 - 'lin' for using the linear mapping. 170 - 'exp' for using exponential mapping with linear shift. 171 """ 172 173 preserve_frame_confidence: bool = False 174 preserve_token_confidence: bool = False 175 preserve_word_confidence: bool = False 176 exclude_blank: bool = True 177 aggregation: str = "min" 178 method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() 179 180 def __post_init__(self): 181 # OmegaConf.structured ensures that post_init check is always executed 182 self.method_cfg = OmegaConf.structured( ... [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of scripts/confidence_ensembles/build_ensemble.py] ... 61 <all arguments as in the previous examples> 62 tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set 63 tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set 64 65 You can check the dataclasses in this file for the full list of supported 66 arguments and their default values. 67 """ 68 69 import atexit 70 71 # using default logging to be able to silence unnecessary messages from nemo 72 import logging 73 import os 74 import random 75 import sys 76 import tempfile 77 from copy import deepcopy 78 from dataclasses import dataclass 79 from pathlib import Path 80 from typing import Dict, List, Optional, Tuple 81 82 import joblib ... ... 195 class_weight: Tuple = (None, "balanced") 196 197 # increase if getting many warnings that algorithm didn't converge 198 max_iter: int = 1000 199 200 201 @dataclass 202 class BuildEnsembleConfig: 203 # where to save the resulting ensemble model 204 output_path: str = MISSING 205 206 # each model specification 207 ensemble: List[EnsembleConfig] = MISSING 208 209 random_seed: int = 0 # for reproducibility 210 211 # default confidence, can override 212 confidence: ConfidenceConfig = ConfidenceConfig( 213 # we keep frame confidences and apply aggregation manually to get full-utterance confidence 214 preserve_frame_confidence=True, 215 exclude_blank=True, 216 aggregation="mean", 217 method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False ... ... 215 exclude_blank=True, 216 aggregation="mean", 217 method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False 229 # used to specify what to tune over. By default runs tuning over some 230 # reasonalbe grid, so that it does not take forever. 231 # Can be changed as needed 232 tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() 233 234 # very fast to tune and can be important in case of imbalanced datasets 235 # will automatically set to False if dev data is not available 236 tune_logistic_regression: bool = True 237 tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() 238 239 def __post_init__(self): 240 """Checking that if any dev data is provided, all are provided. 241 ... [end of scripts/confidence_ensembles/build_ensemble.py] [start of examples/asr/experimental/k2/align_speech_parallel.py] ... 60 61 You may control the aligner's config by setting the aligner_args: 62 aligner_args.alignment_type=argmax \ 63 aligner_args.word_output=False \ 64 aligner_args.cpu_decoding=True \ 65 aligner_args.decode_batch_size=8 \ 66 aligner_args.ctc_cfg.prob_suppress_index=-1 \ 67 aligner_args.ctc_cfg.prob_suppress_value=0.5 \ 68 aligner_args.rnnt_cfg.predictor_window_size=10 \ 69 aligner_args.decoder_module_cfg.intersect_pruned=true \ 70 aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ 71 ... 72 73 """ 74 75 76 import os 77 from dataclasses import dataclass, is_dataclass 78 from typing import Optional 79 80 import pytorch_lightning as ptl 81 import torch ... ... 80 import pytorch_lightning as ptl 81 import torch 82 from omegaconf import MISSING, OmegaConf 83 84 from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter 85 from nemo.collections.asr.models import ASRModel 86 from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig 87 from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig 88 from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel 89 from nemo.core.config import TrainerConfig, hydra_runner 90 from nemo.utils import logging 91 from nemo.utils.get_rank import is_global_rank_zero 92 93 94 @dataclass 95 class ParallelAlignmentConfig: 96 model: Optional[str] = None # name 97 predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) 98 aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() 99 output_path: str = MISSING 100 model_stride: int = 8 101 102 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") 103 104 # there arguments will be ignored 105 return_predictions: bool = False 106 use_cer: bool = False ... [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] ... 93 94 95 # fmt: on 96 97 98 @dataclass 99 class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): 100 # Model global arguments 101 sample_rate: int = 16000 102 repeat: int = 1 103 dropout: float = 0.0 104 separable: bool = True 105 kernel_size_factor: float = 1.0 106 timesteps: int = 128 107 labels: List[str] = MISSING 108 109 # Dataset configs 110 train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 111 manifest_filepath=None, shuffle=True, trim_silence=False 112 ) 113 validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 114 manifest_filepath=None, shuffle=False 115 ) 116 test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 117 manifest_filepath=None, shuffle=False 118 ) 119 120 # Optimizer / Scheduler config 121 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 122 123 # Model general component configs 124 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) 125 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( 126 freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 127 ) 128 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 129 audio_length=128 130 ) 131 132 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 133 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 134 135 136 @dataclass 137 class MatchboxNetVADModelConfig(MatchboxNetModelConfig): ... [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] ... ... 201 of ``model.class_labels`` files.""" 202 203 204 @dataclass 205 class PunctuationCapitalizationModelConfig: 206 """ 207 A configuration of 208 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 209 model. 210 211 See an example of model config in 212 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 213 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 214 215 This config is a part of :class:`~PunctuationCapitalizationConfig`. 216 """ 217 218 class_labels: ClassLabelsConfig = ClassLabelsConfig() 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" ... ... 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" 229 230 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 231 """A configuration for creating validation datasets and data loaders.""" 232 233 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 234 """A configuration for creating test datasets and data loaders.""" 235 236 punct_head: HeadConfig = HeadConfig() 237 """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" 238 239 capit_head: HeadConfig = HeadConfig() 240 """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" 241 242 tokenizer: Any = TokenizerConfig() 243 """A configuration for source text tokenizer.""" 244 245 language_model: LanguageModelConfig = LanguageModelConfig() 246 """A configuration of a BERT-like language model which serves as a model body.""" 247 248 optim: Optional[Any] = None 249 """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For ... ... 297 """ 298 299 pretrained_model: Optional[str] = None 300 """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options 301 by calling method 302 :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. 303 """ 304 305 name: Optional[str] = 'Punctuation_and_Capitalization' 306 """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" 307 308 do_training: bool = True 309 """Whether to perform training of the model.""" 310 311 do_testing: bool = False 312 """Whether ot perform testing of the model.""" 313 314 model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() 315 """A configuration for the 316 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 317 model.""" 318 319 trainer: Optional[TrainerConfig] = TrainerConfig() 320 """Contains ``Trainer`` Lightning class constructor parameters.""" 321 322 exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) 323 """A configuration with various NeMo training options such as output directories, resuming from checkpoint, 324 tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" 325 326 327 @dataclass 328 class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): 329 model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() 330 331 332 def is_legacy_model_config(model_cfg: DictConfig) -> bool: 333 """ ... [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of tools/nemo_forced_aligner/align.py] ... 135 viterbi_device: Optional[str] = None 136 batch_size: int = 1 137 use_local_attention: bool = True 138 additional_segment_grouping_separator: Optional[str] = None 139 audio_filepath_parts_in_utt_id: int = 1 140 141 # Buffered chunked streaming configs 142 use_buffered_chunked_streaming: bool = False 143 chunk_len_in_secs: float = 1.6 144 total_buffer_in_secs: float = 4.0 145 chunk_batch_size: int = 32 146 147 # Cache aware streaming configs 148 simulate_cache_aware_streaming: Optional[bool] = False 149 150 # Output file configs 151 save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) 152 ctm_file_config: CTMFileConfig = CTMFileConfig() 153 ass_file_config: ASSFileConfig = ASSFileConfig() 154 155 156 @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) 157 def main(cfg: AlignmentConfig): ... [end of tools/nemo_forced_aligner/align.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] ... ... 160 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 161 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 162 ] 163 return config 164 # fmt: on 165 166 167 @dataclass 168 class JasperModelConfig(ctc_cfg.EncDecCTCConfig): 169 # Model global arguments 170 sample_rate: int = 16000 171 repeat: int = 1 172 dropout: float = 0.0 173 separable: bool = False 174 labels: List[str] = MISSING 175 176 # Dataset configs 177 train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( 178 manifest_filepath=None, shuffle=True, trim_silence=True 179 ) 180 validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 181 test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 182 183 # Optimizer / Scheduler config 184 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 185 186 # Model general component configs 187 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 188 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 189 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 190 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 191 192 193 @dataclass 194 class QuartzNetModelConfig(JasperModelConfig): ... [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] ... ... 53 return ' '.join(tokens) 54 55 def normalize(self, text) -> str: 56 # Normalization doesn't handle Japanese periods correctly; 57 # 'ใ€‚'becomes '.'. 58 if self.lang_id == 'en': 59 return self.normalizer.normalize(text) 60 else: 61 return text 62 63 64 class JaMecabProcessor: 65 """ 66 Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English 67 """ 68 69 def __init__(self): 70 self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") 71 72 def detokenize(self, text: List[str]) -> str: 73 RE_WS_IN_FW = re.compile( ... [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import re 17 from abc import abstractmethod 18 from dataclasses import dataclass, is_dataclass 19 from typing import Callable, Dict, List, Optional, Tuple, Union 20 21 import editdistance 22 import numpy as np ... ... 1285 1286 def compute(self): 1287 wer = self.scores.float() / self.words 1288 return wer, self.scores.detach(), self.words.detach() 1289 1290 1291 @dataclass 1292 class RNNTDecodingConfig: 1293 model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" 1294 strategy: str = "greedy_batch" 1295 1296 compute_hypothesis_token_set: bool = False 1297 1298 # preserve decoding alignments 1299 preserve_alignments: Optional[bool] = None 1300 1301 # confidence config 1302 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 ... ... 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 1307 # compute RNNT time stamps 1308 compute_timestamps: Optional[bool] = None 1309 1310 # compute language IDs 1311 compute_langs: bool = False 1312 1313 # token representing word seperator 1314 word_seperator: str = " " 1315 1316 # type of timestamps to calculate 1317 rnnt_timestamp_type: str = "all" # can be char, word or all for both 1318 1319 # greedy decoding config 1320 greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() 1321 1322 # beam decoding config 1323 beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) 1324 1325 # can be used to change temperature for decoding 1326 temperature: float = 1.0 1327 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] ... 620 return augmented_spec 621 622 623 class CropOrPadSpectrogramAugmentation(NeuralModule): 624 """ 625 Pad or Crop the incoming Spectrogram to a certain shape. 626 627 Args: 628 audio_length (int): the final number of timesteps that is required. 629 The signal will be either padded or cropped temporally to this 630 size. 631 """ 632 633 def __init__(self, audio_length): 634 super(CropOrPadSpectrogramAugmentation, self).__init__() 635 self.audio_length = audio_length 636 637 @typecheck() 638 @torch.no_grad() 639 def forward(self, input_signal, length): 640 image = input_signal ... [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] ... ... 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig 18 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig 19 from nemo.core.config.modelPT import NemoConfig 20 21 22 @dataclass 23 class GraphModuleConfig: 24 criterion_type: str = "ml" 25 loss_type: str = "ctc" 26 split_batch_size: int = 0 27 dec_type: str = "topo" 28 transcribe_training: bool = True 29 backend_cfg: BackendConfig = BackendConfig() 30 31 32 @dataclass 33 class EncDecK2SeqConfig(EncDecCTCConfig): 34 graph_module_cfg: GraphModuleConfig = GraphModuleConfig() 35 36 37 @dataclass 38 class EncDecK2SeqModelConfig(NemoConfig): 39 model: EncDecK2SeqConfig = EncDecK2SeqConfig() 40 [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] ... ... 110 pretrained_name: Optional[str] = None # Name of a pretrained model 111 dataset_manifest: str = MISSING 112 output_dir: str = MISSING 113 114 # General configs 115 batch_size: int = 32 116 num_workers: int = 4 117 118 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA 119 # device anyway, and do inference on CPU only if CUDA device is not found. 120 # If `cuda` is a negative number, inference will be on CPU only. 121 cuda: Optional[int] = None 122 amp: bool = False 123 audio_type: str = "wav" 124 125 # Confidence configs 126 target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) 127 confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) 128 grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on 129 130 131 @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) ... [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] ... ... 32 from nemo.core.config.modelPT import OptimConfig, SchedConfig 33 34 35 @dataclass 36 class MTSchedConfig(SchedConfig): 37 name: str = 'InverseSquareRootAnnealing' 38 warmup_ratio: Optional[float] = None 39 last_epoch: int = -1 40 41 42 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). 43 @dataclass 44 class MTOptimConfig(OptimConfig): 45 name: str = 'adam' 46 lr: float = 1e-3 47 betas: Tuple[float, float] = (0.9, 0.98) 48 weight_decay: float = 0.0 49 sched: Optional[MTSchedConfig] = MTSchedConfig() 50 51 52 @dataclass 53 class MTEncDecModelConfig(EncDecNLPModelConfig): ... ... 60 len_pen: float = 0.0 61 src_language: Any = 'en' # Any = str or List[str] 62 tgt_language: Any = 'en' # Any = str or List[str] 63 find_unused_parameters: Optional[bool] = True 64 shared_tokenizer: Optional[bool] = True 65 multilingual: Optional[bool] = False 66 preproc_out_dir: Optional[str] = None 67 validate_input_ids: Optional[bool] = True 68 shared_embeddings: bool = False 69 70 # network architecture configuration 71 encoder_tokenizer: Any = MISSING 72 encoder: Any = MISSING 73 74 decoder_tokenizer: Any = MISSING 75 decoder: Any = MISSING 76 77 head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) 78 79 # dataset configurations 80 train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 81 src_file_name=MISSING, 82 tgt_file_name=MISSING, 83 tokens_in_batch=512, 84 clean=True, 85 shuffle=True, 86 cache_ids=False, 87 use_cache=False, 88 ) 89 validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 90 src_file_name=MISSING, 91 tgt_file_name=MISSING, 92 tokens_in_batch=512, 93 clean=False, 94 shuffle=False, 95 cache_ids=False, 96 use_cache=False, 97 ) 98 test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 99 src_file_name=MISSING, 100 tgt_file_name=MISSING, 101 tokens_in_batch=512, 102 clean=False, 103 shuffle=False, 104 cache_ids=False, 105 use_cache=False, 106 ) 107 optim: Optional[OptimConfig] = MTOptimConfig() 108 109 110 @dataclass 111 class AAYNBaseConfig(MTEncDecModelConfig): 112 113 # Attention is All You Need Base Configuration 114 encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 115 decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 116 117 encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( 118 library='nemo', 119 model_name=None, 120 pretrained=False, 121 hidden_size=512, 122 inner_size=2048, 123 num_layers=6, 124 num_attention_heads=8, 125 ffn_dropout=0.1, 126 attn_score_dropout=0.1, 127 attn_layer_dropout=0.1, 128 ) 129 130 decoder: NeMoTransformerConfig = NeMoTransformerConfig( 131 library='nemo', 132 model_name=None, 133 pretrained=False, 134 hidden_size=512, 135 inner_size=2048, 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass ... ... 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass 145 class MTBottleneckModelConfig(AAYNBaseConfig): 146 model_type: str = 'nll' 147 min_logv: float = -6 148 latent_size: int = -1 # -1 will take value of encoder hidden 149 non_recon_warmup_batches: int = 200000 150 recon_per_token: bool = True 151 log_timing: bool = True 152 153 encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( 154 library='nemo', 155 model_name=None, 156 pretrained=False, 157 hidden_size=512, 158 inner_size=2048, 159 num_layers=6, 160 num_attention_heads=8, 161 ffn_dropout=0.1, 162 attn_score_dropout=0.1, 163 attn_layer_dropout=0.1, 164 arch='seq2seq', 165 hidden_steps=32, 166 hidden_blocks=1, 167 hidden_init_method='params', 168 ) 169 170 decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( 171 library='nemo', 172 model_name=None, 173 pretrained=False, 174 inner_size=2048, 175 num_layers=6, 176 num_attention_heads=8, 177 ffn_dropout=0.1, 178 attn_score_dropout=0.1, 179 attn_layer_dropout=0.1, 180 arch='seq2seq', 181 ) 182 [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/utils/exp_manager.py] ... 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import glob 16 import os 17 import subprocess 18 import sys 19 import time 20 import warnings 21 from dataclasses import dataclass 22 from datetime import timedelta 23 from pathlib import Path 24 from shutil import copy, move 25 from typing import Any, Dict, List, Optional, Tuple, Union ... ... 132 133 # Log dir creation parameters 134 explicit_log_dir: Optional[str] = None 135 exp_dir: Optional[str] = None 136 name: Optional[str] = None 137 version: Optional[str] = None 138 use_datetime_version: Optional[bool] = True 139 resume_if_exists: Optional[bool] = False 140 resume_past_end: Optional[bool] = False 141 resume_ignore_no_checkpoint: Optional[bool] = False 142 resume_from_checkpoint: Optional[str] = None 143 # Logging parameters 144 create_tensorboard_logger: Optional[bool] = True 145 summary_writer_kwargs: Optional[Dict[Any, Any]] = None 146 create_wandb_logger: Optional[bool] = False 147 wandb_logger_kwargs: Optional[Dict[Any, Any]] = None 148 create_mlflow_logger: Optional[bool] = False 149 mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() 150 create_dllogger_logger: Optional[bool] = False 151 dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() 152 create_clearml_logger: Optional[bool] = False 153 clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() 154 # Checkpointing parameters 155 create_checkpoint_callback: Optional[bool] = True 156 checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() 157 create_early_stopping_callback: Optional[bool] = False 158 early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() 159 create_preemption_callback: Optional[bool] = True 160 # Additional exp_manager arguments 161 files_to_copy: Optional[List[str]] = None 162 # logs timing of train/val/test steps 163 log_step_timing: Optional[bool] = True 164 step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() 165 # Configures creation of log files for different ranks 166 log_local_rank_0_only: Optional[bool] = False 167 log_global_rank_0_only: Optional[bool] = False 168 # disable initial validation when resuming from a checkpoint saved during validation 169 disable_validation_on_resume: Optional[bool] = True 170 ema: Optional[EMAParams] = EMAParams() 171 # Wall clock time limit 172 max_time_per_run: Optional[str] = None 173 # time to sleep non 0 ranks during initialization 174 seconds_to_sleep: float = 5 ... [end of nemo/utils/exp_manager.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] ... ... 588 beam_threshold: float = 20.0 589 unk_weight: float = -math.inf 590 sil_weight: float = 0.0 591 592 593 @dataclass 594 class BeamCTCInferConfig: 595 beam_size: int 596 search_type: str = 'default' 597 preserve_alignments: bool = False 598 compute_timestamps: bool = False 599 return_best_hypothesis: bool = True 600 601 beam_alpha: float = 1.0 602 beam_beta: float = 0.0 603 kenlm_path: Optional[str] = None 604 605 flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() 606 pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() 607 [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
15db83ec4a65e649d83b61d7a4a58d911586e853
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-10-03T19:14:38Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,7 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,7 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed @@ -2201,7 +2201,7 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -175,7 +175,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() + method_cfg: ConfidenceMethodConfig = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -118,7 +118,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
slackapi__python-slack-events-api-71
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): </issue> <code> [start of README.rst] 1 Slack Events API adapter for Python 2 =================================== 3 4 .. image:: https://badge.fury.io/py/slackeventsapi.svg 5 :target: https://pypi.org/project/slackeventsapi/ 6 .. image:: https://travis-ci.org/slackapi/python-slack-events-api.svg?branch=master 7 :target: https://travis-ci.org/slackapi/python-slack-events-api 8 .. image:: https://codecov.io/gh/slackapi/python-slack-events-api/branch/master/graph/badge.svg 9 :target: https://codecov.io/gh/slackapi/python-slack-events-api 10 11 12 The Slack Events Adapter is a Python-based solution to receive and parse events 13 from Slackโ€™s Events API. This library uses an event emitter framework to allow 14 you to easily process Slack events by simply attaching functions 15 to event listeners. 16 17 This adapter enhances and simplifies Slack's Events API by incorporating useful best practices, patterns, and opportunities to abstract out common tasks. 18 19 ๐Ÿ’ก We wrote a `blog post which explains how`_ the Events API can help you, why we built these tools, and how you can use them to build production-ready Slack apps. 20 21 .. _blog post which explains how: https://medium.com/@SlackAPI/enhancing-slacks-events-api-7535827829ab 22 23 24 ๐Ÿค– Installation 25 ---------------- 26 27 .. code:: shell 28 29 pip install slackeventsapi 30 31 ๐Ÿค– App Setup 32 -------------------- 33 34 Before you can use the `Events API`_ you must 35 `create a Slack App`_, and turn on 36 `Event Subscriptions`_. 37 38 ๐Ÿ’ก When you add the Request URL to your app's Event Subscription settings, 39 Slack will send a request containing a `challenge` code to verify that your 40 server is alive. This package handles that URL Verification event for you, so 41 all you need to do is start the example app, start ngrok and configure your 42 URL accordingly. 43 44 โœ… Once you have your `Request URL` verified, your app is ready to start 45 receiving Team Events. 46 47 ๐Ÿ”‘ Your server will begin receiving Events from Slack's Events API as soon as a 48 user has authorized your app. 49 50 ๐Ÿค– Development workflow: 51 =========================== 52 53 (1) Create a Slack app on https://api.slack.com/apps 54 (2) Add a `bot user` for your app 55 (3) Start the example app on your **Request URL** endpoint 56 (4) Start ngrok and copy the **HTTPS** URL 57 (5) Add your **Request URL** and subscribe your app to events 58 (6) Go to your ngrok URL (e.g. https://myapp12.ngrok.com/) and auth your app 59 60 **๐ŸŽ‰ Once your app has been authorized, you will begin receiving Slack Events** 61 62 โš ๏ธ Ngrok is a great tool for developing Slack apps, but we don't recommend using ngrok 63 for production apps. 64 65 ๐Ÿค– Usage 66 ---------- 67 **โš ๏ธ Keep your app's credentials safe!** 68 69 - For development, keep them in virtualenv variables. 70 71 - For production, use a secure data store. 72 73 - Never post your app's credentials to github. 74 75 .. code:: python 76 77 SLACK_SIGNING_SECRET = os.environ["SLACK_SIGNING_SECRET"] 78 79 Create a Slack Event Adapter for receiving actions via the Events API 80 ----------------------------------------------------------------------- 81 **Using the built-in Flask server:** 82 83 .. code:: python 84 85 from slackeventsapi import SlackEventAdapter 86 87 88 slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, endpoint="/slack/events") 89 90 91 # Create an event listener for "reaction_added" events and print the emoji name 92 @slack_events_adapter.on("reaction_added") 93 def reaction_added(event_data): 94 emoji = event_data["event"]["reaction"] 95 print(emoji) 96 97 98 # Start the server on port 3000 99 slack_events_adapter.start(port=3000) 100 101 102 **Using your existing Flask instance:** 103 104 105 .. code:: python 106 107 from flask import Flask 108 from slackeventsapi import SlackEventAdapter 109 110 111 # This `app` represents your existing Flask app 112 app = Flask(__name__) 113 114 115 # An example of one of your Flask app's routes 116 @app.route("/") 117 def hello(): 118 return "Hello there!" 119 120 121 # Bind the Events API route to your existing Flask app by passing the server 122 # instance as the last param, or with `server=app`. 123 slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, "/slack/events", app) 124 125 126 # Create an event listener for "reaction_added" events and print the emoji name 127 @slack_events_adapter.on("reaction_added") 128 def reaction_added(event_data): 129 emoji = event_data["event"]["reaction"] 130 print(emoji) 131 132 133 # Start the server on port 3000 134 if __name__ == "__main__": 135 app.run(port=3000) 136 137 For a comprehensive list of available Slack `Events` and more information on 138 `Scopes`, see https://api.slack.com/events-api 139 140 ๐Ÿค– Example event listeners 141 ----------------------------- 142 143 See `example.py`_ for usage examples. This example also utilizes the 144 SlackClient Web API client. 145 146 .. _example.py: /example/ 147 148 ๐Ÿค” Support 149 ----------- 150 151 Need help? Join `Slack Community`_ and talk to us in `#slack-api`_. 152 153 You can also `create an Issue`_ right here on GitHub. 154 155 .. _Events API: https://api.slack.com/events-api 156 .. _create a Slack App: https://api.slack.com/apps/new 157 .. _Event Subscriptions: https://api.slack.com/events-api#subscriptions 158 .. _Slack Community: http://slackcommunity.com/ 159 .. _#slack-api: https://dev4slack.slack.com/messages/slack-api/ 160 .. _create an Issue: https://github.com/slackapi/python-slack-events-api/issues/new 161 [end of README.rst] [start of slackeventsapi/server.py] ... ... 4 import sys 5 import hmac 6 import hashlib 7 from time import time 8 from .version import __version__ 9 10 11 class SlackServer(Flask): 12 def __init__(self, signing_secret, endpoint, emitter, server): 13 self.signing_secret = signing_secret 14 self.emitter = emitter 15 self.endpoint = endpoint 16 self.package_info = self.get_package_info() 17 18 # If a server is passed in, bind the event handler routes to it, 19 # otherwise create a new Flask instance. 20 if server: 21 if isinstance(server, Flask) or isinstance(server, Blueprint): 22 self.bind_route(server) 23 else: 24 raise TypeError("Server must be an instance of Flask or Blueprint") 25 else: 26 Flask.__init__(self, __name__) 27 self.bind_route(self) 28 ... [end of slackeventsapi/server.py] [start of /dev/null] ... [end of /dev/null] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
slackapi/python-slack-events-api
0c0ce604b502508622fb14c278a0d64841fa32e3
Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s):
2020-06-12T06:58:10Z
<patch> <patch> diff --git a/example/current_app/main.py b/example/current_app/main.py new file mode 100644 --- /dev/null +++ b/example/current_app/main.py @@ -0,0 +1,49 @@ +# ------------------ +# Only for running this script here +import sys +from os.path import dirname +sys.path.insert(1, f"{dirname(__file__)}/../..") +# ------------------ + +import os +from slack import WebClient +import logging +logging.basicConfig(level=logging.DEBUG) + +from flask import Flask + +app = Flask(__name__) + +with app.app_context(): + from test_module.slack_app import slack_events_adapter + + slack_bot_token = os.environ["SLACK_BOT_TOKEN"] + slack_client = WebClient(slack_bot_token) + + + @slack_events_adapter.on("message") + def handle_message(event_data): + message = event_data["event"] + if message.get("subtype") is None and "hi" in message.get('text'): + channel = message["channel"] + message = "Hi <@%s>! :tada:" % message["user"] + slack_client.chat_postMessage(channel=channel, text=message) + + + @slack_events_adapter.on("error") + def error_handler(err): + print("ERROR: " + str(err)) + +# (Terminal A) +# source env/bin/activate +# (env) $ export SLACK_BOT_TOKEN=xoxb-*** +# (env) $ export SLACK_SIGNING_SECRET=** +# (env) $ cd example/current_app +# (env) $ FLASK_APP=main.py FLASK_ENV=development flask run --port 3000 + +# (Terminal B) +# ngrok http 3000 + +# in Slack +# /invite @{your app's bot user} +# post a message "hi" in the channel diff --git a/slackeventsapi/server.py b/slackeventsapi/server.py --- a/slackeventsapi/server.py +++ b/slackeventsapi/server.py @@ -1,10 +1,13 @@ -from flask import Flask, request, make_response, Blueprint +import hashlib +import hmac import json import platform import sys -import hmac -import hashlib from time import time + +from flask import Flask, request, make_response, Blueprint +from werkzeug.local import LocalProxy + from .version import __version__ @@ -18,10 +21,10 @@ def __init__(self, signing_secret, endpoint, emitter, server): # If a server is passed in, bind the event handler routes to it, # otherwise create a new Flask instance. if server: - if isinstance(server, Flask) or isinstance(server, Blueprint): + if isinstance(server, (Flask, Blueprint, LocalProxy)): self.bind_route(server) else: - raise TypeError("Server must be an instance of Flask or Blueprint") + raise TypeError("Server must be an instance of Flask, Blueprint, or LocalProxy") else: Flask.__init__(self, __name__) self.bind_route(self) </patch> </s> </patch>
diff --git a/example/current_app/test_module/__init__.py b/example/current_app/test_module/__init__.py new file mode 100644 diff --git a/example/current_app/test_module/slack_app.py b/example/current_app/test_module/slack_app.py new file mode 100644 --- /dev/null +++ b/example/current_app/test_module/slack_app.py @@ -0,0 +1,16 @@ +# ------------------ +# Only for running this script here +import logging +import sys +from os.path import dirname + +sys.path.insert(1, f"{dirname(__file__)}/../../..") +logging.basicConfig(level=logging.DEBUG) +# ------------------ + +from flask import current_app as app +from slackeventsapi import SlackEventAdapter +import os + +slack_signing_secret = os.environ["SLACK_SIGNING_SECRET"] +slack_events_adapter = SlackEventAdapter(slack_signing_secret, "/slack/events", app) diff --git a/tests/test_server.py b/tests/test_server.py --- a/tests/test_server.py +++ b/tests/test_server.py @@ -18,7 +18,7 @@ def test_server_not_flask(): with pytest.raises(TypeError) as e: invalid_flask = "I am not a Flask" SlackEventAdapter("SIGNING_SECRET", "/slack/events", invalid_flask) - assert e.value.args[0] == 'Server must be an instance of Flask or Blueprint' + assert e.value.args[0] == 'Server must be an instance of Flask, Blueprint, or LocalProxy' def test_blueprint_server():
1.0
celery__celery-2598
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ``` </issue> <code> [start of README.rst] 1 ================================= 2 celery - Distributed Task Queue 3 ================================= 4 5 .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png 6 7 |build-status| |coverage-status| 8 9 :Version: 3.2.0a1 (Cipater) 10 :Web: http://celeryproject.org/ 11 :Download: http://pypi.python.org/pypi/celery/ 12 :Source: http://github.com/celery/celery/ 13 :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, 14 python, webhooks, queue, distributed 15 16 -- 17 18 What is a Task Queue? 19 ===================== 20 21 Task queues are used as a mechanism to distribute work across threads or 22 machines. 23 24 A task queue's input is a unit of work, called a task, dedicated worker 25 processes then constantly monitor the queue for new work to perform. 26 27 Celery communicates via messages, usually using a broker 28 to mediate between clients and workers. To initiate a task a client puts a 29 message on the queue, the broker then delivers the message to a worker. 30 31 A Celery system can consist of multiple workers and brokers, giving way 32 to high availability and horizontal scaling. 33 34 Celery is a library written in Python, but the protocol can be implemented in 35 any language. So far there's RCelery_ for the Ruby programming language, and a 36 `PHP client`, but language interoperability can also be achieved 37 by using webhooks. 38 39 .. _RCelery: https://github.com/leapfrogonline/rcelery 40 .. _`PHP client`: https://github.com/gjedeer/celery-php 41 .. _`using webhooks`: 42 http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html 43 44 What do I need? 45 =============== 46 47 Celery version 3.0 runs on, 48 49 - Python (2.6, 2.7, 3.3, 3.4) 50 - PyPy (1.8, 1.9) 51 - Jython (2.5, 2.7). 52 53 This is the last version to support Python 2.5, 54 and from Celery 3.1, Python 2.6 or later is required. 55 The last version to support Python 2.4 was Celery series 2.2. 56 57 *Celery* is usually used with a message broker to send and receive messages. 58 The RabbitMQ, Redis transports are feature complete, 59 but there's also experimental support for a myriad of other solutions, including 60 using SQLite for local development. 61 62 *Celery* can run on a single machine, on multiple machines, or even 63 across datacenters. 64 65 Get Started 66 =========== 67 68 If this is the first time you're trying to use Celery, or you are 69 new to Celery 3.0 coming from previous versions then you should read our 70 getting started tutorials: 71 72 - `First steps with Celery`_ 73 74 Tutorial teaching you the bare minimum needed to get started with Celery. 75 76 - `Next steps`_ 77 78 A more complete overview, showing more features. 79 80 .. _`First steps with Celery`: 81 http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html 82 83 .. _`Next steps`: 84 http://docs.celeryproject.org/en/latest/getting-started/next-steps.html 85 86 Celery is... 87 ========== 88 89 - **Simple** 90 91 Celery is easy to use and maintain, and does *not need configuration files*. 92 93 It has an active, friendly community you can talk to for support, 94 including a `mailing-list`_ and and an IRC channel. 95 96 Here's one of the simplest applications you can make:: 97 98 from celery import Celery 99 100 app = Celery('hello', broker='amqp://guest@localhost//') 101 102 @app.task 103 def hello(): 104 return 'hello world' 105 106 - **Highly Available** 107 108 Workers and clients will automatically retry in the event 109 of connection loss or failure, and some brokers support 110 HA in way of *Master/Master* or *Master/Slave* replication. 111 112 - **Fast** 113 114 A single Celery process can process millions of tasks a minute, 115 with sub-millisecond round-trip latency (using RabbitMQ, 116 py-librabbitmq, and optimized settings). 117 118 - **Flexible** 119 120 Almost every part of *Celery* can be extended or used on its own, 121 Custom pool implementations, serializers, compression schemes, logging, 122 schedulers, consumers, producers, autoscalers, broker transports and much more. 123 124 It supports... 125 ============ 126 127 - **Message Transports** 128 129 - RabbitMQ_, Redis_, 130 - MongoDB_ (experimental), Amazon SQS (experimental), 131 - CouchDB_ (experimental), SQLAlchemy_ (experimental), 132 - Django ORM (experimental), `IronMQ`_ 133 - and more... 134 135 - **Concurrency** 136 137 - Prefork, Eventlet_, gevent_, threads/single threaded 138 139 - **Result Stores** 140 141 - AMQP, Redis 142 - memcached, MongoDB 143 - SQLAlchemy, Django ORM 144 - Apache Cassandra, IronCache 145 146 - **Serialization** 147 148 - *pickle*, *json*, *yaml*, *msgpack*. 149 - *zlib*, *bzip2* compression. 150 - Cryptographic message signing. 151 152 .. _`Eventlet`: http://eventlet.net/ 153 .. _`gevent`: http://gevent.org/ 154 155 .. _RabbitMQ: http://rabbitmq.com 156 .. _Redis: http://redis.io 157 .. _MongoDB: http://mongodb.org 158 .. _Beanstalk: http://kr.github.com/beanstalkd 159 .. _CouchDB: http://couchdb.apache.org 160 .. _SQLAlchemy: http://sqlalchemy.org 161 .. _`IronMQ`: http://iron.io 162 163 Framework Integration 164 ===================== 165 166 Celery is easy to integrate with web frameworks, some of which even have 167 integration packages: 168 169 +--------------------+------------------------+ 170 | `Django`_ | not needed | 171 +--------------------+------------------------+ 172 | `Pyramid`_ | `pyramid_celery`_ | 173 +--------------------+------------------------+ 174 | `Pylons`_ | `celery-pylons`_ | 175 +--------------------+------------------------+ 176 | `Flask`_ | not needed | 177 +--------------------+------------------------+ 178 | `web2py`_ | `web2py-celery`_ | 179 +--------------------+------------------------+ 180 | `Tornado`_ | `tornado-celery`_ | 181 +--------------------+------------------------+ 182 183 The integration packages are not strictly necessary, but they can make 184 development easier, and sometimes they add important hooks like closing 185 database connections at ``fork``. 186 187 .. _`Django`: http://djangoproject.com/ 188 .. _`Pylons`: http://www.pylonsproject.org/ 189 .. _`Flask`: http://flask.pocoo.org/ 190 .. _`web2py`: http://web2py.com/ 191 .. _`Bottle`: http://bottlepy.org/ 192 .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html 193 .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ 194 .. _`django-celery`: http://pypi.python.org/pypi/django-celery 195 .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons 196 .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ 197 .. _`Tornado`: http://www.tornadoweb.org/ 198 .. _`tornado-celery`: http://github.com/mher/tornado-celery/ 199 200 .. _celery-documentation: 201 202 Documentation 203 ============= 204 205 The `latest documentation`_ with user guides, tutorials and API reference 206 is hosted at Read The Docs. 207 208 .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ 209 210 .. _celery-installation: 211 212 Installation 213 ============ 214 215 You can install Celery either via the Python Package Index (PyPI) 216 or from source. 217 218 To install using `pip`,:: 219 220 $ pip install -U Celery 221 222 To install using `easy_install`,:: 223 224 $ easy_install -U Celery 225 226 .. _bundles: 227 228 Bundles 229 ------- 230 231 Celery also defines a group of bundles that can be used 232 to install Celery and the dependencies for a given feature. 233 234 You can specify these in your requirements or on the ``pip`` comand-line 235 by using brackets. Multiple bundles can be specified by separating them by 236 commas. 237 :: 238 239 $ pip install "celery[librabbitmq]" 240 241 $ pip install "celery[librabbitmq,redis,auth,msgpack]" 242 243 The following bundles are available: 244 245 Serializers 246 ~~~~~~~~~~~ 247 248 :celery[auth]: 249 for using the auth serializer. 250 251 :celery[msgpack]: 252 for using the msgpack serializer. 253 254 :celery[yaml]: 255 for using the yaml serializer. 256 257 Concurrency 258 ~~~~~~~~~~~ 259 260 :celery[eventlet]: 261 for using the eventlet pool. 262 263 :celery[gevent]: 264 for using the gevent pool. 265 266 :celery[threads]: 267 for using the thread pool. 268 269 Transports and Backends 270 ~~~~~~~~~~~~~~~~~~~~~~~ 271 272 :celery[librabbitmq]: 273 for using the librabbitmq C library. 274 275 :celery[redis]: 276 for using Redis as a message transport or as a result backend. 277 278 :celery[mongodb]: 279 for using MongoDB as a message transport (*experimental*), 280 or as a result backend (*supported*). 281 282 :celery[sqs]: 283 for using Amazon SQS as a message transport (*experimental*). 284 285 :celery[memcache]: 286 for using memcached as a result backend. 287 288 :celery[cassandra]: 289 for using Apache Cassandra as a result backend. 290 291 :celery[couchdb]: 292 for using CouchDB as a message transport (*experimental*). 293 294 :celery[couchbase]: 295 for using CouchBase as a result backend. 296 297 :celery[beanstalk]: 298 for using Beanstalk as a message transport (*experimental*). 299 300 :celery[zookeeper]: 301 for using Zookeeper as a message transport. 302 303 :celery[zeromq]: 304 for using ZeroMQ as a message transport (*experimental*). 305 306 :celery[sqlalchemy]: 307 for using SQLAlchemy as a message transport (*experimental*), 308 or as a result backend (*supported*). 309 310 :celery[pyro]: 311 for using the Pyro4 message transport (*experimental*). 312 313 :celery[slmq]: 314 for using the SoftLayer Message Queue transport (*experimental*). 315 316 .. _celery-installing-from-source: 317 318 Downloading and installing from source 319 -------------------------------------- 320 321 Download the latest version of Celery from 322 http://pypi.python.org/pypi/celery/ 323 324 You can install it by doing the following,:: 325 326 $ tar xvfz celery-0.0.0.tar.gz 327 $ cd celery-0.0.0 328 $ python setup.py build 329 # python setup.py install 330 331 The last command must be executed as a privileged user if 332 you are not currently using a virtualenv. 333 334 .. _celery-installing-from-git: 335 336 Using the development version 337 ----------------------------- 338 339 With pip 340 ~~~~~~~~ 341 342 The Celery development version also requires the development 343 versions of ``kombu``, ``amqp`` and ``billiard``. 344 345 You can install the latest snapshot of these using the following 346 pip commands:: 347 348 $ pip install https://github.com/celery/celery/zipball/master#egg=celery 349 $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard 350 $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp 351 $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu 352 353 With git 354 ~~~~~~~~ 355 356 Please the Contributing section. 357 358 .. _getting-help: 359 360 Getting Help 361 ============ 362 363 .. _mailing-list: 364 365 Mailing list 366 ------------ 367 368 For discussions about the usage, development, and future of celery, 369 please join the `celery-users`_ mailing list. 370 371 .. _`celery-users`: http://groups.google.com/group/celery-users/ 372 373 .. _irc-channel: 374 375 IRC 376 --- 377 378 Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ 379 network. 380 381 .. _`Freenode`: http://freenode.net 382 383 .. _bug-tracker: 384 385 Bug tracker 386 =========== 387 388 If you have any suggestions, bug reports or annoyances please report them 389 to our issue tracker at http://github.com/celery/celery/issues/ 390 391 .. _wiki: 392 393 Wiki 394 ==== 395 396 http://wiki.github.com/celery/celery/ 397 398 .. _contributing-short: 399 400 Contributing 401 ============ 402 403 Development of `celery` happens at Github: http://github.com/celery/celery 404 405 You are highly encouraged to participate in the development 406 of `celery`. If you don't like Github (for some reason) you're welcome 407 to send regular patches. 408 409 Be sure to also read the `Contributing to Celery`_ section in the 410 documentation. 411 412 .. _`Contributing to Celery`: 413 http://docs.celeryproject.org/en/master/contributing.html 414 415 .. _license: 416 417 License 418 ======= 419 420 This software is licensed under the `New BSD License`. See the ``LICENSE`` 421 file in the top distribution directory for the full license text. 422 423 .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround 424 425 426 .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png 427 :alt: Bitdeli badge 428 :target: https://bitdeli.com/free 429 430 .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master 431 :target: https://travis-ci.org/celery/celery 432 .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg 433 :target: https://coveralls.io/r/celery/celery 434 [end of README.rst] [start of celery/backends/amqp.py] ... 181 return payload 182 else: 183 # no new state, use previous 184 try: 185 return self._cache[task_id] 186 except KeyError: 187 # result probably pending. 188 return {'status': states.PENDING, 'result': None} 189 poll = get_task_meta # XXX compat 190 191 def drain_events(self, connection, consumer, 192 timeout=None, on_interval=None, now=monotonic, wait=None): 193 wait = wait or connection.drain_events 194 results = {} 195 196 def callback(meta, message): 197 if meta['status'] in states.READY_STATES: 198 results[meta['task_id']] = meta 199 200 consumer.callbacks[:] = [callback] 201 time_start = now() 202 ... [end of celery/backends/amqp.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
celery/celery
6592ff64b6b024a4b68abcc53b151888fdf0dee3
CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ```
This is biting me as well. Any news?
2015-04-29T14:52:17Z
<patch> <patch> diff --git a/celery/backends/amqp.py b/celery/backends/amqp.py --- a/celery/backends/amqp.py +++ b/celery/backends/amqp.py @@ -195,7 +195,7 @@ def drain_events(self, connection, consumer, def callback(meta, message): if meta['status'] in states.READY_STATES: - results[meta['task_id']] = meta + results[meta['task_id']] = self.meta_from_decoded(meta) consumer.callbacks[:] = [callback] time_start = now() </patch> </s> </patch>
diff --git a/celery/tests/backends/test_amqp.py b/celery/tests/backends/test_amqp.py --- a/celery/tests/backends/test_amqp.py +++ b/celery/tests/backends/test_amqp.py @@ -13,6 +13,7 @@ from celery.backends.amqp import AMQPBackend from celery.exceptions import TimeoutError from celery.five import Empty, Queue, range +from celery.result import AsyncResult from celery.utils import uuid from celery.tests.case import ( @@ -246,10 +247,20 @@ def test_wait_for(self): with self.assertRaises(TimeoutError): b.wait_for(tid, timeout=0.01, cache=False) - def test_drain_events_remaining_timeouts(self): + def test_drain_events_decodes_exceptions_in_meta(self): + tid = uuid() + b = self.create_backend(serializer="json") + b.store_result(tid, RuntimeError("aap"), states.FAILURE) + result = AsyncResult(tid, backend=b) - class Connection(object): + with self.assertRaises(Exception) as cm: + result.get() + self.assertEqual(cm.exception.__class__.__name__, "RuntimeError") + self.assertEqual(str(cm.exception), "aap") + + def test_drain_events_remaining_timeouts(self): + class Connection(object): def drain_events(self, timeout=None): pass
1.0
celery__celery-2840
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost. </issue> <code> [start of README.rst] ================================= celery - Distributed Task Queue ================================= .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png |build-status| |coverage-status| :Version: 3.2.0a1 (Cipater) :Web: http://celeryproject.org/ :Download: http://pypi.python.org/pypi/celery/ :Source: http://github.com/celery/celery/ :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, python, webhooks, queue, distributed -- What is a Task Queue? ===================== Task queues are used as a mechanism to distribute work across threads or machines. A task queue's input is a unit of work, called a task, dedicated worker processes then constantly monitor the queue for new work to perform. Celery communicates via messages, usually using a broker to mediate between clients and workers. To initiate a task a client puts a message on the queue, the broker then delivers the message to a worker. A Celery system can consist of multiple workers and brokers, giving way to high availability and horizontal scaling. Celery is a library written in Python, but the protocol can be implemented in any language. So far there's RCelery_ for the Ruby programming language, and a `PHP client`, but language interoperability can also be achieved by using webhooks. .. _RCelery: https://github.com/leapfrogonline/rcelery .. _`PHP client`: https://github.com/gjedeer/celery-php .. _`using webhooks`: http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html What do I need? =============== Celery version 3.0 runs on, - Python (2.6, 2.7, 3.3, 3.4) - PyPy (1.8, 1.9) - Jython (2.5, 2.7). This is the last version to support Python 2.5, and from Celery 3.1, Python 2.6 or later is required. The last version to support Python 2.4 was Celery series 2.2. *Celery* is usually used with a message broker to send and receive messages. The RabbitMQ, Redis transports are feature complete, but there's also experimental support for a myriad of other solutions, including using SQLite for local development. *Celery* can run on a single machine, on multiple machines, or even across datacenters. Get Started =========== If this is the first time you're trying to use Celery, or you are new to Celery 3.0 coming from previous versions then you should read our getting started tutorials: - `First steps with Celery`_ Tutorial teaching you the bare minimum needed to get started with Celery. - `Next steps`_ A more complete overview, showing more features. .. _`First steps with Celery`: http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html .. _`Next steps`: http://docs.celeryproject.org/en/latest/getting-started/next-steps.html Celery is... ========== - **Simple** Celery is easy to use and maintain, and does *not need configuration files*. It has an active, friendly community you can talk to for support, including a `mailing-list`_ and and an IRC channel. Here's one of the simplest applications you can make:: from celery import Celery app = Celery('hello', broker='amqp://guest@localhost//') @app.task def hello(): return 'hello world' - **Highly Available** Workers and clients will automatically retry in the event of connection loss or failure, and some brokers support HA in way of *Master/Master* or *Master/Slave* replication. - **Fast** A single Celery process can process millions of tasks a minute, with sub-millisecond round-trip latency (using RabbitMQ, py-librabbitmq, and optimized settings). - **Flexible** Almost every part of *Celery* can be extended or used on its own, Custom pool implementations, serializers, compression schemes, logging, schedulers, consumers, producers, autoscalers, broker transports and much more. It supports... ============ - **Message Transports** - RabbitMQ_, Redis_, - MongoDB_ (experimental), Amazon SQS (experimental), - CouchDB_ (experimental), SQLAlchemy_ (experimental), - Django ORM (experimental), `IronMQ`_ - and more... - **Concurrency** - Prefork, Eventlet_, gevent_, threads/single threaded - **Result Stores** - AMQP, Redis - memcached, MongoDB - SQLAlchemy, Django ORM - Apache Cassandra, IronCache - **Serialization** - *pickle*, *json*, *yaml*, *msgpack*. - *zlib*, *bzip2* compression. - Cryptographic message signing. .. _`Eventlet`: http://eventlet.net/ .. _`gevent`: http://gevent.org/ .. _RabbitMQ: http://rabbitmq.com .. _Redis: http://redis.io .. _MongoDB: http://mongodb.org .. _Beanstalk: http://kr.github.com/beanstalkd .. _CouchDB: http://couchdb.apache.org .. _SQLAlchemy: http://sqlalchemy.org .. _`IronMQ`: http://iron.io Framework Integration ===================== Celery is easy to integrate with web frameworks, some of which even have integration packages: +--------------------+----------------------------------------------------+ | `Django`_ | not needed | +--------------------+----------------------------------------------------+ | `Pyramid`_ | `pyramid_celery`_ | +--------------------+----------------------------------------------------+ | `Pylons`_ | `celery-pylons`_ | +--------------------+----------------------------------------------------+ | `Flask`_ | not needed | +--------------------+----------------------------------------------------+ | `web2py`_ | `web2py-celery`_ | +--------------------+----------------------------------------------------+ | `Tornado`_ | `tornado-celery`_ | `another tornado-celery`_ | +--------------------+----------------------------------------------------+ The integration packages are not strictly necessary, but they can make development easier, and sometimes they add important hooks like closing database connections at ``fork``. .. _`Django`: http://djangoproject.com/ .. _`Pylons`: http://www.pylonsproject.org/ .. _`Flask`: http://flask.pocoo.org/ .. _`web2py`: http://web2py.com/ .. _`Bottle`: http://bottlepy.org/ .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ .. _`django-celery`: http://pypi.python.org/pypi/django-celery .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ .. _`Tornado`: http://www.tornadoweb.org/ .. _`tornado-celery`: http://github.com/mher/tornado-celery/ .. _`another tornado-celery`: https://github.com/mayflaver/tornado-celery .. _celery-documentation: Documentation ============= The `latest documentation`_ with user guides, tutorials and API reference is hosted at Read The Docs. .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ .. _celery-installation: Installation ============ You can install Celery either via the Python Package Index (PyPI) or from source. To install using `pip`,:: $ pip install -U Celery To install using `easy_install`,:: $ easy_install -U Celery .. _bundles: Bundles ------- Celery also defines a group of bundles that can be used to install Celery and the dependencies for a given feature. You can specify these in your requirements or on the ``pip`` comand-line by using brackets. Multiple bundles can be specified by separating them by commas. :: $ pip install "celery[librabbitmq]" $ pip install "celery[librabbitmq,redis,auth,msgpack]" The following bundles are available: Serializers ~~~~~~~~~~~ :celery[auth]: for using the auth serializer. :celery[msgpack]: for using the msgpack serializer. :celery[yaml]: for using the yaml serializer. Concurrency ~~~~~~~~~~~ :celery[eventlet]: for using the eventlet pool. :celery[gevent]: for using the gevent pool. :celery[threads]: for using the thread pool. Transports and Backends ~~~~~~~~~~~~~~~~~~~~~~~ :celery[librabbitmq]: for using the librabbitmq C library. :celery[redis]: for using Redis as a message transport or as a result backend. :celery[mongodb]: for using MongoDB as a message transport (*experimental*), or as a result backend (*supported*). :celery[sqs]: for using Amazon SQS as a message transport (*experimental*). :celery[memcache]: for using memcached as a result backend. :celery[cassandra]: for using Apache Cassandra as a result backend. :celery[couchdb]: for using CouchDB as a message transport (*experimental*). :celery[couchbase]: for using CouchBase as a result backend. :celery[beanstalk]: for using Beanstalk as a message transport (*experimental*). :celery[zookeeper]: for using Zookeeper as a message transport. :celery[zeromq]: for using ZeroMQ as a message transport (*experimental*). :celery[sqlalchemy]: for using SQLAlchemy as a message transport (*experimental*), or as a result backend (*supported*). :celery[pyro]: for using the Pyro4 message transport (*experimental*). :celery[slmq]: for using the SoftLayer Message Queue transport (*experimental*). .. _celery-installing-from-source: Downloading and installing from source -------------------------------------- Download the latest version of Celery from http://pypi.python.org/pypi/celery/ You can install it by doing the following,:: $ tar xvfz celery-0.0.0.tar.gz $ cd celery-0.0.0 $ python setup.py build # python setup.py install The last command must be executed as a privileged user if you are not currently using a virtualenv. .. _celery-installing-from-git: Using the development version ----------------------------- With pip ~~~~~~~~ The Celery development version also requires the development versions of ``kombu``, ``amqp`` and ``billiard``. You can install the latest snapshot of these using the following pip commands:: $ pip install https://github.com/celery/celery/zipball/master#egg=celery $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu With git ~~~~~~~~ Please the Contributing section. .. _getting-help: Getting Help ============ .. _mailing-list: Mailing list ------------ For discussions about the usage, development, and future of celery, please join the `celery-users`_ mailing list. .. _`celery-users`: http://groups.google.com/group/celery-users/ .. _irc-channel: IRC --- Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ network. .. _`Freenode`: http://freenode.net .. _bug-tracker: Bug tracker =========== If you have any suggestions, bug reports or annoyances please report them to our issue tracker at http://github.com/celery/celery/issues/ .. _wiki: Wiki ==== http://wiki.github.com/celery/celery/ .. _maintainers: Maintainers =========== - `@ask`_ (primary maintainer) - `@thedrow`_ - `@chrisgogreen`_ - `@PMickael`_ - `@malinoff`_ - And you? We really need more: https://github.com/celery/celery/issues/2534 .. _`@ask`: http://github.com/ask .. _`@thedrow`: http://github.com/thedrow .. _`@chrisgogreen`: http://github.com/chrisgogreen .. _`@PMickael`: http://github.com/PMickael .. _`@malinoff`: http://github.com/malinoff .. _contributing-short: Contributing ============ Development of `celery` happens at Github: http://github.com/celery/celery You are highly encouraged to participate in the development of `celery`. If you don't like Github (for some reason) you're welcome to send regular patches. Be sure to also read the `Contributing to Celery`_ section in the documentation. .. _`Contributing to Celery`: http://docs.celeryproject.org/en/master/contributing.html .. _license: License ======= This software is licensed under the `New BSD License`. See the ``LICENSE`` file in the top distribution directory for the full license text. .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png :alt: Bitdeli badge :target: https://bitdeli.com/free .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master :target: https://travis-ci.org/celery/celery .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg :target: https://coveralls.io/r/celery/celery [end of README.rst] [start of celery/app/defaults.py] # -*- coding: utf-8 -*- """ celery.app.defaults ~~~~~~~~~~~~~~~~~~~ Configuration introspection and defaults. """ from __future__ import absolute_import import sys from collections import deque, namedtuple from datetime import timedelta from celery.five import items from celery.utils import strtobool from celery.utils.functional import memoize __all__ = ['Option', 'NAMESPACES', 'flatten', 'find'] is_jython = sys.platform.startswith('java') is_pypy = hasattr(sys, 'pypy_version_info') DEFAULT_POOL = 'prefork' if is_jython: DEFAULT_POOL = 'threads' elif is_pypy: if sys.pypy_version_info[0:3] < (1, 5, 0): DEFAULT_POOL = 'solo' else: DEFAULT_POOL = 'prefork' DEFAULT_ACCEPT_CONTENT = ['json', 'pickle', 'msgpack', 'yaml'] DEFAULT_PROCESS_LOG_FMT = """ [%(asctime)s: %(levelname)s/%(processName)s] %(message)s """.strip() DEFAULT_LOG_FMT = '[%(asctime)s: %(levelname)s] %(message)s' DEFAULT_TASK_LOG_FMT = """[%(asctime)s: %(levelname)s/%(processName)s] \ %(task_name)s[%(task_id)s]: %(message)s""" _BROKER_OLD = {'deprecate_by': '2.5', 'remove_by': '4.0', 'alt': 'BROKER_URL setting'} _REDIS_OLD = {'deprecate_by': '2.5', 'remove_by': '4.0', 'alt': 'URL form of CELERY_RESULT_BACKEND'} searchresult = namedtuple('searchresult', ('namespace', 'key', 'type')) class Option(object): alt = None deprecate_by = None remove_by = None typemap = dict(string=str, int=int, float=float, any=lambda v: v, bool=strtobool, dict=dict, tuple=tuple) def __init__(self, default=None, *args, **kwargs): self.default = default self.type = kwargs.get('type') or 'string' for attr, value in items(kwargs): setattr(self, attr, value) def to_python(self, value): return self.typemap[self.type](value) def __repr__(self): return '<Option: type->{0} default->{1!r}>'.format(self.type, self.default) NAMESPACES = { 'BROKER': { 'URL': Option(None, type='string'), 'CONNECTION_TIMEOUT': Option(4, type='float'), 'CONNECTION_RETRY': Option(True, type='bool'), 'CONNECTION_MAX_RETRIES': Option(100, type='int'), 'FAILOVER_STRATEGY': Option(None, type='string'), 'HEARTBEAT': Option(None, type='int'), 'HEARTBEAT_CHECKRATE': Option(3.0, type='int'), 'LOGIN_METHOD': Option(None, type='string'), 'POOL_LIMIT': Option(10, type='int'), 'USE_SSL': Option(False, type='bool'), 'TRANSPORT': Option(type='string'), 'TRANSPORT_OPTIONS': Option({}, type='dict'), 'HOST': Option(type='string', **_BROKER_OLD), 'PORT': Option(type='int', **_BROKER_OLD), 'USER': Option(type='string', **_BROKER_OLD), 'PASSWORD': Option(type='string', **_BROKER_OLD), 'VHOST': Option(type='string', **_BROKER_OLD), }, 'CASSANDRA': { 'COLUMN_FAMILY': Option(type='string'), 'DETAILED_MODE': Option(False, type='bool'), 'KEYSPACE': Option(type='string'), 'READ_CONSISTENCY': Option(type='string'), 'SERVERS': Option(type='list'), 'WRITE_CONSISTENCY': Option(type='string'), }, 'CELERY': { 'ACCEPT_CONTENT': Option(DEFAULT_ACCEPT_CONTENT, type='list'), 'ACKS_LATE': Option(False, type='bool'), 'ALWAYS_EAGER': Option(False, type='bool'), 'ANNOTATIONS': Option(type='any'), 'BROADCAST_QUEUE': Option('celeryctl'), 'BROADCAST_EXCHANGE': Option('celeryctl'), 'BROADCAST_EXCHANGE_TYPE': Option('fanout'), 'CACHE_BACKEND': Option(), 'CACHE_BACKEND_OPTIONS': Option({}, type='dict'), 'CHORD_PROPAGATES': Option(True, type='bool'), 'COUCHBASE_BACKEND_SETTINGS': Option(None, type='dict'), 'CREATE_MISSING_QUEUES': Option(True, type='bool'), 'DEFAULT_RATE_LIMIT': Option(type='string'), 'DISABLE_RATE_LIMITS': Option(False, type='bool'), 'DEFAULT_ROUTING_KEY': Option('celery'), 'DEFAULT_QUEUE': Option('celery'), 'DEFAULT_EXCHANGE': Option('celery'), 'DEFAULT_EXCHANGE_TYPE': Option('direct'), 'DEFAULT_DELIVERY_MODE': Option(2, type='string'), 'EAGER_PROPAGATES_EXCEPTIONS': Option(False, type='bool'), 'ENABLE_UTC': Option(True, type='bool'), 'ENABLE_REMOTE_CONTROL': Option(True, type='bool'), 'EVENT_SERIALIZER': Option('json'), 'EVENT_QUEUE_EXPIRES': Option(60.0, type='float'), 'EVENT_QUEUE_TTL': Option(5.0, type='float'), 'IMPORTS': Option((), type='tuple'), 'INCLUDE': Option((), type='tuple'), 'IGNORE_RESULT': Option(False, type='bool'), 'MAX_CACHED_RESULTS': Option(100, type='int'), 'MESSAGE_COMPRESSION': Option(type='string'), 'MONGODB_BACKEND_SETTINGS': Option(type='dict'), 'REDIS_HOST': Option(type='string', **_REDIS_OLD), 'REDIS_PORT': Option(type='int', **_REDIS_OLD), 'REDIS_DB': Option(type='int', **_REDIS_OLD), 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 'REDIS_MAX_CONNECTIONS': Option(type='int'), 'RESULT_BACKEND': Option(type='string'), 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 'RESULT_DB_TABLENAMES': Option(type='dict'), 'RESULT_DBURI': Option(), 'RESULT_ENGINE_OPTIONS': Option(type='dict'), 'RESULT_EXCHANGE': Option('celeryresults'), 'RESULT_EXCHANGE_TYPE': Option('direct'), 'RESULT_SERIALIZER': Option('json'), 'RESULT_PERSISTENT': Option(None, type='bool'), 'RIAK_BACKEND_SETTINGS': Option(type='dict'), 'ROUTES': Option(type='any'), 'SEND_EVENTS': Option(False, type='bool'), 'SEND_TASK_ERROR_EMAILS': Option(False, type='bool'), 'SEND_TASK_SENT_EVENT': Option(False, type='bool'), 'STORE_ERRORS_EVEN_IF_IGNORED': Option(False, type='bool'), 'TASK_PROTOCOL': Option(1, type='int'), 'TASK_PUBLISH_RETRY': Option(True, type='bool'), 'TASK_PUBLISH_RETRY_POLICY': Option({ 'max_retries': 3, 'interval_start': 0, 'interval_max': 1, 'interval_step': 0.2}, type='dict'), 'TASK_RESULT_EXPIRES': Option(timedelta(days=1), type='float'), 'TASK_SERIALIZER': Option('json'), 'TIMEZONE': Option(type='string'), 'TRACK_STARTED': Option(False, type='bool'), 'REDIRECT_STDOUTS': Option(True, type='bool'), 'REDIRECT_STDOUTS_LEVEL': Option('WARNING'), 'QUEUES': Option(type='dict'), 'QUEUE_HA_POLICY': Option(None, type='string'), 'SECURITY_KEY': Option(type='string'), 'SECURITY_CERTIFICATE': Option(type='string'), 'SECURITY_CERT_STORE': Option(type='string'), 'WORKER_DIRECT': Option(False, type='bool'), }, 'CELERYD': { 'AGENT': Option(None, type='string'), 'AUTOSCALER': Option('celery.worker.autoscale:Autoscaler'), 'AUTORELOADER': Option('celery.worker.autoreload:Autoreloader'), 'CONCURRENCY': Option(0, type='int'), 'TIMER': Option(type='string'), 'TIMER_PRECISION': Option(1.0, type='float'), 'FORCE_EXECV': Option(False, type='bool'), 'HIJACK_ROOT_LOGGER': Option(True, type='bool'), 'CONSUMER': Option('celery.worker.consumer:Consumer', type='string'), 'LOG_FORMAT': Option(DEFAULT_PROCESS_LOG_FMT), 'LOG_COLOR': Option(type='bool'), 'LOG_LEVEL': Option('WARN', deprecate_by='2.4', remove_by='4.0', alt='--loglevel argument'), 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', alt='--logfile argument'), 'MAX_TASKS_PER_CHILD': Option(type='int'), 'POOL': Option(DEFAULT_POOL), 'POOL_PUTLOCKS': Option(True, type='bool'), 'POOL_RESTARTS': Option(False, type='bool'), 'PREFETCH_MULTIPLIER': Option(4, type='int'), 'STATE_DB': Option(), 'TASK_LOG_FORMAT': Option(DEFAULT_TASK_LOG_FMT), 'TASK_SOFT_TIME_LIMIT': Option(type='float'), 'TASK_TIME_LIMIT': Option(type='float'), 'WORKER_LOST_WAIT': Option(10.0, type='float') }, 'CELERYBEAT': { 'SCHEDULE': Option({}, type='dict'), 'SCHEDULER': Option('celery.beat:PersistentScheduler'), 'SCHEDULE_FILENAME': Option('celerybeat-schedule'), 'SYNC_EVERY': Option(0, type='int'), 'MAX_LOOP_INTERVAL': Option(0, type='float'), 'LOG_LEVEL': Option('INFO', deprecate_by='2.4', remove_by='4.0', alt='--loglevel argument'), 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', alt='--logfile argument'), }, 'CELERYMON': { 'LOG_LEVEL': Option('INFO', deprecate_by='2.4', remove_by='4.0', alt='--loglevel argument'), 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', alt='--logfile argument'), 'LOG_FORMAT': Option(DEFAULT_LOG_FMT), }, 'EMAIL': { 'HOST': Option('localhost'), 'PORT': Option(25, type='int'), 'HOST_USER': Option(), 'HOST_PASSWORD': Option(), 'TIMEOUT': Option(2, type='float'), 'USE_SSL': Option(False, type='bool'), 'USE_TLS': Option(False, type='bool'), 'CHARSET': Option('us-ascii'), }, 'SERVER_EMAIL': Option('celery@localhost'), 'ADMINS': Option((), type='tuple'), } def flatten(d, ns=''): stack = deque([(ns, d)]) while stack: name, space = stack.popleft() for key, value in items(space): if isinstance(value, dict): stack.append((name + key + '_', value)) else: yield name + key, value DEFAULTS = {key: value.default for key, value in flatten(NAMESPACES)} def find_deprecated_settings(source): from celery.utils import warn_deprecated for name, opt in flatten(NAMESPACES): if (opt.deprecate_by or opt.remove_by) and getattr(source, name, None): warn_deprecated(description='The {0!r} setting'.format(name), deprecation=opt.deprecate_by, removal=opt.remove_by, alternative='Use the {0.alt} instead'.format(opt)) return source @memoize(maxsize=None) def find(name, namespace='celery'): # - Try specified namespace first. namespace = namespace.upper() try: return searchresult( namespace, name.upper(), NAMESPACES[namespace][name.upper()], ) except KeyError: # - Try all the other namespaces. for ns, keys in items(NAMESPACES): if ns.upper() == name.upper(): return searchresult(None, ns, keys) elif isinstance(keys, dict): try: return searchresult(ns, name.upper(), keys[name.upper()]) except KeyError: pass # - See if name is a qualname last. return searchresult(None, name.upper(), DEFAULTS[name.upper()]) [end of celery/app/defaults.py] [start of celery/app/task.py] # -*- coding: utf-8 -*- """ celery.app.task ~~~~~~~~~~~~~~~ Task Implementation: Task request context, and the base task class. """ from __future__ import absolute_import import sys from billiard.einfo import ExceptionInfo from celery import current_app, group from celery import states from celery._state import _task_stack from celery.canvas import signature from celery.exceptions import Ignore, MaxRetriesExceededError, Reject, Retry from celery.five import class_property, items from celery.result import EagerResult from celery.utils import abstract from celery.utils import uuid, maybe_reraise from celery.utils.functional import mattrgetter, maybe_list from celery.utils.imports import instantiate from celery.utils.mail import ErrorMail from .annotations import resolve_all as resolve_all_annotations from .registry import _unpickle_task_v2 from .utils import appstr __all__ = ['Context', 'Task'] #: extracts attributes related to publishing a message from an object. extract_exec_options = mattrgetter( 'queue', 'routing_key', 'exchange', 'priority', 'expires', 'serializer', 'delivery_mode', 'compression', 'time_limit', 'soft_time_limit', 'immediate', 'mandatory', # imm+man is deprecated ) # We take __repr__ very seriously around here ;) R_BOUND_TASK = '<class {0.__name__} of {app}{flags}>' R_UNBOUND_TASK = '<unbound {0.__name__}{flags}>' R_SELF_TASK = '<@task {0.name} bound to other {0.__self__}>' R_INSTANCE = '<@task: {0.name} of {app}{flags}>' #: Here for backwards compatibility as tasks no longer use a custom metaclass. TaskType = type def _strflags(flags, default=''): if flags: return ' ({0})'.format(', '.join(flags)) return default def _reprtask(task, fmt=None, flags=None): flags = list(flags) if flags is not None else [] flags.append('v2 compatible') if task.__v2_compat__ else None if not fmt: fmt = R_BOUND_TASK if task._app else R_UNBOUND_TASK return fmt.format( task, flags=_strflags(flags), app=appstr(task._app) if task._app else None, ) class Context(object): # Default context logfile = None loglevel = None hostname = None id = None args = None kwargs = None retries = 0 eta = None expires = None is_eager = False headers = None delivery_info = None reply_to = None root_id = None parent_id = None correlation_id = None taskset = None # compat alias to group group = None chord = None utc = None called_directly = True callbacks = None errbacks = None timelimit = None _children = None # see property _protected = 0 def __init__(self, *args, **kwargs): self.update(*args, **kwargs) def update(self, *args, **kwargs): return self.__dict__.update(*args, **kwargs) def clear(self): return self.__dict__.clear() def get(self, key, default=None): return getattr(self, key, default) def __repr__(self): return '<Context: {0!r}>'.format(vars(self)) @property def children(self): # children must be an empy list for every thread if self._children is None: self._children = [] return self._children class Task(object): """Task base class. When called tasks apply the :meth:`run` method. This method must be defined by all tasks (that is unless the :meth:`__call__` method is overridden). """ __trace__ = None __v2_compat__ = False # set by old base in celery.task.base ErrorMail = ErrorMail MaxRetriesExceededError = MaxRetriesExceededError #: Execution strategy used, or the qualified name of one. Strategy = 'celery.worker.strategy:default' #: This is the instance bound to if the task is a method of a class. __self__ = None #: The application instance associated with this task class. _app = None #: Name of the task. name = None #: If :const:`True` the task is an abstract base class. abstract = True #: Maximum number of retries before giving up. If set to :const:`None`, #: it will **never** stop retrying. max_retries = 3 #: Default time in seconds before a retry of the task should be #: executed. 3 minutes by default. default_retry_delay = 3 * 60 #: Rate limit for this task type. Examples: :const:`None` (no rate #: limit), `'100/s'` (hundred tasks a second), `'100/m'` (hundred tasks #: a minute),`'100/h'` (hundred tasks an hour) rate_limit = None #: If enabled the worker will not store task state and return values #: for this task. Defaults to the :setting:`CELERY_IGNORE_RESULT` #: setting. ignore_result = None #: If enabled the request will keep track of subtasks started by #: this task, and this information will be sent with the result #: (``result.children``). trail = True #: When enabled errors will be stored even if the task is otherwise #: configured to ignore results. store_errors_even_if_ignored = None #: If enabled an email will be sent to :setting:`ADMINS` whenever a task #: of this type fails. send_error_emails = None #: The name of a serializer that are registered with #: :mod:`kombu.serialization.registry`. Default is `'pickle'`. serializer = None #: Hard time limit. #: Defaults to the :setting:`CELERYD_TASK_TIME_LIMIT` setting. time_limit = None #: Soft time limit. #: Defaults to the :setting:`CELERYD_TASK_SOFT_TIME_LIMIT` setting. soft_time_limit = None #: The result store backend used for this task. backend = None #: If disabled this task won't be registered automatically. autoregister = True #: If enabled the task will report its status as 'started' when the task #: is executed by a worker. Disabled by default as the normal behaviour #: is to not report that level of granularity. Tasks are either pending, #: finished, or waiting to be retried. #: #: Having a 'started' status can be useful for when there are long #: running tasks and there is a need to report which task is currently #: running. #: #: The application default can be overridden using the #: :setting:`CELERY_TRACK_STARTED` setting. track_started = None #: When enabled messages for this task will be acknowledged **after** #: the task has been executed, and not *just before* which is the #: default behavior. #: #: Please note that this means the task may be executed twice if the #: worker crashes mid execution (which may be acceptable for some #: applications). #: #: The application default can be overridden with the #: :setting:`CELERY_ACKS_LATE` setting. acks_late = None #: Tuple of expected exceptions. #: #: These are errors that are expected in normal operation #: and that should not be regarded as a real error by the worker. #: Currently this means that the state will be updated to an error #: state, but the worker will not log the event as an error. throws = () #: Default task expiry time. expires = None #: Task request stack, the current request will be the topmost. request_stack = None #: Some may expect a request to exist even if the task has not been #: called. This should probably be deprecated. _default_request = None _exec_options = None __bound__ = False from_config = ( ('send_error_emails', 'CELERY_SEND_TASK_ERROR_EMAILS'), ('serializer', 'CELERY_TASK_SERIALIZER'), ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), ('track_started', 'CELERY_TRACK_STARTED'), ('acks_late', 'CELERY_ACKS_LATE'), ('ignore_result', 'CELERY_IGNORE_RESULT'), ('store_errors_even_if_ignored', 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), ) #: ignored accept_magic_kwargs = False _backend = None # set by backend property. __bound__ = False # - Tasks are lazily bound, so that configuration is not set # - until the task is actually used @classmethod def bind(self, app): was_bound, self.__bound__ = self.__bound__, True self._app = app conf = app.conf self._exec_options = None # clear option cache for attr_name, config_name in self.from_config: if getattr(self, attr_name, None) is None: setattr(self, attr_name, conf[config_name]) # decorate with annotations from config. if not was_bound: self.annotate() from celery.utils.threads import LocalStack self.request_stack = LocalStack() # PeriodicTask uses this to add itself to the PeriodicTask schedule. self.on_bound(app) return app @classmethod def on_bound(self, app): """This method can be defined to do additional actions when the task class is bound to an app.""" pass @classmethod def _get_app(self): if self._app is None: self._app = current_app if not self.__bound__: # The app property's __set__ method is not called # if Task.app is set (on the class), so must bind on use. self.bind(self._app) return self._app app = class_property(_get_app, bind) @classmethod def annotate(self): for d in resolve_all_annotations(self.app.annotations, self): for key, value in items(d): if key.startswith('@'): self.add_around(key[1:], value) else: setattr(self, key, value) @classmethod def add_around(self, attr, around): orig = getattr(self, attr) if getattr(orig, '__wrapped__', None): orig = orig.__wrapped__ meth = around(orig) meth.__wrapped__ = orig setattr(self, attr, meth) def __call__(self, *args, **kwargs): _task_stack.push(self) self.push_request(args=args, kwargs=kwargs) try: # add self if this is a bound task if self.__self__ is not None: return self.run(self.__self__, *args, **kwargs) return self.run(*args, **kwargs) finally: self.pop_request() _task_stack.pop() def __reduce__(self): # - tasks are pickled into the name of the task only, and the reciever # - simply grabs it from the local registry. # - in later versions the module of the task is also included, # - and the receiving side tries to import that module so that # - it will work even if the task has not been registered. mod = type(self).__module__ mod = mod if mod and mod in sys.modules else None return (_unpickle_task_v2, (self.name, mod), None) def run(self, *args, **kwargs): """The body of the task executed by workers.""" raise NotImplementedError('Tasks must define the run method.') def start_strategy(self, app, consumer, **kwargs): return instantiate(self.Strategy, self, app, consumer, **kwargs) def delay(self, *args, **kwargs): """Star argument version of :meth:`apply_async`. Does not support the extra options enabled by :meth:`apply_async`. :param \*args: positional arguments passed on to the task. :param \*\*kwargs: keyword arguments passed on to the task. :returns :class:`celery.result.AsyncResult`: """ return self.apply_async(args, kwargs) def apply_async(self, args=None, kwargs=None, task_id=None, producer=None, link=None, link_error=None, shadow=None, **options): """Apply tasks asynchronously by sending a message. :keyword args: The positional arguments to pass on to the task (a :class:`list` or :class:`tuple`). :keyword kwargs: The keyword arguments to pass on to the task (a :class:`dict`) :keyword countdown: Number of seconds into the future that the task should execute. Defaults to immediate execution. :keyword eta: A :class:`~datetime.datetime` object describing the absolute time and date of when the task should be executed. May not be specified if `countdown` is also supplied. :keyword expires: Either a :class:`int`, describing the number of seconds, or a :class:`~datetime.datetime` object that describes the absolute time and date of when the task should expire. The task will not be executed after the expiration time. :keyword shadow: Override task name used in logs/monitoring (default from :meth:`shadow_name`). :keyword connection: Re-use existing broker connection instead of establishing a new one. :keyword retry: If enabled sending of the task message will be retried in the event of connection loss or failure. Default is taken from the :setting:`CELERY_TASK_PUBLISH_RETRY` setting. Note that you need to handle the producer/connection manually for this to work. :keyword retry_policy: Override the retry policy used. See the :setting:`CELERY_TASK_PUBLISH_RETRY_POLICY` setting. :keyword routing_key: Custom routing key used to route the task to a worker server. If in combination with a ``queue`` argument only used to specify custom routing keys to topic exchanges. :keyword queue: The queue to route the task to. This must be a key present in :setting:`CELERY_QUEUES`, or :setting:`CELERY_CREATE_MISSING_QUEUES` must be enabled. See :ref:`guide-routing` for more information. :keyword exchange: Named custom exchange to send the task to. Usually not used in combination with the ``queue`` argument. :keyword priority: The task priority, a number between 0 and 9. Defaults to the :attr:`priority` attribute. :keyword serializer: A string identifying the default serialization method to use. Can be `pickle`, `json`, `yaml`, `msgpack` or any custom serialization method that has been registered with :mod:`kombu.serialization.registry`. Defaults to the :attr:`serializer` attribute. :keyword compression: A string identifying the compression method to use. Can be one of ``zlib``, ``bzip2``, or any custom compression methods registered with :func:`kombu.compression.register`. Defaults to the :setting:`CELERY_MESSAGE_COMPRESSION` setting. :keyword link: A single, or a list of tasks to apply if the task exits successfully. :keyword link_error: A single, or a list of tasks to apply if an error occurs while executing the task. :keyword producer: :class:`kombu.Producer` instance to use. :keyword add_to_parent: If set to True (default) and the task is applied while executing another task, then the result will be appended to the parent tasks ``request.children`` attribute. Trailing can also be disabled by default using the :attr:`trail` attribute :keyword publisher: Deprecated alias to ``producer``. :keyword headers: Message headers to be sent in the task (a :class:`dict`) :rtype :class:`celery.result.AsyncResult`: if :setting:`CELERY_ALWAYS_EAGER` is not set, otherwise :class:`celery.result.EagerResult`: Also supports all keyword arguments supported by :meth:`kombu.Producer.publish`. .. note:: If the :setting:`CELERY_ALWAYS_EAGER` setting is set, it will be replaced by a local :func:`apply` call instead. """ try: check_arguments = self.__header__ except AttributeError: pass else: check_arguments(*(args or ()), **(kwargs or {})) app = self._get_app() if app.conf.CELERY_ALWAYS_EAGER: return self.apply(args, kwargs, task_id=task_id or uuid(), link=link, link_error=link_error, **options) # add 'self' if this is a "task_method". if self.__self__ is not None: args = args if isinstance(args, tuple) else tuple(args or ()) args = (self.__self__,) + args shadow = shadow or self.shadow_name(args, kwargs, options) preopts = self._get_exec_options() options = dict(preopts, **options) if options else preopts return app.send_task( self.name, args, kwargs, task_id=task_id, producer=producer, link=link, link_error=link_error, result_cls=self.AsyncResult, shadow=shadow, **options ) def shadow_name(self, args, kwargs, options): """Override for custom task name in worker logs/monitoring. :param args: Task positional arguments. :param kwargs: Task keyword arguments. :param options: Task execution options. **Example**: .. code-block:: python from celery.utils.imports import qualname def shadow_name(task, args, kwargs, options): return qualname(args[0]) @app.task(shadow_name=shadow_name, serializer='pickle') def apply_function_async(fun, *args, **kwargs): return fun(*args, **kwargs) """ pass def signature_from_request(self, request=None, args=None, kwargs=None, queue=None, **extra_options): request = self.request if request is None else request args = request.args if args is None else args kwargs = request.kwargs if kwargs is None else kwargs limit_hard, limit_soft = request.timelimit or (None, None) options = { 'task_id': request.id, 'link': request.callbacks, 'link_error': request.errbacks, 'group_id': request.group, 'chord': request.chord, 'soft_time_limit': limit_soft, 'time_limit': limit_hard, 'reply_to': request.reply_to, 'headers': request.headers, } options.update( {'queue': queue} if queue else (request.delivery_info or {}), ) return self.signature( args, kwargs, options, type=self, **extra_options ) subtask_from_request = signature_from_request def retry(self, args=None, kwargs=None, exc=None, throw=True, eta=None, countdown=None, max_retries=None, **options): """Retry the task. :param args: Positional arguments to retry with. :param kwargs: Keyword arguments to retry with. :keyword exc: Custom exception to report when the max restart limit has been exceeded (default: :exc:`~@MaxRetriesExceededError`). If this argument is set and retry is called while an exception was raised (``sys.exc_info()`` is set) it will attempt to reraise the current exception. If no exception was raised it will raise the ``exc`` argument provided. :keyword countdown: Time in seconds to delay the retry for. :keyword eta: Explicit time and date to run the retry at (must be a :class:`~datetime.datetime` instance). :keyword max_retries: If set, overrides the default retry limit. A value of :const:`None`, means "use the default", so if you want infinite retries you would have to set the :attr:`max_retries` attribute of the task to :const:`None` first. :keyword time_limit: If set, overrides the default time limit. :keyword soft_time_limit: If set, overrides the default soft time limit. :keyword \*\*options: Any extra options to pass on to meth:`apply_async`. :keyword throw: If this is :const:`False`, do not raise the :exc:`~@Retry` exception, that tells the worker to mark the task as being retried. Note that this means the task will be marked as failed if the task raises an exception, or successful if it returns. :raises celery.exceptions.Retry: To tell the worker that the task has been re-sent for retry. This always happens, unless the `throw` keyword argument has been explicitly set to :const:`False`, and is considered normal operation. **Example** .. code-block:: pycon >>> from imaginary_twitter_lib import Twitter >>> from proj.celery import app >>> @app.task(bind=True) ... def tweet(self, auth, message): ... twitter = Twitter(oauth=auth) ... try: ... twitter.post_status_update(message) ... except twitter.FailWhale as exc: ... # Retry in 5 minutes. ... raise self.retry(countdown=60 * 5, exc=exc) Although the task will never return above as `retry` raises an exception to notify the worker, we use `raise` in front of the retry to convey that the rest of the block will not be executed. """ request = self.request retries = request.retries + 1 max_retries = self.max_retries if max_retries is None else max_retries # Not in worker or emulated by (apply/always_eager), # so just raise the original exception. if request.called_directly: maybe_reraise() # raise orig stack if PyErr_Occurred raise exc or Retry('Task can be retried', None) if not eta and countdown is None: countdown = self.default_retry_delay is_eager = request.is_eager S = self.signature_from_request( request, args, kwargs, countdown=countdown, eta=eta, retries=retries, **options ) if max_retries is not None and retries > max_retries: if exc: # first try to reraise the original exception maybe_reraise() # or if not in an except block then raise the custom exc. raise exc raise self.MaxRetriesExceededError( "Can't retry {0}[{1}] args:{2} kwargs:{3}".format( self.name, request.id, S.args, S.kwargs)) ret = Retry(exc=exc, when=eta or countdown) if is_eager: # if task was executed eagerly using apply(), # then the retry must also be executed eagerly. S.apply().get() if throw: raise ret return ret try: S.apply_async() except Exception as exc: raise Reject(exc, requeue=False) if throw: raise ret return ret def apply(self, args=None, kwargs=None, link=None, link_error=None, **options): """Execute this task locally, by blocking until the task returns. :param args: positional arguments passed on to the task. :param kwargs: keyword arguments passed on to the task. :keyword throw: Re-raise task exceptions. Defaults to the :setting:`CELERY_EAGER_PROPAGATES_EXCEPTIONS` setting. :rtype :class:`celery.result.EagerResult`: """ # trace imports Task, so need to import inline. from celery.app.trace import build_tracer app = self._get_app() args = args or () # add 'self' if this is a bound method. if self.__self__ is not None: args = (self.__self__,) + tuple(args) kwargs = kwargs or {} task_id = options.get('task_id') or uuid() retries = options.get('retries', 0) throw = app.either('CELERY_EAGER_PROPAGATES_EXCEPTIONS', options.pop('throw', None)) # Make sure we get the task instance, not class. task = app._tasks[self.name] request = {'id': task_id, 'retries': retries, 'is_eager': True, 'logfile': options.get('logfile'), 'loglevel': options.get('loglevel', 0), 'callbacks': maybe_list(link), 'errbacks': maybe_list(link_error), 'headers': options.get('headers'), 'delivery_info': {'is_eager': True}} tb = None tracer = build_tracer( task.name, task, eager=True, propagate=throw, app=self._get_app(), ) ret = tracer(task_id, args, kwargs, request) retval = ret.retval if isinstance(retval, ExceptionInfo): retval, tb = retval.exception, retval.traceback state = states.SUCCESS if ret.info is None else ret.info.state return EagerResult(task_id, retval, state, traceback=tb) def AsyncResult(self, task_id, **kwargs): """Get AsyncResult instance for this kind of task. :param task_id: Task id to get result for. """ return self._get_app().AsyncResult(task_id, backend=self.backend, task_name=self.name, **kwargs) def signature(self, args=None, *starargs, **starkwargs): """Return :class:`~celery.signature` object for this task, wrapping arguments and execution options for a single task invocation.""" starkwargs.setdefault('app', self.app) return signature(self, args, *starargs, **starkwargs) subtask = signature def s(self, *args, **kwargs): """``.s(*a, **k) -> .signature(a, k)``""" return self.signature(args, kwargs) def si(self, *args, **kwargs): """``.si(*a, **k) -> .signature(a, k, immutable=True)``""" return self.signature(args, kwargs, immutable=True) def chunks(self, it, n): """Creates a :class:`~celery.canvas.chunks` task for this task.""" from celery import chunks return chunks(self.s(), it, n, app=self.app) def map(self, it): """Creates a :class:`~celery.canvas.xmap` task from ``it``.""" from celery import xmap return xmap(self.s(), it, app=self.app) def starmap(self, it): """Creates a :class:`~celery.canvas.xstarmap` task from ``it``.""" from celery import xstarmap return xstarmap(self.s(), it, app=self.app) def send_event(self, type_, **fields): req = self.request with self.app.events.default_dispatcher(hostname=req.hostname) as d: return d.send(type_, uuid=req.id, **fields) def replace(self, sig): """Replace the current task, with a new task inheriting the same task id. :param sig: :class:`@signature` Note: This will raise :exc:`~@Ignore`, so the best practice is to always use ``raise self.replace(...)`` to convey to the reader that the task will not continue after being replaced. :param: Signature of new task. """ chord = self.request.chord if isinstance(sig, group): sig |= self.app.tasks['celery.accumulate'].s(index=0).set( chord=chord, ) chord = None sig.freeze(self.request.id, group_id=self.request.group, chord=chord, root_id=self.request.root_id) sig.delay() raise Ignore('Chord member replaced by new task') def add_to_chord(self, sig, lazy=False): """Add signature to the chord the current task is a member of. :param sig: Signature to extend chord with. :param lazy: If enabled the new task will not actually be called, and ``sig.delay()`` must be called manually. Currently only supported by the Redis result backend when ``?new_join=1`` is enabled. """ if not self.request.chord: raise ValueError('Current task is not member of any chord') result = sig.freeze(group_id=self.request.group, chord=self.request.chord, root_id=self.request.root_id) self.backend.add_to_chord(self.request.group, result) return sig.delay() if not lazy else sig def update_state(self, task_id=None, state=None, meta=None): """Update task state. :keyword task_id: Id of the task to update, defaults to the id of the current task :keyword state: New state (:class:`str`). :keyword meta: State metadata (:class:`dict`). """ if task_id is None: task_id = self.request.id self.backend.store_result(task_id, meta, state) def on_success(self, retval, task_id, args, kwargs): """Success handler. Run by the worker if the task executes successfully. :param retval: The return value of the task. :param task_id: Unique id of the executed task. :param args: Original arguments for the executed task. :param kwargs: Original keyword arguments for the executed task. The return value of this handler is ignored. """ pass def on_retry(self, exc, task_id, args, kwargs, einfo): """Retry handler. This is run by the worker when the task is to be retried. :param exc: The exception sent to :meth:`retry`. :param task_id: Unique id of the retried task. :param args: Original arguments for the retried task. :param kwargs: Original keyword arguments for the retried task. :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` instance, containing the traceback. The return value of this handler is ignored. """ pass def on_failure(self, exc, task_id, args, kwargs, einfo): """Error handler. This is run by the worker when the task fails. :param exc: The exception raised by the task. :param task_id: Unique id of the failed task. :param args: Original arguments for the task that failed. :param kwargs: Original keyword arguments for the task that failed. :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` instance, containing the traceback. The return value of this handler is ignored. """ pass def after_return(self, status, retval, task_id, args, kwargs, einfo): """Handler called after the task returns. :param status: Current task state. :param retval: Task return value/exception. :param task_id: Unique id of the task. :param args: Original arguments for the task. :param kwargs: Original keyword arguments for the task. :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` instance, containing the traceback (if any). The return value of this handler is ignored. """ pass def send_error_email(self, context, exc, **kwargs): if self.send_error_emails and \ not getattr(self, 'disable_error_emails', None): self.ErrorMail(self, **kwargs).send(context, exc) def add_trail(self, result): if self.trail: self.request.children.append(result) return result def push_request(self, *args, **kwargs): self.request_stack.push(Context(*args, **kwargs)) def pop_request(self): self.request_stack.pop() def __repr__(self): """`repr(task)`""" return _reprtask(self, R_SELF_TASK if self.__self__ else R_INSTANCE) def _get_request(self): """Get current request object.""" req = self.request_stack.top if req is None: # task was not called, but some may still expect a request # to be there, perhaps that should be deprecated. if self._default_request is None: self._default_request = Context() return self._default_request return req request = property(_get_request) def _get_exec_options(self): if self._exec_options is None: self._exec_options = extract_exec_options(self) return self._exec_options @property def backend(self): backend = self._backend if backend is None: return self.app.backend return backend @backend.setter def backend(self, value): # noqa self._backend = value @property def __name__(self): return self.__class__.__name__ abstract.CallableTask.register(Task) BaseTask = Task # compat alias [end of celery/app/task.py] [start of celery/worker/request.py] # -*- coding: utf-8 -*- """ celery.worker.request ~~~~~~~~~~~~~~~~~~~~~ This module defines the :class:`Request` class, which specifies how tasks are executed. """ from __future__ import absolute_import, unicode_literals import logging import socket import sys from datetime import datetime from weakref import ref from kombu.utils.encoding import safe_repr, safe_str from celery import signals from celery.app.trace import trace_task, trace_task_ret from celery.exceptions import ( Ignore, TaskRevokedError, InvalidTaskError, SoftTimeLimitExceeded, TimeLimitExceeded, WorkerLostError, Terminated, Retry, Reject, ) from celery.five import string from celery.platforms import signals as _signals from celery.utils.functional import noop from celery.utils.log import get_logger from celery.utils.timeutils import maybe_iso8601, timezone, maybe_make_aware from celery.utils.serialization import get_pickled_exception from . import state __all__ = ['Request'] IS_PYPY = hasattr(sys, 'pypy_version_info') logger = get_logger(__name__) debug, info, warn, error = (logger.debug, logger.info, logger.warning, logger.error) _does_info = False _does_debug = False def __optimize__(): # this is also called by celery.app.trace.setup_worker_optimizations global _does_debug global _does_info _does_debug = logger.isEnabledFor(logging.DEBUG) _does_info = logger.isEnabledFor(logging.INFO) __optimize__() # Localize tz_utc = timezone.utc tz_or_local = timezone.tz_or_local send_revoked = signals.task_revoked.send task_accepted = state.task_accepted task_ready = state.task_ready revoked_tasks = state.revoked class Request(object): """A request for task execution.""" acknowledged = False time_start = None worker_pid = None time_limits = (None, None) _already_revoked = False _terminate_on_ack = None _apply_result = None _tzlocal = None if not IS_PYPY: # pragma: no cover __slots__ = ( 'app', 'type', 'name', 'id', 'on_ack', 'body', 'hostname', 'eventer', 'connection_errors', 'task', 'eta', 'expires', 'request_dict', 'on_reject', 'utc', 'content_type', 'content_encoding', '__weakref__', '__dict__', ) def __init__(self, message, on_ack=noop, hostname=None, eventer=None, app=None, connection_errors=None, request_dict=None, task=None, on_reject=noop, body=None, headers=None, decoded=False, utc=True, maybe_make_aware=maybe_make_aware, maybe_iso8601=maybe_iso8601, **opts): if headers is None: headers = message.headers if body is None: body = message.body self.app = app self.message = message self.body = body self.utc = utc if decoded: self.content_type = self.content_encoding = None else: self.content_type, self.content_encoding = ( message.content_type, message.content_encoding, ) self.id = headers['id'] type = self.type = self.name = headers['task'] if 'shadow' in headers: self.name = headers['shadow'] if 'timelimit' in headers: self.time_limits = headers['timelimit'] self.on_ack = on_ack self.on_reject = on_reject self.hostname = hostname or socket.gethostname() self.eventer = eventer self.connection_errors = connection_errors or () self.task = task or self.app.tasks[type] # timezone means the message is timezone-aware, and the only timezone # supported at this point is UTC. eta = headers.get('eta') if eta is not None: try: eta = maybe_iso8601(eta) except (AttributeError, ValueError, TypeError) as exc: raise InvalidTaskError( 'invalid eta value {0!r}: {1}'.format(eta, exc)) self.eta = maybe_make_aware(eta, self.tzlocal) else: self.eta = None expires = headers.get('expires') if expires is not None: try: expires = maybe_iso8601(expires) except (AttributeError, ValueError, TypeError) as exc: raise InvalidTaskError( 'invalid expires value {0!r}: {1}'.format(expires, exc)) self.expires = maybe_make_aware(expires, self.tzlocal) else: self.expires = None delivery_info = message.delivery_info or {} properties = message.properties or {} headers.update({ 'reply_to': properties.get('reply_to'), 'correlation_id': properties.get('correlation_id'), 'delivery_info': { 'exchange': delivery_info.get('exchange'), 'routing_key': delivery_info.get('routing_key'), 'priority': delivery_info.get('priority'), 'redelivered': delivery_info.get('redelivered'), } }) self.request_dict = headers @property def delivery_info(self): return self.request_dict['delivery_info'] def execute_using_pool(self, pool, **kwargs): """Used by the worker to send this task to the pool. :param pool: A :class:`celery.concurrency.base.TaskPool` instance. :raises celery.exceptions.TaskRevokedError: if the task was revoked and ignored. """ task_id = self.id task = self.task if self.revoked(): raise TaskRevokedError(task_id) time_limit, soft_time_limit = self.time_limits time_limit = time_limit or task.time_limit soft_time_limit = soft_time_limit or task.soft_time_limit result = pool.apply_async( trace_task_ret, args=(self.type, task_id, self.request_dict, self.body, self.content_type, self.content_encoding), accept_callback=self.on_accepted, timeout_callback=self.on_timeout, callback=self.on_success, error_callback=self.on_failure, soft_timeout=soft_time_limit, timeout=time_limit, correlation_id=task_id, ) # cannot create weakref to None self._apply_result = ref(result) if result is not None else result return result def execute(self, loglevel=None, logfile=None): """Execute the task in a :func:`~celery.app.trace.trace_task`. :keyword loglevel: The loglevel used by the task. :keyword logfile: The logfile used by the task. """ if self.revoked(): return # acknowledge task as being processed. if not self.task.acks_late: self.acknowledge() request = self.request_dict args, kwargs, embed = self.message.payload request.update({'loglevel': loglevel, 'logfile': logfile, 'hostname': self.hostname, 'is_eager': False, 'args': args, 'kwargs': kwargs}, **embed or {}) retval = trace_task(self.task, self.id, args, kwargs, request, hostname=self.hostname, loader=self.app.loader, app=self.app)[0] self.acknowledge() return retval def maybe_expire(self): """If expired, mark the task as revoked.""" if self.expires: now = datetime.now(self.expires.tzinfo) if now > self.expires: revoked_tasks.add(self.id) return True def terminate(self, pool, signal=None): signal = _signals.signum(signal or 'TERM') if self.time_start: pool.terminate_job(self.worker_pid, signal) self._announce_revoked('terminated', True, signal, False) else: self._terminate_on_ack = pool, signal if self._apply_result is not None: obj = self._apply_result() # is a weakref if obj is not None: obj.terminate(signal) def _announce_revoked(self, reason, terminated, signum, expired): task_ready(self) self.send_event('task-revoked', terminated=terminated, signum=signum, expired=expired) if self.store_errors: self.task.backend.mark_as_revoked(self.id, reason, request=self) self.acknowledge() self._already_revoked = True send_revoked(self.task, request=self, terminated=terminated, signum=signum, expired=expired) def revoked(self): """If revoked, skip task and mark state.""" expired = False if self._already_revoked: return True if self.expires: expired = self.maybe_expire() if self.id in revoked_tasks: info('Discarding revoked task: %s[%s]', self.name, self.id) self._announce_revoked( 'expired' if expired else 'revoked', False, None, expired, ) return True return False def send_event(self, type, **fields): if self.eventer and self.eventer.enabled: self.eventer.send(type, uuid=self.id, **fields) def on_accepted(self, pid, time_accepted): """Handler called when task is accepted by worker pool.""" self.worker_pid = pid self.time_start = time_accepted task_accepted(self) if not self.task.acks_late: self.acknowledge() self.send_event('task-started') if _does_debug: debug('Task accepted: %s[%s] pid:%r', self.name, self.id, pid) if self._terminate_on_ack is not None: self.terminate(*self._terminate_on_ack) def on_timeout(self, soft, timeout): """Handler called if the task times out.""" task_ready(self) if soft: warn('Soft time limit (%ss) exceeded for %s[%s]', soft, self.name, self.id) exc = SoftTimeLimitExceeded(soft) else: error('Hard time limit (%ss) exceeded for %s[%s]', timeout, self.name, self.id) exc = TimeLimitExceeded(timeout) if self.store_errors: self.task.backend.mark_as_failure(self.id, exc, request=self) if self.task.acks_late: self.acknowledge() def on_success(self, failed__retval__runtime, **kwargs): """Handler called if the task was successfully processed.""" failed, retval, runtime = failed__retval__runtime if failed: if isinstance(retval.exception, (SystemExit, KeyboardInterrupt)): raise retval.exception return self.on_failure(retval, return_ok=True) task_ready(self) if self.task.acks_late: self.acknowledge() self.send_event('task-succeeded', result=retval, runtime=runtime) def on_retry(self, exc_info): """Handler called if the task should be retried.""" if self.task.acks_late: self.acknowledge() self.send_event('task-retried', exception=safe_repr(exc_info.exception.exc), traceback=safe_str(exc_info.traceback)) def on_failure(self, exc_info, send_failed_event=True, return_ok=False): """Handler called if the task raised an exception.""" task_ready(self) if isinstance(exc_info.exception, MemoryError): raise MemoryError('Process got: %s' % (exc_info.exception,)) elif isinstance(exc_info.exception, Reject): return self.reject(requeue=exc_info.exception.requeue) elif isinstance(exc_info.exception, Ignore): return self.acknowledge() exc = exc_info.exception if isinstance(exc, Retry): return self.on_retry(exc_info) # These are special cases where the process would not have had # time to write the result. if self.store_errors: if isinstance(exc, Terminated): self._announce_revoked( 'terminated', True, string(exc), False) send_failed_event = False # already sent revoked event elif isinstance(exc, WorkerLostError) or not return_ok: self.task.backend.mark_as_failure( self.id, exc, request=self, ) # (acks_late) acknowledge after result stored. if self.task.acks_late: self.acknowledge() if send_failed_event: self.send_event( 'task-failed', exception=safe_repr(get_pickled_exception(exc_info.exception)), traceback=exc_info.traceback, ) if not return_ok: error('Task handler raised error: %r', exc, exc_info=exc_info.exc_info) def acknowledge(self): """Acknowledge task.""" if not self.acknowledged: self.on_ack(logger, self.connection_errors) self.acknowledged = True def reject(self, requeue=False): if not self.acknowledged: self.on_reject(logger, self.connection_errors, requeue) self.acknowledged = True def info(self, safe=False): return {'id': self.id, 'name': self.name, 'type': self.type, 'body': self.body, 'hostname': self.hostname, 'time_start': self.time_start, 'acknowledged': self.acknowledged, 'delivery_info': self.delivery_info, 'worker_pid': self.worker_pid} def __str__(self): return ' '.join([ self.humaninfo(), ' eta:[{0}]'.format(self.eta) if self.eta else '', ' expires:[{0}]'.format(self.expires) if self.expires else '', ]) shortinfo = __str__ def humaninfo(self): return '{0.name}[{0.id}]'.format(self) def __repr__(self): return '<{0}: {1}>'.format(type(self).__name__, self.humaninfo()) @property def tzlocal(self): if self._tzlocal is None: self._tzlocal = self.app.conf.CELERY_TIMEZONE return self._tzlocal @property def store_errors(self): return (not self.task.ignore_result or self.task.store_errors_even_if_ignored) @property def task_id(self): # XXX compat return self.id @task_id.setter # noqa def task_id(self, value): self.id = value @property def task_name(self): # XXX compat return self.name @task_name.setter # noqa def task_name(self, value): self.name = value @property def reply_to(self): # used by rpc backend when failures reported by parent process return self.request_dict['reply_to'] @property def correlation_id(self): # used similarly to reply_to return self.request_dict['correlation_id'] def create_request_cls(base, task, pool, hostname, eventer, ref=ref, revoked_tasks=revoked_tasks, task_ready=task_ready): from celery.app.trace import trace_task_ret as trace default_time_limit = task.time_limit default_soft_time_limit = task.soft_time_limit apply_async = pool.apply_async acks_late = task.acks_late events = eventer and eventer.enabled class Request(base): def execute_using_pool(self, pool, **kwargs): task_id = self.id if (self.expires or task_id in revoked_tasks) and self.revoked(): raise TaskRevokedError(task_id) time_limit, soft_time_limit = self.time_limits time_limit = time_limit or default_time_limit soft_time_limit = soft_time_limit or default_soft_time_limit result = apply_async( trace, args=(self.type, task_id, self.request_dict, self.body, self.content_type, self.content_encoding), accept_callback=self.on_accepted, timeout_callback=self.on_timeout, callback=self.on_success, error_callback=self.on_failure, soft_timeout=soft_time_limit, timeout=time_limit, correlation_id=task_id, ) # cannot create weakref to None self._apply_result = ref(result) if result is not None else result return result def on_success(self, failed__retval__runtime, **kwargs): failed, retval, runtime = failed__retval__runtime if failed: if isinstance(retval.exception, ( SystemExit, KeyboardInterrupt)): raise retval.exception return self.on_failure(retval, return_ok=True) task_ready(self) if acks_late: self.acknowledge() if events: self.send_event( 'task-succeeded', result=retval, runtime=runtime, ) return Request [end of celery/worker/request.py] </code> I need you to solve this issue by regenerating the full files in the code base that you would like to change. You can change as many files as you like. Please respond with a list of files and their revised contents in the following format. <example> [start of /src/this_file.py] import os def euclidean(a, b): if b == 0: return a return euclidean(b, a % b) [end of /src/this_file.py] [start of /src/another_file.py] def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) x, y = x0, y0 sx = -1 if x0 > x1 else 1 sy = -1 if y0 > y1 else 1 if dx > dy: err = dx / 2.0 while x != x1: points.append((x, y)) err -= dy if err < 0: y += sy err += dx x += sx else: err = dy / 2.0 while y != y1: points.append((x err -= dx if err < 0: x += sx err += dy y += sy points.append((x, y)) return points [end of /src/another_file.py] </example>
celery/celery
045b52f1450d6d5cc500e0057a4b498250dc5692
Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost.
This is deliberate as if a task is killed it may mean that the next invocation will also cause the same to happen. If the task is redelivered it may cause a loop where the same conditions occur again and again. Also, sadly you cannot distinguish processes killed by OOM from processes killed by other means, and if an administrator kills -9 a task going amok, you usually don't want that task to be called again. There could be a configuration option for not acking terminated tasks, but I'm not sure how useful that would be. A better solution could be to use `basic_reject(requeue=False)` instead of `basic_ack`, that way you can configure a dead letter queue so that the killed tasks will be sent to a queue for manual inspection. I must say, regardless of the status of this feature request, the documentation is misleading. Specifically, [this FAQ makes it seem that process failures would NOT acknowledge messages](http://celery.readthedocs.org/en/latest/faq.html#faq-acks-late-vs-retry). And [this FAQ boldface states](http://celery.readthedocs.org/en/latest/faq.html#id54) that in the event of a kill signal (9), that acks_late will allow the task to re-run (which again, is patently wrong based on this poorly documented behavior). Nowhere in the docs have I found that if the process _dies_, the message will be acknowledged, regardless of acks_late or not. (for instance, I have a set of 10k+ tasks, and some 1% of tasks wind up acknowledged but incomplete when a WorkerLostError is thrown in connection with the worker, although there are no other errors of any kind in any of my logs related to that task). TL;DR at the least, appropriately document the current state when describing the functionality and limitations of acks_late. A work-around would be helpful -- I'm not sure I understand the solution of using `basic_reject`, although I'll keep looking into it. The docs are referring to killing the worker process with KILL, not the child processes. The term worker will always refer to the worker instance, not the pool processes. The section within about acks_late is probably not very helpful and should be removed
2015-10-06T05:34:34Z
<patch> <patch> diff --git a/celery/app/defaults.py b/celery/app/defaults.py --- a/celery/app/defaults.py +++ b/celery/app/defaults.py @@ -132,6 +132,7 @@ def __repr__(self): 'REDIS_DB': Option(type='int', **_REDIS_OLD), 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 'REDIS_MAX_CONNECTIONS': Option(type='int'), + 'REJECT_ON_WORKER_LOST': Option(type='bool'), 'RESULT_BACKEND': Option(type='string'), 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 'RESULT_DB_TABLENAMES': Option(type='dict'), diff --git a/celery/app/task.py b/celery/app/task.py --- a/celery/app/task.py +++ b/celery/app/task.py @@ -220,6 +220,12 @@ class Task(object): #: :setting:`CELERY_ACKS_LATE` setting. acks_late = None + #: When CELERY_ACKS_LATE is set to True, the default behavior to + #: handle worker crash is to acknowledge the message. Setting + #: this to true allows the message to be rejected and requeued so + #: it will be executed again by another worker. + reject_on_worker_lost = None + #: Tuple of expected exceptions. #: #: These are errors that are expected in normal operation @@ -248,6 +254,7 @@ class Task(object): ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), ('track_started', 'CELERY_TRACK_STARTED'), ('acks_late', 'CELERY_ACKS_LATE'), + ('reject_on_worker_lost', 'CELERY_REJECT_ON_WORKER_LOST'), ('ignore_result', 'CELERY_IGNORE_RESULT'), ('store_errors_even_if_ignored', 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), diff --git a/celery/worker/request.py b/celery/worker/request.py --- a/celery/worker/request.py +++ b/celery/worker/request.py @@ -326,7 +326,6 @@ def on_retry(self, exc_info): def on_failure(self, exc_info, send_failed_event=True, return_ok=False): """Handler called if the task raised an exception.""" task_ready(self) - if isinstance(exc_info.exception, MemoryError): raise MemoryError('Process got: %s' % (exc_info.exception,)) elif isinstance(exc_info.exception, Reject): @@ -352,7 +351,13 @@ def on_failure(self, exc_info, send_failed_event=True, return_ok=False): ) # (acks_late) acknowledge after result stored. if self.task.acks_late: - self.acknowledge() + reject_and_requeue = (self.task.reject_on_worker_lost and + isinstance(exc, WorkerLostError) and + self.delivery_info.get('redelivered', False) is False) + if reject_and_requeue: + self.reject(requeue=True) + else: + self.acknowledge() if send_failed_event: self.send_event( </patch> </s> </patch>
diff --git a/celery/tests/worker/test_request.py b/celery/tests/worker/test_request.py --- a/celery/tests/worker/test_request.py +++ b/celery/tests/worker/test_request.py @@ -325,6 +325,20 @@ def test_on_failure_Reject_rejects_with_requeue(self): req_logger, req.connection_errors, True, ) + def test_on_failure_WrokerLostError_rejects_with_requeue(self): + einfo = None + try: + raise WorkerLostError() + except: + einfo = ExceptionInfo(internal=True) + req = self.get_request(self.add.s(2, 2)) + req.task.acks_late = True + req.task.reject_on_worker_lost = True + req.delivery_info['redelivered'] = False + req.on_failure(einfo) + req.on_reject.assert_called_with(req_logger, + req.connection_errors, True) + def test_tzlocal_is_cached(self): req = self.get_request(self.add.s(2, 2)) req._tzlocal = 'foo'
1.0
NVIDIA__NeMo-473
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ``` </issue> <code> [start of README.rst] .. image:: http://www.repostatus.org/badges/latest/active.svg :target: http://www.repostatus.org/#active :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. .. image:: https://img.shields.io/badge/documentation-github.io-blue.svg :target: https://nvidia.github.io/NeMo/ :alt: NeMo documentation on GitHub pages .. image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE :alt: NeMo core license and license for collections in this repo .. image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python :alt: Language grade: Python .. image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ :alt: Total alerts .. image:: https://img.shields.io/badge/code%20style-black-000000.svg :target: https://github.com/psf/black :alt: Code style: black NVIDIA Neural Modules: NeMo =========================== NeMo is a toolkit for defining and building `Conversational AI <https://developer.nvidia.com/conversational-ai#started>`_ applications. Goal of the NeMo toolkit is to make it possible for researchers to easily compose complex neural network architectures for conversational AI using reusable components. Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. **Neural Modules** are conceptual blocks of neural networks that take *typed* inputs and produce *typed* outputs. Such modules typically represent data layers, encoders, decoders, language models, loss functions, or methods of combining activations. The toolkit comes with extendable collections of pre-built modules for automatic speech recognition (ASR), natural language processing (NLP) and text synthesis (TTS). **Introduction** * Watch `this video <https://nvidia.github.io/NeMo/>`_ for a quick walk-through. * Documentation (latest released version): https://nvidia.github.io/NeMo/ * Read NVIDIA `Developer Blog for example applications <https://devblogs.nvidia.com/how-to-build-domain-specific-automatic-speech-recognition-models-on-gpus/>`_ * Read NVIDIA `Developer Blog for Quartznet ASR model <https://devblogs.nvidia.com/develop-smaller-speech-recognition-models-with-nvidias-nemo-framework/>`_ * Recommended version to install is **0.9.0** via ``pip install nemo-toolkit`` * Recommended NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ * Pretrained models are available on NVIDIA `NGC Model repository <https://ngc.nvidia.com/catalog/models?orderBy=modifiedDESC&query=nemo&quickFilter=models&filters=>`_ Getting started ~~~~~~~~~~~~~~~ THE LATEST STABLE VERSION OF NeMo is **0.9.0** (Available via PIP). **Requirements** 1) Python 3.6 or 3.7 2) PyTorch 1.4.* with GPU support 3) (optional, for best performance) NVIDIA APEX. Install from here: https://github.com/NVIDIA/apex **NeMo Docker Container** NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ is now available. * Pull the docker: ``docker pull nvcr.io/nvidia/nemo:v0.9`` * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/nemo:v0.9`` If you are using the NVIDIA `NGC PyTorch container <https://ngc.nvidia.com/catalog/containers/nvidia:pytorch>`_ follow these instructions * Pull the docker: ``docker pull nvcr.io/nvidia/pytorch:20.01-py3`` * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/pytorch:20.01-py3`` * ``apt-get update && apt-get install -y libsndfile1`` * ``pip install nemo_toolkit`` NeMo core * ``pip install nemo_asr`` NeMo ASR (Speech Recognition) collection * ``pip install nemo_nlp`` NeMo NLP (Natural Language Processing) collection * ``pip install nemo_tts`` NeMo TTS (Speech Synthesis) collection See `examples/start_here` to get started with the simplest example. The folder `examples` contains several examples to get you started with various tasks in NLP and ASR. **Tutorials** * `Speech recognition <https://nvidia.github.io/NeMo/asr/intro.html>`_ * `Natural language processing <https://nvidia.github.io/NeMo/nlp/intro.html>`_ * `Speech Synthesis <https://nvidia.github.io/NeMo/tts/intro.html>`_ DEVELOPMENT ~~~~~~~~~~~ If you'd like to use master branch and/or develop NeMo you can run "reinstall.sh" script. `Documentation (master branch) <http://nemo-master-docs.s3-website.us-east-2.amazonaws.com/>`_. **Installing From Github** If you prefer to use NeMo's latest development version (from GitHub) follow the steps below: 1) Clone the repository ``git clone https://github.com/NVIDIA/NeMo.git`` 2) Go to NeMo folder and re-install the toolkit with collections: .. code-block:: bash ./reinstall.sh **Style tests** .. code-block:: bash python setup.py style # Checks overall project code style and output issues with diff. python setup.py style --fix # Tries to fix error in-place. python setup.py style --scope=tests # Operates within certain scope (dir of file). **Unittests** This command runs unittests: .. code-block:: bash ./reinstall.sh python pytest tests Citation ~~~~~~~~ If you are using NeMo please cite the following publication .. code-block:: tex @misc{nemo2019, title={NeMo: a toolkit for building AI applications using Neural Modules}, author={Oleksii Kuchaiev and Jason Li and Huyen Nguyen and Oleksii Hrinchuk and Ryan Leary and Boris Ginsburg and Samuel Kriman and Stanislav Beliaev and Vitaly Lavrukhin and Jack Cook and Patrice Castonguay and Mariya Popova and Jocelyn Huang and Jonathan M. Cohen}, year={2019}, eprint={1909.09577}, archivePrefix={arXiv}, primaryClass={cs.LG} } [end of README.rst] [start of nemo/backends/pytorch/actions.py] # Copyright (c) 2019 NVIDIA Corporation import copy import importlib import itertools import json import os from collections import defaultdict from contextlib import ExitStack from pathlib import Path from typing import List, Optional import torch import torch.distributed as dist import torch.nn as nn import torch.optim as optim from torch.nn.parallel import DistributedDataParallel as DDP from nemo import logging from nemo.backends.pytorch.module_wrapper import TrainableNeuralModuleWrapper from nemo.backends.pytorch.nm import DataLayerNM, TrainableNM from nemo.backends.pytorch.optimizers import AdamW, Novograd, master_params from nemo.core import DeploymentFormat, DeviceType, NeuralModule, NmTensor from nemo.core.callbacks import ActionCallback, EvaluatorCallback, SimpleLossLoggerCallback from nemo.core.neural_factory import Actions, ModelMode, Optimization from nemo.core.neural_types import * from nemo.utils.helpers import get_checkpoint_from_dir # these imports will happen on as-needed basis amp = None # convert_syncbn = None # create_syncbn_process_group = None LARC = None FusedLAMB = None FusedAdam = None FusedNovoGrad = None AmpOptimizations = { Optimization.mxprO0: "O0", Optimization.mxprO1: "O1", Optimization.mxprO2: "O2", Optimization.mxprO3: "O3", } _float_2_half_req = { Optimization.mxprO1, Optimization.mxprO2, Optimization.mxprO3, } class PtActions(Actions): def __init__( self, local_rank=None, global_rank=None, tb_writer=None, optimization_level=Optimization.mxprO0, ): need_apex = local_rank is not None or optimization_level != Optimization.mxprO0 if need_apex: try: apex = importlib.import_module('apex') if optimization_level != Optimization.mxprO0: global amp amp = importlib.import_module('apex.amp') if local_rank is not None: # global convert_syncbn # global create_syncbn_process_group global LARC global FusedLAMB global FusedAdam global FusedNovoGrad parallel = importlib.import_module('apex.parallel') apex_optimizer = importlib.import_module('apex.optimizers') # convert_syncbn = parallel.convert_syncbn_model # create_syncbn_process_group = parallel.create_syncbn_process_group LARC = parallel.LARC FusedLAMB = apex_optimizer.FusedLAMB FusedAdam = apex_optimizer.FusedAdam FusedNovoGrad = apex_optimizer.FusedNovoGrad except ImportError: raise ImportError( "NVIDIA Apex is necessary for distributed training and" "mixed precision training. It only works on GPUs." "Please install Apex from " "https://www.github.com/nvidia/apex" ) super(PtActions, self).__init__( local_rank=local_rank, global_rank=global_rank, optimization_level=optimization_level, ) # will be [unique_instance_id -> (NMModule, PTModule)] self.module_reference_table = {} self.step = 0 self.epoch_num = 0 self.optimizers = [] self.tb_writer = tb_writer self._modules = set() self.cache = None self.amp_initialized = False @property def modules(self): return self._modules def __get_top_sorted_modules_and_dataloader(self, hook): """ Constructs DAG leading to hook and creates its topological order. It also populates self.module_reference_table. Args: hook: an NmTensor or a list of NmTensors representing leaf nodes in DAG Returns: list of modules with their call arguments and outputs, and dataset """ def create_node(producer, producer_args): if producer_args is None: return tuple((producer, ())) else: return tuple((producer, tuple([(k, v) for k, v in producer_args.items()]),)) def is_in_degree_zero(node, processed_nodes): """A node has in degree of zero""" if node[1] == (): return True for portname, nmtensor in node[1]: nd = create_node(nmtensor.producer, nmtensor.producer_args) if nd not in processed_nodes: return False return True hooks = hook if isinstance(hook, list) else [hook] # ensures that no tensors are processed twice processed_nmtensors = set() indices_to_remove = [] # Check for duplicates in hook for i, nmtensor in enumerate(hook): if nmtensor in processed_nmtensors: indices_to_remove.append(i) else: processed_nmtensors.add(nmtensor) for i in reversed(indices_to_remove): hook.pop(i) _top_sorted_modules = [] all_nodes = {} # extract all nodes to all_nodes set hooks_lst = list(hooks) while len(hooks_lst) > 0: # take nmtensor from the end of the list nmtensor = hooks_lst.pop() node = create_node(nmtensor.producer, nmtensor.producer_args) # Store nmtensor as an output of its producer # first make sure all keys are present per output port # and nm is inside all_nodes if node not in all_nodes: all_nodes[node] = {k: None for k in nmtensor.producer.output_ports} # second, populate output port with current nmtensor # where applicable all_nodes[node][nmtensor.name] = nmtensor processed_nmtensors.add(nmtensor) if nmtensor.producer_args is not None and nmtensor.producer_args != {}: for _, new_nmtensor in nmtensor.producer_args.items(): if new_nmtensor not in processed_nmtensors: # put in the start of list hooks_lst.insert(0, new_nmtensor) all_node_with_output = [] # Iterate over all_nodes to create new nodes that include its output # now all nodes have (module, input tensors, output tensors) for node in all_nodes: all_node_with_output.append(tuple((node[0], node[1], all_nodes[node]))) processed_nodes = [] while len(all_node_with_output) > 0: for node in all_node_with_output.copy(): # if node's in_degree is zero it can be added to # _top_sorted_modules # this will also reduce in_degree of its children if is_in_degree_zero(node, processed_nodes): _top_sorted_modules.append(node) processed_nodes.append((node[0], node[1])) all_node_with_output.remove(node) # Create top_sorted_modules aka callchain top_sorted_modules = [] for i, m in enumerate(_top_sorted_modules): top_sorted_modules.append((m[0], dict(m[1]), m[2])) # Ensure that there is only one dataset in callchain if i > 0 and isinstance(m[0], DataLayerNM): raise ValueError("There were more than one DataLayer NeuralModule inside your DAG.") if not isinstance(top_sorted_modules[0][0], DataLayerNM): raise ValueError("The first module in your DAG was not a DataLayer NeuralModule.") tdataset = top_sorted_modules[0][0].dataset # populate self.module_reference_table for m in top_sorted_modules: if m[0].factory is None and self._local_rank is not None: raise ValueError( "Neural module {0} was created without " "NeuralModuleFactory, but you are trying to" "run in distributed mode. Please instantiate" "NeuralModuleFactory first and pass its " "instance as `factory` parameter to all your" "Neural Module objects." "".format(str(m[0])) ) key = m[0].unique_instance_id if key not in self.module_reference_table: if isinstance(m[0], TrainableNeuralModuleWrapper): self.module_reference_table[key] = (m[0], m[0]._pt_module) else: self.module_reference_table[key] = (m[0], m[0]) return top_sorted_modules, tdataset def create_optimizer(self, optimizer, things_to_optimize, optimizer_params=None): """ Wrapper function around __setup_optimizer() Args: optimizer : A instantiated PyTorch optimizer or string. For currently supported strings, see __setup_optimizer(). things_to_optimize (list): Must be a list of Neural Modules and/or parameters. If a Neural Module is passed, all trainable parameters are extracted and passed to the optimizer. optimizer_params (dict): Optional parameters dictionary. Returns: Optimizer """ optimizer_instance = None optimizer_class = None if isinstance(optimizer, str): optimizer_class = optimizer elif isinstance(optimizer, torch.optim.Optimizer): optimizer_instance = optimizer else: raise ValueError("`optimizer` must be a string or an instance of torch.optim.Optimizer") modules_to_optimize = [] tensors_to_optimize = [] if not isinstance(things_to_optimize, list): things_to_optimize = [things_to_optimize] for thing in things_to_optimize: if isinstance(thing, NeuralModule): modules_to_optimize.append(thing) elif isinstance(thing, NmTensor): tensors_to_optimize.append(thing) else: raise ValueError( "{} passed to create_optimizer() was neither a neural module nor a neural module tensor" ) if tensors_to_optimize: call_chain, _ = self.__get_top_sorted_modules_and_dataloader(tensors_to_optimize) for module in call_chain: if module[0] not in modules_to_optimize: modules_to_optimize.append(module[0]) # Extract trainable weights which will be optimized params_list = [p.parameters() for p in modules_to_optimize if isinstance(p, TrainableNM) or p.is_trainable()] params_to_optimize = itertools.chain(*params_list) if optimizer_params is None: optimizer_params = {} # Init amp optimizer = self.__setup_optimizer( optimizer_instance=optimizer_instance, optimizer_class=optimizer_class, optimization_params=optimizer_params, params_to_optimize=params_to_optimize, ) self.optimizers.append(optimizer) return optimizer @staticmethod def __setup_optimizer( optimizer_instance, optimizer_class, optimization_params, params_to_optimize, ): if optimizer_instance is None: # Setup optimizer instance, by default it is SGD lr = optimization_params["lr"] if optimizer_class.lower() == "sgd": optimizer = optim.SGD( params_to_optimize, lr=lr, momentum=optimization_params.get("momentum", 0.9), weight_decay=optimization_params.get("weight_decay", 0.0), ) elif optimizer_class.lower() == "adam": optimizer = optim.Adam( params=params_to_optimize, lr=lr, betas=optimization_params.get("betas", (0.9, 0.999)), ) elif optimizer_class.lower() == "fused_adam": optimizer = FusedAdam(params=params_to_optimize, lr=lr) elif optimizer_class.lower() == "adam_w": optimizer = AdamW( params=params_to_optimize, lr=lr, weight_decay=optimization_params.get("weight_decay", 0.0), betas=optimization_params.get("betas", (0.9, 0.999)), ) elif optimizer_class.lower() == "novograd": optimizer = Novograd( params_to_optimize, lr=lr, weight_decay=optimization_params.get("weight_decay", 0.0), luc=optimization_params.get("luc", False), luc_trust=optimization_params.get("luc_eta", 1e-3), betas=optimization_params.get("betas", (0.95, 0.25)), ) elif optimizer_class.lower() == "fused_novograd": optimizer = FusedNovoGrad( params_to_optimize, lr=lr, weight_decay=optimization_params.get("weight_decay", 0.0), reg_inside_moment=True, grad_averaging=False, betas=optimization_params.get("betas", (0.95, 0.25)), ) elif optimizer_class.lower() == "fused_lamb": optimizer = FusedLAMB(params_to_optimize, lr=lr,) else: raise ValueError("Unknown optimizer class: {0}".format(optimizer_class)) if optimization_params.get("larc", False): logging.info("Enabling larc") optimizer = LARC(optimizer, trust_coefficient=optimization_params.get("larc_eta", 2e-2),) else: logging.info("Optimizer instance: {0} is provided.") if optimizer_class is not None and optimizer_class != "": logging.warning("Ignoring `optimizer_class` parameter because `optimizer_instance` is provided") if optimization_params is not None and optimization_params != {}: logging.warning( "Ignoring `optimization_params` parameter for " "optimizer because `optimizer_instance` is provided" ) optimizer = optimizer_instance return optimizer def __initialize_amp( self, optimizer, optim_level, amp_max_loss_scale=2.0 ** 24, amp_min_loss_scale=1.0, ): if optim_level not in AmpOptimizations: raise ValueError(f"__initialize_amp() was called with unknown optim_level={optim_level}") # in this case, nothing to do here if optim_level == Optimization.mxprO0: return optimizer if len(self.modules) < 1: raise ValueError("There were no modules to initialize") pt_modules = [] for module in self.modules: if isinstance(module, nn.Module): pt_modules.append(module) elif isinstance(module, TrainableNeuralModuleWrapper): pt_modules.append(module._pt_module) _, optimizer = amp.initialize( max_loss_scale=amp_max_loss_scale, min_loss_scale=amp_min_loss_scale, models=pt_modules, optimizers=optimizer, opt_level=AmpOptimizations[optim_level], ) self.amp_initialized = True return optimizer def __nm_graph_forward_pass( self, call_chain, registered_tensors, mode=ModelMode.train, use_cache=False, ): for ind in range(1, len(call_chain)): if use_cache: in_cache = True for tensor in call_chain[ind][2].values(): if tensor is None: # NM has an output tensor that is not used in the # current call chain, so we don't care if it's not in # cache continue if tensor.unique_name not in registered_tensors: in_cache = False if in_cache: continue call_args = call_chain[ind][1] # module = call_chain[ind][0] m_id = call_chain[ind][0].unique_instance_id pmodule = self.module_reference_table[m_id][1] # if self._local_rank is not None: # if isinstance(pmodule, DDP): # if disable_allreduce: # pmodule.disable_allreduce() # else: # pmodule.enable_allreduce() if mode == ModelMode.train: # if module.is_trainable(): if isinstance(pmodule, nn.Module): pmodule.train() elif mode == ModelMode.eval: # if module.is_trainable(): if isinstance(pmodule, nn.Module): pmodule.eval() else: raise ValueError("Unknown ModelMode") # prepare call signature for `module` call_set = {} for tensor_name, nmtensor in call_args.items(): # _add_uuid_2_name(nmtensor.name, nmtensor.producer._uuid) key = nmtensor.unique_name call_set[tensor_name] = registered_tensors[key] # actual PyTorch module call with signature if isinstance(self.module_reference_table[m_id][0], TrainableNeuralModuleWrapper,): new_tensors = pmodule(**call_set) else: new_tensors = pmodule(force_pt=True, **call_set) if not isinstance(new_tensors, List): if not isinstance(new_tensors, tuple): new_tensors = [new_tensors] else: new_tensors = list(new_tensors) for t_tensor, nm_tensor in zip(new_tensors, call_chain[ind][2].values()): if nm_tensor is None: continue t_name = nm_tensor.unique_name if t_name not in registered_tensors: registered_tensors[t_name] = t_tensor else: raise ValueError("A NMTensor was produced twice in " f"the same DAG. {t_name}") @staticmethod def pad_tensor(t: torch.Tensor, target_size: torch.Size): padded_shape = target_size.cpu().data.numpy().tolist() padded_t = torch.zeros(padded_shape).cuda().type_as(t) t_size = t.size() if len(t_size) == 0: padded_t = t elif len(t_size) == 1: padded_t[: t_size[0]] = t elif len(t_size) == 2: padded_t[: t_size[0], : t_size[1]] = t elif len(t_size) == 3: padded_t[: t_size[0], : t_size[1], : t_size[2]] = t elif len(t_size) == 4: padded_t[: t_size[0], : t_size[1], : t_size[2], : t_size[3]] = t else: raise NotImplementedError return padded_t @staticmethod def depad_tensor(t: torch.Tensor, original_size: torch.Size): t_size = original_size if len(t_size) == 0: depadded_t = t elif len(t_size) == 1: depadded_t = t[: t_size[0]] elif len(t_size) == 2: depadded_t = t[: t_size[0], : t_size[1]] elif len(t_size) == 3: depadded_t = t[: t_size[0], : t_size[1], : t_size[2]] elif len(t_size) == 4: depadded_t = t[: t_size[0], : t_size[1], : t_size[2], : t_size[3]] else: raise NotImplementedError return depadded_t def _eval(self, tensors_2_evaluate, callback, step, verbose=False): """ Evaluation process. WARNING THIS function assumes that all tensors_2_evaluate are based on a single datalayer Args: tensors_2_evaluate: list of NmTensors to evaluate callback: instance of EvaluatorCallback step: current training step, used for logging Returns: None """ with torch.no_grad(): # each call chain corresponds to a tensor in tensors_2_evaluate call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors_2_evaluate) # "Retrieve" data layer from call chain. dl_nm = call_chain[0][0] # Prepare eval_dataloader # For distributed training it should have disjoint subsets of # all data on every worker is_distributed = False world_size = None if dl_nm.placement == DeviceType.AllGpu: assert dist.is_initialized() is_distributed = True world_size = torch.distributed.get_world_size() # logging.info( # "Doing distributed evaluation. Rank {0} of {1}".format( # self.local_rank, world_size # ) # ) if dl_nm.dataset is not None: sampler = torch.utils.data.distributed.DistributedSampler( dataset=dl_nm.dataset, shuffle=dl_nm.shuffle ) eval_dataloader = torch.utils.data.DataLoader( dataset=dl_nm.dataset, sampler=sampler, num_workers=dl_nm.num_workers, batch_size=dl_nm.batch_size, shuffle=False, ) else: eval_dataloader = dl_nm.data_iterator if hasattr(eval_dataloader, 'sampler'): eval_dataloader.sampler.set_epoch(0) else: # Not distributed if dl_nm.dataset is not None: # Todo: remove local_parameters eval_dataloader = torch.utils.data.DataLoader( dataset=dl_nm.dataset, sampler=None, # not distributed sampler num_workers=dl_nm.num_workers, batch_size=dl_nm.batch_size, shuffle=dl_nm.shuffle, ) else: eval_dataloader = dl_nm.data_iterator # after this eval_dataloader is ready to be used # reset global_var_dict - results of evaluation will be stored # there callback.clear_global_var_dict() dl_device = dl_nm._device # Evaluation mini-batch for loop num_batches = None if hasattr(eval_dataloader, "__len__"): num_batches = len(eval_dataloader) for epoch_i, data in enumerate(eval_dataloader, 0): if ( verbose and num_batches is not None and (num_batches < 10 or (epoch_i % int(num_batches / 10) == 0)) ): logging.info(f"Evaluating batch {epoch_i} out of {num_batches}") tensors = [] if isinstance(data, torch.Tensor): data = (data,) for d in data: if isinstance(d, torch.Tensor): tensors.append(d.to(dl_device)) else: tensors.append(d) registered_e_tensors = { t.unique_name: d for t, d in zip(call_chain[0][2].values(), tensors) if t is not None } self.__nm_graph_forward_pass( call_chain=call_chain, registered_tensors=registered_e_tensors, mode=ModelMode.eval, ) if not is_distributed or self.global_rank == 0: values_dict = {} # If distributed. For the outer loop, we need to ensure that # all processes loop through the elements in the same order for t2e in tensors_2_evaluate: key = t2e.unique_name if key not in registered_e_tensors.keys(): logging.info("WARNING: Tensor {} was not found during eval".format(key)) continue if is_distributed: # where we will all_gather results from all workers tensors_list = [] # where we will all_gather tensor sizes tensor_on_worker = registered_e_tensors[key] if tensor_on_worker.shape != torch.Size([]): tensor_on_worker_size_as_tensor = torch.tensor(tensor_on_worker.shape).cuda() sizes = [] for ind in range(world_size): sizes.append(torch.empty_like(tensor_on_worker_size_as_tensor)) dist.all_gather(sizes, tensor_on_worker_size_as_tensor) mx_dim, _ = torch.max(torch.stack(sizes), dim=0) else: # this is a singleton. For example, loss value sizes = [torch.Size([])] * world_size mx_dim = None for ind in range(world_size): # we have to use max shape for all_gather if mx_dim is None: # singletons tensors_list.append(torch.tensor(2).cuda().type_as(tensor_on_worker)) else: # non-singletons tensors_list.append( torch.empty(mx_dim.cpu().data.numpy().tolist()).cuda().type_as(tensor_on_worker) ) if mx_dim is not None: t_to_send = self.pad_tensor(tensor_on_worker, mx_dim) else: t_to_send = tensor_on_worker dist.all_gather(tensors_list, t_to_send) tensors_list = [self.depad_tensor(t, size) for t, size in zip(tensors_list, sizes)] if self.global_rank == 0: values_dict["IS_FROM_DIST_EVAL"] = True values_dict[key] = tensors_list else: # NON-DISTRIBUTED TRAINING values_dict["IS_FROM_DIST_EVAL"] = False values_dict[key] = [registered_e_tensors[key]] if callback.user_iter_callback and (self.global_rank is None or self.global_rank == 0): # values_dict will contain results from all workers callback.user_iter_callback(values_dict, callback._global_var_dict) # final aggregation (over minibatches) and logging of results # should happend only on one worker if callback.user_done_callback and (self.global_rank is None or self.global_rank == 0): vals_to_log = callback.user_done_callback(callback._global_var_dict) # log results to Tensorboard or Weights & Biases if vals_to_log is not None: if hasattr(callback, 'swriter') and callback.swriter is not None: if hasattr(callback, 'tb_writer_func') and callback.tb_writer_func is not None: callback.tb_writer_func(callback.swriter, vals_to_log, step) else: for key, val in vals_to_log.items(): callback.swriter.add_scalar(key, val, step) if hasattr(callback, 'wandb_log'): callback.wandb_log(vals_to_log) def _infer( self, tensors_to_return, verbose=False, cache=False, use_cache=False, offload_to_cpu=True, ): """ Does the same as _eval() just with tensors instead of eval callback. """ # Checking that cache is used properly if cache and use_cache: raise ValueError( "cache and use_cache were both set. However cache must first be created prior to using it." ) if cache: if self.cache is not None: raise ValueError("cache was set but was not empty") self.cache = [] if use_cache: if not self.cache: raise ValueError("use_cache was set, but cache was empty") with torch.no_grad(): # each call chain corresponds to a tensor in tensors_2_evaluate dl_nm = None call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors_to_return) dl_nm = call_chain[0][0] # Prepare eval_dataloader # For distributed training it should have disjoint subsets of # all data on every worker is_distributed = False world_size = None if dl_nm.placement == DeviceType.AllGpu: if self.cache or use_cache: raise NotImplementedError("Caching is not available for distributed training.") assert dist.is_initialized() is_distributed = True world_size = torch.distributed.get_world_size() # logging.info( # "Doing distributed evaluation. Rank {0} of {1}".format( # self.local_rank, world_size # ) # ) if dl_nm.dataset is not None: sampler = torch.utils.data.distributed.DistributedSampler( dataset=dl_nm.dataset, shuffle=dl_nm.shuffle ) eval_dataloader = torch.utils.data.DataLoader( dataset=dl_nm.dataset, sampler=sampler, num_workers=dl_nm.num_workers, batch_size=dl_nm.batch_size, shuffle=False, ) else: eval_dataloader = dl_nm.data_iterator eval_dataloader.sampler.set_epoch(0) elif not use_cache: # Not distributed and not using cache # Dataloaders are only used if use_cache is False # When caching, the DAG must cache all outputs from dataloader if dl_nm.dataset is not None: # Todo: remove local_parameters eval_dataloader = torch.utils.data.DataLoader( dataset=dl_nm.dataset, sampler=None, # not distributed sampler num_workers=dl_nm.num_workers, batch_size=dl_nm.batch_size, shuffle=dl_nm.shuffle, ) else: eval_dataloader = dl_nm.data_iterator # after this eval_dataloader is ready to be used # reset global_var_dict - results of evaluation will be stored # there if not is_distributed or self.global_rank == 0: values_dict = {} for t in tensors_to_return: values_dict[t.unique_name] = [] dl_device = dl_nm._device # Evaluation mini-batch for loop if use_cache: num_batches = len(self.cache) loop_iterator = self.cache else: num_batches = len(eval_dataloader) loop_iterator = eval_dataloader for epoch_i, data in enumerate(loop_iterator, 0): logging.debug(torch.cuda.memory_allocated()) if verbose and (num_batches < 10 or (epoch_i % int(num_batches / 10) == 0)): logging.info(f"Evaluating batch {epoch_i} out of {num_batches}") tensors = [] if use_cache: registered_e_tensors = data # delete tensors_to_return for t in tensors_to_return: if t.unique_name in registered_e_tensors: del registered_e_tensors[t.unique_name] # Need to check for device type mismatch for t in registered_e_tensors: registered_e_tensors[t].to(dl_device) else: if isinstance(data, torch.Tensor): data = (data,) for d in data: if isinstance(d, torch.Tensor): tensors.append(d.to(dl_device)) else: tensors.append(d) registered_e_tensors = { t.unique_name: d for t, d in zip(call_chain[0][2].values(), tensors) if t is not None } self.__nm_graph_forward_pass( call_chain=call_chain, registered_tensors=registered_e_tensors, mode=ModelMode.eval, use_cache=use_cache, ) # if offload_to_cpu: # # Take all cuda tensors and save them to value_dict as # # cpu tensors to save GPU memory # for name, tensor in registered_e_tensors.items(): # if isinstance(tensor, torch.Tensor): # registered_e_tensors[name] = tensor.cpu() if cache: self.append_to_cache(registered_e_tensors, offload_to_cpu) # If distributed. For the outer loop, we need to ensure that # all processes loop through the elements in the same order for t2e in tensors_to_return: key = t2e.unique_name if key not in registered_e_tensors.keys(): logging.info("WARNING: Tensor {} was not found during eval".format(key)) continue if is_distributed: # where we will all_gather results from all workers tensors_list = [] # where we will all_gather tensor sizes tensor_on_worker = registered_e_tensors[key] if tensor_on_worker.shape != torch.Size([]): tensor_on_worker_size_as_tensor = torch.tensor(tensor_on_worker.shape).cuda() sizes = [] for ind in range(world_size): sizes.append(torch.empty_like(tensor_on_worker_size_as_tensor)) dist.all_gather(sizes, tensor_on_worker_size_as_tensor) mx_dim, _ = torch.max(torch.stack(sizes), dim=0) else: # this is a singleton. For example, loss value sizes = [torch.Size([])] * world_size mx_dim = None for ind in range(world_size): # we have to use max shape for all_gather if mx_dim is None: # singletons tensors_list.append(torch.tensor(2).cuda().type_as(tensor_on_worker)) else: # non-singletons tensors_list.append( torch.empty(mx_dim.cpu().data.numpy().tolist()).cuda().type_as(tensor_on_worker) ) if mx_dim is not None: t_to_send = self.pad_tensor(tensor_on_worker, mx_dim) else: t_to_send = tensor_on_worker dist.all_gather(tensors_list, t_to_send) tensors_list = [self.depad_tensor(t, size) for t, size in zip(tensors_list, sizes)] if offload_to_cpu: tensors_list = [t.cpu() for t in tensors_list] if self.global_rank == 0: values_dict[key] += tensors_list else: # NON-DISTRIBUTED TRAINING tensor = registered_e_tensors[key] if offload_to_cpu and isinstance(tensor, torch.Tensor): tensor = tensor.cpu() values_dict[key] += [tensor] if not is_distributed or self.global_rank == 0: inferred_tensors = [] for t in tensors_to_return: inferred_tensors.append(values_dict[t.unique_name]) return inferred_tensors # For all other ranks return None def append_to_cache(self, registered_tensors: dict, offload_to_cpu): """Simpler helper function to add results of __nm_graph_forward_pass to current cache. """ if offload_to_cpu: for t in registered_tensors: registered_tensors[t] = registered_tensors[t].cpu() self.cache.append(registered_tensors) def clear_cache(self): """ Simple helpful function to clear cache by setting self.cache to None """ self.cache = None def save_state_to(self, path: str): """ Saves current state such as step, epoch and optimizer parameters Args: path: Returns: """ state = { "step": self.step, "epoch_num": self.epoch_num, "optimizer_state": [opt.state_dict() for opt in self.optimizers], } torch.save(state, path) def restore_state_from(self, path: str): """ Restores state such as step, epoch and optimizer parameters Args: path: Returns: """ if os.path.isfile(path): # map_location could be cuda:<device_id> but cpu seems to be more # general since we are also saving step and epoch_num # load_state_dict should move the variables to the relevant device checkpoint = torch.load(path, map_location="cpu") self.step = checkpoint["step"] self.epoch_num = checkpoint["epoch_num"] if checkpoint["optimizer_state"]: for opt, opt_chkpt in zip(self.optimizers, checkpoint["optimizer_state"]): opt.load_state_dict(opt_chkpt) else: raise FileNotFoundError("Could not find checkpoint file: {0}".format(path)) @staticmethod def _check_all_tensors(list_of_tensors): """Method that checks if the passed list contains all NmTensors """ if not isinstance(list_of_tensors, list): return False for tensor in list_of_tensors: if not isinstance(tensor, NmTensor): return False return True @staticmethod def _check_tuples(list_of_tuples): """Method that checks if the passed tuple contains an optimizer in the first element, and a list of NmTensors in the second. """ for tup in list_of_tuples: if not (isinstance(tup[0], torch.optim.Optimizer) and PtActions._check_all_tensors(tup[1])): return False return True def _get_all_modules(self, training_loop, callbacks, logging_callchain=None): """Gets all neural modules that will be used by train() and eval() via EvaluatorCallbacks. Saves all modules to self.modules """ # If there is a SimpleLossLoggerCallback, create an logger_callchain # with all callchains from training_loop and # SimpleLossLoggerCallback.tensors if logging_callchain: for module in logging_callchain: self.modules.add(module[0]) # Else grab all callchains from training_loop else: for step in training_loop: for module in step[2]: self.modules.add(module[0]) # Lastly, grab all eval modules if callbacks is not None: for callback in callbacks: if isinstance(callback, EvaluatorCallback): (callchain, _,) = self.__get_top_sorted_modules_and_dataloader(hook=callback.eval_tensors) for module in callchain: self.modules.add(module[0]) @staticmethod def __module_export(module, output, d_format: DeploymentFormat, input_example=None, output_example=None): # Check if output already exists destination = Path(output) if destination.exists(): raise FileExistsError(f"Destination {output} already exists. " f"Aborting export.") input_names = list(module.input_ports.keys()) output_names = list(module.output_ports.keys()) dynamic_axes = defaultdict(list) def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defaultdict): if ntype.axes: for ind, axis in enumerate(ntype.axes): if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: dynamic_axes[port_name].append(ind) # This is a hack for Jasper to Jarvis export -- need re-design for this inputs_to_drop = set() outputs_to_drop = set() if type(module).__name__ == "JasperEncoder": logging.info( "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " "deployment" ) inputs_to_drop.add("length") outputs_to_drop.add("encoded_lengths") # for input_ports for port_name, ntype in module.input_ports.items(): if port_name in inputs_to_drop: input_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) # for output_ports for port_name, ntype in module.output_ports.items(): if port_name in outputs_to_drop: output_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) if len(dynamic_axes) == 0: dynamic_axes = None # Make a deep copy of init parameters. init_params_copy = copy.deepcopy(module._init_params) # Remove NeMo-related things from the module # We need to change __call__ method. Note that this will change the # whole class, not just this object! Which is why we need to repair it # in the finally block type(module).__call__ = torch.nn.Module.__call__ # Reset standard instance field - making the file (probably) lighter. module._init_params = None module._placement = None module._factory = None module._device = None module.eval() try: if d_format == DeploymentFormat.TORCHSCRIPT: if input_example is None: # Route 1 - via torch.jit.script traced_m = torch.jit.script(module) traced_m.save(output) else: # Route 2 - via tracing traced_m = torch.jit.trace(module, input_example) traced_m.save(output) elif d_format == DeploymentFormat.ONNX or d_format == DeploymentFormat.TRTONNX: if input_example is None: raise ValueError(f'Example input is None, but ONNX tracing was' f' attempted') if output_example is None: if isinstance(input_example, tuple): output_example = module.forward(*input_example) else: output_example = module.forward(input_example) with torch.jit.optimized_execution(True): jitted_model = torch.jit.trace(module, input_example) torch.onnx.export( jitted_model, input_example, output, input_names=input_names, output_names=output_names, verbose=False, export_params=True, do_constant_folding=True, dynamic_axes=dynamic_axes, opset_version=11, example_outputs=output_example, ) # fn = output + ".readable" # with open(fn, 'w') as f: # tempModel = onnx.load(output) # onnx.save(tempModel, output + ".copy") # onnx.checker.check_model(tempModel) # pgraph = onnx.helper.printable_graph(tempModel.graph) # f.write(pgraph) elif d_format == DeploymentFormat.PYTORCH: torch.save(module.state_dict(), output) with open(output + ".json", 'w') as outfile: json.dump(init_params_copy, outfile) else: raise NotImplementedError(f"Not supported deployment format: {d_format}") except Exception as e: # nopep8 logging.error(f'module export failed for {module} ' f'with exception {e}') finally: def __old_call__(self, force_pt=False, *input, **kwargs): pt_call = len(input) > 0 or force_pt if pt_call: return nn.Module.__call__(self, *input, **kwargs) else: return NeuralModule.__call__(self, **kwargs) type(module).__call__ = __old_call__ @staticmethod def deployment_export(module, output: str, d_format: DeploymentFormat, input_example=None, output_example=None): """Exports Neural Module instance for deployment. Args: module: neural module to export output (str): where export results should be saved d_format (DeploymentFormat): which deployment format to use input_example: sometimes tracing will require input examples output_example: Should match inference on input_example amp_max_loss_scale (float): Max value for amp loss scaling. Defaults to 2.0**24. """ with torch.no_grad(): PtActions.__module_export( module=module, output=output, d_format=d_format, input_example=input_example, output_example=output_example, ) def train( self, tensors_to_optimize, optimizer=None, optimization_params=None, callbacks: Optional[List[ActionCallback]] = None, lr_policy=None, batches_per_step=None, stop_on_nan_loss=False, synced_batchnorm=False, synced_batchnorm_groupsize=0, gradient_predivide=False, amp_max_loss_scale=2.0 ** 24, ): if gradient_predivide: logging.error( "gradient_predivide is currently disabled, and is under consideration for removal in future versions. " "If this functionality is needed, please raise a github issue." ) if not optimization_params: optimization_params = {} num_epochs = optimization_params.get("num_epochs", None) max_steps = optimization_params.get("max_steps", None) if num_epochs is None and max_steps is None: raise ValueError("You must specify either max_steps or num_epochs") grad_norm_clip = optimization_params.get('grad_norm_clip', None) if batches_per_step is None: batches_per_step = 1 # this is necessary because we average gradients over batch bps_scale = torch.FloatTensor([1.0 / batches_per_step]).squeeze() if tensors_to_optimize is None: # This is Evaluation Mode self._init_callbacks(callbacks) # Do action start callbacks self._perform_on_action_end(callbacks=callbacks) return # Check if tensors_to_optimize is just a list of NmTensors elif tensors_to_optimize is not None and ( isinstance(tensors_to_optimize[0], NmTensor) and PtActions._check_all_tensors(tensors_to_optimize) ): # Parse graph into a topologically sorted sequence of neural # modules' calls (opt_call_chain, t_dataset,) = self.__get_top_sorted_modules_and_dataloader(hook=tensors_to_optimize) # Extract trainable weights which will be optimized params_list = [ p[0].parameters() for p in opt_call_chain if isinstance(p[0], TrainableNM) or p[0].is_trainable() ] params_to_optimize = itertools.chain(*params_list) # Setup optimizer instance. By default it is SGD optimizer_instance = None optimizer_class = None if isinstance(optimizer, str): optimizer_class = optimizer elif isinstance(optimizer, torch.optim.Optimizer): optimizer_instance = optimizer else: raise ValueError("optimizer was not understood") optimizer = self.__setup_optimizer( optimizer_instance=optimizer_instance, optimizer_class=optimizer_class, optimization_params=optimization_params, params_to_optimize=params_to_optimize, ) training_loop = [(optimizer, tensors_to_optimize, opt_call_chain)] self.optimizers.append(optimizer) assert ( len(self.optimizers) == 1 ), "There was more than one optimizer, was create_optimizer() called before train()?" elif PtActions._check_tuples(tensors_to_optimize): if batches_per_step != 1: raise ValueError("Gradient accumlation with multiple optimizers is not supported") datasets = [] training_loop = [] for step in tensors_to_optimize: (step_call_chain, dataset,) = self.__get_top_sorted_modules_and_dataloader(hook=step[1]) datasets.append(dataset) training_loop.append((step[0], step[1], step_call_chain)) t_dataset = datasets[0] for dataset in datasets: if type(dataset) is not type(t_dataset): raise ValueError("There were two training datasets, we only support 1.") else: raise ValueError("tensors_to_optimize was not understood") logging_callchain = None # callbacks setup if callbacks is not None: for callback in callbacks: if not isinstance(callback, ActionCallback): raise ValueError("A callback was received that was not a child of ActionCallback") elif isinstance(callback, SimpleLossLoggerCallback): if logging_callchain: raise ValueError("We only support one logger callback but more than one were found") logger_step_freq = callback._step_freq logging_tensors = callback.tensors all_tensors = logging_tensors for step in training_loop: all_tensors = all_tensors + step[1] (logging_callchain, _,) = self.__get_top_sorted_modules_and_dataloader(hook=all_tensors) self._get_all_modules(training_loop, callbacks, logging_callchain) # Intialize Amp if needed if self._optim_level in AmpOptimizations: # Store mapping of self.optimizers to optimizer in callchain training_loop_opts = [] for opt in training_loop: training_loop_opts.append(self.optimizers.index(opt[0])) self.optimizers = self.__initialize_amp( optimizer=self.optimizers, optim_level=self._optim_level, amp_max_loss_scale=amp_max_loss_scale, amp_min_loss_scale=optimization_params.get('amp_min_loss_scale', 1.0), ) # Use stored mapping to map amp_init opts to training loop for i, step in enumerate(training_loop): training_loop[i] = ( self.optimizers[training_loop_opts[i]], step[1], step[2], ) dataNM = training_loop[0][2][0][0] if dataNM.placement == DeviceType.AllGpu: # if len(training_loop) > 1: # raise NotImplementedError( # "Distributed training does nor work with multiple " # "optimizers") logging.info("Doing distributed training") if t_dataset is not None: train_sampler = torch.utils.data.distributed.DistributedSampler( dataset=t_dataset, shuffle=dataNM.shuffle ) train_dataloader = torch.utils.data.DataLoader( dataset=t_dataset, sampler=train_sampler, num_workers=dataNM.num_workers, batch_size=dataNM.batch_size, shuffle=False, ) else: train_dataloader = dataNM.data_iterator if hasattr(train_dataloader, 'sampler'): train_sampler = train_dataloader.sampler else: train_sampler = None for train_iter in training_loop: call_chain = train_iter[2] for i in range(1, len(call_chain) - 1): key = call_chain[i][0].unique_instance_id pmodule = self.module_reference_table[key][1] if not isinstance(pmodule, DDP) and isinstance(pmodule, torch.nn.Module): # gpf = 1 # if gradient_predivide: # gpf = dist.get_world_size() # pmodule = DDP(pmodule, gradient_predivide_factor=gpf) # Old Apex Method # Per pytorch docs, convert sync bn prior to DDP if synced_batchnorm: world_size = dist.get_world_size() sync_batchnorm_group = None if synced_batchnorm_groupsize > 0: if world_size % synced_batchnorm_groupsize != 0: raise ValueError( f"Synchronized batch norm group size ({synced_batchnorm_groupsize}) must be 0" f" or divide total number of GPUs ({world_size})." ) # Find ranks of other nodes in the same batchnorm group rank = torch.distributed.get_rank() group = rank // synced_batchnorm_groupsize group_rank_ids = range( group * synced_batchnorm_groupsize, (group + 1) * synced_batchnorm_groupsize ) sync_batchnorm_group = torch.distributed.new_group(group_rank_ids) pmodule = nn.SyncBatchNorm.convert_sync_batchnorm( pmodule, process_group=sync_batchnorm_group ) # By default, disable broadcast_buffers. This disables batch norm synchronization on forward # pass pmodule = DDP( pmodule, device_ids=[self.local_rank], broadcast_buffers=False, find_unused_parameters=True ) # # Convert batchnorm modules to synced if applicable # if synced_batchnorm and isinstance(pmodule, torch.nn.Module): # world_size = dist.get_world_size() # if synced_batchnorm_groupsize > 0 and world_size % synced_batchnorm_groupsize != 0: # raise ValueError( # f"Synchronized batch norm group size" # f" ({synced_batchnorm_groupsize}) must be 0" # f" or divide total number of GPUs" # f" ({world_size})." # ) # process_group = create_syncbn_process_group(synced_batchnorm_groupsize) # pmodule = convert_syncbn(pmodule, process_group=process_group) self.module_reference_table[key] = ( self.module_reference_table[key][0], pmodule, ) # single GPU/CPU training else: if t_dataset is not None: train_sampler = None train_dataloader = torch.utils.data.DataLoader( dataset=t_dataset, sampler=None, num_workers=dataNM.num_workers, batch_size=dataNM.batch_size, shuffle=dataNM.shuffle, ) else: train_dataloader = dataNM.data_iterator train_sampler = None self._init_callbacks(callbacks) # Do action start callbacks self._perform_on_action_start(callbacks=callbacks) # MAIN TRAINING LOOP # iteration over epochs while num_epochs is None or self.epoch_num < num_epochs: if train_sampler is not None: train_sampler.set_epoch(self.epoch_num) if max_steps is not None and self.step >= max_steps: break # Register epochs start with callbacks self._perform_on_epoch_start(callbacks=callbacks) # iteration over batches in epoch batch_counter = 0 for _, data in enumerate(train_dataloader, 0): if max_steps is not None and self.step >= max_steps: break if batch_counter == 0: # Started step, zero gradients curr_optimizer = training_loop[self.step % len(training_loop)][0] curr_optimizer.zero_grad() # Register iteration start with callbacks self._perform_on_iteration_start(callbacks=callbacks) # set learning rate policy if lr_policy is not None: adjusted_lr = lr_policy(optimization_params["lr"], self.step, self.epoch_num) for param_group in curr_optimizer.param_groups: param_group["lr"] = adjusted_lr if self.tb_writer is not None: value = curr_optimizer.param_groups[0]['lr'] self.tb_writer.add_scalar('param/lr', value, self.step) if callbacks is not None: for callback in callbacks: callback.learning_rate = curr_optimizer.param_groups[0]['lr'] # registered_tensors will contain created tensors # named by output port and uuid of module which created them # Get and properly name tensors returned by data layer curr_call_chain = training_loop[self.step % len(training_loop)][2] dl_device = curr_call_chain[0][0]._device if logging_callchain and self.step % logger_step_freq == 0: curr_call_chain = logging_callchain tensors = [] if isinstance(data, torch.Tensor): data = (data,) for d in data: if isinstance(d, torch.Tensor): tensors.append(d.to(dl_device)) else: tensors.append(d) registered_tensors = { t.unique_name: d for t, d in zip(curr_call_chain[0][2].values(), tensors) if t is not None } disable_allreduce = batch_counter < (batches_per_step - 1) self.__nm_graph_forward_pass( call_chain=curr_call_chain, registered_tensors=registered_tensors, ) curr_tensors_to_optimize = training_loop[self.step % len(training_loop)][1] final_loss = 0 nan = False for tensor in curr_tensors_to_optimize: if ( torch.isnan(registered_tensors[tensor.unique_name]).any() or torch.isinf(registered_tensors[tensor.unique_name]).any() ): if stop_on_nan_loss: raise ValueError('Loss is NaN or inf - exiting') logging.warning('Loss is NaN or inf') curr_optimizer.zero_grad() nan = True break final_loss += registered_tensors[tensor.unique_name] if nan: continue if self._optim_level in AmpOptimizations and self._optim_level != Optimization.mxprO0: with amp.scale_loss(final_loss, curr_optimizer, delay_unscale=disable_allreduce) as scaled_loss: if torch.isnan(scaled_loss).any() or torch.isinf(scaled_loss).any(): if stop_on_nan_loss: raise ValueError('Loss is NaN or inf -' ' exiting') logging.warning('WARNING: Loss is NaN or inf') curr_optimizer.zero_grad() continue if disable_allreduce: with ExitStack() as stack: for mod in self.get_DDP_modules(curr_call_chain): stack.enter_context(mod.no_sync()) scaled_loss.backward(bps_scale.to(scaled_loss.get_device())) else: scaled_loss.backward(bps_scale.to(scaled_loss.get_device())) # no AMP optimizations needed else: # multi-GPU, float32 if self._local_rank is not None: if disable_allreduce: with ExitStack() as stack: for mod in self.get_DDP_modules(curr_call_chain): stack.enter_context(mod.no_sync()) final_loss.backward(bps_scale.to(final_loss.get_device())) else: final_loss.backward(bps_scale.to(final_loss.get_device())) # single device (CPU or GPU) else: # Fix (workaround?) enabling to backpropagate gradiens on CPUs. if final_loss.get_device() < 0: final_loss.backward(bps_scale) else: final_loss.backward(bps_scale.to(final_loss.get_device())) batch_counter += 1 if batch_counter == batches_per_step: # Ended step. Do optimizer update if grad_norm_clip is not None: torch.nn.utils.clip_grad_norm_(master_params(curr_optimizer), grad_norm_clip) curr_optimizer.step() batch_counter = 0 # Register iteration end with callbacks self._update_callbacks( callbacks=callbacks, registered_tensors=registered_tensors, ) self._perform_on_iteration_end(callbacks=callbacks) self.step += 1 # End of epoch for loop # Register epochs end with callbacks self._perform_on_epoch_end(callbacks=callbacks) self.epoch_num += 1 self._perform_on_action_end(callbacks=callbacks) def infer( self, tensors, checkpoint_dir=None, ckpt_pattern='', verbose=True, cache=False, use_cache=False, offload_to_cpu=True, modules_to_restore=None, ): """See NeuralModuleFactory.infer() """ call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors) if checkpoint_dir: # Find all modules that need to be restored if modules_to_restore is None: modules_to_restore = [] modules_to_restore_name = [] for op in call_chain: if op[0].num_weights > 0: modules_to_restore.append(op[0]) if not isinstance(modules_to_restore, list): modules_to_restore = [modules_to_restore] modules_to_restore_name = [] for mod in modules_to_restore: if not isinstance(mod, NeuralModule): raise ValueError("Found something that was not a Neural Module inside modules_to_restore") elif mod.num_weights == 0: raise ValueError("Found a Neural Module with 0 weights inside modules_to_restore") modules_to_restore_name.append(str(mod)) module_checkpoints = get_checkpoint_from_dir(modules_to_restore_name, checkpoint_dir, ckpt_pattern) for mod, checkpoint in zip(modules_to_restore, module_checkpoints): logging.info(f"Restoring {mod} from {checkpoint}") mod.restore_from(checkpoint, self._local_rank) # Init Amp if ( self._optim_level in AmpOptimizations and self._optim_level != Optimization.mxprO0 and not self.amp_initialized ): pt_modules = [] for i in range(len(call_chain)): if isinstance(call_chain[i][0], nn.Module): pt_modules.append(call_chain[i][0]) elif isinstance(call_chain[i][0], TrainableNeuralModuleWrapper): pt_modules.append(call_chain[i][0]._pt_module) amp.initialize( min_loss_scale=1.0, models=pt_modules, optimizers=None, opt_level=AmpOptimizations[self._optim_level], ) self.amp_initialized = True # Run infer return self._infer( tensors_to_return=tensors, verbose=verbose, cache=cache, use_cache=use_cache, offload_to_cpu=offload_to_cpu, ) def get_DDP_modules(self, call_chain): modules = [] for ind in range(1, len(call_chain)): m_id = call_chain[ind][0].unique_instance_id module = self.module_reference_table[m_id][1] if isinstance(module, DDP): modules.append(module) return modules [end of nemo/backends/pytorch/actions.py] [start of nemo/collections/asr/jasper.py] # Copyright (c) 2019 NVIDIA Corporation from typing import Optional import torch import torch.nn as nn import torch.nn.functional as F import nemo from .parts.jasper import JasperBlock, init_weights, jasper_activations from nemo.backends.pytorch.nm import TrainableNM from nemo.core.neural_types import * from nemo.utils.decorators import add_port_docs logging = nemo.logging class JasperEncoder(TrainableNM): """ Jasper Encoder creates the pre-processing (prologue), Jasper convolution block, and the first 3 post-processing (epilogue) layers as described in Jasper (https://arxiv.org/abs/1904.03288) Args: jasper (list): A list of dictionaries. Each element in the list represents the configuration of one Jasper Block. Each element should contain:: { # Required parameters 'filters' (int) # Number of output channels, 'repeat' (int) # Number of sub-blocks, 'kernel' (int) # Size of conv kernel, 'stride' (int) # Conv stride 'dilation' (int) # Conv dilation 'dropout' (float) # Dropout probability 'residual' (bool) # Whether to use residual or not. # Optional parameters 'residual_dense' (bool) # Whether to use Dense Residuals # or not. 'residual' must be True for 'residual_dense' # to be enabled. # Defaults to False. 'separable' (bool) # Whether to use separable convolutions. # Defaults to False 'groups' (int) # Number of groups in each conv layer. # Defaults to 1 'heads' (int) # Sharing of separable filters # Defaults to -1 'tied' (bool) # Whether to use the same weights for all # sub-blocks. # Defaults to False 'se' (bool) # Whether to add Squeeze and Excitation # sub-blocks. # Defaults to False 'se_reduction_ratio' (int) # The reduction ratio of the Squeeze # sub-module. # Must be an integer > 1. # Defaults to 16 'kernel_size_factor' (float) # Conv kernel size multiplier # Can be either an int or float # Kernel size is recomputed as below: # new_kernel_size = int(max(1, (kernel_size * kernel_width))) # to prevent kernel sizes than 1. # Note: If rescaled kernel size is an even integer, # adds 1 to the rescaled kernel size to allow "same" # padding. } activation (str): Activation function used for each sub-blocks. Can be one of ["hardtanh", "relu", "selu"]. feat_in (int): Number of channels being input to this module normalization_mode (str): Normalization to be used in each sub-block. Can be one of ["batch", "layer", "instance", "group"] Defaults to "batch". residual_mode (str): Type of residual connection. Can be "add" or "max". Defaults to "add". norm_groups (int): Number of groups for "group" normalization type. If set to -1, number of channels is used. Defaults to -1. conv_mask (bool): Controls the use of sequence length masking prior to convolutions. Defaults to True. frame_splicing (int): Defaults to 1. init_mode (str): Describes how neural network parameters are initialized. Options are ['xavier_uniform', 'xavier_normal', 'kaiming_uniform','kaiming_normal']. Defaults to "xavier_uniform". """ length: Optional[torch.Tensor] @property @add_port_docs() def input_ports(self): """Returns definitions of module input ports. """ return { # "audio_signal": NeuralType( # {0: AxisType(BatchTag), 1: AxisType(SpectrogramSignalTag), 2: AxisType(ProcessedTimeTag),} # ), # "length": NeuralType({0: AxisType(BatchTag)}), "audio_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), "length": NeuralType(tuple('B'), LengthsType()), } @property @add_port_docs() def output_ports(self): """Returns definitions of module output ports. """ return { # "outputs": NeuralType( # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag),} # ), # "encoded_lengths": NeuralType({0: AxisType(BatchTag)}), "outputs": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), "encoded_lengths": NeuralType(tuple('B'), LengthsType()), } @property def disabled_deployment_input_ports(self): return set(["length"]) @property def disabled_deployment_output_ports(self): return set(["encoded_lengths"]) def prepare_for_deployment(self): m_count = 0 for m in self.modules(): if type(m).__name__ == "MaskedConv1d": m.use_mask = False m_count += 1 logging.warning(f"Turned off {m_count} masked convolutions") def __init__( self, jasper, activation, feat_in, normalization_mode="batch", residual_mode="add", norm_groups=-1, conv_mask=True, frame_splicing=1, init_mode='xavier_uniform', ): super().__init__() activation = jasper_activations[activation]() feat_in = feat_in * frame_splicing residual_panes = [] encoder_layers = [] self.dense_residual = False for lcfg in jasper: dense_res = [] if lcfg.get('residual_dense', False): residual_panes.append(feat_in) dense_res = residual_panes self.dense_residual = True groups = lcfg.get('groups', 1) separable = lcfg.get('separable', False) heads = lcfg.get('heads', -1) se = lcfg.get('se', False) se_reduction_ratio = lcfg.get('se_reduction_ratio', 16) kernel_size_factor = lcfg.get('kernel_size_factor', 1.0) encoder_layers.append( JasperBlock( feat_in, lcfg['filters'], repeat=lcfg['repeat'], kernel_size=lcfg['kernel'], stride=lcfg['stride'], dilation=lcfg['dilation'], dropout=lcfg['dropout'], residual=lcfg['residual'], groups=groups, separable=separable, heads=heads, residual_mode=residual_mode, normalization=normalization_mode, norm_groups=norm_groups, activation=activation, residual_panes=dense_res, conv_mask=conv_mask, se=se, se_reduction_ratio=se_reduction_ratio, kernel_size_factor=kernel_size_factor, ) ) feat_in = lcfg['filters'] self.encoder = nn.Sequential(*encoder_layers) self.apply(lambda x: init_weights(x, mode=init_mode)) self.to(self._device) def forward(self, audio_signal, length=None): # type: (Tensor, Optional[Tensor]) -> Tensor, Optional[Tensor] s_input, length = self.encoder(([audio_signal], length)) if length is None: return s_input[-1] return s_input[-1], length class JasperDecoderForCTC(TrainableNM): """ Jasper Decoder creates the final layer in Jasper that maps from the outputs of Jasper Encoder to the vocabulary of interest. Args: feat_in (int): Number of channels being input to this module num_classes (int): Number of characters in ASR model's vocab/labels. This count should not include the CTC blank symbol. init_mode (str): Describes how neural network parameters are initialized. Options are ['xavier_uniform', 'xavier_normal', 'kaiming_uniform','kaiming_normal']. Defaults to "xavier_uniform". """ @property @add_port_docs() def input_ports(self): """Returns definitions of module input ports. """ return { # "encoder_output": NeuralType( # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag),} # ) "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()) } @property @add_port_docs() def output_ports(self): """Returns definitions of module output ports. """ # return {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag), 2: AxisType(ChannelTag),})} return {"output": NeuralType(('B', 'T', 'D'), LogprobsType())} def __init__(self, feat_in, num_classes, init_mode="xavier_uniform"): super().__init__() self._feat_in = feat_in # Add 1 for blank char self._num_classes = num_classes + 1 self.decoder_layers = nn.Sequential(nn.Conv1d(self._feat_in, self._num_classes, kernel_size=1, bias=True)) self.apply(lambda x: init_weights(x, mode=init_mode)) self.to(self._device) def forward(self, encoder_output): return F.log_softmax(self.decoder_layers(encoder_output).transpose(1, 2), dim=-1) class JasperDecoderForClassification(TrainableNM): """ Jasper Decoder creates the final layer in Jasper that maps from the outputs of Jasper Encoder to one class label. Args: feat_in (int): Number of channels being input to this module num_classes (int): Number of characters in ASR model's vocab/labels. This count should not include the CTC blank symbol. init_mode (str): Describes how neural network parameters are initialized. Options are ['xavier_uniform', 'xavier_normal', 'kaiming_uniform','kaiming_normal']. Defaults to "xavier_uniform". """ @property def input_ports(self): """Returns definitions of module input ports. """ return { # "encoder_output": NeuralType( # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag)} # ) "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()) } @property def output_ports(self): """Returns definitions of module output ports. """ # return {"logits": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} return {"logits": NeuralType(('B', 'D'), LogitsType())} def __init__( self, *, feat_in, num_classes, init_mode="xavier_uniform", return_logits=True, pooling_type='avg', **kwargs ): TrainableNM.__init__(self, **kwargs) self._feat_in = feat_in self._return_logits = return_logits self._num_classes = num_classes if pooling_type == 'avg': self.pooling = nn.AdaptiveAvgPool1d(1) elif pooling_type == 'max': self.pooling = nn.AdaptiveMaxPool1d(1) else: raise ValueError('Pooling type chosen is not valid. Must be either `avg` or `max`') self.decoder_layers = nn.Sequential(nn.Linear(self._feat_in, self._num_classes, bias=True)) self.apply(lambda x: init_weights(x, mode=init_mode)) self.to(self._device) def forward(self, encoder_output): batch, in_channels, timesteps = encoder_output.size() encoder_output = self.pooling(encoder_output).view(batch, in_channels) # [B, C] logits = self.decoder_layers(encoder_output) # [B, num_classes] if self._return_logits: return logits return F.softmax(logits, dim=-1) [end of nemo/collections/asr/jasper.py] [start of nemo/core/neural_factory.py] # ! /usr/bin/python # -*- coding: utf-8 -*- # Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. __all__ = [ 'Backend', 'ModelMode', 'Optimization', 'DeviceType', 'Actions', 'NeuralModuleFactory', 'DeploymentFormat', ] import random from abc import ABC, abstractmethod from enum import Enum from typing import List, Optional import numpy as np import nemo from ..utils import ExpManager from .callbacks import ActionCallback, EvaluatorCallback from .neural_types import * from nemo.utils.decorators import deprecated logging = nemo.logging class DeploymentFormat(Enum): """Which format to use when exporting a Neural Module for deployment""" AUTO = 0 PYTORCH = 1 TORCHSCRIPT = 2 ONNX = 3 TRTONNX = 4 class Backend(Enum): """Supported backends. For now, it is only PyTorch.""" PyTorch = 1 NotSupported = 2 class ModelMode(Enum): """Training Mode or Evaluation/Inference""" train = 0 eval = 1 class Optimization(Enum): """Various levels of Apex/amp Optimization. WARNING: This might have effect on model accuracy.""" mxprO0 = 0 mxprO1 = 1 mxprO2 = 2 mxprO3 = 3 class DeviceType(Enum): """Device types where Neural Modules can be placed.""" GPU = 1 CPU = 2 AllGpu = 3 class Actions(ABC): """Basic actions allowed on graphs of Neural Modules""" def __init__(self, local_rank, global_rank, optimization_level=Optimization.mxprO0): self._local_rank = local_rank self._global_rank = global_rank self._optim_level = optimization_level self.step = None self.epoch_num = None @property def local_rank(self): """Local rank during distributed execution. None if single GPU/CPU Returns: (int) rank or worker or None if not in distributed model """ return self._local_rank @property def global_rank(self): """Global rank during distributed execution. None if single GPU/CPU Returns: (int) rank or worker or None if not in distributed model """ return self._global_rank @abstractmethod def train( self, tensors_to_optimize: List[NmTensor], callbacks: Optional[List[ActionCallback]], lr_policy=None, batches_per_step=None, stop_on_nan_loss=False, ): """This action executes training and (optionally) evaluation. Args: tensors_to_optimize: which tensors to optimize. Typically this is single loss tesnor. callbacks: list of callback objects lr_policy: function which should take (initial_lr, step, epoch) and return learning rate batches_per_step: number of mini-batches to process before one optimizer step. (default: None, same as 1). Use this to simulate larger batch sizes on hardware which could not fit larger batch in memory otherwise. Effectively, this will make "algorithmic" batch size per GPU/worker = batches_per_step* batch_size stop_on_nan_loss: (default: False) If set to True, the training will stop if loss=nan. If set to False, the training will continue, but the gradients will be zeroed before next mini-batch. Returns: None """ pass @abstractmethod def infer(self, tensors: List[NmTensor]): """This action executes inference. Nothing is optimized. Args: tensors: which tensors to evaluate. Returns: None """ pass @abstractmethod def save_state_to(self, path: str): """ Saves current state such as step, epoch and optimizer parameters Args: path: Returns: """ pass @abstractmethod def restore_state_from(self, path: str): """ Restores state such as step, epoch and optimizer parameters Args: path: Returns: """ pass @abstractmethod def create_optimizer(self, optimizer, things_to_optimize, optimizer_params): """ Creates an optimizer object to be use in the train() method. Args: optimizer: Specifies which optimizer to use. things_to_optimize: A list of neural modules or tensors to be optimized. optimizer_params: Specifies the parameters of the optimizer Returns: Optimizer """ pass def _perform_on_iteration_start(self, callbacks): # TODO: Most of these checks can be relaxed since we enforce callbacks # to be a list of ActionCallback objects if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback.on_iteration_start() def _perform_on_iteration_end(self, callbacks): if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback.on_iteration_end() def _perform_on_action_start(self, callbacks): if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback.on_action_start() def _perform_on_action_end(self, callbacks): if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback.on_action_end() def _perform_on_epoch_start(self, callbacks): if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback.on_epoch_start() def _perform_on_epoch_end(self, callbacks): if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback.on_epoch_end() def _init_callbacks(self, callbacks): if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback.action = self def _update_callbacks( self, callbacks=None, registered_tensors=None, ): # if self.local_rank is None or self.local_rank == 0: if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: for callback in callbacks: callback._registered_tensors = registered_tensors def _str_to_opt_level(opt_str: str) -> Optimization: number = int(opt_str[1:]) if number not in Optimization._value2member_map_: raise ValueError(f"Unknown optimization value {opt_str}") return Optimization(number) class NeuralModuleFactory(object): _DEFAULT = None """ Neural Module Factory instance is used to create neural modules and trainers Args: backend (Backend): Currently only Backend.PyTorch is supported local_rank (int): Process rank. Should be set by distributed runner optimization_level (Optimization): Level of optimization to use. Will be passed to neural modules and actions created by this factory. placement (DeviceType: where to place NeuralModule instances by default cudnn_benchmark (bool): (default False) If set to True it will use cudnnFind method to find the best kernels instead of using heuristics. If the shapes of your inputs are constant this should help, for various shapes it can slow things down. Give it few iterations to warmup if set to True. Currently only supported by PyTorch backend. random_seed (int): (default None) Sets random seed to control for randomness. This should be used for debugging purposes as it might have negative impact on performance. Can't be used when `cudnn_benchmark=True`. master_process (bool): (default True) Flag for master process indication set_default (bool): (default True) True if should set this instance as default factory for modules instantiating. """ def __init__( self, backend=Backend.PyTorch, local_rank=None, optimization_level=Optimization.mxprO0, placement=None, cudnn_benchmark=False, random_seed=None, set_default=True, log_dir=None, checkpoint_dir=None, tensorboard_dir=None, create_tb_writer=False, files_to_copy=None, add_time_to_log_dir=False, ): self._local_rank = local_rank self._global_rank = None if isinstance(optimization_level, str): optimization_level = _str_to_opt_level(optimization_level) self._optim_level = optimization_level if placement is None: if local_rank is not None: device = DeviceType.AllGpu else: device = DeviceType.GPU self._placement = device else: self._placement = placement self._backend = backend self._world_size = 1 broadcast_func = None if backend == Backend.PyTorch: # TODO: Move all framework specific code from this file import torch if self._placement != DeviceType.CPU: if not torch.cuda.is_available(): raise ValueError( "You requested to use GPUs but CUDA is " "not installed. You can try running using" " CPU-only. To do this, instantiate your" " factory with placement=DeviceType.CPU" "\n" "Note that this is slow and is not " "well supported." ) torch.backends.cudnn.benchmark = cudnn_benchmark if random_seed is not None and cudnn_benchmark: raise ValueError("cudnn_benchmark can not be set to True when random_seed is not None.") if random_seed is not None: torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False torch.manual_seed(random_seed) np.random.seed(random_seed) random.seed(random_seed) if self._local_rank is not None: torch.distributed.init_process_group(backend="nccl", init_method="env://") cuda_set = True # Try to set cuda device. This should fail if self._local_rank # is greater than the number of available GPUs try: torch.cuda.set_device(self._local_rank) except RuntimeError: # Note in this case, all tensors are now sent to GPU 0 # who could crash because of OOM. Thus init_process_group() # must be done before any cuda tensors are allocated cuda_set = False cuda_set_t = torch.cuda.IntTensor([cuda_set]) # Do an all_reduce to ensure all workers obtained a GPU # For the strangest reason, BAND doesn't work so I am resorting # to MIN. torch.distributed.all_reduce(cuda_set_t, op=torch.distributed.ReduceOp.MIN) if cuda_set_t.item() == 0: raise RuntimeError( "There was an error initializing distributed training." " Perhaps you specified more gpus than you have " "available" ) del cuda_set_t torch.cuda.empty_cache() # Remove test tensor from memory self._world_size = torch.distributed.get_world_size() self._global_rank = torch.distributed.get_rank() def torch_broadcast_wrapper(str_len=None, string=None, src=0): """Wrapper function to broadcast string values across all workers """ # Create byte cuda torch tensor if string is not None: string_tensor = torch.tensor(list(string.encode()), dtype=torch.uint8).cuda() else: string_tensor = torch.tensor([0] * str_len, dtype=torch.uint8).cuda() # Run broadcast torch.distributed.broadcast(string_tensor, src) # turn byte tensor back to string return_string = string_tensor.cpu().numpy() return_string = b''.join(return_string).decode() return return_string broadcast_func = torch_broadcast_wrapper else: raise NotImplementedError("Only Pytorch backend is currently supported.") # Create ExpManager # if log_dir is None, only create logger self._exp_manager = ExpManager( work_dir=log_dir, ckpt_dir=checkpoint_dir, use_tb=create_tb_writer, tb_dir=tensorboard_dir, local_rank=local_rank, global_rank=self._global_rank, files_to_copy=files_to_copy, add_time=add_time_to_log_dir, exist_ok=True, broadcast_func=broadcast_func, ) self._tb_writer = self._exp_manager.tb_writer # Create trainer self._trainer = self._get_trainer(tb_writer=self._tb_writer) if set_default: NeuralModuleFactory.set_default_factory(self) @classmethod def get_default_factory(cls): return cls._DEFAULT @classmethod def set_default_factory(cls, factory): cls._DEFAULT = factory @classmethod def reset_default_factory(cls): cls._DEFAULT = None @staticmethod def __name_import(name): components = name.split(".") mod = __import__(components[0]) for comp in components[1:]: mod = getattr(mod, comp) return mod @deprecated(version=0.11) def __get_pytorch_module(self, name, collection, params, pretrained): # TK: "factory" is not passed as parameter anymore. # params["factory"] = self if collection == "toys" or collection == "tutorials" or collection == "other": constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.tutorials." + name) elif collection == "nemo_nlp": constructor = NeuralModuleFactory.__name_import("nemo_nlp." + name) if name == "BERT" and pretrained is True: params["pretrained"] = True elif collection == "nemo_asr": constructor = NeuralModuleFactory.__name_import("nemo_asr." + name) elif collection == "nemo_lpr": constructor = NeuralModuleFactory.__name_import("nemo_lpr." + name) elif collection == 'common': constructor = NeuralModuleFactory.__name_import('nemo.backends.pytorch.common.' + name) elif collection == "torchvision": import torchvision.models as tv_models import nemo.backends.pytorch.module_wrapper as mw import torch.nn as nn if name == "ImageFolderDataLayer": constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.torchvision.data." + name) instance = constructor(**params) return instance else: _nm_name = name.lower() if _nm_name == "resnet18": input_ports = { "x": NeuralType( { 0: AxisType(BatchTag), 1: AxisType(ChannelTag), 2: AxisType(HeightTag, 224), 3: AxisType(WidthTag, 224), } ) } output_ports = {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} pt_model = tv_models.resnet18(pretrained=pretrained) num_classes = params.get("num_classes", None) if num_classes is not None: pt_model.fc = nn.Linear(512, params["num_classes"]) return mw.TrainableNeuralModuleWrapper( pt_nn_module=pt_model, input_ports_dict=input_ports, output_ports_dict=output_ports, ) elif _nm_name == "resnet50": input_ports = { "x": NeuralType( { 0: AxisType(BatchTag), 1: AxisType(ChannelTag), 2: AxisType(HeightTag, 224), 3: AxisType(WidthTag, 224), } ) } output_ports = {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} pt_model = tv_models.resnet50(pretrained=pretrained) num_classes = params.get("num_classes", None) if num_classes is not None: pt_model.fc = nn.Linear(2048, params["num_classes"]) return mw.TrainableNeuralModuleWrapper( pt_nn_module=pt_model, input_ports_dict=input_ports, output_ports_dict=output_ports, ) else: collection_path = "nemo.collections." + collection + "." + name constructor = NeuralModuleFactory.__name_import(collection_path) if name == "BERT" and pretrained is True: params["pretrained"] = True # TK: "placement" is not passed as parameter anymore. # if "placement" not in params: # params["placement"] = self._placement instance = constructor(**params) return instance @deprecated(version=0.11) def get_module(self, name, collection, params, pretrained=False): """ Creates NeuralModule instance Args: name (str): name of NeuralModule which instance should be returned. params (dict): local parameters which should be passed to NeuralModule's constructor. collection (str): in which collection to look for `neural_module_name` pretrained (bool): return pre-trained instance or randomly initialized (default) Returns: NeuralModule instance """ # TK: "optimization_level" is not passed as parameter anymore. # if params is not None and "optimization_level" in params: # if params["optimization_level"] != self._optim_level: # logging.warning( # "Module's {0} requested optimization level {1} is" # "different from the one specified by factory - {2}." # "Using: {3} for this module".format( # name, params["optimization_level"], self._optim_level, params["optimization_level"], # ) # ) # else: # if params is None: # params = {} # params["optimization_level"] = self._optim_level if self._backend == Backend.PyTorch: return self.__get_pytorch_module(name=name, collection=collection, params=params, pretrained=pretrained,) else: return None def create_optimizer(self, optimizer, things_to_optimize, optimizer_params): return self._trainer.create_optimizer( optimizer=optimizer, things_to_optimize=things_to_optimize, optimizer_params=optimizer_params, ) def train( self, tensors_to_optimize, optimizer=None, optimization_params=None, callbacks: Optional[List[ActionCallback]] = None, lr_policy=None, batches_per_step=None, stop_on_nan_loss=False, synced_batchnorm=False, synced_batchnorm_groupsize=0, gradient_predivide=False, amp_max_loss_scale=2.0 ** 24, reset=False, ): if reset: self.reset_trainer() return self._trainer.train( tensors_to_optimize=tensors_to_optimize, optimizer=optimizer, optimization_params=optimization_params, callbacks=callbacks, lr_policy=lr_policy, batches_per_step=batches_per_step, stop_on_nan_loss=stop_on_nan_loss, synced_batchnorm=synced_batchnorm, synced_batchnorm_groupsize=synced_batchnorm_groupsize, gradient_predivide=gradient_predivide, amp_max_loss_scale=amp_max_loss_scale, ) def eval(self, callbacks: List[EvaluatorCallback]): if callbacks is None or len(callbacks) == 0: raise ValueError(f"You need to provide at lease one evaluation" f"callback to eval") for callback in callbacks: if not isinstance(callback, EvaluatorCallback): raise TypeError(f"All callbacks passed to the eval action must" f"be inherited from EvaluatorCallback") self.train( tensors_to_optimize=None, optimizer='sgd', callbacks=callbacks, optimization_params={'num_epochs': 1}, ) def deployment_export( self, module, output: str, d_format: DeploymentFormat, input_example=None, output_example=None ): """Exports Neural Module instance for deployment. Args: module: neural module to export output (str): where export results should be saved d_format (DeploymentFormat): which deployment format to use input_example: sometimes tracing will require input examples output_example: Should match inference on input_example """ module.prepare_for_deployment() return self._trainer.deployment_export( module=module, output=output, d_format=d_format, input_example=input_example, output_example=output_example, ) def infer( self, tensors: List[NmTensor], checkpoint_dir=None, ckpt_pattern='', verbose=True, cache=False, use_cache=False, offload_to_cpu=True, modules_to_restore=None, ): """Runs inference to obtain values for tensors Args: tensors (list[NmTensor]): List of NeMo tensors that we want to get values of. checkpoint_dir (str): Path to checkpoint directory. Default is None which does not load checkpoints. ckpt_pattern (str): Pattern used to check for checkpoints inside checkpoint_dir. Default is '' which matches any checkpoints inside checkpoint_dir. verbose (bool): Controls printing. Defaults to True. cache (bool): If True, cache all `tensors` and intermediate tensors so that future calls that have use_cache set will avoid computation. Defaults to False. use_cache (bool): Values from `tensors` will be always re-computed. It will re-use intermediate tensors from the DAG leading to `tensors`. If you want something to be re-computed, put it into `tensors` list. Defaults to False. offload_to_cpu (bool): If True, all evaluated tensors are moved to cpu memory after each inference batch. Defaults to True. modules_to_restore (list): Defaults to None, in which case all NMs inside callchain with weights will be restored. If specified only the modules inside this list will be restored. Returns: List of evaluated tensors. Each element in the list is also a list where each element is now a batch of tensor values. """ return self._trainer.infer( tensors=tensors, checkpoint_dir=checkpoint_dir, ckpt_pattern=ckpt_pattern, verbose=verbose, cache=cache, use_cache=use_cache, offload_to_cpu=offload_to_cpu, modules_to_restore=modules_to_restore, ) def clear_cache(self): """Helper function to clean inference cache.""" self._trainer.clear_cache() @deprecated(version="future") def _get_trainer(self, tb_writer=None): if self._backend == Backend.PyTorch: constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.PtActions") instance = constructor( local_rank=self._local_rank, global_rank=self._global_rank, tb_writer=tb_writer, optimization_level=self._optim_level, ) return instance else: raise ValueError("Only PyTorch backend is currently supported.") @deprecated( version="future", explanation="Please use .train(...), .eval(...), .infer(...) and " f".create_optimizer(...) of the NeuralModuleFactory instance directly.", ) def get_trainer(self, tb_writer=None): if self._trainer: logging.warning( "The trainer instance was created during initialization of " "Neural factory, using the already created instance." ) return self._trainer return self._get_trainer(tb_writer) def reset_trainer(self): del self._trainer self._trainer = self._get_trainer(tb_writer=self._tb_writer) def sync_all_processes(self, status=True): """ Helper function for testing that allows proccess 0 to inform all other processes of failures. Does nothing if not using distributed training. Usage example can be seen in examples/asr/jasper_an4.py Args: status (bool): Defaults to True. If any proccess passes False, it will trigger a graceful exit on all other processes. It is assumed that the process that passed False will print an error message on its own and exit """ if self._world_size == 1: logging.info("sync_all_processes does nothing if there is one process") return if self._backend == Backend.PyTorch: import torch status_tensor = torch.cuda.IntTensor([status]) torch.distributed.all_reduce(status_tensor, op=torch.distributed.ReduceOp.MIN) if status_tensor.item() == 0: logging.error("At least one process had a failure") if status: raise ValueError( f"Process with global rank {self._global_rank} entered" " sync_all_processes with a passing status, but " "another process indicated a failure" ) @property def world_size(self): return self._world_size @property def tb_writer(self): return self._tb_writer @property def placement(self): return self._placement @property def optim_level(self): return self._optim_level @property @deprecated(version=0.11, explanation="Please use ``nemo.logging instead``") def logger(self): return nemo.logging @property def checkpoint_dir(self): return self._exp_manager.ckpt_dir @property def work_dir(self): return self._exp_manager.work_dir @property def global_rank(self): return self._global_rank [end of nemo/core/neural_factory.py] [start of nemo/core/neural_modules.py] # ! /usr/bin/python # -*- coding: utf-8 -*- # Copyright (c) 2019-, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """This file contains NeuralModule and NmTensor classes.""" __all__ = ['WeightShareTransform', 'NeuralModule'] import collections import uuid from abc import ABC, abstractmethod from collections import namedtuple from enum import Enum from inspect import getargvalues, getfullargspec, stack from os import path from typing import Dict, List, Optional, Set, Tuple from ruamel.yaml import YAML from .neural_types import ( CanNotInferResultNeuralType, NeuralPortNameMismatchError, NeuralPortNmTensorMismatchError, NeuralType, NeuralTypeComparisonResult, NmTensor, ) from nemo import logging from nemo.core import NeuralModuleFactory from nemo.package_info import __version__ as nemo_version from nemo.utils.decorators.deprecated import deprecated YAML = YAML(typ='safe') class WeightShareTransform(Enum): """When sharing parameters, what kind of transform to apply.""" SAME = 0 TRANSPOSE = 1 PretrainedModelInfo = namedtuple( "PretrainedModleInfo", ("pretrained_model_name", "description", "parameters", "location"), ) class NeuralModule(ABC): """Abstract class that every Neural Module must inherit from. """ def __init__(self): # Get default factory. self._factory = NeuralModuleFactory.get_default_factory() # Set module properties from factory else use defaults self._placement = self._factory.placement # If one needs to change that should override it manually. # Optimization level. self._opt_level = self._factory.optim_level # Get object UUID. self._uuid = str(uuid.uuid4()) # Retrieve dictionary of parameters (keys, values) passed to init. self._init_params = self.__extract_init_params() # Pint the types of the values. # for key, value in self._init_params.items(): # print("{}: {} ({})".format(key, value, type(value))) # Validate the parameters. # self._validate_params(self._init_params) @property def init_params(self) -> Optional[Dict]: """ Property returning parameters used to instantiate the module. Returns: Dictionary containing parameters used to instantiate the module. """ return self._init_params def __extract_init_params(self): """ Retrieves the dictionary of of parameters (keys, values) passed to constructor of a class derived (also indirectly) from the Neural Module class. Returns: Dictionary containing parameters passed to init(). """ # Get names of arguments of the original module init method. init_keys = getfullargspec(type(self).__init__).args # Remove self. if "self" in init_keys: init_keys.remove("self") # Create list of params. init_params = {}.fromkeys(init_keys) # Retrieve values of those params from the call list. for frame in stack()[1:]: localvars = getargvalues(frame[0]).locals # print("localvars: ", localvars) for key in init_keys: # Found the variable! if key in localvars.keys(): # Save the value. init_params[key] = localvars[key] # Return parameters. return init_params def __validate_params(self, params): """ Checks whether dictionary contains parameters being primitive types (string, int, float etc.) or (lists of)+ primitive types. Args: params: dictionary of parameters. Returns: True if all parameters were ok, False otherwise. """ ok = True # Iterate over parameters and check them one by one. for key, variable in params.items(): if not self.__is_of_allowed_type(variable): logging.warning( "Parameter '{}' contains a variable '{}' of type '{}' which is not allowed.".format( key, variable, type(variable) ) ) ok = False # Return the result. return ok def __is_of_allowed_type(self, var): """ A recursive function that checks if a given variable is of allowed type. Args: pretrained_model_name (str): name of pretrained model to use in order. Returns: True if all parameters were ok, False otherwise. """ # Special case: None is also allowed. if var is None: return True var_type = type(var) # If this is list - check its elements. if var_type == list: for list_var in var: if not self.__is_of_allowed_type(list_var): return False # If this is dict - check its elements. elif var_type == dict: for _, dict_var in var.items(): if not self.__is_of_allowed_type(dict_var): return False elif var_type not in (str, int, float, bool): return False # Well, seems that everything is ok. return True def _create_config_header(self): """ A protected method that create a header stored later in the configuration file. """ # Get module "full specification". module_full_spec = str(self.__module__) + "." + str(self.__class__.__qualname__) module_class_name = type(self).__name__ # print(module_full_spec) # Check whether module belongs to a collection. spec_list = module_full_spec.split(".") # Do not check Neural Modules from unit tests. if spec_list[0] == "tests": # Set collection variables. collection_type = "tests" collection_version = None else: # Check if component belongs to any collection if len(spec_list) < 3 or (spec_list[0] != "nemo" and spec_list[1] != "collection"): logging.warning( "Module `{}` does not belong to any collection. This won't be allowed in the next release.".format( module_class_name ) ) collection_type = "unknown" collection_version = None else: # Ok, set collection. collection_type = spec_list[2] collection_version = None # TODO: to be SET! # print(getattr("nemo.collections.nlp", __version__)) # Create a "header" with module "specification". header = { "nemo_core_version": nemo_version, "collection_type": collection_type, "collection_version": collection_version, # "class": module_class_name, # Operating only on full_spec now. "full_spec": module_full_spec, } return header def export_to_config(self, config_file): """ A function that exports module "configuration" (i.e. init parameters) to a YAML file. Raises a ValueError exception in case then parameters coudn't be exported. Args: config_file: path (absolute or relative) and name of the config file (YML) """ # Check if generic export will work. if not self.__validate_params(self._init_params): raise ValueError( "Generic configuration export enables to use of parameters of primitive types (string, int, float) " F"or (lists of/dicts of) primitive types. Please implement your own custom `export_to_config()` and " F"`import_from_config()` methods for your custom Module class." ) # Greate an absolute path. abs_path_file = path.expanduser(config_file) # Create the dictionary to be exported. to_export = {} # Add "header" with module "specification". to_export["header"] = self._create_config_header() # Add init parameters. to_export["init_params"] = self._init_params # print(to_export) # All parameters are ok, let's export. with open(abs_path_file, 'w') as outfile: YAML.dump(to_export, outfile) logging.info( "Configuration of module {} ({}) exported to {}".format(self._uuid, type(self).__name__, abs_path_file) ) @classmethod def _validate_config_file(cls, config_file, section_name=None): """ Class method validating whether the config file has a proper content (sections, specification etc.). Raises an ImportError exception when config file is invalid or incompatible (when called from a particular class). Args: config_file: path (absolute or relative) and name of the config file (YML) section_name: section in the configuration file storing module configuration (optional, DEFAULT: None) Returns: A loaded configuration file (dictionary). """ # Greate an absolute path. abs_path_file = path.expanduser(config_file) # Open the config file. with open(abs_path_file, 'r') as stream: loaded_config = YAML.load(stream) # Check section. if section_name is not None: if section_name not in loaded_config: raise ImportError( "The loaded config `{}` doesn't contain the indicated `{}` section".format( config_file, section_name ) ) # Section exists - use only it for configuration. loaded_config = loaded_config[section_name] # Make sure that the config is valid. if "header" not in loaded_config: raise ImportError("The loaded config `{}` doesn't contain the `header` section".format(config_file)) if "init_params" not in loaded_config: raise ImportError("The loaded config `{}` doesn't contain the `init_params` section".format(config_file)) # Parse the "full specification". spec_list = loaded_config["header"]["full_spec"].split(".") # Check if config contains data of a compatible class. if cls.__name__ != "NeuralModule" and spec_list[-1] != cls.__name__: txt = "The loaded file `{}` contains configuration of ".format(config_file) txt = txt + "`{}` thus cannot be used for instantiation of an object of type `{}`".format( spec_list[-1], cls.__name__ ) raise ImportError(txt) # Success - return configuration. return loaded_config @classmethod def import_from_config(cls, config_file, section_name=None, overwrite_params={}): """ Class method importing the configuration file. Raises an ImportError exception when config file is invalid or incompatible (when called from a particular class). Args: config_file: path (absolute or relative) and name of the config file (YML) section_name: section in the configuration file storing module configuration (optional, DEFAULT: None) overwrite_params: Dictionary containing parameters that will be added to or overwrite (!) the default parameters loaded from the configuration file Returns: Instance of the created NeuralModule object. """ # Validate the content of the configuration file (its header). loaded_config = cls._validate_config_file(config_file, section_name) # Parse the "full specification". spec_list = loaded_config["header"]["full_spec"].split(".") # Get object class from "full specification". mod_obj = __import__(spec_list[0]) for spec in spec_list[1:]: mod_obj = getattr(mod_obj, spec) # print(mod_obj) # Get init parameters. init_params = loaded_config["init_params"] # Update parameters with additional ones. init_params.update(overwrite_params) # Create and return the object. obj = mod_obj(**init_params) logging.info( "Instantiated a new Neural Module of type `{}` using configuration loaded from the `{}` file".format( spec_list[-1], config_file ) ) return obj @deprecated(version=0.11) @staticmethod def create_ports(**kwargs): """ Deprecated method, to be remoted in the next release.""" raise Exception( 'Deprecated method. Please implement ``inputs`` and ``outputs`` \ properties to define module ports instead' ) @property @abstractmethod def input_ports(self) -> Optional[Dict[str, NeuralType]]: """Returns definitions of module input ports Returns: A (dict) of module's input ports names to NeuralTypes mapping """ @property @abstractmethod def output_ports(self) -> Optional[Dict[str, NeuralType]]: """Returns definitions of module output ports Returns: A (dict) of module's output ports names to NeuralTypes mapping """ @property def disabled_deployment_input_ports(self) -> Optional[Set[str]]: """Returns names of input ports that will not be included in an export Returns: A (set) of module's input port names that are not exportable """ return set([]) @property def disabled_deployment_output_ports(self) -> Optional[Set[str]]: """Returns names of output ports that will not be included in an export Returns: A (set) of module's output port names that are not exportable """ return set([]) def prepare_for_deployment(self) -> None: """Patch the module if required to prepare for deployment """ return @staticmethod def pretrained_storage(): return '' def __call__(self, **kwargs): """This method allows objects to be called with their port names Args: kwargs: Input ports and their values. For example: ... mymodule1 = Subclass1_of_NeuralModule(...) mymodule2 = Subclass2_of_NeuralModule(...) ... out_port1, out_port2 = mymodule1(input_port1=value1, input_port2=value2, input_port3=value3) out_port11 = mymodule2(input_port1=out_port2) ... Returns: NmTensor object or tuple of NmTensor objects """ # Get input and output ports definitions. input_port_defs = self.input_ports output_port_defs = self.output_ports first_input_nmtensor_type = None input_nmtensors_are_of_same_type = True for port_name, tgv in kwargs.items(): # make sure that passed arguments correspond to input port names if port_name not in input_port_defs.keys(): raise NeuralPortNameMismatchError("Wrong input port name: {0}".format(port_name)) input_port = input_port_defs[port_name] type_comatibility = input_port.compare(tgv) if ( type_comatibility != NeuralTypeComparisonResult.SAME and type_comatibility != NeuralTypeComparisonResult.GREATER ): raise NeuralPortNmTensorMismatchError( "\n\nIn {0}. \n" "Port: {1} and a NmTensor it was fed are \n" "of incompatible neural types:\n\n{2} \n\n and \n\n{3}" "\n\nType comparison result: {4}".format( self.__class__.__name__, port_name, input_port_defs[port_name], tgv, type_comatibility, ) ) # if first_input_nmtensor_type is None: # first_input_nmtensor_type = NeuralType(tgv._axis2type) # else: # if first_input_nmtensor_type._axis2type is None: # input_nmtensors_are_of_same_type = True # else: # input_nmtensors_are_of_same_type = first_input_nmtensor_type.compare( # tgv # ) == NeuralTypeComparisonResult.SAME and len(first_input_nmtensor_type._axis2type) # if not ( # type_comatibility == NeuralTypeComparisonResult.SAME # or type_comatibility == NeuralTypeComparisonResult.GREATER # ): # raise NeuralPortNmTensorMismatchError( # "\n\nIn {0}. \n" # "Port: {1} and a NmTensor it was fed are \n" # "of incompatible neural types:\n\n{2} \n\n and \n\n{3}" # "\n\nType comparison result: {4}".format( # self.__class__.__name__, port_name, input_port_defs[port_name], tgv, type_comatibility, # ) # ) # if type_comatibility == NeuralTypeComparisonResult.LESS: # print('Types were raised') if len(output_port_defs) == 1: out_name = list(output_port_defs)[0] out_type = output_port_defs[out_name] if out_type is None: if input_nmtensors_are_of_same_type: out_type = first_input_nmtensor_type else: raise CanNotInferResultNeuralType( "Can't infer output neural type. Likely your inputs are of different type." ) return NmTensor(producer=self, producer_args=kwargs, name=out_name, ntype=out_type,) else: result = [] for out_port, n_type in output_port_defs.items(): out_type = n_type if out_type is None: if input_nmtensors_are_of_same_type: out_type = first_input_nmtensor_type else: raise CanNotInferResultNeuralType( "Can't infer output neural type. Likely your inputs are of different type." ) result.append(NmTensor(producer=self, producer_args=kwargs, name=out_port, ntype=out_type,)) # Creating ad-hoc class for returning from module's forward pass. output_class_name = f'{self.__class__.__name__}Output' field_names = list(output_port_defs) result_type = collections.namedtuple(typename=output_class_name, field_names=field_names,) # Tie tuple of output tensors with corresponding names. result = result_type(*result) return result def __str__(self): return self.__class__.__name__ @abstractmethod def get_weights(self) -> Optional[Dict[(str, bool)]]: """Returns NeuralModule's weights copy. Returns: Dictionary of name -> (weights, trainable)""" pass @abstractmethod def set_weights( self, name2weight: Dict[(str, Tuple[str, bool])], name2name_and_transform: Dict[(str, Tuple[str, WeightShareTransform])] = None, ): """Sets weight from given values. For every named weight in name2weight, if weight with the same name is found in the model, it will be set to found value. WARNING: This will NOT tie weights. It will copy values. If ``name2name_and_transform`` is provided then if will set weights using name mapping and transform. For example, suppose ``objec1.X = 3x5 weight``. Then, if ``name2name_and_transform['X']=('Y', WeightShareTransform.TRANSPOSE)`` and ``Y`` is 5x3 weight and ``name2weight['Y']=Y. Then: ``object1.set_weights(name2weight, name2name_and_transform)`` will set object1.X=transpose(Y). Args: name2weight (dict): dictionary of name to (weight, trainable). Typically this is output of get_weights method. name2name_and_transform: mapping from name -> (name, transform) """ pass @staticmethod def list_pretrained_models() -> Optional[List[PretrainedModelInfo]]: """List all available pre-trained models (e.g. weights) for this NM. Returns: A list of PretrainedModelInfo tuples. The pretrained_model_name field of the tuple can be used to retrieve pre-trained model's weights (pass it as pretrained_model_name argument to the module's constructor) """ return None def get_config_dict_and_checkpoint(self, pretrained_model_name): """WARNING: This part is work in progress""" return None @abstractmethod def tie_weights_with( self, module, weight_names=List[str], name2name_and_transform: Dict[(str, Tuple[str, WeightShareTransform])] = None, ): """Ties weights between self and module. For every weight name in weight_names, if weight with the same name is found in self, it will be tied with a same weight from ``module``. WARNING: Once weights are tied, updates to one weights's weights will affect other module's weights. If ``name2name_and_transform`` is provided then if will set weights using name mapping and transform. For example, suppose ``objec1.X = 3x5 weights`` and ``object2.Y = 5x3 weights``. Then these weights can be tied like this: .. code-block:: python object1.tie_weights_with(object2, weight_names=['X'], name2name_and_transform = { 'X': ('Y', WeightShareTransform.TRANSPOSE)}) Args: module: with which module to tie weights weight_names (List[str]): list of self weights' names name2name_and_transform: mapping from name -> (name, transform) """ pass def is_trainable(self) -> bool: """ Checks if NeuralModule is trainable. A NeuralModule is trainable IFF it contains at least one trainable weight Returns: True if module has trainable weights, False otherwise """ weights = self.get_weights() if weights is None: return False for name, w in weights.items(): if w[1]: return True return False @abstractmethod def save_to(self, path: str): """Save module state to file. Args: path (string): path to while where to save. """ pass @abstractmethod def restore_from(self, path: str): """Restore module's state from file. Args: path (string): path to where to restore from. """ pass @abstractmethod def freeze(self, weights: Set[str] = None): """Freeze weights Args: weights (set): set of weight names to freeze If None, all weights are freezed. """ pass @abstractmethod def unfreeze(self, weights: Set[str] = None): """Unfreeze weights Args: weights (set): set of weight names to unfreeze If None, all weights are unfreezed. """ pass @property def placement(self): """Module's placement. Currently CPU or GPU. DataParallel and ModelParallel will come later. Returns: (DeviceType) Device where NM's weights are located """ return self._placement @property @deprecated(version=0.11) def local_parameters(self) -> Optional[Dict]: """Get module's parameters Returns: module's parameters """ return self._init_params # return self._local_parameters @property def unique_instance_id(self): """A unique instance id for this object Returns: A uniq uuid which can be used to identify this object """ return self._uuid @property def factory(self): """ Neural module factory which created this module Returns: NeuralModuleFactory instance or None """ return self._factory @property @abstractmethod def num_weights(self): """Number of module's weights """ pass [end of nemo/core/neural_modules.py] </code> I need you to solve this issue by regenerating the full files in the code base that you would like to change. You can change as many files as you like. Please respond with a list of files and their revised contents in the following format. <example> [start of /src/this_file.py] import os def euclidean(a, b): if b == 0: return a return euclidean(b, a % b) [end of /src/this_file.py] [start of /src/another_file.py] def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) x, y = x0, y0 sx = -1 if x0 > x1 else 1 sy = -1 if y0 > y1 else 1 if dx > dy: err = dx / 2.0 while x != x1: points.append((x, y)) err -= dy if err < 0: y += sy err += dx x += sx else: err = dy / 2.0 while y != y1: points.append((x err -= dx if err < 0: x += sx err += dy y += sy points.append((x, y)) return points [end of /src/another_file.py] </example>
NVIDIA/NeMo
ba4616f1f011d599de87f0cb3315605e715d402a
Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ```
2020-03-10T03:03:23Z
<patch> <patch> diff --git a/nemo/backends/pytorch/actions.py b/nemo/backends/pytorch/actions.py --- a/nemo/backends/pytorch/actions.py +++ b/nemo/backends/pytorch/actions.py @@ -937,26 +937,16 @@ def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defa if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: dynamic_axes[port_name].append(ind) - # This is a hack for Jasper to Jarvis export -- need re-design for this - inputs_to_drop = set() - outputs_to_drop = set() - if type(module).__name__ == "JasperEncoder": - logging.info( - "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " - "deployment" - ) - inputs_to_drop.add("length") - outputs_to_drop.add("encoded_lengths") - + # extract dynamic axes and remove unnecessary inputs/outputs # for input_ports for port_name, ntype in module.input_ports.items(): - if port_name in inputs_to_drop: + if port_name in module._disabled_deployment_input_ports: input_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) # for output_ports for port_name, ntype in module.output_ports.items(): - if port_name in outputs_to_drop: + if port_name in module._disabled_deployment_output_ports: output_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) diff --git a/nemo/collections/asr/jasper.py b/nemo/collections/asr/jasper.py --- a/nemo/collections/asr/jasper.py +++ b/nemo/collections/asr/jasper.py @@ -118,14 +118,14 @@ def output_ports(self): } @property - def disabled_deployment_input_ports(self): + def _disabled_deployment_input_ports(self): return set(["length"]) @property - def disabled_deployment_output_ports(self): + def _disabled_deployment_output_ports(self): return set(["encoded_lengths"]) - def prepare_for_deployment(self): + def _prepare_for_deployment(self): m_count = 0 for m in self.modules(): if type(m).__name__ == "MaskedConv1d": diff --git a/nemo/core/neural_factory.py b/nemo/core/neural_factory.py --- a/nemo/core/neural_factory.py +++ b/nemo/core/neural_factory.py @@ -610,7 +610,7 @@ def deployment_export( input_example: sometimes tracing will require input examples output_example: Should match inference on input_example """ - module.prepare_for_deployment() + module._prepare_for_deployment() return self._trainer.deployment_export( module=module, diff --git a/nemo/core/neural_modules.py b/nemo/core/neural_modules.py --- a/nemo/core/neural_modules.py +++ b/nemo/core/neural_modules.py @@ -393,7 +393,7 @@ def output_ports(self) -> Optional[Dict[str, NeuralType]]: """ @property - def disabled_deployment_input_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_input_ports(self) -> Optional[Set[str]]: """Returns names of input ports that will not be included in an export Returns: @@ -402,7 +402,7 @@ def disabled_deployment_input_ports(self) -> Optional[Set[str]]: return set([]) @property - def disabled_deployment_output_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_output_ports(self) -> Optional[Set[str]]: """Returns names of output ports that will not be included in an export Returns: @@ -410,7 +410,7 @@ def disabled_deployment_output_ports(self) -> Optional[Set[str]]: """ return set([]) - def prepare_for_deployment(self) -> None: + def _prepare_for_deployment(self) -> None: """Patch the module if required to prepare for deployment """ </patch> </s> </patch>
diff --git a/tests/unit/core/test_deploy_export.py b/tests/unit/core/test_deploy_export.py --- a/tests/unit/core/test_deploy_export.py +++ b/tests/unit/core/test_deploy_export.py @@ -46,9 +46,11 @@ import nemo.collections.nlp.nm.trainables.common.token_classification_nm from nemo import logging +TRT_ONNX_DISABLED = False + # Check if the required libraries and runtimes are installed. +# Only initialize GPU after this runner is activated. try: - # Only initialize GPU after this runner is activated. import pycuda.autoinit # This import causes pycuda to automatically manage CUDA context creation and cleanup. @@ -63,16 +65,17 @@ ) from .tensorrt_runner import TensorRTRunnerV2 except: - # Skip tests. - pytestmark = pytest.mark.skip + TRT_ONNX_DISABLED = True @pytest.mark.usefixtures("neural_factory") class TestDeployExport(TestCase): - def setUp(self): - logging.setLevel(logging.WARNING) - device = nemo.core.DeviceType.GPU - self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) + # def setUp(self): + # super().setUp() + + # logging.setLevel(logging.WARNING) + # device = nemo.core.DeviceType.GPU + # self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) def __test_export_route(self, module, out_name, mode, input_example=None): out = Path(out_name) @@ -112,7 +115,13 @@ def __test_export_route(self, module, out_name, mode, input_example=None): loader_cache = DataLoaderCache(data_loader) profile_shapes = OrderedDict() names = list(module.input_ports) + list(module.output_ports) - + names = list( + filter( + lambda x: x + not in (module._disabled_deployment_input_ports | module._disabled_deployment_output_ports), + names, + ) + ) if isinstance(input_example, tuple): si = [tuple(input_example[i].shape) for i in range(len(input_example))] elif isinstance(input_example, OrderedDict): @@ -152,7 +161,7 @@ def __test_export_route(self, module, out_name, mode, input_example=None): input_names = list(input_metadata.keys()) for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: + if input_name in module._disabled_deployment_input_ports: continue inputs[input_name] = ( input_example[input_name].cpu().numpy() @@ -209,8 +218,8 @@ def __test_export_route(self, module, out_name, mode, input_example=None): ort_inputs = ort_session.get_inputs() for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: - input_name = ort_inputs[i].name + if input_name in module._disabled_deployment_input_ports: + continue inputs[input_name] = ( input_example[input_name].cpu().numpy() if isinstance(input_example, OrderedDict) @@ -263,9 +272,10 @@ def __test_export_route(self, module, out_name, mode, input_example=None): def __test_export_route_all(self, module, out_name, input_example=None): if input_example is not None: - self.__test_export_route( - module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example - ) + if not TRT_ONNX_DISABLED: + self.__test_export_route( + module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example + ) self.__test_export_route(module, out_name + '.onnx', nemo.core.DeploymentFormat.ONNX, input_example) self.__test_export_route(module, out_name + '.pt', nemo.core.DeploymentFormat.PYTORCH, input_example) self.__test_export_route(module, out_name + '.ts', nemo.core.DeploymentFormat.TORCHSCRIPT, input_example) @@ -323,9 +333,7 @@ def test_jasper_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="jasper_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randn(256).cuda()), + module=jasper_encoder, out_name="jasper_encoder", input_example=torch.randn(16, 64, 256).cuda(), ) @pytest.mark.unit @@ -343,7 +351,5 @@ def test_quartz_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="quartz_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randint(20, (16,)).cuda()), + module=jasper_encoder, out_name="quartz_encoder", input_example=torch.randn(16, 64, 256).cuda(), )
1.0
NVIDIA__NeMo-3632
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> ./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`. </issue> <code> [start of README.rst] |status| |documentation| |license| |lgtm_grade| |lgtm_alerts| |black| .. |status| image:: http://www.repostatus.org/badges/latest/active.svg :target: http://www.repostatus.org/#active :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main :alt: Documentation :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE :alt: NeMo core license and license for collections in this repo .. |lgtm_grade| image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python :alt: Language grade: Python .. |lgtm_alerts| image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ :alt: Total alerts .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg :target: https://github.com/psf/black :alt: Code style: black .. _main-readme: **NVIDIA NeMo** =============== Introduction ------------ NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), natural language processing (NLP), and text-to-speech synthesis (TTS). The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. `Pre-trained NeMo models. <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_ `Introductory video. <https://www.youtube.com/embed/wBgpMf_KQVw>`_ Key Features ------------ * Speech processing * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ * Supported models: Jasper, QuartzNet, CitriNet, Conformer-CTC, Conformer-Transducer, ContextNet, ... * Supports CTC and Transducer/RNNT losses/decoders * Beam Search decoding * `Language Modelling for ASR <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer * Streaming and Buffered ASR (CTC/Transdcer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/main/examples/asr/asr_chunked_inference>`_ * `Speech Classification and Speech Command Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition) * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ * Natural Language Processing * `Compatible with Hugging Face Transformers and NVIDIA Megatron <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html>`_ * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation.html>`_ * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ * `BERT pre-training <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/bert_pretraining.html>`_ * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ * `Neural Duplex Text Normalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization.html>`_ * `Prompt Tuning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html#prompt-tuning>`_ * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ * `Speech synthesis (TTS) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ * Spectrogram generation: Tacotron2, GlowTTS, TalkNet, FastPitch, FastSpeech2, Mixer-TTS, Mixer-TTS-X * Vocoders: WaveGlow, SqueezeWave, UniGlow, MelGAN, HiFiGAN, UnivNet * End-to-end speech generation: FastPitch_HifiGan_E2E, FastSpeech2_HifiGan_E2E * `NGC collection of pre-trained TTS models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ * `Tools <https://github.com/NVIDIA/NeMo/tree/main/tools>`_ * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/text_processing_deployment.html>`_ * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. Requirements ------------ 1) Python 3.6, 3.7 or 3.8 2) Pytorch 1.10.0 or above 3) NVIDIA GPU for training Documentation ------------- .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main :alt: Documentation Status :scale: 100% :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable :alt: Documentation Status :scale: 100% :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ | Version | Status | Description | +=========+=============+==========================================================================================================================================+ | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ Tutorials --------- A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. Getting help with NeMo ---------------------- FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. Installation ------------ Pip ~~~ Use this installation mode if you want the latest released version. .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg pip install Cython pip install nemo_toolkit['all'] .. note:: Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. Pip from source ~~~~~~~~~~~~~~~ Use this installation mode if you want the a version from particular GitHub branch (e.g main). .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg pip install Cython python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] From source ~~~~~~~~~~~ Use this installation mode if you are contributing to NeMo. .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg git clone https://github.com/NVIDIA/NeMo cd NeMo ./reinstall.sh .. note:: If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` with ``pip install -e .`` when your PWD is the root of the NeMo repository. RNNT ~~~~ Note that RNNT requires numba to be installed from conda. .. code-block:: bash conda remove numba pip uninstall numba conda install -c conda-forge numba Megatron GPT ~~~~~~~~~~~~ Megatron GPT training requires NVIDIA Apex to be installed. .. code-block:: bash git clone https://github.com/NVIDIA/apex cd apex git checkout c8bcc98176ad8c3a0717082600c70c907891f9cb pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" ./ Docker containers: ~~~~~~~~~~~~~~~~~~ To build a nemo container with Dockerfile from a branch, please run .. code-block:: bash DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 22.01-py3 and then installing from GitHub. .. code-block:: bash docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:22.01-py3 Examples -------- Many examples can be found under `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. Contributing ------------ We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. Publications ------------ We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/blob/main/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! Citation -------- .. code-block:: bash @article{kuchaiev2019nemo, title={Nemo: a toolkit for building ai applications using neural modules}, author={Kuchaiev, Oleksii and Li, Jason and Nguyen, Huyen and Hrinchuk, Oleksii and Leary, Ryan and Ginsburg, Boris and Kriman, Samuel and Beliaev, Stanislav and Lavrukhin, Vitaly and Cook, Jack and others}, journal={arXiv preprint arXiv:1909.09577}, year={2019} } License ------- NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. [end of README.rst] [start of /dev/null] [end of /dev/null] [start of nemo_text_processing/text_normalization/__init__.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from nemo.utils import logging try: import pynini PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): logging.warning( "`pynini` is not installed ! \n" "Please run the `nemo_text_processing/setup.sh` script" "prior to usage of this toolkit." ) PYNINI_AVAILABLE = False [end of nemo_text_processing/text_normalization/__init__.py] [start of nemo_text_processing/text_normalization/en/graph_utils.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # Copyright 2015 and onwards Google, Inc. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import string from pathlib import Path from typing import Dict from nemo_text_processing.text_normalization.en.utils import get_abs_path try: import pynini from pynini import Far from pynini.export import export from pynini.examples import plurals from pynini.lib import byte, pynutil, utf8 NEMO_CHAR = utf8.VALID_UTF8_CHAR NEMO_DIGIT = byte.DIGIT NEMO_LOWER = pynini.union(*string.ascii_lowercase).optimize() NEMO_UPPER = pynini.union(*string.ascii_uppercase).optimize() NEMO_ALPHA = pynini.union(NEMO_LOWER, NEMO_UPPER).optimize() NEMO_ALNUM = pynini.union(NEMO_DIGIT, NEMO_ALPHA).optimize() NEMO_HEX = pynini.union(*string.hexdigits).optimize() NEMO_NON_BREAKING_SPACE = u"\u00A0" NEMO_SPACE = " " NEMO_WHITE_SPACE = pynini.union(" ", "\t", "\n", "\r", u"\u00A0").optimize() NEMO_NOT_SPACE = pynini.difference(NEMO_CHAR, NEMO_WHITE_SPACE).optimize() NEMO_NOT_QUOTE = pynini.difference(NEMO_CHAR, r'"').optimize() NEMO_PUNCT = pynini.union(*map(pynini.escape, string.punctuation)).optimize() NEMO_GRAPH = pynini.union(NEMO_ALNUM, NEMO_PUNCT).optimize() NEMO_SIGMA = pynini.closure(NEMO_CHAR) delete_space = pynutil.delete(pynini.closure(NEMO_WHITE_SPACE)) insert_space = pynutil.insert(" ") delete_extra_space = pynini.cross(pynini.closure(NEMO_WHITE_SPACE, 1), " ") delete_preserve_order = pynini.closure( pynutil.delete(" preserve_order: true") | (pynutil.delete(" field_order: \"") + NEMO_NOT_QUOTE + pynutil.delete("\"")) ) suppletive = pynini.string_file(get_abs_path("data/suppletive.tsv")) # _v = pynini.union("a", "e", "i", "o", "u") _c = pynini.union( "b", "c", "d", "f", "g", "h", "j", "k", "l", "m", "n", "p", "q", "r", "s", "t", "v", "w", "x", "y", "z" ) _ies = NEMO_SIGMA + _c + pynini.cross("y", "ies") _es = NEMO_SIGMA + pynini.union("s", "sh", "ch", "x", "z") + pynutil.insert("es") _s = NEMO_SIGMA + pynutil.insert("s") graph_plural = plurals._priority_union( suppletive, plurals._priority_union(_ies, plurals._priority_union(_es, _s, NEMO_SIGMA), NEMO_SIGMA), NEMO_SIGMA ).optimize() SINGULAR_TO_PLURAL = graph_plural PLURAL_TO_SINGULAR = pynini.invert(graph_plural) TO_LOWER = pynini.union(*[pynini.cross(x, y) for x, y in zip(string.ascii_uppercase, string.ascii_lowercase)]) TO_UPPER = pynini.invert(TO_LOWER) PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): # Create placeholders NEMO_CHAR = None NEMO_DIGIT = None NEMO_LOWER = None NEMO_UPPER = None NEMO_ALPHA = None NEMO_ALNUM = None NEMO_HEX = None NEMO_NON_BREAKING_SPACE = u"\u00A0" NEMO_SPACE = " " NEMO_WHITE_SPACE = None NEMO_NOT_SPACE = None NEMO_NOT_QUOTE = None NEMO_PUNCT = None NEMO_GRAPH = None NEMO_SIGMA = None delete_space = None insert_space = None delete_extra_space = None delete_preserve_order = None suppletive = None # _v = pynini.union("a", "e", "i", "o", "u") _c = None _ies = None _es = None _s = None graph_plural = None SINGULAR_TO_PLURAL = None PLURAL_TO_SINGULAR = None TO_LOWER = None TO_UPPER = None PYNINI_AVAILABLE = False def generator_main(file_name: str, graphs: Dict[str, 'pynini.FstLike']): """ Exports graph as OpenFst finite state archive (FAR) file with given file name and rule name. Args: file_name: exported file name graphs: Mapping of a rule name and Pynini WFST graph to be exported """ exporter = export.Exporter(file_name) for rule, graph in graphs.items(): exporter[rule] = graph.optimize() exporter.close() print(f'Created {file_name}') def get_plurals(fst): """ Given singular returns plurals Args: fst: Fst Returns plurals to given singular forms """ return SINGULAR_TO_PLURAL @ fst def get_singulars(fst): """ Given plural returns singulars Args: fst: Fst Returns singulars to given plural forms """ return PLURAL_TO_SINGULAR @ fst def convert_space(fst) -> 'pynini.FstLike': """ Converts space to nonbreaking space. Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. Args: fst: input fst Returns output fst where breaking spaces are converted to non breaking spaces """ return fst @ pynini.cdrewrite(pynini.cross(NEMO_SPACE, NEMO_NON_BREAKING_SPACE), "", "", NEMO_SIGMA) class GraphFst: """ Base class for all grammar fsts. Args: name: name of grammar class kind: either 'classify' or 'verbalize' deterministic: if True will provide a single transduction option, for False multiple transduction are generated (used for audio-based normalization) """ def __init__(self, name: str, kind: str, deterministic: bool = True): self.name = name self.kind = str self._fst = None self.deterministic = deterministic self.far_path = Path(os.path.dirname(__file__) + '/grammars/' + kind + '/' + name + '.far') if self.far_exist(): self._fst = Far(self.far_path, mode="r", arc_type="standard", far_type="default").get_fst() def far_exist(self) -> bool: """ Returns true if FAR can be loaded """ return self.far_path.exists() @property def fst(self) -> 'pynini.FstLike': return self._fst @fst.setter def fst(self, fst): self._fst = fst def add_tokens(self, fst) -> 'pynini.FstLike': """ Wraps class name around to given fst Args: fst: input fst Returns: Fst: fst """ return pynutil.insert(f"{self.name} {{ ") + fst + pynutil.insert(" }") def delete_tokens(self, fst) -> 'pynini.FstLike': """ Deletes class name wrap around output of given fst Args: fst: input fst Returns: Fst: fst """ res = ( pynutil.delete(f"{self.name}") + delete_space + pynutil.delete("{") + delete_space + fst + delete_space + pynutil.delete("}") ) return res @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) [end of nemo_text_processing/text_normalization/en/graph_utils.py] [start of nemo_text_processing/text_normalization/en/taggers/punctuation.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # Copyright 2015 and onwards Google, Inc. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import sys from unicodedata import category from nemo_text_processing.text_normalization.en.graph_utils import GraphFst try: import pynini from pynini.lib import pynutil PYNINI_AVAILABLE = False except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False class PunctuationFst(GraphFst): """ Finite state transducer for classifying punctuation e.g. a, -> tokens { name: "a" } tokens { name: "," } Args: deterministic: if True will provide a single transduction option, for False multiple transduction are generated (used for audio-based normalization) """ def __init__(self, deterministic: bool = True): super().__init__(name="punctuation", kind="classify", deterministic=deterministic) s = "!#%&\'()*+,-./:;<=>?@^_`{|}~\"" punct_unicode = [chr(i) for i in range(sys.maxunicode) if category(chr(i)).startswith("P")] punct_unicode.remove('[') punct_unicode.remove(']') punct = pynini.union(*s) | pynini.union(*punct_unicode) self.graph = punct self.fst = (pynutil.insert("name: \"") + self.graph + pynutil.insert("\"")).optimize() [end of nemo_text_processing/text_normalization/en/taggers/punctuation.py] [start of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # Copyright 2015 and onwards Google, Inc. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: import pynini from pynini.lib import pynutil PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False class WhiteListFst(GraphFst): """ Finite state transducer for verbalizing whitelist e.g. tokens { name: "misses" } } -> misses Args: deterministic: if True will provide a single transduction option, for False multiple transduction are generated (used for audio-based normalization) """ def __init__(self, deterministic: bool = True): super().__init__(name="whitelist", kind="verbalize", deterministic=deterministic) graph = ( pynutil.delete("name:") + delete_space + pynutil.delete("\"") + pynini.closure(NEMO_CHAR - " ", 1) + pynutil.delete("\"") ) graph = graph @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) self.fst = graph.optimize() [end of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] [start of nemo_text_processing/text_normalization/en/verbalizers/word.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # Copyright 2015 and onwards Google, Inc. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: import pynini from pynini.lib import pynutil PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False class WordFst(GraphFst): """ Finite state transducer for verbalizing word e.g. tokens { name: "sleep" } -> sleep Args: deterministic: if True will provide a single transduction option, for False multiple transduction are generated (used for audio-based normalization) """ def __init__(self, deterministic: bool = True): super().__init__(name="word", kind="verbalize", deterministic=deterministic) chars = pynini.closure(NEMO_CHAR - " ", 1) char = pynutil.delete("name:") + delete_space + pynutil.delete("\"") + chars + pynutil.delete("\"") graph = char @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) self.fst = graph.optimize() [end of nemo_text_processing/text_normalization/en/verbalizers/word.py] [start of nemo_text_processing/text_normalization/normalize.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import itertools import os import re from argparse import ArgumentParser from collections import OrderedDict from math import factorial from typing import Dict, List, Union from nemo_text_processing.text_normalization.data_loader_utils import get_installation_msg, pre_process from nemo_text_processing.text_normalization.token_parser import PRESERVE_ORDER_KEY, TokenParser from tqdm import tqdm try: import pynini PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False try: from nemo.collections.common.tokenizers.moses_tokenizers import MosesProcessor from nemo.collections.nlp.data.text_normalization.utils import post_process_punct NLP_AVAILABLE = True except (ModuleNotFoundError, ImportError): NLP_AVAILABLE = False SPACE_DUP = re.compile(' {2,}') class Normalizer: """ Normalizer class that converts text from written to spoken form. Useful for TTS preprocessing. Args: input_case: expected input capitalization lang: language specifying the TN rules, by default: English cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. overwrite_cache: set to True to overwrite .far files whitelist: path to a file with whitelist replacements """ def __init__( self, input_case: str, lang: str = 'en', deterministic: bool = True, cache_dir: str = None, overwrite_cache: bool = False, whitelist: str = None, ): assert input_case in ["lower_cased", "cased"] if not PYNINI_AVAILABLE: raise ImportError(get_installation_msg()) if lang == 'en' and deterministic: from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst elif lang == 'en' and not deterministic: from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify_with_audio import ClassifyFst from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst elif lang == 'ru': # Ru TN only support non-deterministic cases and produces multiple normalization options # use normalize_with_audio.py from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst elif lang == 'de': # Ru TN only support non-deterministic cases and produces multiple normalization options # use normalize_with_audio.py from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst self.tagger = ClassifyFst( input_case=input_case, deterministic=deterministic, cache_dir=cache_dir, overwrite_cache=overwrite_cache, whitelist=whitelist, ) self.verbalizer = VerbalizeFinalFst(deterministic=deterministic) self.parser = TokenParser() self.lang = lang if NLP_AVAILABLE: self.processor = MosesProcessor(lang_id=lang) else: self.processor = None print("NeMo NLP is not available. Moses de-tokenization will be skipped.") def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: """ NeMo text normalizer Args: texts: list of input strings verbose: whether to print intermediate meta information Returns converted list input strings """ res = [] for input in tqdm(texts): try: text = self.normalize(input, verbose=verbose, punct_post_process=punct_post_process) except: print(input) raise Exception res.append(text) return res def _estimate_number_of_permutations_in_nested_dict( self, token_group: Dict[str, Union[OrderedDict, str, bool]] ) -> int: num_perms = 1 for k, inner in token_group.items(): if isinstance(inner, dict): num_perms *= self._estimate_number_of_permutations_in_nested_dict(inner) num_perms *= factorial(len(token_group)) return num_perms def _split_tokens_to_reduce_number_of_permutations( self, tokens: List[dict], max_number_of_permutations_per_split: int = 729 ) -> List[List[dict]]: """ Splits a sequence of tokens in a smaller sequences of tokens in a way that maximum number of composite tokens permutations does not exceed ``max_number_of_permutations_per_split``. For example, .. code-block:: python tokens = [ {"tokens": {"date": {"year": "twenty eighteen", "month": "december", "day": "thirty one"}}}, {"tokens": {"date": {"year": "twenty eighteen", "month": "january", "day": "eight"}}}, ] split = normalizer._split_tokens_to_reduce_number_of_permutations( tokens, max_number_of_permutations_per_split=6 ) assert split == [ [{"tokens": {"date": {"year": "twenty eighteen", "month": "december", "day": "thirty one"}}}], [{"tokens": {"date": {"year": "twenty eighteen", "month": "january", "day": "eight"}}}], ] Date tokens contain 3 items each which gives 6 permutations for every date. Since there are 2 dates, total number of permutations would be ``6 * 6 == 36``. Parameter ``max_number_of_permutations_per_split`` equals 6, so input sequence of tokens is split into 2 smaller sequences. Args: tokens (:obj:`List[dict]`): a list of dictionaries, possibly nested. max_number_of_permutations_per_split (:obj:`int`, `optional`, defaults to :obj:`243`): a maximum number of permutations which can be generated from input sequence of tokens. Returns: :obj:`List[List[dict]]`: a list of smaller sequences of tokens resulting from ``tokens`` split. """ splits = [] prev_end_of_split = 0 current_number_of_permutations = 1 for i, token_group in enumerate(tokens): n = self._estimate_number_of_permutations_in_nested_dict(token_group) if n * current_number_of_permutations > max_number_of_permutations_per_split: splits.append(tokens[prev_end_of_split:i]) prev_end_of_split = i current_number_of_permutations = 1 if n > max_number_of_permutations_per_split: raise ValueError( f"Could not split token list with respect to condition that every split can generate number of " f"permutations less or equal to " f"`max_number_of_permutations_per_split={max_number_of_permutations_per_split}`. " f"There is an unsplittable token group that generates more than " f"{max_number_of_permutations_per_split} permutations. Try to increase " f"`max_number_of_permutations_per_split` parameter." ) current_number_of_permutations *= n splits.append(tokens[prev_end_of_split:]) assert sum([len(s) for s in splits]) == len(tokens) return splits def normalize( self, text: str, verbose: bool = False, punct_pre_process: bool = False, punct_post_process: bool = False ) -> str: """ Main function. Normalizes tokens from written to spoken form e.g. 12 kg -> twelve kilograms Args: text: string that may include semiotic classes verbose: whether to print intermediate meta information punct_pre_process: whether to perform punctuation pre-processing, for example, [25] -> [ 25 ] punct_post_process: whether to normalize punctuation Returns: spoken form """ original_text = text if punct_pre_process: text = pre_process(text) text = text.strip() if not text: if verbose: print(text) return text text = pynini.escape(text) tagged_lattice = self.find_tags(text) tagged_text = self.select_tag(tagged_lattice) if verbose: print(tagged_text) self.parser(tagged_text) tokens = self.parser.parse() split_tokens = self._split_tokens_to_reduce_number_of_permutations(tokens) output = "" for s in split_tokens: tags_reordered = self.generate_permutations(s) verbalizer_lattice = None for tagged_text in tags_reordered: tagged_text = pynini.escape(tagged_text) verbalizer_lattice = self.find_verbalizer(tagged_text) if verbalizer_lattice.num_states() != 0: break if verbalizer_lattice is None: raise ValueError(f"No permutations were generated from tokens {s}") output += ' ' + self.select_verbalizer(verbalizer_lattice) output = SPACE_DUP.sub(' ', output[1:]) if punct_post_process: # do post-processing based on Moses detokenizer if self.processor: output = self.processor.moses_detokenizer.detokenize([output], unescape=False) output = post_process_punct(input=original_text, normalized_text=output) else: print("NEMO_NLP collection is not available: skipping punctuation post_processing") return output def _permute(self, d: OrderedDict) -> List[str]: """ Creates reorderings of dictionary elements and serializes as strings Args: d: (nested) dictionary of key value pairs Return permutations of different string serializations of key value pairs """ l = [] if PRESERVE_ORDER_KEY in d.keys(): d_permutations = [d.items()] else: d_permutations = itertools.permutations(d.items()) for perm in d_permutations: subl = [""] for k, v in perm: if isinstance(v, str): subl = ["".join(x) for x in itertools.product(subl, [f"{k}: \"{v}\" "])] elif isinstance(v, OrderedDict): rec = self._permute(v) subl = ["".join(x) for x in itertools.product(subl, [f" {k} {{ "], rec, [f" }} "])] elif isinstance(v, bool): subl = ["".join(x) for x in itertools.product(subl, [f"{k}: true "])] else: raise ValueError() l.extend(subl) return l def generate_permutations(self, tokens: List[dict]): """ Generates permutations of string serializations of list of dictionaries Args: tokens: list of dictionaries Returns string serialization of list of dictionaries """ def _helper(prefix: str, tokens: List[dict], idx: int): """ Generates permutations of string serializations of given dictionary Args: tokens: list of dictionaries prefix: prefix string idx: index of next dictionary Returns string serialization of dictionary """ if idx == len(tokens): yield prefix return token_options = self._permute(tokens[idx]) for token_option in token_options: yield from _helper(prefix + token_option, tokens, idx + 1) return _helper("", tokens, 0) def find_tags(self, text: str) -> 'pynini.FstLike': """ Given text use tagger Fst to tag text Args: text: sentence Returns: tagged lattice """ lattice = text @ self.tagger.fst return lattice def select_tag(self, lattice: 'pynini.FstLike') -> str: """ Given tagged lattice return shortest path Args: tagged_text: tagged text Returns: shortest path """ tagged_text = pynini.shortestpath(lattice, nshortest=1, unique=True).string() return tagged_text def find_verbalizer(self, tagged_text: str) -> 'pynini.FstLike': """ Given tagged text creates verbalization lattice This is context-independent. Args: tagged_text: input text Returns: verbalized lattice """ lattice = tagged_text @ self.verbalizer.fst return lattice def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: """ Given verbalized lattice return shortest path Args: lattice: verbalization lattice Returns: shortest path """ output = pynini.shortestpath(lattice, nshortest=1, unique=True).string() return output def parse_args(): parser = ArgumentParser() parser.add_argument("input_string", help="input string", type=str) parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) parser.add_argument( "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) parser.add_argument("--verbose", help="print info for debugging", action='store_true') parser.add_argument( "--punct_post_process", help="set to True to enable punctuation post processing", action="store_true" ) parser.add_argument( "--punct_pre_process", help="set to True to enable punctuation pre processing", action="store_true" ) parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") parser.add_argument("--whitelist", help="path to a file with with whitelist", default=None, type=str) parser.add_argument( "--cache_dir", help="path to a dir with .far grammar file. Set to None to avoid using cache", default=None, type=str, ) return parser.parse_args() if __name__ == "__main__": args = parse_args() whitelist = os.path.abspath(args.whitelist) if args.whitelist else None normalizer = Normalizer( input_case=args.input_case, cache_dir=args.cache_dir, overwrite_cache=args.overwrite_cache, whitelist=whitelist, lang=args.language, ) print( normalizer.normalize( args.input_string, verbose=args.verbose, punct_pre_process=args.punct_pre_process, punct_post_process=args.punct_post_process, ) ) [end of nemo_text_processing/text_normalization/normalize.py] [start of nemo_text_processing/text_normalization/normalize_with_audio.py] # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os import time from argparse import ArgumentParser from glob import glob from typing import List, Tuple from joblib import Parallel, delayed from nemo_text_processing.text_normalization.normalize import Normalizer from tqdm import tqdm try: from nemo.collections.asr.metrics.wer import word_error_rate from nemo.collections.asr.models import ASRModel ASR_AVAILABLE = True except (ModuleNotFoundError, ImportError): ASR_AVAILABLE = False try: import pynini from pynini.lib import rewrite PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False try: from nemo.collections.nlp.data.text_normalization.utils import post_process_punct from nemo_text_processing.text_normalization.data_loader_utils import pre_process NLP_AVAILABLE = True except (ModuleNotFoundError, ImportError): NLP_AVAILABLE = False """ The script provides multiple normalization options and chooses the best one that minimizes CER of the ASR output (most of the semiotic classes use deterministic=False flag). To run this script with a .json manifest file, the manifest file should contain the following fields: "audio_data" - path to the audio file "text" - raw text "pred_text" - ASR model prediction See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions When the manifest is ready, run: python normalize_with_audio.py \ --audio_data PATH/TO/MANIFEST.JSON \ --language en To run with a single audio file, specify path to audio and text with: python normalize_with_audio.py \ --audio_data PATH/TO/AUDIO.WAV \ --language en \ --text raw text OR PATH/TO/.TXT/FILE --model QuartzNet15x5Base-En \ --verbose To see possible normalization options for a text input without an audio file (could be used for debugging), run: python python normalize_with_audio.py --text "RAW TEXT" Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference """ class NormalizerWithAudio(Normalizer): """ Normalizer class that converts text from written to spoken form. Useful for TTS preprocessing. Args: input_case: expected input capitalization lang: language cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. overwrite_cache: set to True to overwrite .far files whitelist: path to a file with whitelist replacements """ def __init__( self, input_case: str, lang: str = 'en', cache_dir: str = None, overwrite_cache: bool = False, whitelist: str = None, ): super().__init__( input_case=input_case, lang=lang, deterministic=False, cache_dir=cache_dir, overwrite_cache=overwrite_cache, whitelist=whitelist, ) def normalize(self, text: str, n_tagged: int, punct_post_process: bool = True, verbose: bool = False,) -> str: """ Main function. Normalizes tokens from written to spoken form e.g. 12 kg -> twelve kilograms Args: text: string that may include semiotic classes n_tagged: number of tagged options to consider, -1 - to get all possible tagged options punct_post_process: whether to normalize punctuation verbose: whether to print intermediate meta information Returns: normalized text options (usually there are multiple ways of normalizing a given semiotic class) """ original_text = text if self.lang == "en": text = pre_process(text) text = text.strip() if not text: if verbose: print(text) return text text = pynini.escape(text) if n_tagged == -1: if self.lang == "en": try: tagged_texts = rewrite.rewrites(text, self.tagger.fst_no_digits) except pynini.lib.rewrite.Error: tagged_texts = rewrite.rewrites(text, self.tagger.fst) else: tagged_texts = rewrite.rewrites(text, self.tagger.fst) else: if self.lang == "en": try: tagged_texts = rewrite.top_rewrites(text, self.tagger.fst_no_digits, nshortest=n_tagged) except pynini.lib.rewrite.Error: tagged_texts = rewrite.top_rewrites(text, self.tagger.fst, nshortest=n_tagged) else: tagged_texts = rewrite.top_rewrites(text, self.tagger.fst, nshortest=n_tagged) # non-deterministic Eng normalization uses tagger composed with verbalizer, no permutation in between if self.lang == "en": normalized_texts = tagged_texts else: normalized_texts = [] for tagged_text in tagged_texts: self._verbalize(tagged_text, normalized_texts, verbose=verbose) if len(normalized_texts) == 0: raise ValueError() if punct_post_process: # do post-processing based on Moses detokenizer if self.processor: normalized_texts = [self.processor.detokenize([t]) for t in normalized_texts] normalized_texts = [ post_process_punct(input=original_text, normalized_text=t) for t in normalized_texts ] normalized_texts = set(normalized_texts) return normalized_texts def _verbalize(self, tagged_text: str, normalized_texts: List[str], verbose: bool = False): """ Verbalizes tagged text Args: tagged_text: text with tags normalized_texts: list of possible normalization options verbose: if true prints intermediate classification results """ def get_verbalized_text(tagged_text): return rewrite.rewrites(tagged_text, self.verbalizer.fst) self.parser(tagged_text) tokens = self.parser.parse() tags_reordered = self.generate_permutations(tokens) for tagged_text_reordered in tags_reordered: try: tagged_text_reordered = pynini.escape(tagged_text_reordered) normalized_texts.extend(get_verbalized_text(tagged_text_reordered)) if verbose: print(tagged_text_reordered) except pynini.lib.rewrite.Error: continue def select_best_match( self, normalized_texts: List[str], input_text: str, pred_text: str, verbose: bool = False, remove_punct: bool = False, ): """ Selects the best normalization option based on the lowest CER Args: normalized_texts: normalized text options input_text: input text pred_text: ASR model transcript of the audio file corresponding to the normalized text verbose: whether to print intermediate meta information remove_punct: whether to remove punctuation before calculating CER Returns: normalized text with the lowest CER and CER value """ if pred_text == "": return input_text, 1000 normalized_texts_cer = calculate_cer(normalized_texts, pred_text, remove_punct) normalized_texts_cer = sorted(normalized_texts_cer, key=lambda x: x[1]) normalized_text, cer = normalized_texts_cer[0] if verbose: print('-' * 30) for option in normalized_texts: print(option) print('-' * 30) return normalized_text, cer def calculate_cer(normalized_texts: List[str], pred_text: str, remove_punct=False) -> List[Tuple[str, float]]: """ Calculates character error rate (CER) Args: normalized_texts: normalized text options pred_text: ASR model output Returns: normalized options with corresponding CER """ normalized_options = [] for text in normalized_texts: text_clean = text.replace('-', ' ').lower() if remove_punct: for punct in "!?:;,.-()*+-/<=>@^_": text_clean = text_clean.replace(punct, "") cer = round(word_error_rate([pred_text], [text_clean], use_cer=True) * 100, 2) normalized_options.append((text, cer)) return normalized_options def get_asr_model(asr_model): """ Returns ASR Model Args: asr_model: NeMo ASR model """ if os.path.exists(args.model): asr_model = ASRModel.restore_from(asr_model) elif args.model in ASRModel.get_available_model_names(): asr_model = ASRModel.from_pretrained(asr_model) else: raise ValueError( f'Provide path to the pretrained checkpoint or choose from {ASRModel.get_available_model_names()}' ) return asr_model def parse_args(): parser = ArgumentParser() parser.add_argument("--text", help="input string or path to a .txt file", default=None, type=str) parser.add_argument( "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) parser.add_argument( "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str ) parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") parser.add_argument( '--model', type=str, default='QuartzNet15x5Base-En', help='Pre-trained model name or path to model checkpoint' ) parser.add_argument( "--n_tagged", type=int, default=30, help="number of tagged options to consider, -1 - return all possible tagged options", ) parser.add_argument("--verbose", help="print info for debugging", action="store_true") parser.add_argument( "--no_remove_punct_for_cer", help="Set to True to NOT remove punctuation before calculating CER", action="store_true", ) parser.add_argument( "--no_punct_post_process", help="set to True to disable punctuation post processing", action="store_true" ) parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") parser.add_argument("--whitelist", help="path to a file with with whitelist", default=None, type=str) parser.add_argument( "--cache_dir", help="path to a dir with .far grammar file. Set to None to avoid using cache", default=None, type=str, ) parser.add_argument("--n_jobs", default=-2, type=int, help="The maximum number of concurrently running jobs") parser.add_argument("--batch_size", default=200, type=int, help="Number of examples for each process") return parser.parse_args() def _normalize_line(normalizer: NormalizerWithAudio, n_tagged, verbose, line: str, remove_punct, punct_post_process): line = json.loads(line) pred_text = line["pred_text"] normalized_texts = normalizer.normalize( text=line["text"], verbose=verbose, n_tagged=n_tagged, punct_post_process=punct_post_process, ) normalized_text, cer = normalizer.select_best_match( normalized_texts=normalized_texts, input_text=line["text"], pred_text=pred_text, verbose=verbose, remove_punct=remove_punct, ) line["nemo_normalized"] = normalized_text line["CER_nemo_normalized"] = cer return line def normalize_manifest( normalizer, audio_data: str, n_jobs: int, n_tagged: int, remove_punct: bool, punct_post_process: bool, batch_size: int, ): """ Args: args.audio_data: path to .json manifest file. """ def __process_batch(batch_idx, batch, dir_name): normalized_lines = [ _normalize_line( normalizer, n_tagged, verbose=False, line=line, remove_punct=remove_punct, punct_post_process=punct_post_process, ) for line in tqdm(batch) ] with open(f"{dir_name}/{batch_idx}.json", "w") as f_out: for line in normalized_lines: f_out.write(json.dumps(line, ensure_ascii=False) + '\n') print(f"Batch -- {batch_idx} -- is complete") return normalized_lines manifest_out = audio_data.replace('.json', '_normalized.json') with open(audio_data, 'r') as f: lines = f.readlines() print(f'Normalizing {len(lines)} lines of {audio_data}...') # to save intermediate results to a file batch = min(len(lines), batch_size) tmp_dir = manifest_out.replace(".json", "_parts") os.makedirs(tmp_dir, exist_ok=True) Parallel(n_jobs=n_jobs)( delayed(__process_batch)(idx, lines[i : i + batch], tmp_dir) for idx, i in enumerate(range(0, len(lines), batch)) ) # aggregate all intermediate files with open(manifest_out, "w") as f_out: for batch_f in sorted(glob(f"{tmp_dir}/*.json")): with open(batch_f, "r") as f_in: lines = f_in.read() f_out.write(lines) print(f'Normalized version saved at {manifest_out}') if __name__ == "__main__": args = parse_args() if not ASR_AVAILABLE and args.audio_data: raise ValueError("NeMo ASR collection is not installed.") start = time.time() args.whitelist = os.path.abspath(args.whitelist) if args.whitelist else None if args.text is not None: normalizer = NormalizerWithAudio( input_case=args.input_case, lang=args.language, cache_dir=args.cache_dir, overwrite_cache=args.overwrite_cache, whitelist=args.whitelist, ) if os.path.exists(args.text): with open(args.text, 'r') as f: args.text = f.read().strip() normalized_texts = normalizer.normalize( text=args.text, verbose=args.verbose, n_tagged=args.n_tagged, punct_post_process=not args.no_punct_post_process, ) if args.audio_data: asr_model = get_asr_model(args.model) pred_text = asr_model.transcribe([args.audio_data])[0] normalized_text, cer = normalizer.select_best_match( normalized_texts=normalized_texts, pred_text=pred_text, input_text=args.text, verbose=args.verbose, remove_punct=not args.no_remove_punct_for_cer, ) print(f"Transcript: {pred_text}") print(f"Normalized: {normalized_text}") else: print("Normalization options:") for norm_text in normalized_texts: print(norm_text) elif not os.path.exists(args.audio_data): raise ValueError(f"{args.audio_data} not found.") elif args.audio_data.endswith('.json'): normalizer = NormalizerWithAudio( input_case=args.input_case, lang=args.language, cache_dir=args.cache_dir, overwrite_cache=args.overwrite_cache, whitelist=args.whitelist, ) normalize_manifest( normalizer=normalizer, audio_data=args.audio_data, n_jobs=args.n_jobs, n_tagged=args.n_tagged, remove_punct=not args.no_remove_punct_for_cer, punct_post_process=not args.no_punct_post_process, batch_size=args.batch_size, ) else: raise ValueError( "Provide either path to .json manifest in '--audio_data' OR " + "'--audio_data' path to audio file and '--text' path to a text file OR" "'--text' string text (for debugging without audio)" ) print(f'Execution time: {round((time.time() - start)/60, 2)} min.') [end of nemo_text_processing/text_normalization/normalize_with_audio.py] [start of tools/text_processing_deployment/pynini_export.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # Copyright 2015 and onwards Google, Inc. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import time from argparse import ArgumentParser from nemo.utils import logging try: import pynini from nemo_text_processing.text_normalization.en.graph_utils import generator_main PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): logging.warning( "`pynini` is not installed ! \n" "Please run the `nemo_text_processing/setup.sh` script" "prior to usage of this toolkit." ) PYNINI_AVAILABLE = False # This script exports compiled grammars inside nemo_text_processing into OpenFst finite state archive files # tokenize_and_classify.far and verbalize.far for production purposes def itn_grammars(**kwargs): d = {} d['classify'] = { 'TOKENIZE_AND_CLASSIFY': ITNClassifyFst( cache_dir=kwargs["cache_dir"], overwrite_cache=kwargs["overwrite_cache"] ).fst } d['verbalize'] = {'ALL': ITNVerbalizeFst().fst, 'REDUP': pynini.accep("REDUP")} return d def tn_grammars(**kwargs): d = {} d['classify'] = { 'TOKENIZE_AND_CLASSIFY': TNClassifyFst( input_case=kwargs["input_case"], deterministic=True, cache_dir=kwargs["cache_dir"], overwrite_cache=kwargs["overwrite_cache"], ).fst } d['verbalize'] = {'ALL': TNVerbalizeFst(deterministic=True).fst, 'REDUP': pynini.accep("REDUP")} return d def export_grammars(output_dir, grammars): """ Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. Args: output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. grammars: grammars to be exported """ for category, graphs in grammars.items(): out_dir = os.path.join(output_dir, category) if not os.path.exists(out_dir): os.makedirs(out_dir) time.sleep(1) if category == "classify": category = "tokenize_and_classify" generator_main(f"{out_dir}/{category}.far", graphs) def parse_args(): parser = ArgumentParser() parser.add_argument("--output_dir", help="output directory for grammars", required=True, type=str) parser.add_argument( "--language", help="language", choices=["en", "de", "es", "ru", 'fr', 'vi'], type=str, default='en' ) parser.add_argument( "--grammars", help="grammars to be exported", choices=["tn_grammars", "itn_grammars"], type=str, required=True ) parser.add_argument( "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") parser.add_argument( "--cache_dir", help="path to a dir with .far grammar file. Set to None to avoid using cache", default=None, type=str, ) return parser.parse_args() if __name__ == '__main__': args = parse_args() if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') if args.language == 'en': from nemo_text_processing.inverse_text_normalization.en.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, ) from nemo_text_processing.inverse_text_normalization.en.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify import ( ClassifyFst as TNClassifyFst, ) from nemo_text_processing.text_normalization.en.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst elif args.language == 'de': from nemo_text_processing.inverse_text_normalization.de.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, ) from nemo_text_processing.inverse_text_normalization.de.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ( ClassifyFst as TNClassifyFst, ) from nemo_text_processing.text_normalization.de.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst elif args.language == 'ru': from nemo_text_processing.inverse_text_normalization.ru.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, ) from nemo_text_processing.inverse_text_normalization.ru.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) elif args.language == 'es': from nemo_text_processing.inverse_text_normalization.es.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, ) from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) elif args.language == 'fr': from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, ) from nemo_text_processing.inverse_text_normalization.fr.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) elif args.language == 'vi': from nemo_text_processing.inverse_text_normalization.vi.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, ) from nemo_text_processing.inverse_text_normalization.vi.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) output_dir = os.path.join(args.output_dir, args.language) export_grammars( output_dir=output_dir, grammars=locals()[args.grammars]( input_case=args.input_case, cache_dir=args.cache_dir, overwrite_cache=args.overwrite_cache ), ) [end of tools/text_processing_deployment/pynini_export.py] </code> I need you to solve this issue by regenerating the full files in the code base that you would like to change. You can change as many files as you like. Please respond with a list of files and their revised contents in the following format. <example> [start of /src/this_file.py] import os def euclidean(a, b): if b == 0: return a return euclidean(b, a % b) [end of /src/this_file.py] [start of /src/another_file.py] def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) x, y = x0, y0 sx = -1 if x0 > x1 else 1 sy = -1 if y0 > y1 else 1 if dx > dy: err = dx / 2.0 while x != x1: points.append((x, y)) err -= dy if err < 0: y += sy err += dx x += sx else: err = dy / 2.0 while y != y1: points.append((x err -= dx if err < 0: x += sx err += dy y += sy points.append((x, y)) return points [end of /src/another_file.py] </example>
NVIDIA/NeMo
022f0292aecbc98d591d49423d5045235394f793
./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`.
2022-02-09T05:12:31Z
<patch> <patch> diff --git a/nemo_text_processing/text_normalization/__init__.py b/nemo_text_processing/text_normalization/__init__.py --- a/nemo_text_processing/text_normalization/__init__.py +++ b/nemo_text_processing/text_normalization/__init__.py @@ -21,7 +21,7 @@ except (ModuleNotFoundError, ImportError): logging.warning( "`pynini` is not installed ! \n" - "Please run the `nemo_text_processing/setup.sh` script" + "Please run the `nemo_text_processing/setup.sh` script " "prior to usage of this toolkit." ) diff --git a/nemo_text_processing/text_normalization/en/graph_utils.py b/nemo_text_processing/text_normalization/en/graph_utils.py --- a/nemo_text_processing/text_normalization/en/graph_utils.py +++ b/nemo_text_processing/text_normalization/en/graph_utils.py @@ -159,7 +159,7 @@ def convert_space(fst) -> 'pynini.FstLike': """ Converts space to nonbreaking space. Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" - This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. + This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. Args: fst: input fst @@ -208,9 +208,9 @@ def add_tokens(self, fst) -> 'pynini.FstLike': """ Wraps class name around to given fst - Args: + Args: fst: input fst - + Returns: Fst: fst """ diff --git a/nemo_text_processing/text_normalization/en/taggers/punctuation.py b/nemo_text_processing/text_normalization/en/taggers/punctuation.py --- a/nemo_text_processing/text_normalization/en/taggers/punctuation.py +++ b/nemo_text_processing/text_normalization/en/taggers/punctuation.py @@ -22,7 +22,7 @@ import pynini from pynini.lib import pynutil - PYNINI_AVAILABLE = False + PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py --- a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py @@ -12,8 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -21,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/word.py b/nemo_text_processing/text_normalization/en/verbalizers/word.py --- a/nemo_text_processing/text_normalization/en/verbalizers/word.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/word.py @@ -12,7 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -20,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/es/__init__.py b/nemo_text_processing/text_normalization/es/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/__init__.py @@ -0,0 +1,15 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +LOCALIZATION = "eu" # Set to am for alternate formatting diff --git a/nemo_text_processing/text_normalization/es/data/__init__.py b/nemo_text_processing/text_normalization/es/data/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/dates/__init__.py b/nemo_text_processing/text_normalization/es/data/dates/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/dates/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/electronic/__init__.py b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/fractions/__init__.py b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/measures/__init__.py b/nemo_text_processing/text_normalization/es/data/measures/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/measures/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/money/__init__.py b/nemo_text_processing/text_normalization/es/data/money/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/money/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/numbers/__init__.py b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/roman/__init__.py b/nemo_text_processing/text_normalization/es/data/roman/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/roman/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/time/__init__.py b/nemo_text_processing/text_normalization/es/data/time/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/time/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/graph_utils.py b/nemo_text_processing/text_normalization/es/graph_utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/graph_utils.py @@ -0,0 +1,179 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, NEMO_SPACE +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + digits = pynini.project(pynini.string_file(get_abs_path("data/numbers/digit.tsv")), "input") + tens = pynini.project(pynini.string_file(get_abs_path("data/numbers/ties.tsv")), "input") + teens = pynini.project(pynini.string_file(get_abs_path("data/numbers/teen.tsv")), "input") + twenties = pynini.project(pynini.string_file(get_abs_path("data/numbers/twenties.tsv")), "input") + hundreds = pynini.project(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv")), "input") + + accents = pynini.string_map([("รก", "a"), ("รฉ", "e"), ("รญ", "i"), ("รณ", "o"), ("รบ", "u")]) + + if LOCALIZATION == "am": # Setting localization for central and northern america formatting + cardinal_separator = pynini.string_map([",", NEMO_SPACE]) + decimal_separator = pynini.accep(".") + else: + cardinal_separator = pynini.string_map([".", NEMO_SPACE]) + decimal_separator = pynini.accep(",") + + ones = pynini.union("un", "รบn") + fem_ones = pynini.union(pynini.cross("un", "una"), pynini.cross("รบn", "una"), pynini.cross("uno", "una")) + one_to_one_hundred = pynini.union(digits, tens, teens, twenties, tens + pynini.accep(" y ") + digits) + fem_hundreds = hundreds @ pynini.cdrewrite(pynini.cross("ientos", "ientas"), "", "", NEMO_SIGMA) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digits = None + tens = None + teens = None + twenties = None + hundreds = None + + accents = None + + cardinal_separator = None + decimal_separator = None + + ones = None + fem_ones = None + one_to_one_hundred = None + fem_hundreds = None + + PYNINI_AVAILABLE = False + + +def strip_accent(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Converts all accented vowels to non-accented equivalents + + Args: + fst: Any fst. Composes vowel conversion onto fst's output strings + """ + return fst @ pynini.cdrewrite(accents, "", "", NEMO_SIGMA) + + +def shift_cardinal_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Applies gender conversion rules to a cardinal string. These include: rendering all masculine forms of "uno" (including apocopated forms) as "una" and + Converting all gendered numbers in the hundreds series (200,300,400...) to feminine equivalent (e.g. "doscientos" -> "doscientas"). Converssion only applies + to value place for <1000 and multiple of 1000. (e.g. "doscientos mil doscientos" -> "doscientas mil doscientas".) For place values greater than the thousands, there + is no gender shift as the higher powers of ten ("millones", "billones") are masculine nouns and any conversion would be formally + ungrammatical. + e.g. + "doscientos" -> "doscientas" + "doscientos mil" -> "doscientas mil" + "doscientos millones" -> "doscientos millones" + "doscientos mil millones" -> "doscientos mil millones" + "doscientos millones doscientos mil doscientos" -> "doscientos millones doscientas mil doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + before_mil = ( + NEMO_SPACE + + (pynini.accep("mil") | pynini.accep("milรฉsimo")) + + pynini.closure(NEMO_SPACE + hundreds, 0, 1) + + pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + + pynini.union(pynini.accep("[EOS]"), pynini.accep("\""), decimal_separator) + ) + before_double_digits = pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + pynini.union( + pynini.accep("[EOS]"), pynini.accep("\"") + ) + + fem_allign = pynini.cdrewrite(fem_hundreds, "", before_mil, NEMO_SIGMA) # doscientas mil dosciento + fem_allign @= pynini.cdrewrite(fem_hundreds, "", before_double_digits, NEMO_SIGMA) # doscientas mil doscienta + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union("[EOS]", "\"", decimal_separator), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def shift_number_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Performs gender conversion on all verbalized numbers in output. All values in the hundreds series (200,300,400) are changed to + feminine gender (e.g. "doscientos" -> "doscientas") and all forms of "uno" (including apocopated forms) are converted to "una". + This has no boundary restriction and will perform shift across all values in output string. + e.g. + "doscientos" -> "doscientas" + "doscientos millones" -> "doscientas millones" + "doscientos millones doscientos" -> "doscientas millones doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + fem_allign = pynini.cdrewrite(fem_hundreds, "", "", NEMO_SIGMA) + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union(NEMO_SPACE, pynini.accep("[EOS]"), pynini.accep("\"")), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def strip_cardinal_apocope(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Reverts apocope on cardinal strings in line with formation rules. e.g. "un" -> "uno". Due to cardinal formation rules, this in effect only + affects strings where the final value is a variation of "un". + e.g. + "un" -> "uno" + "veintiรบn" -> "veintiuno" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + # Since cardinals use apocope by default for large values (e.g. "millรณn"), this only needs to act on the last instance of one + strip = pynini.cross("un", "uno") | pynini.cross("รบn", "uno") + strip = pynini.cdrewrite(strip, "", pynini.union("[EOS]", "\""), NEMO_SIGMA) + return fst @ strip + + +def roman_to_int(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Alters given fst to convert Roman integers (lower and upper cased) into Arabic numerals. Valid for values up to 1000. + e.g. + "V" -> "5" + "i" -> "1" + + Args: + fst: Any fst. Composes fst onto Roman conversion outputs. + """ + + def _load_roman(file: str): + roman = load_labels(get_abs_path(file)) + roman_numerals = [(x, y) for x, y in roman] + [(x.upper(), y) for x, y in roman] + return pynini.string_map(roman_numerals) + + digit = _load_roman("data/roman/digit.tsv") + ties = _load_roman("data/roman/ties.tsv") + hundreds = _load_roman("data/roman/hundreds.tsv") + + graph = ( + digit + | ties + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + | ( + hundreds + + (ties | pynutil.add_weight(pynutil.insert("0"), 0.01)) + + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + ) + ).optimize() + + return graph @ fst diff --git a/nemo_text_processing/text_normalization/es/taggers/__init__.py b/nemo_text_processing/text_normalization/es/taggers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/taggers/cardinal.py b/nemo_text_processing/text_normalization/es/taggers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/cardinal.py @@ -0,0 +1,190 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import cardinal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + teen = pynini.invert(pynini.string_file(get_abs_path("data/numbers/teen.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/ties.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/twenties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + zero = None + digit = None + teen = None + ties = None + twenties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def filter_punctuation(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Helper function for parsing number strings. Converts common cardinal strings (groups of three digits delineated by 'cardinal_separator' - see graph_utils) + and converts to a string of digits: + "1 000" -> "1000" + "1.000.000" -> "1000000" + Args: + fst: Any pynini.FstLike object. Function composes fst onto string parser fst + + Returns: + fst: A pynini.FstLike object + """ + exactly_three_digits = NEMO_DIGIT ** 3 # for blocks of three + up_to_three_digits = pynini.closure(NEMO_DIGIT, 1, 3) # for start of string + + cardinal_string = pynini.closure( + NEMO_DIGIT, 1 + ) # For string w/o punctuation (used for page numbers, thousand series) + + cardinal_string |= ( + up_to_three_digits + + pynutil.delete(cardinal_separator) + + pynini.closure(exactly_three_digits + pynutil.delete(cardinal_separator)) + + exactly_three_digits + ) + + return cardinal_string @ fst + + +class CardinalFst(GraphFst): + """ + Finite state transducer for classifying cardinals, e.g. + "1000" -> cardinal { integer: "mil" } + "2.000.000" -> cardinal { integer: "dos millones" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="classify", deterministic=deterministic) + + # Any single digit + graph_digit = digit + digits_no_one = (NEMO_DIGIT - "1") @ graph_digit + + # Any double digit + graph_tens = teen + graph_tens |= ties + (pynutil.delete('0') | (pynutil.insert(" y ") + graph_digit)) + graph_tens |= twenties + + self.tens = graph_tens.optimize() + + self.two_digit_non_zero = pynini.union( + graph_digit, graph_tens, (pynini.cross("0", NEMO_SPACE) + graph_digit) + ).optimize() + + # Three digit strings + graph_hundreds = hundreds + pynini.union( + pynutil.delete("00"), (insert_space + graph_tens), (pynini.cross("0", NEMO_SPACE) + graph_digit) + ) + graph_hundreds |= pynini.cross("100", "cien") + graph_hundreds |= ( + pynini.cross("1", "ciento") + insert_space + pynini.union(graph_tens, pynutil.delete("0") + graph_digit) + ) + + self.hundreds = graph_hundreds.optimize() + + # For all three digit strings with leading zeroes (graph appends '0's to manage place in string) + graph_hundreds_component = pynini.union(graph_hundreds, pynutil.delete("0") + graph_tens) + + graph_hundreds_component_at_least_one_none_zero_digit = graph_hundreds_component | ( + pynutil.delete("00") + graph_digit + ) + graph_hundreds_component_at_least_one_none_zero_digit_no_one = graph_hundreds_component | ( + pynutil.delete("00") + digits_no_one + ) + + graph_thousands_component_at_least_one_none_zero_digit = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_thousands_component_at_least_one_none_zero_digit_no_one = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit_no_one, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_million = pynutil.add_weight(pynini.cross("000001", "un millรณn"), -0.001) + graph_million |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" millones") + graph_million |= pynutil.delete("000000") + graph_million += insert_space + + graph_billion = pynutil.add_weight(pynini.cross("000001", "un billรณn"), -0.001) + graph_billion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" billones") + graph_billion |= pynutil.delete("000000") + graph_billion += insert_space + + graph_trillion = pynutil.add_weight(pynini.cross("000001", "un trillรณn"), -0.001) + graph_trillion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" trillones") + graph_trillion |= pynutil.delete("000000") + graph_trillion += insert_space + + graph = ( + graph_trillion + + graph_billion + + graph_million + + (graph_thousands_component_at_least_one_none_zero_digit | pynutil.delete("000000")) + ) + + self.graph = ( + ((NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 0)) + @ pynini.cdrewrite(pynini.closure(pynutil.insert("0")), "[BOS]", "", NEMO_SIGMA) + @ NEMO_DIGIT ** 24 + @ graph + @ pynini.cdrewrite(delete_space, "[BOS]", "", NEMO_SIGMA) + @ pynini.cdrewrite(delete_space, "", "[EOS]", NEMO_SIGMA) + @ pynini.cdrewrite( + pynini.cross(pynini.closure(NEMO_WHITE_SPACE, 2), NEMO_SPACE), NEMO_ALPHA, NEMO_ALPHA, NEMO_SIGMA + ) + ) + self.graph |= zero + + self.graph = filter_punctuation(self.graph).optimize() + + optional_minus_graph = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + final_graph = optional_minus_graph + pynutil.insert("integer: \"") + self.graph + pynutil.insert("\"") + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/date.py b/nemo_text_processing/text_normalization/es/taggers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/date.py @@ -0,0 +1,107 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_DIGIT, NEMO_SPACE, GraphFst, delete_extra_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + articles = pynini.union("de", "del", "el", "del aรฑo") + delete_leading_zero = (pynutil.delete("0") | (NEMO_DIGIT - "0")) + NEMO_DIGIT + month_numbers = pynini.string_file(get_abs_path("data/dates/months.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + articles = None + delete_leading_zero = None + month_numbers = None + + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for classifying date, e.g. + "01.04.2010" -> date { day: "un" month: "enero" year: "dos mil diez" preserve_order: true } + "marzo 4 2000" -> date { month: "marzo" day: "cuatro" year: "dos mil" } + "1990-20-01" -> date { year: "mil novecientos noventa" day: "veinte" month: "enero" } + + Args: + cardinal: cardinal GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool): + super().__init__(name="date", kind="classify", deterministic=deterministic) + + number_to_month = month_numbers.optimize() + month_graph = pynini.project(number_to_month, "output") + + numbers = cardinal.graph + optional_leading_zero = delete_leading_zero | NEMO_DIGIT + + # 01, 31, 1 + digit_day = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 32)]) @ numbers + day = (pynutil.insert("day: \"") + digit_day + pynutil.insert("\"")).optimize() + + digit_month = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 13)]) + number_to_month = digit_month @ number_to_month + + month_name = (pynutil.insert("month: \"") + month_graph + pynutil.insert("\"")).optimize() + month_number = (pynutil.insert("month: \"") + number_to_month + pynutil.insert("\"")).optimize() + + # prefer cardinal over year + year = (NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 1, 3) # 90, 990, 1990 + year @= numbers + self.year = year + + year_only = pynutil.insert("year: \"") + year + pynutil.insert("\"") + year_with_articles = ( + pynutil.insert("year: \"") + pynini.closure(articles + NEMO_SPACE, 0, 1) + year + pynutil.insert("\"") + ) + + graph_dmy = ( + day + + pynini.closure(pynutil.delete(" de")) + + NEMO_SPACE + + month_name + + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + graph_mdy = ( # English influences on language + month_name + delete_extra_space + day + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + separators = [".", "-", "/"] + for sep in separators: + year_optional = pynini.closure(pynini.cross(sep, NEMO_SPACE) + year_only, 0, 1) + new_graph = day + pynini.cross(sep, NEMO_SPACE) + month_number + year_optional + graph_dmy |= new_graph + if not deterministic: + new_graph = month_number + pynini.cross(sep, NEMO_SPACE) + day + year_optional + graph_mdy |= new_graph + + dash = "-" + day_optional = pynini.closure(pynini.cross(dash, NEMO_SPACE) + day, 0, 1) + graph_ymd = NEMO_DIGIT ** 4 @ year_only + pynini.cross(dash, NEMO_SPACE) + month_number + day_optional + + final_graph = graph_dmy + pynutil.insert(" preserve_order: true") + final_graph |= graph_ymd + final_graph |= graph_mdy + + self.final_graph = final_graph.optimize() + self.fst = self.add_tokens(self.final_graph).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/decimals.py b/nemo_text_processing/text_normalization/es/taggers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/decimals.py @@ -0,0 +1,138 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + cardinal_separator, + decimal_separator, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + quantities = pynini.string_file(get_abs_path("data/numbers/quantities.tsv")) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + quantities = None + digit = None + zero = None + + PYNINI_AVAILABLE = False + + +def get_quantity(decimal_graph: 'pynini.FstLike', cardinal_graph: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Returns FST that transforms either a cardinal or decimal followed by a quantity into a numeral, + e.g. 2 millones -> integer_part: "dos" quantity: "millones" + e.g. 2,4 millones -> integer_part: "dos" fractional_part: "quatro" quantity: "millones" + e.g. 2,400 millones -> integer_part: "dos mil cuatrocientos" fractional_part: "quatro" quantity: "millones" + + Args: + decimal_graph: DecimalFST + cardinal_graph: CardinalFST + """ + numbers = pynini.closure(NEMO_DIGIT, 1, 6) @ cardinal_graph + numbers = pynini.cdrewrite(pynutil.delete(cardinal_separator), "", "", NEMO_SIGMA) @ numbers + + res = ( + pynutil.insert("integer_part: \"") + + numbers # The cardinal we're passing only produces 'un' for one, so gender agreement is safe (all quantities are masculine). Limit to 10^6 power. + + pynutil.insert("\"") + + NEMO_SPACE + + pynutil.insert("quantity: \"") + + quantities + + pynutil.insert("\"") + ) + res |= decimal_graph + NEMO_SPACE + pynutil.insert("quantity: \"") + quantities + pynutil.insert("\"") + return res + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + -11,4006 billones -> decimal { negative: "true" integer_part: "once" fractional_part: "cuatro cero cero seis" quantity: "billones" preserve_order: true } + 1 billรณn -> decimal { integer_part: "un" quantity: "billรณn" preserve_order: true } + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + graph_digit = digit | zero + + if not deterministic: + graph = pynini.union(graph_digit, cardinal.hundreds, cardinal.tens) + graph += pynini.closure(insert_space + graph) + + else: + # General pattern seems to be 1-3 digits: map as cardinal, default to digits otherwise \ + graph = pynini.union( + graph_digit, + cardinal.tens, + cardinal.hundreds, + graph_digit + pynini.closure(insert_space + graph_digit, 3), + zero + + pynini.closure(insert_space + zero) + + pynini.closure(insert_space + graph_digit), # For cases such as "1,010" + ) + + # Need to strip apocope everywhere BUT end of string + reverse_apocope = pynini.string_map([("un", "uno"), ("รบn", "uno")]) + apply_reverse_apocope = pynini.cdrewrite(reverse_apocope, "", NEMO_SPACE, NEMO_SIGMA) + graph @= apply_reverse_apocope + + # Technically decimals should be space delineated groups of three, e.g. (1,333 333). This removes any possible spaces + strip_formatting = pynini.cdrewrite(delete_space, "", "", NEMO_SIGMA) + graph = strip_formatting @ graph + + self.graph = graph.optimize() + + graph_separator = pynutil.delete(decimal_separator) + optional_graph_negative = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + self.graph_fractional = pynutil.insert("fractional_part: \"") + self.graph + pynutil.insert("\"") + + # Integer graph maintains apocope except for ones place + graph_integer = ( + strip_cardinal_apocope(cardinal.graph) + if deterministic + else pynini.union(cardinal.graph, strip_cardinal_apocope(cardinal.graph)) + ) # Gives us forms w/ and w/o apocope + self.graph_integer = pynutil.insert("integer_part: \"") + graph_integer + pynutil.insert("\"") + final_graph_wo_sign = self.graph_integer + graph_separator + insert_space + self.graph_fractional + + self.final_graph_wo_negative = ( + final_graph_wo_sign | get_quantity(final_graph_wo_sign, cardinal.graph).optimize() + ) + final_graph = optional_graph_negative + self.final_graph_wo_negative + + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/electronic.py b/nemo_text_processing/text_normalization/es/taggers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/electronic.py @@ -0,0 +1,84 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_ALPHA, NEMO_DIGIT, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + common_domains = [x[0] for x in load_labels(get_abs_path("data/electronic/domain.tsv"))] + symbols = [x[0] for x in load_labels(get_abs_path("data/electronic/symbols.tsv"))] + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + common_domains = None + symbols = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for classifying electronic: email addresses + e.g. "abc@hotmail.com" -> electronic { username: "abc" domain: "hotmail.com" preserve_order: true } + e.g. "www.abc.com/123" -> electronic { protocol: "www." domain: "abc.com/123" preserve_order: true } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="classify", deterministic=deterministic) + + dot = pynini.accep(".") + accepted_common_domains = pynini.union(*common_domains) + accepted_symbols = pynini.union(*symbols) - dot + accepted_characters = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols) + acceepted_characters_with_dot = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols | dot) + + # email + username = ( + pynutil.insert("username: \"") + + acceepted_characters_with_dot + + pynutil.insert("\"") + + pynini.cross('@', ' ') + ) + domain_graph = accepted_characters + dot + accepted_characters + domain_graph = pynutil.insert("domain: \"") + domain_graph + pynutil.insert("\"") + domain_common_graph = ( + pynutil.insert("domain: \"") + + accepted_characters + + accepted_common_domains + + pynini.closure((accepted_symbols | dot) + pynini.closure(accepted_characters, 1), 0, 1) + + pynutil.insert("\"") + ) + graph = (username + domain_graph) | domain_common_graph + + # url + protocol_start = pynini.accep("https://") | pynini.accep("http://") + protocol_end = ( + pynini.accep("www.") + if deterministic + else pynini.accep("www.") | pynini.cross("www.", "doble ve doble ve doble ve.") + ) + protocol = protocol_start | protocol_end | (protocol_start + protocol_end) + protocol = pynutil.insert("protocol: \"") + protocol + pynutil.insert("\"") + graph |= protocol + insert_space + (domain_graph | domain_common_graph) + self.graph = graph + + final_graph = self.add_tokens(self.graph + pynutil.insert(" preserve_order: true")) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/fraction.py b/nemo_text_processing/text_normalization/es/taggers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/fraction.py @@ -0,0 +1,124 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + ordinal_exceptions = pynini.string_file(get_abs_path("data/fractions/ordinal_exceptions.tsv")) + higher_powers_of_ten = pynini.string_file(get_abs_path("data/fractions/powers_of_ten.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + ordinal_exceptions = None + higher_powers_of_ten = None + + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for classifying fraction + "23 4/5" -> + tokens { fraction { integer: "veintitrรฉs" numerator: "cuatro" denominator: "quinto" mophosyntactic_features: "ordinal" } } + + Args: + cardinal: CardinalFst + ordinal: OrdinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, ordinal: GraphFst, deterministic: bool = True): + super().__init__(name="fraction", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + ordinal_graph = ordinal.graph + + # 2-10 are all ordinals + three_to_ten = pynini.string_map(["2", "3", "4", "5", "6", "7", "8", "9", "10",]) + block_three_to_ten = pynutil.delete(three_to_ten) # To block cardinal productions + if not deterministic: # Multiples of tens are sometimes rendered as ordinals + three_to_ten |= pynini.string_map(["20", "30", "40", "50", "60", "70", "80", "90",]) + graph_three_to_ten = three_to_ten @ ordinal_graph + graph_three_to_ten @= pynini.cdrewrite(ordinal_exceptions, "", "", NEMO_SIGMA) + + # Higher powers of tens (and multiples) are converted to ordinals. + hundreds = pynini.string_map(["100", "200", "300", "400", "500", "600", "700", "800", "900",]) + graph_hundreds = hundreds @ ordinal_graph + + multiples_of_thousand = ordinal.multiples_of_thousand # So we can have X milรฉsimos + + graph_higher_powers_of_ten = ( + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + + pynini.closure("mil ", 0, 1) + + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + ) # x millones / x mil millones / x mil z millones + graph_higher_powers_of_ten += higher_powers_of_ten + graph_higher_powers_of_ten = cardinal_graph @ graph_higher_powers_of_ten + graph_higher_powers_of_ten @= pynini.cdrewrite( + pynutil.delete("un "), pynini.accep("[BOS]"), pynini.project(higher_powers_of_ten, "output"), NEMO_SIGMA + ) # we drop 'un' from these ordinals (millionths, not one-millionths) + + graph_higher_powers_of_ten = multiples_of_thousand | graph_hundreds | graph_higher_powers_of_ten + block_higher_powers_of_ten = pynutil.delete( + pynini.project(graph_higher_powers_of_ten, "input") + ) # For cardinal graph + + graph_fractions_ordinals = graph_higher_powers_of_ten | graph_three_to_ten + graph_fractions_ordinals += pynutil.insert( + "\" morphosyntactic_features: \"ordinal\"" + ) # We note the root for processing later + + # Blocking the digits and hundreds from Cardinal graph + graph_fractions_cardinals = pynini.cdrewrite( + block_three_to_ten | block_higher_powers_of_ten, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fractions_cardinals @= NEMO_CHAR.plus @ pynini.cdrewrite( + pynutil.delete("0"), pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) # Empty characters become '0' for NEMO_CHAR fst, so ned to block + graph_fractions_cardinals @= cardinal_graph + graph_fractions_cardinals += pynutil.insert( + "\" morphosyntactic_features: \"add_root\"" + ) # blocking these entries to reduce erroneous possibilities in debugging + + if deterministic: + graph_fractions_cardinals = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ graph_fractions_cardinals + ) # Past hundreds the conventional scheme can be hard to read. For determinism we stop here + + graph_denominator = pynini.union( + graph_fractions_ordinals, + graph_fractions_cardinals, + pynutil.add_weight(cardinal_graph + pynutil.insert("\""), 0.001), + ) # Last form is simply recording the cardinal. Weighting so last resort + + integer = pynutil.insert("integer_part: \"") + cardinal_graph + pynutil.insert("\"") + NEMO_SPACE + numerator = ( + pynutil.insert("numerator: \"") + cardinal_graph + (pynini.cross("/", "\" ") | pynini.cross(" / ", "\" ")) + ) + denominator = pynutil.insert("denominator: \"") + graph_denominator + + self.graph = pynini.closure(integer, 0, 1) + numerator + denominator + + final_graph = self.add_tokens(self.graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/measure.py b/nemo_text_processing/text_normalization/es/taggers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/measure.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_NON_BREAKING_SPACE, + NEMO_SPACE, + GraphFst, + convert_space, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit = pynini.string_file(get_abs_path("data/measures/measurements.tsv")) + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit = None + unit_plural_fem = None + unit_plural_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for classifying measure, e.g. + "2,4 g" -> measure { cardinal { integer_part: "dos" fractional_part: "cuatro" units: "gramos" preserve_order: true } } + "1 g" -> measure { cardinal { integer: "un" units: "gramo" preserve_order: true } } + "1 millรณn g" -> measure { cardinal { integer: "un quantity: "millรณn" units: "gramos" preserve_order: true } } + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + This class also converts words containing numbers and letters + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + + + Args: + cardinal: CardinalFst + decimal: DecimalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, fraction: GraphFst, deterministic: bool = True): + super().__init__(name="measure", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + + unit_singular = unit + unit_plural = unit_singular @ (unit_plural_fem | unit_plural_masc) + + graph_unit_singular = convert_space(unit_singular) + graph_unit_plural = convert_space(unit_plural) + + optional_graph_negative = pynini.closure("-", 0, 1) + + graph_unit_denominator = ( + pynini.cross("/", "por") + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_singular + ) + + optional_unit_denominator = pynini.closure( + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_denominator, 0, 1, + ) + + unit_plural = ( + pynutil.insert("units: \"") + + ((graph_unit_plural + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + unit_singular_graph = ( + pynutil.insert("units: \"") + + ((graph_unit_singular + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + subgraph_decimal = decimal.fst + insert_space + pynini.closure(NEMO_SPACE, 0, 1) + unit_plural + + subgraph_cardinal = ( + (optional_graph_negative + (pynini.closure(NEMO_DIGIT) - "1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_plural + ) + + subgraph_cardinal |= ( + (optional_graph_negative + pynini.accep("1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_singular_graph + ) + + subgraph_fraction = fraction.fst + insert_space + pynini.closure(delete_space, 0, 1) + unit_plural + + decimal_times = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_times = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.insert("\" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_dash_alpha = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.delete('-') + + pynutil.insert("\" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + decimal_dash_alpha = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.delete('-') + + pynutil.insert(" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + alpha_dash_cardinal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" cardinal { integer: \"") + + cardinal_graph + + pynutil.insert("\" } preserve_order: true") + ) + + alpha_dash_decimal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } preserve_order: true") + ) + + final_graph = ( + subgraph_decimal + | subgraph_cardinal + | subgraph_fraction + | cardinal_dash_alpha + | alpha_dash_cardinal + | decimal_dash_alpha + | decimal_times + | cardinal_times + | alpha_dash_decimal + ) + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/money.py b/nemo_text_processing/text_normalization/es/taggers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/money.py @@ -0,0 +1,194 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import decimal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + maj_singular_labels = load_labels(get_abs_path("data/money/currency_major.tsv")) + maj_singular = pynini.string_file((get_abs_path("data/money/currency_major.tsv"))) + min_singular = pynini.string_file(get_abs_path("data/money/currency_minor.tsv")) + fem_plural = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc_plural = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + maj_singular_labels = None + min_singular = None + maj_singular = None + fem_plural = None + masc_plural = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for classifying money, e.g. + "โ‚ฌ1" -> money { currency_maj: "euro" integer_part: "un"} + "โ‚ฌ1,000" -> money { currency_maj: "euro" integer_part: "un" } + "โ‚ฌ1,001" -> money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un" } + "ยฃ1,4" -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true } + -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "penique" preserve_order: true } + "0,01 ยฃ" -> money { fractional_part: "un" currency_min: "penique" preserve_order: true } + "0,02 ยฃ" -> money { fractional_part: "dos" currency_min: "peniques" preserve_order: true } + "ยฃ0,01 million" -> money { currency_maj: "libra" integer_part: "cero" fractional_part: "cero un" quantity: "million" } + + Args: + cardinal: CardinalFst + decimal: DecimalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + graph_decimal_final = decimal.final_graph_wo_negative + + maj_singular_graph = maj_singular + min_singular_graph = min_singular + maj_plural_graph = maj_singular @ (fem_plural | masc_plural) + min_plural_graph = min_singular @ (fem_plural | masc_plural) + + graph_maj_singular = pynutil.insert("currency_maj: \"") + maj_singular_graph + pynutil.insert("\"") + graph_maj_plural = pynutil.insert("currency_maj: \"") + maj_plural_graph + pynutil.insert("\"") + + graph_integer_one = pynutil.insert("integer_part: \"") + pynini.cross("1", "un") + pynutil.insert("\"") + + decimal_with_quantity = (NEMO_SIGMA + NEMO_ALPHA) @ graph_decimal_final + + graph_decimal_plural = pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, # 1,05 $ + ) + graph_decimal_plural = ( + (NEMO_SIGMA - "1") + decimal_separator + NEMO_SIGMA + ) @ graph_decimal_plural # Can't have "un euros" + + graph_decimal_singular = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, # 1,05 $ + ) + graph_decimal_singular = (pynini.accep("1") + decimal_separator + NEMO_SIGMA) @ graph_decimal_singular + + graph_decimal = pynini.union( + graph_decimal_singular, + graph_decimal_plural, + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + decimal_with_quantity, + ) + + graph_integer = ( + pynutil.insert("integer_part: \"") + ((NEMO_SIGMA - "1") @ cardinal_graph) + pynutil.insert("\"") + ) + + graph_integer_only = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer_one, + graph_integer_one + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, + ) + graph_integer_only |= pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer, + graph_integer + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, + ) + + graph = graph_integer_only | graph_decimal + + # remove trailing zeros of non zero number in the first 2 digits and fill up to 2 digits + # e.g. 2000 -> 20, 0200->02, 01 -> 01, 10 -> 10 + # not accepted: 002, 00, 0, + two_digits_fractional_part = ( + pynini.closure(NEMO_DIGIT) + (NEMO_DIGIT - "0") + pynini.closure(pynutil.delete("0")) + ) @ ( + (pynutil.delete("0") + (NEMO_DIGIT - "0")) + | ((NEMO_DIGIT - "0") + pynutil.insert("0")) + | ((NEMO_DIGIT - "0") + NEMO_DIGIT) + ) + + graph_min_singular = pynutil.insert("currency_min: \"") + min_singular_graph + pynutil.insert("\"") + graph_min_plural = pynutil.insert("currency_min: \"") + min_plural_graph + pynutil.insert("\"") + + # format ** euro ** cent + decimal_graph_with_minor = None + for curr_symbol, _ in maj_singular_labels: + preserve_order = pynutil.insert(" preserve_order: true") + + integer_plus_maj = pynini.union( + graph_integer + insert_space + pynutil.insert(curr_symbol) @ graph_maj_plural, + graph_integer_one + insert_space + pynutil.insert(curr_symbol) @ graph_maj_singular, + ) + # non zero integer part + integer_plus_maj = (pynini.closure(NEMO_DIGIT) - "0") @ integer_plus_maj + + graph_fractional_one = ( + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ pynini.cross("1", "un") + + pynutil.insert("\"") + ) + + graph_fractional = ( + two_digits_fractional_part @ (pynini.closure(NEMO_DIGIT, 1, 2) - "1") @ cardinal.two_digit_non_zero + ) + graph_fractional = pynutil.insert("fractional_part: \"") + graph_fractional + pynutil.insert("\"") + + fractional_plus_min = pynini.union( + graph_fractional + insert_space + pynutil.insert(curr_symbol) @ graph_min_plural, + graph_fractional_one + insert_space + pynutil.insert(curr_symbol) @ graph_min_singular, + ) + + decimal_graph_with_minor_curr = ( + integer_plus_maj + pynini.cross(decimal_separator, NEMO_SPACE) + fractional_plus_min + ) + decimal_graph_with_minor_curr |= pynutil.add_weight( + integer_plus_maj + + pynini.cross(decimal_separator, NEMO_SPACE) + + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ cardinal.two_digit_non_zero + + pynutil.insert("\""), + weight=0.0001, + ) + + decimal_graph_with_minor_curr |= pynutil.delete("0,") + fractional_plus_min + decimal_graph_with_minor_curr = pynini.union( + pynutil.delete(curr_symbol) + + pynini.closure(delete_space, 0, 1) + + decimal_graph_with_minor_curr + + preserve_order, + decimal_graph_with_minor_curr + + preserve_order + + pynini.closure(delete_space, 0, 1) + + pynutil.delete(curr_symbol), + ) + + decimal_graph_with_minor = ( + decimal_graph_with_minor_curr + if decimal_graph_with_minor is None + else pynini.union(decimal_graph_with_minor, decimal_graph_with_minor_curr) + ) + + final_graph = graph | pynutil.add_weight(decimal_graph_with_minor, -0.001) + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/ordinal.py b/nemo_text_processing/text_normalization/es/taggers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/ordinal.py @@ -0,0 +1,186 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import roman_to_int, strip_accent +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/digit.tsv"))) + teens = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/teen.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/twenties.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/ties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ImportError, ModuleNotFoundError): + digit = None + teens = None + twenties = None + ties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def get_one_to_one_thousand(cardinal: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Produces an acceptor for verbalizations of all numbers from 1 to 1000. Needed for ordinals and fractions. + + Args: + cardinal: CardinalFst + + Returns: + fst: A pynini.FstLike object + """ + numbers = pynini.string_map([str(_) for _ in range(1, 1000)]) @ cardinal + return pynini.project(numbers, "output").optimize() + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for classifying ordinal + "21.ยบ" -> ordinal { integer: "vigรฉsimo primero" morphosyntactic_features: "gender_masc" } + This class converts ordinal up to the millionth (millonรฉsimo) order (exclusive). + + This FST also records the ending of the ordinal (called "morphosyntactic_features"): + either as gender_masc, gender_fem, or apocope. Also introduces plural feature for non-deterministic graphs. + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="ordinal", kind="classify") + cardinal_graph = cardinal.graph + + graph_digit = digit.optimize() + graph_teens = teens.optimize() + graph_ties = ties.optimize() + graph_twenties = twenties.optimize() + graph_hundreds = hundreds.optimize() + + if not deterministic: + # Some alternative derivations + graph_ties = graph_ties | pynini.cross("sesenta", "setuagรฉsimo") + + graph_teens = graph_teens | pynini.cross("once", "decimoprimero") + graph_teens |= pynini.cross("doce", "decimosegundo") + + graph_digit = graph_digit | pynini.cross("nueve", "nono") + graph_digit |= pynini.cross("siete", "sรฉtimo") + + graph_tens_component = ( + graph_teens + | (graph_ties + pynini.closure(pynini.cross(" y ", NEMO_SPACE) + graph_digit, 0, 1)) + | graph_twenties + ) + + graph_hundred_component = pynini.union( + graph_hundreds + pynini.closure(NEMO_SPACE + pynini.union(graph_tens_component, graph_digit), 0, 1), + graph_tens_component, + graph_digit, + ) + + # Need to go up to thousands for fractions + self.one_to_one_thousand = get_one_to_one_thousand(cardinal_graph) + + thousands = pynini.cross("mil", "milรฉsimo") + + graph_thousands = ( + strip_accent(self.one_to_one_thousand) + NEMO_SPACE + thousands + ) # Cardinals become prefix for thousands series. Snce accent on the powers of ten we strip accent from leading words + graph_thousands @= pynini.cdrewrite(delete_space, "", "milรฉsimo", NEMO_SIGMA) # merge as a prefix + graph_thousands |= thousands + + self.multiples_of_thousand = (cardinal_graph @ graph_thousands).optimize() + + if ( + not deterministic + ): # Formally the words preceding the power of ten should be a prefix, but some maintain word boundaries. + graph_thousands |= (self.one_to_one_thousand @ graph_hundred_component) + NEMO_SPACE + thousands + + graph_thousands += pynini.closure(NEMO_SPACE + graph_hundred_component, 0, 1) + + ordinal_graph = graph_thousands | graph_hundred_component + ordinal_graph = cardinal_graph @ ordinal_graph + + if not deterministic: + # The 10's and 20's series can also be two words + split_words = pynini.cross("decimo", "dรฉcimo ") | pynini.cross("vigesimo", "vigรฉsimo ") + split_words = pynini.cdrewrite(split_words, "", NEMO_CHAR, NEMO_SIGMA) + ordinal_graph |= ordinal_graph @ split_words + + # If "octavo" is preceeded by the "o" within string, it needs deletion + ordinal_graph @= pynini.cdrewrite(pynutil.delete("o"), "", "octavo", NEMO_SIGMA) + + self.graph = ordinal_graph.optimize() + + masc = pynini.accep("gender_masc") + fem = pynini.accep("gender_fem") + apocope = pynini.accep("apocope") + + delete_period = pynini.closure(pynutil.delete("."), 0, 1) # Sometimes the period is omitted f + + accept_masc = delete_period + pynini.cross("ยบ", masc) + accep_fem = delete_period + pynini.cross("ยช", fem) + accep_apocope = delete_period + pynini.cross("แต‰สณ", apocope) + + # Managing Romanization + graph_roman = pynutil.insert("integer: \"") + roman_to_int(ordinal_graph) + pynutil.insert("\"") + if not deterministic: + # Introduce plural + plural = pynini.closure(pynutil.insert("/plural"), 0, 1) + accept_masc += plural + accep_fem += plural + + # Romanizations have no morphology marker, so in non-deterministic case we provide option for all + insert_morphology = pynutil.insert(pynini.union(masc, fem)) + plural + insert_morphology |= pynutil.insert(apocope) + insert_morphology = ( + pynutil.insert(" morphosyntactic_features: \"") + insert_morphology + pynutil.insert("\"") + ) + + graph_roman += insert_morphology + + else: + # We assume masculine gender as default + graph_roman += pynutil.insert(" morphosyntactic_features: \"gender_masc\"") + + # Rest of graph + convert_abbreviation = accept_masc | accep_fem | accep_apocope + + graph = ( + pynutil.insert("integer: \"") + + ordinal_graph + + pynutil.insert("\"") + + pynutil.insert(" morphosyntactic_features: \"") + + convert_abbreviation + + pynutil.insert("\"") + ) + graph = pynini.union(graph, graph_roman) + + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/telephone.py b/nemo_text_processing/text_normalization/es/taggers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/telephone.py @@ -0,0 +1,156 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.graph_utils import ones +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + graph_digit = pynini.string_file(get_abs_path("data/numbers/digit.tsv")) + graph_ties = pynini.string_file(get_abs_path("data/numbers/ties.tsv")) + graph_teen = pynini.string_file(get_abs_path("data/numbers/teen.tsv")) + graph_twenties = pynini.string_file(get_abs_path("data/numbers/twenties.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + graph_digit = None + graph_ties = None + graph_teen = None + graph_twenties = None + + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for classifying telephone numbers, e.g. + 123-123-5678 -> { number_part: "uno dos tres uno dos tres cinco seis siete ocho" }. + In Spanish, digits are generally read individually, or as 2-digit numbers, + eg. "123" = "uno dos tres", + "1234" = "doce treinta y cuatro". + This will verbalize sequences of 10 (3+3+4 e.g. 123-456-7890). + 9 (3+3+3 e.g. 123-456-789) and 8 (4+4 e.g. 1234-5678) digits. + + (we ignore more complicated cases such as "doscientos y dos" or "tres nueves"). + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="classify") + + # create `single_digits` and `double_digits` graphs as these will be + # the building blocks of possible telephone numbers + single_digits = pynini.invert(graph_digit).optimize() | pynini.cross("0", "cero") + + double_digits = pynini.union( + graph_twenties, + graph_teen, + (graph_ties + pynutil.delete("0")), + (graph_ties + insert_space + pynutil.insert("y") + insert_space + graph_digit), + ) + double_digits = pynini.invert(double_digits) + + # define `ten_digit_graph`, `nine_digit_graph`, `eight_digit_graph` + # which produces telephone numbers spoken (1) only with single digits, + # or (2) spoken with double digits (and sometimes single digits) + + # 10-digit option (1): all single digits + ten_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + # 9-digit option (1): all single digits + nine_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 2, 2) + + single_digits + ) + + # 8-digit option (1): all single digits + eight_digit_graph = ( + pynini.closure(single_digits + insert_space, 4, 4) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + if not deterministic: + # 10-digit option (2): (1+2) + (1+2) + (2+2) digits + ten_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + # 9-digit option (2): (1+2) + (1+2) + (1+2) digits + nine_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + ) + + # 8-digit option (2): (2+2) + (2+2) digits + eight_digit_graph |= ( + double_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + number_part = pynini.union(ten_digit_graph, nine_digit_graph, eight_digit_graph) + number_part @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", "", NEMO_SIGMA) + + number_part = pynutil.insert("number_part: \"") + number_part + pynutil.insert("\"") + + graph = number_part + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/time.py b/nemo_text_processing/text_normalization/es/taggers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/time.py @@ -0,0 +1,218 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + time_zone_graph = pynini.string_file(get_abs_path("data/time/time_zone.tsv")) + suffix = pynini.string_file(get_abs_path("data/time/time_suffix.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + time_zone_graph = None + suffix = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for classifying time, e.g. + "02:15 est" -> time { hours: "dos" minutes: "quince" zone: "e s t"} + "2 h" -> time { hours: "dos" } + "9 h" -> time { hours: "nueve" } + "02:15:10 h" -> time { hours: "dos" minutes: "quince" seconds: "diez"} + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="time", kind="classify", deterministic=deterministic) + + delete_time_delimiter = pynutil.delete(pynini.union(".", ":")) + + one = pynini.string_map([("un", "una"), ("รบn", "una")]) + change_one = pynini.cdrewrite(one, "", "", NEMO_SIGMA) + cardinal_graph = cardinal.graph @ change_one + + day_suffix = pynutil.insert("suffix: \"") + suffix + pynutil.insert("\"") + day_suffix = delete_space + insert_space + day_suffix + + delete_hora_suffix = delete_space + insert_space + pynutil.delete("h") + delete_minute_suffix = delete_space + insert_space + pynutil.delete("min") + delete_second_suffix = delete_space + insert_space + pynutil.delete("s") + + labels_hour_24 = [ + str(x) for x in range(0, 25) + ] # Can see both systems. Twelve hour requires am/pm for ambiguity resolution + labels_hour_12 = [str(x) for x in range(1, 13)] + labels_minute_single = [str(x) for x in range(1, 10)] + labels_minute_double = [str(x) for x in range(10, 60)] + + delete_leading_zero_to_double_digit = ( + pynini.closure(pynutil.delete("0") | (NEMO_DIGIT - "0"), 0, 1) + NEMO_DIGIT + ) + + graph_24 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_24) + ) + graph_12 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_12) + ) + + graph_hour_24 = graph_24 @ cardinal_graph + graph_hour_12 = graph_12 @ cardinal_graph + + graph_minute_single = delete_leading_zero_to_double_digit @ pynini.union(*labels_minute_single) + graph_minute_double = pynini.union(*labels_minute_double) + + graph_minute = pynini.union(graph_minute_single, graph_minute_double) @ cardinal_graph + + final_graph_hour_only_24 = ( + pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + delete_hora_suffix + ) + final_graph_hour_only_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + day_suffix + + final_graph_hour_24 = pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + final_graph_hour_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + + final_graph_minute = pynutil.insert("minutes: \"") + graph_minute + pynutil.insert("\"") + final_graph_second = pynutil.insert("seconds: \"") + graph_minute + pynutil.insert("\"") + final_time_zone_optional = pynini.closure( + delete_space + insert_space + pynutil.insert("zone: \"") + time_zone_graph + pynutil.insert("\""), 0, 1, + ) + + # 02.30 h + graph_hm = ( + final_graph_hour_24 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 h + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_24 + + delete_hora_suffix + + delete_space + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + delete_minute_suffix + + pynini.closure( + delete_space + + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)) + + delete_second_suffix, + 0, + 1, + ) # For seconds + + final_time_zone_optional + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_12 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 a. m. + + day_suffix + + final_time_zone_optional + ) + + graph_h = ( + pynini.union(final_graph_hour_only_24, final_graph_hour_only_12) + final_time_zone_optional + ) # Should always have a time indicator, else we'll pass to cardinals + + if not deterministic: + # This includes alternate vocalization (hour menos min, min para hour), here we shift the times and indicate a `style` tag + hour_shift_24 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_24.tsv"))) + hour_shift_12 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_12.tsv"))) + minute_shift = pynini.string_file(get_abs_path("data/time/minute_to.tsv")) + + graph_hour_to_24 = graph_24 @ hour_shift_24 @ cardinal_graph + graph_hour_to_12 = graph_12 @ hour_shift_12 @ cardinal_graph + + graph_minute_to = pynini.union(graph_minute_single, graph_minute_double) @ minute_shift @ cardinal_graph + + final_graph_hour_to_24 = pynutil.insert("hours: \"") + graph_hour_to_24 + pynutil.insert("\"") + final_graph_hour_to_12 = pynutil.insert("hours: \"") + graph_hour_to_12 + pynutil.insert("\"") + + final_graph_minute_to = pynutil.insert("minutes: \"") + graph_minute_to + pynutil.insert("\"") + + graph_menos = pynutil.insert(" style: \"1\"") + graph_para = pynutil.insert(" style: \"2\"") + + final_graph_style = graph_menos | graph_para + + # 02.30 h (omitting seconds since a bit awkward) + graph_hm |= ( + final_graph_hour_to_24 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + + final_graph_style + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_to_24 + + delete_hora_suffix + + delete_space + + insert_space + + final_graph_minute_to + + delete_minute_suffix + + final_time_zone_optional + + final_graph_style + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_to_12 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + day_suffix + + final_time_zone_optional + + final_graph_style + ) + + final_graph = graph_hm | graph_h + if deterministic: + final_graph = final_graph + pynutil.insert(" preserve_order: true") + final_graph = final_graph.optimize() + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py @@ -0,0 +1,157 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os + +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_space, + generator_main, +) +from nemo_text_processing.text_normalization.en.taggers.punctuation import PunctuationFst +from nemo_text_processing.text_normalization.es.taggers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.taggers.date import DateFst +from nemo_text_processing.text_normalization.es.taggers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.taggers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.taggers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.taggers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.taggers.money import MoneyFst +from nemo_text_processing.text_normalization.es.taggers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.taggers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.taggers.time import TimeFst +from nemo_text_processing.text_normalization.es.taggers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.taggers.word import WordFst + +from nemo.utils import logging + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class ClassifyFst(GraphFst): + """ + Final class that composes all other classification grammars. This class can process an entire sentence, that is lower cased. + For deployment, this grammar will be compiled and exported to OpenFst Finate State aRchive (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. + overwrite_cache: set to True to overwrite .far files + whitelist: path to a file with whitelist replacements + """ + + def __init__( + self, + input_case: str, + deterministic: bool = False, + cache_dir: str = None, + overwrite_cache: bool = False, + whitelist: str = None, + ): + super().__init__(name="tokenize_and_classify", kind="classify", deterministic=deterministic) + far_file = None + if cache_dir is not None and cache_dir != "None": + os.makedirs(cache_dir, exist_ok=True) + whitelist_file = os.path.basename(whitelist) if whitelist else "" + far_file = os.path.join( + cache_dir, f"_{input_case}_es_tn_{deterministic}_deterministic{whitelist_file}.far" + ) + if not overwrite_cache and far_file and os.path.exists(far_file): + self.fst = pynini.Far(far_file, mode="r")["tokenize_and_classify"] + logging.info(f"ClassifyFst.fst was restored from {far_file}.") + else: + logging.info(f"Creating ClassifyFst grammars. This might take some time...") + + self.cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = self.cardinal.fst + + self.ordinal = OrdinalFst(cardinal=self.cardinal, deterministic=deterministic) + ordinal_graph = self.ordinal.fst + + self.decimal = DecimalFst(cardinal=self.cardinal, deterministic=deterministic) + decimal_graph = self.decimal.fst + + self.fraction = FractionFst(cardinal=self.cardinal, ordinal=self.ordinal, deterministic=deterministic) + fraction_graph = self.fraction.fst + self.measure = MeasureFst( + cardinal=self.cardinal, decimal=self.decimal, fraction=self.fraction, deterministic=deterministic + ) + measure_graph = self.measure.fst + self.date = DateFst(cardinal=self.cardinal, deterministic=deterministic) + date_graph = self.date.fst + word_graph = WordFst(deterministic=deterministic).fst + self.time = TimeFst(self.cardinal, deterministic=deterministic) + time_graph = self.time.fst + self.telephone = TelephoneFst(deterministic=deterministic) + telephone_graph = self.telephone.fst + self.electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = self.electronic.fst + self.money = MoneyFst(cardinal=self.cardinal, decimal=self.decimal, deterministic=deterministic) + money_graph = self.money.fst + self.whitelist = WhiteListFst(input_case=input_case, deterministic=deterministic, input_file=whitelist) + whitelist_graph = self.whitelist.fst + punct_graph = PunctuationFst(deterministic=deterministic).fst + + classify = ( + pynutil.add_weight(whitelist_graph, 1.01) + | pynutil.add_weight(time_graph, 1.09) + | pynutil.add_weight(measure_graph, 1.08) + | pynutil.add_weight(cardinal_graph, 1.1) + | pynutil.add_weight(fraction_graph, 1.09) + | pynutil.add_weight(date_graph, 1.1) + | pynutil.add_weight(ordinal_graph, 1.1) + | pynutil.add_weight(decimal_graph, 1.1) + | pynutil.add_weight(money_graph, 1.1) + | pynutil.add_weight(telephone_graph, 1.1) + | pynutil.add_weight(electronic_graph, 1.1) + | pynutil.add_weight(word_graph, 200) + ) + punct = pynutil.insert("tokens { ") + pynutil.add_weight(punct_graph, weight=2.1) + pynutil.insert(" }") + punct = pynini.closure( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct), + 1, + ) + token = pynutil.insert("tokens { ") + classify + pynutil.insert(" }") + token_plus_punct = ( + pynini.closure(punct + pynutil.insert(" ")) + token + pynini.closure(pynutil.insert(" ") + punct) + ) + + graph = token_plus_punct + pynini.closure( + ( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct + pynutil.insert(" ")) + ) + + token_plus_punct + ) + + graph = delete_space + graph + delete_space + graph |= punct + + self.fst = graph.optimize() + + if far_file: + generator_main(far_file, {"tokenize_and_classify": self.fst}) + logging.info(f"ClassifyFst grammars are saved to {far_file}.") diff --git a/nemo_text_processing/text_normalization/es/taggers/whitelist.py b/nemo_text_processing/text_normalization/es/taggers/whitelist.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/whitelist.py @@ -0,0 +1,69 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, convert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WhiteListFst(GraphFst): + """ + Finite state transducer for classifying whitelist, e.g. + "sr." -> tokens { name: "seรฑor" } + This class has highest priority among all classifier grammars. Whitelisted tokens are defined and loaded from "data/whitelist.tsv". + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + input_file: path to a file with whitelist replacements + """ + + def __init__(self, input_case: str, deterministic: bool = True, input_file: str = None): + super().__init__(name="whitelist", kind="classify", deterministic=deterministic) + + def _get_whitelist_graph(input_case, file): + whitelist = load_labels(file) + if input_case == "lower_cased": + whitelist = [[x[0].lower()] + x[1:] for x in whitelist] + graph = pynini.string_map(whitelist) + return graph + + graph = _get_whitelist_graph(input_case, get_abs_path("data/whitelist.tsv")) + if not deterministic and input_case != "lower_cased": + graph |= pynutil.add_weight( + _get_whitelist_graph("lower_cased", get_abs_path("data/whitelist.tsv")), weight=0.0001 + ) + + if input_file: + whitelist_provided = _get_whitelist_graph(input_case, input_file) + if not deterministic: + graph |= whitelist_provided + else: + graph = whitelist_provided + + if not deterministic: + units_graph = _get_whitelist_graph(input_case, file=get_abs_path("data/measures/measurements.tsv")) + graph |= units_graph + + self.graph = graph + self.final_graph = convert_space(self.graph).optimize() + self.fst = (pynutil.insert("name: \"") + self.final_graph + pynutil.insert("\"")).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/word.py b/nemo_text_processing/text_normalization/es/taggers/word.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/word.py @@ -0,0 +1,39 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_SPACE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WordFst(GraphFst): + """ + Finite state transducer for classifying word. + e.g. dormir -> tokens { name: "dormir" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="word", kind="classify") + word = pynutil.insert("name: \"") + pynini.closure(NEMO_NOT_SPACE, 1) + pynutil.insert("\"") + self.fst = word.optimize() diff --git a/nemo_text_processing/text_normalization/es/utils.py b/nemo_text_processing/text_normalization/es/utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/utils.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import csv +import os + + +def get_abs_path(rel_path): + """ + Get absolute path + + Args: + rel_path: relative path to this file + + Returns absolute path + """ + return os.path.dirname(os.path.abspath(__file__)) + '/' + rel_path + + +def load_labels(abs_path): + """ + loads relative path file as dictionary + + Args: + abs_path: absolute path + + Returns dictionary of mappings + """ + label_tsv = open(abs_path) + labels = list(csv.reader(label_tsv, delimiter="\t")) + return labels diff --git a/nemo_text_processing/text_normalization/es/verbalizers/__init__.py b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py @@ -0,0 +1,57 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_cardinal_gender, strip_cardinal_apocope + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class CardinalFst(GraphFst): + """ + Finite state transducer for verbalizing cardinals + e.g. cardinal { integer: "dos" } -> "dos" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="verbalize", deterministic=deterministic) + optional_sign = pynini.closure(pynini.cross("negative: \"true\" ", "menos "), 0, 1) + self.optional_sign = optional_sign + + integer = pynini.closure(NEMO_NOT_QUOTE, 1) + self.integer = pynutil.delete(" \"") + integer + pynutil.delete("\"") + + integer = pynutil.delete("integer:") + self.integer + self.numbers = integer + graph = optional_sign + self.numbers + + if not deterministic: + # For alternate renderings + no_adjust = graph + fem_adjust = shift_cardinal_gender(graph) + apocope_adjust = strip_cardinal_apocope(graph) + graph = no_adjust | fem_adjust | apocope_adjust + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/date.py b/nemo_text_processing/text_normalization/es/verbalizers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/date.py @@ -0,0 +1,86 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.taggers.date import articles + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for verbalizing date, e.g. + date { day: "treinta y uno" month: "marzo" year: "dos mil" } -> "treinta y uno de marzo de dos mil" + date { day: "uno" month: "mayo" year: "del mil novecientos noventa" } -> "primero de mayo del mil novecientos noventa" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="date", kind="verbalize", deterministic=deterministic) + + day_cardinal = pynutil.delete("day: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + day = strip_cardinal_apocope(day_cardinal) + + primero = pynini.cdrewrite(pynini.cross("uno", "primero"), "[BOS]", "[EOS]", NEMO_SIGMA) + day = ( + (day @ primero) if deterministic else pynini.union(day, day @ primero) + ) # Primero for first day is traditional, but will vary depending on region + + month = pynutil.delete("month: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + year = ( + pynutil.delete("year: \"") + + articles + + NEMO_SPACE + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # Insert preposition if wasn't originally with the year. This would mean a space was present + year = pynutil.add_weight(year, -0.001) + year |= ( + pynutil.delete("year: \"") + + pynutil.insert("de ") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # day month year + graph_dmy = day + pynini.cross(NEMO_SPACE, " de ") + month + pynini.closure(pynini.accep(" ") + year, 0, 1) + + graph_mdy = month + NEMO_SPACE + day + pynini.closure(NEMO_SPACE + year, 0, 1) + if deterministic: + graph_mdy += pynutil.delete(" preserve_order: true") # Only accepts this if was explicitly passed + + self.graph = graph_dmy | graph_mdy + final_graph = self.graph + delete_preserve_order + + delete_tokens = self.delete_tokens(final_graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/decimals.py b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py @@ -0,0 +1,87 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + decimal { negative: "true" integer_part: "dos" fractional_part: "cuatro cero" quantity: "billones" } -> menos dos coma quatro cero billones + decimal { integer_part: "un" quantity: "billรณn" } -> un billรณn + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + + self.optional_sign = pynini.closure(pynini.cross("negative: \"true\"", "menos ") + delete_space, 0, 1) + self.integer = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + self.fractional_default = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + conjunction = pynutil.insert(" punto ") if LOCALIZATION == "am" else pynutil.insert(" coma ") + if not deterministic: + conjunction |= pynutil.insert(pynini.union(" con ", " y ")) + self.fractional_default |= strip_cardinal_apocope(self.fractional_default) + self.fractional = conjunction + self.fractional_default + + self.quantity = ( + delete_space + + insert_space + + pynutil.delete("quantity: \"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + self.optional_quantity = pynini.closure(self.quantity, 0, 1) + + graph = self.optional_sign + pynini.union( + (self.integer + self.quantity), (self.integer + delete_space + self.fractional + self.optional_quantity) + ) + + self.numbers = graph.optimize() + self.numbers_no_quantity = self.integer + delete_space + self.fractional + self.optional_quantity + + if not deterministic: + graph |= self.optional_sign + ( + shift_cardinal_gender(self.integer + delete_space) + shift_number_gender(self.fractional) + ) + + graph += delete_preserve_order + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/electronic.py b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py @@ -0,0 +1,91 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit_no_zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + graph_symbols = pynini.string_file(get_abs_path("data/electronic/symbols.tsv")) + server_common = pynini.string_file(get_abs_path("data/electronic/server_name.tsv")) + domain_common = pynini.string_file(get_abs_path("data/electronic/domain.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digit_no_zero = None + zero = None + + graph_symbols = None + server_common = None + domain_common = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for verbalizing electronic + e.g. electronic { username: "abc" domain: "hotmail.com" } -> "a b c arroba hotmail punto com" + -> "a b c arroba h o t m a i l punto c o m" + -> "a b c arroba hotmail punto c o m" + -> "a b c at h o t m a i l punto com" + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="verbalize", deterministic=deterministic) + + graph_digit_no_zero = ( + digit_no_zero @ pynini.cdrewrite(pynini.cross("un", "uno"), "", "", NEMO_SIGMA).optimize() + ) + graph_digit = graph_digit_no_zero | zero + + def add_space_after_char(): + return pynini.closure(NEMO_NOT_QUOTE - pynini.accep(" ") + insert_space) + ( + NEMO_NOT_QUOTE - pynini.accep(" ") + ) + + verbalize_characters = pynini.cdrewrite(graph_symbols | graph_digit, "", "", NEMO_SIGMA) + + user_name = pynutil.delete("username: \"") + add_space_after_char() + pynutil.delete("\"") + user_name @= verbalize_characters + + convert_defaults = pynutil.add_weight(NEMO_NOT_QUOTE, weight=0.0001) | domain_common | server_common + domain = convert_defaults + pynini.closure(insert_space + convert_defaults) + domain @= verbalize_characters + + domain = pynutil.delete("domain: \"") + domain + pynutil.delete("\"") + protocol = ( + pynutil.delete("protocol: \"") + + add_space_after_char() @ pynini.cdrewrite(graph_symbols, "", "", NEMO_SIGMA) + + pynutil.delete("\"") + ) + self.graph = (pynini.closure(protocol + pynini.accep(" "), 0, 1) + domain) | ( + user_name + pynini.accep(" ") + pynutil.insert("arroba ") + domain + ) + delete_tokens = self.delete_tokens(self.graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/fraction.py b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_NOT_QUOTE, + NEMO_NOT_SPACE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + accents, + shift_cardinal_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for verbalizing fraction + e.g. tokens { fraction { integer: "treinta y tres" numerator: "cuatro" denominator: "quinto" } } -> + treinta y tres y cuatro quintos + + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="fraction", kind="verbalize", deterministic=deterministic) + + # Derivational strings append 'avo' as a suffix. Adding space for processing aid + fraction_stem = pynutil.insert(" avo") + plural = pynutil.insert("s") + + integer = ( + pynutil.delete("integer_part: \"") + + strip_cardinal_apocope(pynini.closure(NEMO_NOT_QUOTE)) + + pynutil.delete("\"") + ) + + numerator_one = pynutil.delete("numerator: \"") + pynini.accep("un") + pynutil.delete("\" ") + numerator = ( + pynutil.delete("numerator: \"") + + pynini.difference(pynini.closure(NEMO_NOT_QUOTE), "un") + + pynutil.delete("\" ") + ) + + denominator_add_stem = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + + fraction_stem + + pynutil.delete("\" morphosyntactic_features: \"add_root\"") + ) + denominator_ordinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\" morphosyntactic_features: \"ordinal\"") + ) + denominator_cardinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + ) + + denominator_singular = pynini.union(denominator_add_stem, denominator_ordinal) + denominator_plural = denominator_singular + plural + + if not deterministic: + # Occasional exceptions + denominator_singular |= denominator_add_stem @ pynini.string_map( + [("once avo", "undรฉcimo"), ("doce avo", "duodรฉcimo")] + ) + + # Merging operations + merge = pynini.cdrewrite( + pynini.cross(" y ", "i"), "", "", NEMO_SIGMA + ) # The denominator must be a single word, with the conjunction "y" replaced by i + merge @= pynini.cdrewrite(delete_space, "", pynini.difference(NEMO_CHAR, "parte"), NEMO_SIGMA) + + # The merger can produce duplicate vowels. This is not allowed in orthography + delete_duplicates = pynini.string_map([("aa", "a"), ("oo", "o")]) # Removes vowels + delete_duplicates = pynini.cdrewrite(delete_duplicates, "", "", NEMO_SIGMA) + + remove_accents = pynini.cdrewrite( + accents, + pynini.union(NEMO_SPACE, pynini.accep("[BOS]")) + pynini.closure(NEMO_NOT_SPACE), + pynini.closure(NEMO_NOT_SPACE) + pynini.union("avo", "ava", "รฉsimo", "รฉsima"), + NEMO_SIGMA, + ) + merge_into_single_word = merge @ remove_accents @ delete_duplicates + + fraction_default = numerator + delete_space + insert_space + (denominator_plural @ merge_into_single_word) + fraction_with_one = ( + numerator_one + delete_space + insert_space + (denominator_singular @ merge_into_single_word) + ) + + fraction_with_cardinal = strip_cardinal_apocope(numerator | numerator_one) + fraction_with_cardinal += ( + delete_space + pynutil.insert(" sobre ") + strip_cardinal_apocope(denominator_cardinal) + ) + + conjunction = pynutil.insert(" y ") + + if not deterministic: + # There is an alternative rendering where ordinals act as adjectives for 'parte'. This requires use of the feminine + # Other rules will manage use of "un" at end, so just worry about endings + exceptions = pynini.string_map([("tercia", "tercera")]) + apply_exceptions = pynini.cdrewrite(exceptions, "", "", NEMO_SIGMA) + vowel_change = pynini.cdrewrite(pynini.cross("o", "a"), "", pynini.accep("[EOS]"), NEMO_SIGMA) + + denominator_singular_fem = shift_cardinal_gender(denominator_singular) @ vowel_change @ apply_exceptions + denominator_plural_fem = denominator_singular_fem + plural + + numerator_one_fem = shift_cardinal_gender(numerator_one) + numerator_fem = shift_cardinal_gender(numerator) + + fraction_with_cardinal |= ( + (numerator_one_fem | numerator_fem) + + delete_space + + pynutil.insert(" sobre ") + + shift_cardinal_gender(denominator_cardinal) + ) + + # Still need to manage stems + merge_stem = pynini.cdrewrite( + delete_space, "", pynini.union("avo", "ava", "avos", "avas"), NEMO_SIGMA + ) # For managing alternative spacing + merge_stem @= remove_accents @ delete_duplicates + + fraction_with_one_fem = numerator_one_fem + delete_space + insert_space + fraction_with_one_fem += pynini.union( + denominator_singular_fem @ merge_stem, denominator_singular_fem @ merge_into_single_word + ) # Both forms exists + fraction_with_one_fem @= pynini.cdrewrite(pynini.cross("una media", "media"), "", "", NEMO_SIGMA) + fraction_with_one_fem += pynutil.insert(" parte") + + fraction_default_fem = numerator_fem + delete_space + insert_space + fraction_default_fem += pynini.union( + denominator_plural_fem @ merge_stem, denominator_plural_fem @ merge_into_single_word + ) + fraction_default_fem += pynutil.insert(" partes") + + fraction_default |= ( + numerator + delete_space + insert_space + denominator_plural @ merge_stem + ) # Case of no merger + fraction_default |= fraction_default_fem + + fraction_with_one |= numerator_one + delete_space + insert_space + denominator_singular @ merge_stem + fraction_with_one |= fraction_with_one_fem + + # Integers are influenced by dominant noun, need to allow feminine forms as well + integer |= shift_cardinal_gender(integer) + + # Remove 'un medio' + fraction_with_one @= pynini.cdrewrite(pynini.cross("un medio", "medio"), "", "", NEMO_SIGMA) + + integer = pynini.closure(integer + delete_space + conjunction, 0, 1) + + fraction = fraction_with_one | fraction_default | fraction_with_cardinal + + graph = integer + fraction + + self.graph = graph + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/measure.py b/nemo_text_processing/text_normalization/es/verbalizers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/measure.py @@ -0,0 +1,110 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ones, shift_cardinal_gender +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + unit_singular_fem = pynini.project(unit_plural_fem, "input") + unit_singular_masc = pynini.project(unit_plural_masc, "input") + + unit_plural_fem = pynini.project(unit_plural_fem, "output") + unit_plural_masc = pynini.project(unit_plural_masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit_plural_fem = None + unit_plural_masc = None + + unit_singular_fem = None + unit_singular_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for verbalizing measure, e.g. + measure { cardinal { integer: "dos" units: "gramos" } } -> "dos gramos" + measure { cardinal { integer_part: "dos" quantity: "millones" units: "gramos" } } -> "dos millones de gramos" + + Args: + decimal: DecimalFst + cardinal: CardinalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, cardinal: GraphFst, fraction: GraphFst, deterministic: bool): + super().__init__(name="measure", kind="verbalize", deterministic=deterministic) + + graph_decimal = decimal.fst + graph_cardinal = cardinal.fst + graph_fraction = fraction.fst + + unit_masc = (unit_plural_masc | unit_singular_masc) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_masc |= "por" + pynini.closure(NEMO_NOT_QUOTE, 1) + unit_masc = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_masc) + pynutil.delete("\"") + + unit_fem = (unit_plural_fem | unit_singular_fem) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_fem = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_fem) + pynutil.delete("\"") + + graph_masc = (graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_masc + graph_fem = ( + shift_cardinal_gender(graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_fem + ) + graph = graph_masc | graph_fem + + graph = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph + ) # billones de xyz + + graph @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", NEMO_WHITE_SPACE + "por", NEMO_SIGMA) + + # To manage alphanumeric combonations ("a-8, 5x"), we let them use a weighted default path. + alpha_num_unit = pynutil.delete("units: \"") + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + graph_alpha_num = pynini.union( + (graph_cardinal | graph_decimal) + NEMO_SPACE + alpha_num_unit, + alpha_num_unit + delete_extra_space + (graph_cardinal | graph_decimal), + ) + + graph |= pynutil.add_weight(graph_alpha_num, 0.01) + + graph += delete_preserve_order + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/money.py b/nemo_text_processing/text_normalization/es/verbalizers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/money.py @@ -0,0 +1,195 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + fem = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + fem_singular = pynini.project(fem, "input") + masc_singular = pynini.project(masc, "input") + + fem_plural = pynini.project(fem, "output") + masc_plural = pynini.project(masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + fem_plural = None + masc_plural = None + + fem_singular = None + masc_singular = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for verbalizing money, e.g. + money { currency_maj: "euro" integer_part: "un"} -> "un euro" + money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un"} -> "uno coma cero cero uno euros" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true} -> "una libra cuarenta" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "peniques" preserve_order: true} -> "una libra con cuarenta peniques" + money { fractional_part: "un" currency_min: "penique" preserve_order: true} -> "un penique" + + Args: + decimal: GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="verbalize", deterministic=deterministic) + + maj_singular_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + maj_singular_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + maj_plural_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + maj_plural_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + maj_masc = maj_plural_masc | maj_singular_masc # Tagger kept quantity resolution stable + maj_fem = maj_plural_fem | maj_singular_fem + + min_singular_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + min_singular_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + min_plural_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + min_plural_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + min_masc = min_plural_masc | min_singular_masc + min_fem = min_plural_fem | min_singular_fem + + fractional_part = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + integer_part = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + optional_add_and = pynini.closure(pynutil.insert(pynini.union("con ", "y ")), 0, 1) + + # *** currency_maj + graph_integer_masc = integer_part + NEMO_SPACE + maj_masc + graph_integer_fem = shift_cardinal_gender(integer_part) + NEMO_SPACE + maj_fem + graph_integer = graph_integer_fem | graph_integer_masc + + # *** currency_maj + (***) | ((con) *** current_min) + graph_integer_with_minor_masc = ( + integer_part + + NEMO_SPACE + + maj_masc + + NEMO_SPACE + + pynini.union( + optional_add_and + strip_cardinal_apocope(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor_fem = ( + shift_cardinal_gender(integer_part) + + NEMO_SPACE + + maj_fem + + NEMO_SPACE + + pynini.union( + optional_add_and + shift_cardinal_gender(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor = graph_integer_with_minor_fem | graph_integer_with_minor_masc + + # *** coma *** currency_maj + graph_decimal_masc = decimal.numbers + NEMO_SPACE + maj_masc + + # Need to fix some of the inner parts, so don't use decimal here (note: quantities covered by masc) + graph_decimal_fem = ( + pynini.accep("integer_part: \"") + + shift_cardinal_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SPACE + + pynini.accep("fractional_part: \"") + + shift_number_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SIGMA + ) + graph_decimal_fem @= decimal.numbers_no_quantity + graph_decimal_fem += NEMO_SPACE + maj_fem + + graph_decimal = graph_decimal_fem | graph_decimal_masc + graph_decimal = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph_decimal + ) # formally it's millones/billones de *** + + # *** current_min + graph_minor_masc = fractional_part + NEMO_SPACE + min_masc + delete_preserve_order + graph_minor_fem = shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem + delete_preserve_order + graph_minor = graph_minor_fem | graph_minor_masc + + graph = graph_integer | graph_integer_with_minor | graph_decimal | graph_minor + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py @@ -0,0 +1,76 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, NEMO_SIGMA, NEMO_SPACE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_number_gender + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for verbalizing ordinals + e.g. ordinal { integer: "tercer" } } -> "tercero" + -> "tercera" + -> "tercer" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="ordinal", kind="verbalize", deterministic=deterministic) + + graph = pynutil.delete("integer: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + # masculne gender we leave as is + graph_masc = graph + pynutil.delete(" morphosyntactic_features: \"gender_masc") + + # shift gender + graph_fem_ending = graph @ pynini.cdrewrite( + pynini.cross("o", "a"), "", NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fem = shift_number_gender(graph_fem_ending) + pynutil.delete(" morphosyntactic_features: \"gender_fem") + + # Apocope just changes tercero and primero. May occur if someone wrote 11.er (uncommon) + graph_apocope = ( + pynini.cross("tercero", "tercer") + | pynini.cross("primero", "primer") + | pynini.cross("undรฉcimo", "decimoprimer") + ) # In case someone wrote 11.er with deterministic + graph_apocope = (graph @ pynini.cdrewrite(graph_apocope, "", "", NEMO_SIGMA)) + pynutil.delete( + " morphosyntactic_features: \"apocope" + ) + + graph = graph_apocope | graph_masc | graph_fem + + if not deterministic: + # Plural graph + graph_plural = pynini.cdrewrite( + pynutil.insert("s"), pynini.union("o", "a"), NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + + graph |= (graph @ graph_plural) + pynutil.delete("/plural") + + self.graph = graph + pynutil.delete("\"") + + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/telephone.py b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for verbalizing telephone, e.g. + telephone { number_part: "uno dos tres uno dos tres cinco seis siete ocho" } + -> uno dos tres uno dos tres cinco seis siete ocho + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="verbalize") + + number_part = pynutil.delete("number_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + delete_tokens = self.delete_tokens(number_part) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/time.py b/nemo_text_processing/text_normalization/es/verbalizers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/time.py @@ -0,0 +1,269 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + alt_minutes = pynini.string_file(get_abs_path("data/time/alt_minutes.tsv")) + + morning_times = pynini.string_file(get_abs_path("data/time/morning_times.tsv")) + afternoon_times = pynini.string_file(get_abs_path("data/time/afternoon_times.tsv")) + evening_times = pynini.string_file(get_abs_path("data/time/evening_times.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + alt_minutes = None + + morning_times = None + afternoon_times = None + evening_times = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for verbalizing time, e.g. + time { hours: "doce" minutes: "media" suffix: "a m" } -> doce y media de la noche + time { hours: "doce" } -> twelve o'clock + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="time", kind="verbalize", deterministic=deterministic) + + change_minutes = pynini.cdrewrite(alt_minutes, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA) + + morning_phrases = pynini.cross("am", "de la maรฑana") + afternoon_phrases = pynini.cross("pm", "de la tarde") + evening_phrases = pynini.cross("pm", "de la noche") + + # For the 12's + mid_times = pynini.accep("doce") + mid_phrases = ( + pynini.string_map([("pm", "del mediodรญa"), ("am", "de la noche")]) + if deterministic + else pynini.string_map( + [ + ("pm", "de la maรฑana"), + ("pm", "del dรญa"), + ("pm", "del mediodรญa"), + ("am", "de la noche"), + ("am", "de la medianoche"), + ] + ) + ) + + hour = ( + pynutil.delete("hours:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = ( + pynutil.delete("minutes:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = (minute @ change_minutes) if deterministic else pynini.union(minute, minute @ change_minutes) + + suffix = ( + pynutil.delete("suffix:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + zone = ( + pynutil.delete("zone:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + optional_zone = pynini.closure(delete_space + insert_space + zone, 0, 1) + second = ( + pynutil.delete("seconds:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + graph_hms = ( + hour + + pynutil.insert(" horas ") + + delete_space + + minute + + pynutil.insert(" minutos y ") + + delete_space + + second + + pynutil.insert(" segundos") + ) + + graph_hm = hour + delete_space + pynutil.insert(" y ") + minute + graph_hm |= pynini.union( + (hour @ morning_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases), + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases), + (hour @ evening_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases), + (hour @ mid_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases), + ) + + graph_h = pynini.union( + hour, + (hour @ morning_times) + delete_space + insert_space + (suffix @ morning_phrases), + (hour @ afternoon_times) + delete_space + insert_space + (suffix @ afternoon_phrases), + (hour @ evening_times) + delete_space + insert_space + (suffix @ evening_phrases), + (hour @ mid_times) + delete_space + insert_space + (suffix @ mid_phrases), + ) + + graph = (graph_hms | graph_hm | graph_h) + optional_zone + + if not deterministic: + graph_style_1 = pynutil.delete(" style: \"1\"") + graph_style_2 = pynutil.delete(" style: \"2\"") + + graph_menos = hour + delete_space + pynutil.insert(" menos ") + minute + graph_style_1 + graph_menos |= ( + (hour @ morning_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ evening_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ mid_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_1 + ) + graph_menos += optional_zone + + graph_para = minute + pynutil.insert(" para las ") + delete_space + hour + graph_style_2 + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ morning_times) + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ afternoon_times) + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ evening_times) + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ mid_times) + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_2 + ) + graph_para += optional_zone + graph_para @= pynini.cdrewrite( + pynini.cross(" las ", " la "), "para", "una", NEMO_SIGMA + ) # Need agreement with one + + graph |= graph_menos | graph_para + delete_tokens = self.delete_tokens(graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py @@ -0,0 +1,73 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst +from nemo_text_processing.text_normalization.en.verbalizers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.verbalizers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.verbalizers.date import DateFst +from nemo_text_processing.text_normalization.es.verbalizers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.verbalizers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.verbalizers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.verbalizers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.verbalizers.money import MoneyFst +from nemo_text_processing.text_normalization.es.verbalizers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.verbalizers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.verbalizers.time import TimeFst + + +class VerbalizeFst(GraphFst): + """ + Composes other verbalizer grammars. + For deployment, this grammar will be compiled and exported to OpenFst Finate State Archiv (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize", kind="verbalize", deterministic=deterministic) + cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = cardinal.fst + ordinal = OrdinalFst(deterministic=deterministic) + ordinal_graph = ordinal.fst + decimal = DecimalFst(deterministic=deterministic) + decimal_graph = decimal.fst + fraction = FractionFst(deterministic=deterministic) + fraction_graph = fraction.fst + date = DateFst(deterministic=deterministic) + date_graph = date.fst + measure = MeasureFst(cardinal=cardinal, decimal=decimal, fraction=fraction, deterministic=deterministic) + measure_graph = measure.fst + electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = electronic.fst + whitelist_graph = WhiteListFst(deterministic=deterministic).fst + money_graph = MoneyFst(decimal=decimal, deterministic=deterministic).fst + telephone_graph = TelephoneFst(deterministic=deterministic).fst + time_graph = TimeFst(deterministic=deterministic).fst + + graph = ( + cardinal_graph + | measure_graph + | decimal_graph + | ordinal_graph + | date_graph + | electronic_graph + | money_graph + | fraction_graph + | whitelist_graph + | telephone_graph + | time_graph + ) + self.fst = graph diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py @@ -0,0 +1,52 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, delete_extra_space, delete_space +from nemo_text_processing.text_normalization.en.verbalizers.word import WordFst +from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class VerbalizeFinalFst(GraphFst): + """ + Finite state transducer that verbalizes an entire sentence + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize_final", kind="verbalize", deterministic=deterministic) + verbalize = VerbalizeFst(deterministic=deterministic).fst + word = WordFst(deterministic=deterministic).fst + types = verbalize | word + graph = ( + pynutil.delete("tokens") + + delete_space + + pynutil.delete("{") + + delete_space + + types + + delete_space + + pynutil.delete("}") + ) + graph = delete_space + pynini.closure(graph + delete_extra_space) + graph + delete_space + self.fst = graph diff --git a/nemo_text_processing/text_normalization/normalize.py b/nemo_text_processing/text_normalization/normalize.py --- a/nemo_text_processing/text_normalization/normalize.py +++ b/nemo_text_processing/text_normalization/normalize.py @@ -46,8 +46,8 @@ class Normalizer: """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -83,10 +83,11 @@ def __init__( from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst elif lang == 'de': - # Ru TN only support non-deterministic cases and produces multiple normalization options - # use normalize_with_audio.py from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst + elif lang == 'es': + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ClassifyFst + from nemo_text_processing.text_normalization.es.verbalizers.verbalize_final import VerbalizeFinalFst self.tagger = ClassifyFst( input_case=input_case, deterministic=deterministic, @@ -106,7 +107,7 @@ def __init__( def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: """ - NeMo text normalizer + NeMo text normalizer Args: texts: list of input strings @@ -357,7 +358,7 @@ def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: def parse_args(): parser = ArgumentParser() parser.add_argument("input_string", help="input string", type=str) - parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) + parser.add_argument("--language", help="language", choices=["en", "de", "es"], default="en", type=str) parser.add_argument( "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) diff --git a/nemo_text_processing/text_normalization/normalize_with_audio.py b/nemo_text_processing/text_normalization/normalize_with_audio.py --- a/nemo_text_processing/text_normalization/normalize_with_audio.py +++ b/nemo_text_processing/text_normalization/normalize_with_audio.py @@ -55,15 +55,15 @@ "audio_data" - path to the audio file "text" - raw text "pred_text" - ASR model prediction - + See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions - + When the manifest is ready, run: python normalize_with_audio.py \ --audio_data PATH/TO/MANIFEST.JSON \ - --language en - - + --language en + + To run with a single audio file, specify path to audio and text with: python normalize_with_audio.py \ --audio_data PATH/TO/AUDIO.WAV \ @@ -71,18 +71,18 @@ --text raw text OR PATH/TO/.TXT/FILE --model QuartzNet15x5Base-En \ --verbose - + To see possible normalization options for a text input without an audio file (could be used for debugging), run: python python normalize_with_audio.py --text "RAW TEXT" - + Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference """ class NormalizerWithAudio(Normalizer): """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -282,7 +282,7 @@ def parse_args(): "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) parser.add_argument( - "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str + "--language", help="Select target language", choices=["en", "ru", "de", "es"], default="en", type=str ) parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") parser.add_argument( diff --git a/tools/text_processing_deployment/pynini_export.py b/tools/text_processing_deployment/pynini_export.py --- a/tools/text_processing_deployment/pynini_export.py +++ b/tools/text_processing_deployment/pynini_export.py @@ -67,7 +67,7 @@ def tn_grammars(**kwargs): def export_grammars(output_dir, grammars): """ - Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. + Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. Args: output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. @@ -109,7 +109,7 @@ def parse_args(): if __name__ == '__main__': args = parse_args() - if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': + if args.language in ['ru', 'fr', 'vi'] and args.grammars == 'tn_grammars': raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') if args.language == 'en': @@ -148,6 +148,10 @@ def parse_args(): from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ( + ClassifyFst as TNClassifyFst, + ) + from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst elif args.language == 'fr': from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, </patch> </s> </patch>
diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt @@ -0,0 +1,86 @@ +1~un +2~dos +3~tres +4~cuatro +5~cinco +6~seis +7~siete +8~ocho +9~nueve +10~diez +11~once +12~doce +13~trece +14~catorce +15~quince +16~diecisรฉis +17~diecisiete +18~dieciocho +19~diecinueve +20~veinte +21~veintiรบn +22~veintidรณs +23~veintitrรฉs +24~veinticuatro +25~veinticinco +26~veintisรฉis +27~veintisiete +28~veintiocho +29~veintinueve +30~treinta +31~treinta y un +40~cuarenta +41~cuarenta y un +50~cincuenta +51~cincuenta y un +60~sesenta +70~setenta +80~ochenta +90~noventa +100~cien +101~ciento un +120~ciento veinte +121~ciento veintiรบn +130~ciento treinta +131~ciento treinta y un +200~doscientos +201~doscientos un +300~trescientos +301~trescientos un +1000~mil +1 000~mil +1.000~mil +1001~mil un +1010~mil diez +1020~mil veinte +1021~mil veintiรบn +1100~mil cien +1101~mil ciento un +1110~mil ciento diez +1111~mil ciento once +1234~mil doscientos treinta y cuatro +2000~dos mil +2001~dos mil un +2010~dos mil diez +2020~dos mil veinte +2100~dos mil cien +2101~dos mil ciento un +2110~dos mil ciento diez +2111~dos mil ciento once +2222~dos mil doscientos veintidรณs +10000~diez mil +10 000~diez mil +10.000~diez mil +100000~cien mil +100 000~cien mil +100.000~cien mil +1 000 000~un millรณn +1.000.000~un millรณn +1 234 568~un millรณn doscientos treinta y cuatro mil quinientos sesenta y ocho +2.000.000~dos millones +1.000.000.000~mil millones +2.000.000.000~dos mil millones +3 000 000 000 000~tres billones +3.000.000.000.000~tres billones +100 000 000 000 000 000 000 000~cien mil trillones +100 000 000 000 000 000 000 001~cien mil trillones un diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt @@ -0,0 +1,13 @@ +1 enero~primero de enero +5 febrero~cinco de febrero +20 de marzo~veinte de marzo +abril 30~treinta de abril +31 marzo~treinta y uno de marzo +10 mayo 1990~diez de mayo de mil novecientos noventa +junio 11 2000~once de junio de dos mil +30 julio del 2020~treinta de julio del dos mil veinte +30-2-1990~treinta de febrero de mil novecientos noventa +30/2/1990~treinta de febrero de mil novecientos noventa +30.2.1990~treinta de febrero de mil novecientos noventa +1990-2-30~treinta de febrero de mil novecientos noventa +1990-02-30~treinta de febrero de mil novecientos noventa \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt @@ -0,0 +1,27 @@ +0,1~cero coma un +0,01~cero coma cero un +0,010~cero coma cero uno cero +1,0101~uno coma cero uno cero un +0,0~cero coma cero +1,0~uno coma cero +1,00~uno coma cero cero +1,1~uno coma un +233,32~doscientos treinta y tres coma treinta y dos +32,22 millones~treinta y dos coma veintidรณs millones +320 320,22 millones~trescientos veinte mil trescientos veinte coma veintidรณs millones +5.002,232~cinco mil dos coma doscientos treinta y dos +3,2 trillones~tres coma dos trillones +3 millones~tres millones +3 000 millones~tres mil millones +3000 millones~tres mil millones +3.000 millones~tres mil millones +3.001 millones~tres mil un millones +1 millรณn~un millรณn +1 000 millones~mil millones +1000 millones~mil millones +1.000 millones~mil millones +2,33302 millones~dos coma tres tres tres cero dos millones +1,5332 millรณn~uno coma cinco tres tres dos millรณn +1,53322 millรณn~uno coma cinco tres tres dos dos millรณn +1,53321 millรณn~uno coma cinco tres tres dos un millรณn +101,010101 millones~ciento uno coma cero uno cero uno cero un millones \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt @@ -0,0 +1,12 @@ +a.bc@gmail.com~a punto b c arroba gmail punto com +cdf@abc.edu~c d f arroba a b c punto e d u +abc@gmail.abc~a b c arroba gmail punto a b c +abc@abc.com~a b c arroba a b c punto com +asdf123@abc.com~a s d f uno dos tres arroba a b c punto com +a1b2@abc.com~a uno b dos arroba a b c punto com +ab3.sdd.3@gmail.com~a b tres punto s d d punto tres arroba gmail punto com +https://www.nvidia.com~h t t p s dos puntos barra barra w w w punto nvidia punto com +www.nvidia.com~w w w punto nvidia punto com +www.abc.es/efg~w w w punto a b c punto es barra e f g +www.abc.es~w w w punto a b c punto es +http://www.ourdailynews.com.sm~h t t p dos puntos barra barra w w w punto o u r d a i l y n e w s punto com punto s m \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt @@ -0,0 +1,76 @@ +1/2~medio +1 1/2~uno y medio +3/2~tres medios +1 3/2~uno y tres medios +1/3~un tercio +2/3~dos tercios +1/4~un cuarto +2/4~dos cuartos +1/5~un quinto +2/5~dos quintos +1/6~un sexto +2/6~dos sextos +1/7~un sรฉptimo +2/7~dos sรฉptimos +1/8~un octavo +2/8~dos octavos +1/9~un noveno +2/9~dos novenos +1/10~un dรฉcimo +2/10~dos dรฉcimos +1/11~un onceavo +1/12~un doceavo +1/13~un treceavo +1/14~un catorceavo +1/15~un quinceavo +1/16~un dieciseisavo +1/17~un diecisieteavo +1/18~un dieciochoavo +1/19~un diecinueveavo +1/20~un veinteavo +1/21~un veintiunavo +1/22~un veintidosavo +1/30~un treintavo +1/31~un treintaiunavo +1/40~un cuarentavo +1/41~un cuarentaiunavo +1/50~un cincuentavo +1/60~un sesentavo +1/70~un setentavo +1/80~un ochentavo +1/90~un noventavo +1/100~un centรฉsimo +2/100~dos centรฉsimos +1 2/100~uno y dos centรฉsimos +1/101~uno sobre ciento uno +1/110~uno sobre ciento diez +1/111~uno sobre ciento once +1/112~uno sobre ciento doce +1/123~uno sobre ciento veintitrรฉs +1/134~uno sobre ciento treinta y cuatro +1/200~un ducentรฉsimo +1/201~uno sobre doscientos uno +1/234~uno sobre doscientos treinta y cuatro +1/300~un tricentรฉsimo +1/345~uno sobre trescientos cuarenta y cinco +1/400~un cuadringentรฉsimo +1/456~uno sobre cuatrocientos cincuenta y seis +1/500~un quingentรฉsimo +1/600~un sexcentรฉsimo +1/700~un septingentรฉsimo +1/800~un octingentรฉsimo +1/900~un noningentรฉsimo +1/1000~un milรฉsimo +2/1000~dos milรฉsimos +1 2/1000~uno y dos milรฉsimos +1/1001~uno sobre mil uno +1/1100~uno sobre mil cien +1/1200~uno sobre mil doscientos +1/1234~uno sobre mil doscientos treinta y cuatro +1/2000~un dosmilรฉsimo +1/5000~un cincomilรฉsimo +1/10000~un diezmilรฉsimo +1/100.000~un cienmilรฉsimo +1/1.000.000~un millonรฉsimo +1/100.000.000~un cienmillonรฉsimo +1/1.200.000.000~un mildoscientosmillonรฉsimo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt @@ -0,0 +1,17 @@ +1,2-a~uno coma dos a +a-5~a cinco +200 m~doscientos metros +3 h~tres horas +1 h~una hora +245 mph~doscientas cuarenta y cinco millas por hora +2 kg~dos kilogramos +60,2400 kg~sesenta coma dos cuatro cero cero kilogramos +-60,2400 kg~menos sesenta coma dos cuatro cero cero kilogramos +8,52 %~ocho coma cincuenta y dos por ciento +-8,52 %~menos ocho coma cincuenta y dos por ciento +1 %~uno por ciento +3 cm~tres centรญmetros +4 s~cuatro segundos +5 l~cinco litros +4,51/s~cuatro coma cincuenta y uno por segundo +0,0101 s~cero coma cero uno cero un segundos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt @@ -0,0 +1,24 @@ +$1~un dรณlar +1 $~un dรณlar +$1,50~un dรณlar cincuenta centavos +1,50 $~un dรณlar cincuenta centavos +ยฃ200.000.001~doscientos millones una libras +200.000.001 ยฃ~doscientos millones una libras +2 billones de euros~dos billones de euros +โ‚ฌ2 billones~dos billones de euros +โ‚ฌ 2 billones~dos billones de euros +โ‚ฌ 2,3 billones~dos coma tres billones de euros +2,3 billones de euros~dos coma tres billones de euros +โ‚ฌ5,50~cinco euros cincuenta cรฉntimos +5,50 โ‚ฌ~cinco euros cincuenta cรฉntimos +5,01 โ‚ฌ~cinco euros un cรฉntimo +5,01 ยฃ~cinco libras un penique +21 czk~veintiuna coronas checas +czk21~veintiuna coronas checas +czk21,1 millones~veintiuna coma una millones de coronas checas +czk 5,50 billones~cinco coma cincuenta billones de coronas checas +rs 5,50 billones~cinco coma cincuenta billones de rupias +czk5,50 billones~cinco coma cincuenta billones de coronas checas +0,55 $~cincuenta y cinco centavos +1,01 $~un dรณlar un centavo +ยฅ12,05~doce yenes cinco centavos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt @@ -0,0 +1,120 @@ +~121 +ciento veintiรบn +ciento veintiuno +ciento veintiuna +121 +~200 +doscientos +doscientas +200 +~201 +doscientos un +doscientos uno +doscientas una +201 +~1 +un +uno +una +1 +~550.000.001 +quinientos cincuenta millones un +quinientos cincuenta millones una +quinientos cincuenta millones uno +550.000.001 +~500.501 +quinientos mil quinientos un +quinientos mil quinientos uno +quinientas mil quinientas una +500.501 +~500.001.ยบ +quinientosmilรฉsimo primero +quingentรฉsimo milรฉsimo primero +quinientosmilรฉsimos primeros +quingentรฉsimos milรฉsimos primeros +500.001.ยบ +~500.001.ยช +quinientasmilรฉsima primera +quingentรฉsima milรฉsima primera +quinientasmilรฉsimas primeras +quingentรฉsimas milรฉsimas primeras +500.001.ยช +~11.ยช +dรฉcima primera +decimoprimera +dรฉcimas primeras +decimoprimeras +undรฉcima +undรฉcimas +11.ยช +~11.ยบ +dรฉcimo primero +decimoprimero +dรฉcimos primeros +decimoprimeros +undรฉcimo +undรฉcimos +11.ยบ +~12.ยบ +dรฉcimo segundo +decimosegundo +dรฉcimos segundos +decimosegundos +duodรฉcimo +duodรฉcimos +12.ยบ +~200,0101 +doscientos coma cero uno cero un +doscientos coma cero uno cero uno +doscientas coma cero una cero una +200,0101 +~1.000.200,21 +un millรณn doscientos coma veintiรบn +un millรณn doscientos coma veintiuno +un millรณn doscientas coma veintiuna +un millรณn doscientos coma dos un +un millรณn doscientos coma dos uno +un millรณn doscientas coma dos una +1.000.200,21 +~1/12 +un doceavo +una doceava parte +un duodรฉcimo +una duodรฉcima parte +uno sobre doce +1/12 +~5/200 +cinco ducentรฉsimos +cinco ducentรฉsimas partes +cinco sobre doscientos +5/200 +~1 5/3 +uno y cinco tercios +una y cinco terceras partes +uno y cinco sobre tres +una y cinco sobre tres +~1/5/2020 +primero de mayo de dos mil veinte +uno de mayo de dos mil veinte +cinco de enero de dos mil veinte +~$5,50 +cinco dรณlares con cincuenta +cinco dรณlares y cincuenta +cinco dรณlares cincuenta +cinco dรณlares con cincuenta centavos +cinco dรณlares y cincuenta centavos +cinco dรณlares cincuenta centavos +~2.30 h +dos y treinta +dos y media +tres menos treinta +tres menos media +treinta para las tres +~12.30 a.m. +doce y treinta de la medianoche +doce y treinta de la noche +doce y media de la medianoche +doce y media de la noche +una menos treinta de la maรฑana +una menos media de la maรฑana +treinta para la una de la maรฑana \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt @@ -0,0 +1,137 @@ +1.แต‰สณ~primer +1.ยบ~primero +1.ยช~primera +2.ยบ~segundo +2.ยช~segunda +ii~segundo +II~segundo +3.แต‰สณ~tercer +3.ยบ~tercero +3.ยช~tercera +4.ยบ~cuarto +4.ยช~cuarta +5.ยบ~quinto +5.ยช~quinta +6.ยบ~sexto +6.ยช~sexta +7.ยบ~sรฉptimo +7.ยช~sรฉptima +8.ยบ~octavo +8.ยช~octava +9.ยบ~noveno +9.ยช~novena +10.ยบ~dรฉcimo +10.ยช~dรฉcima +11.แต‰สณ~decimoprimer +11.ยบ~undรฉcimo +11.ยช~undรฉcima +12.ยบ~duodรฉcimo +12.ยช~duodรฉcima +13.แต‰สณ~decimotercer +13.ยบ~decimotercero +13.ยช~decimotercera +14.ยบ~decimocuarto +14.ยช~decimocuarta +15.ยบ~decimoquinto +15.ยช~decimoquinta +16.ยบ~decimosexto +16.ยช~decimosexta +17.ยบ~decimosรฉptimo +17.ยช~decimosรฉptima +18.ยบ~decimoctavo +18.ยช~decimoctava +19.ยบ~decimonoveno +19.ยช~decimonovena +20.ยบ~vigรฉsimo +20.ยช~vigรฉsima +21.แต‰สณ~vigesimoprimer +21.ยบ~vigesimoprimero +21.ยช~vigesimoprimera +30.ยบ~trigรฉsimo +30.ยช~trigรฉsima +31.แต‰สณ~trigรฉsimo primer +31.ยบ~trigรฉsimo primero +31.ยช~trigรฉsima primera +40.ยบ~cuadragรฉsimo +40.ยช~cuadragรฉsima +41.แต‰สณ~cuadragรฉsimo primer +41.ยบ~cuadragรฉsimo primero +41.ยช~cuadragรฉsima primera +50.ยบ~quincuagรฉsimo +50.ยช~quincuagรฉsima +51.แต‰สณ~quincuagรฉsimo primer +51.ยบ~quincuagรฉsimo primero +51.ยช~quincuagรฉsima primera +60.ยบ~sexagรฉsimo +60.ยช~sexagรฉsima +70.ยบ~septuagรฉsimo +70.ยช~septuagรฉsima +80.ยบ~octogรฉsimo +80.ยช~octogรฉsima +90.ยบ~nonagรฉsimo +90.ยช~nonagรฉsima +100.ยบ~centรฉsimo +100.ยช~centรฉsima +101.แต‰สณ~centรฉsimo primer +101.ยบ~centรฉsimo primero +101.ยช~centรฉsima primera +134.ยบ~centรฉsimo trigรฉsimo cuarto +134.ยช~centรฉsima trigรฉsima cuarta +200.ยบ~ducentรฉsimo +200.ยช~ducentรฉsima +300.ยบ~tricentรฉsimo +300.ยช~tricentรฉsima +400.ยบ~cuadringentรฉsimo +400.ยช~cuadringentรฉsima +500.ยบ~quingentรฉsimo +500.ยช~quingentรฉsima +600.ยบ~sexcentรฉsimo +600.ยช~sexcentรฉsima +700.ยบ~septingentรฉsimo +700.ยช~septingentรฉsima +800.ยบ~octingentรฉsimo +800.ยช~octingentรฉsima +900.ยบ~noningentรฉsimo +900.ยช~noningentรฉsima +1000.ยบ~milรฉsimo +1000.ยช~milรฉsima +1001.แต‰สณ~milรฉsimo primer +1 000.ยบ~milรฉsimo +1 000.ยช~milรฉsima +1 001.แต‰สณ~milรฉsimo primer +1.000.ยบ~milรฉsimo +1.000.ยช~milรฉsima +1.001.แต‰สณ~milรฉsimo primer +1248.ยบ~milรฉsimo ducentรฉsimo cuadragรฉsimo octavo +1248.ยช~milรฉsima ducentรฉsima cuadragรฉsima octava +2000.ยบ~dosmilรฉsimo +100 000.ยบ~cienmilรฉsimo +i~primero +I~primero +ii~segundo +II~segundo +iii~tercero +III~tercero +iv~cuarto +IV~cuarto +V~quinto +VI~sexto +VII~sรฉptimo +VIII~octavo +IX~noveno +X~dรฉcimo +XI~undรฉcimo +XII~duodรฉcimo +XIII~decimotercero +XX~vigรฉsimo +XXI~vigesimoprimero +XXX~trigรฉsimo +XL~cuadragรฉsimo +L~quincuagรฉsimo +XC~nonagรฉsimo +C~centรฉsimo +CD~cuadringentรฉsimo +D~quingentรฉsimo +CM~noningentรฉsimo +999.ยบ~noningentรฉsimo nonagรฉsimo noveno +cmxcix~noningentรฉsimo nonagรฉsimo noveno \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt @@ -0,0 +1,3 @@ +123-123-5678~uno dos tres uno dos tres cinco seis siete ocho +123-456-789~uno dos tres cuatro cinco seis siete ocho nueve +1234-5678~uno dos tres cuatro cinco seis siete ocho \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt @@ -0,0 +1,26 @@ +1.00~una +1:00~una +01:00~una +01 h~una +3 h~tres horas +1 h~una hora +1.05 h~una y cinco +01.05 h~una y cinco +1.00 h~una +1.00 a.m.~una de la maรฑana +1.00 a.m~una de la maรฑana +1.00 p.m.~una de la tarde +1.00 p.m est~una de la tarde e s t +1.00 est~una e s t +5:02 est~cinco y dos e s t +5:02 p.m pst~cinco y dos de la noche p s t +5:02 p.m.~cinco y dos de la noche +12.15~doce y cuarto +12.15 a.m.~doce y cuarto de la noche +12.15 p.m.~doce y cuarto del mediodรญa +13.30~trece y media +14.05~catorce y cinco +24:50~veinticuatro y cincuenta +3:02:32 pst~tres horas dos minutos y treinta y dos segundos p s t +00:52~cero y cincuenta y dos +0:52~cero y cincuenta y dos diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt @@ -0,0 +1,3 @@ +el dr.~el doctor +sr. rodriguez~seรฑor rodriguez +182 esq. toledo~ciento ochenta y dos esquina toledo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt @@ -0,0 +1,48 @@ +~ +yahoo!~yahoo! +veinte!~veinte! +โ€”~โ€” +aaa~aaa +aabach~aabach +aabenraa~aabenraa +aabye~aabye +aaccessed~aaccessed +aach~aach +aachen's~aachen's +aadri~aadri +aafia~aafia +aagaard~aagaard +aagadu~aagadu +aagard~aagard +aagathadi~aagathadi +aaghart's~aaghart's +aagnes~aagnes +aagomoni~aagomoni +aagon~aagon +aagoo~aagoo +aagot~aagot +aahar~aahar +aahh~aahh +aahperd~aahperd +aaibinterstate~aaibinterstate +aajab~aajab +aakasa~aakasa +aakervik~aakervik +aakirkeby~aakirkeby +aalam~aalam +aalbaek~aalbaek +aaldiu~aaldiu +aalem~aalem +a'ali~a'ali +aalilaassamthey~aalilaassamthey +aalin~aalin +aaliyan~aaliyan +aaliyan's~aaliyan's +aamadu~aamadu +aamara~aamara +aambala~aambala +aamera~aamera +aamer's~aamer's +aamina~aamina +aaminah~aaminah +aamjiwnaang~aamjiwnaang diff --git a/tests/nemo_text_processing/es/test_cardinal.py b/tests/nemo_text_processing/es/test_cardinal.py --- a/tests/nemo_text_processing/es/test_cardinal.py +++ b/tests/nemo_text_processing/es/test_cardinal.py @@ -22,7 +22,8 @@ class TestCardinal: - inverse_normalizer_es = ( + + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +33,34 @@ class TestCardinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_cardinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_date.py b/tests/nemo_text_processing/es/test_date.py --- a/tests/nemo_text_processing/es/test_date.py +++ b/tests/nemo_text_processing/es/test_date.py @@ -22,7 +22,7 @@ class TestDate: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDate: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_date.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_decimal.py b/tests/nemo_text_processing/es/test_decimal.py --- a/tests/nemo_text_processing/es/test_decimal.py +++ b/tests/nemo_text_processing/es/test_decimal.py @@ -22,7 +22,7 @@ class TestDecimal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDecimal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_decimal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_electronic.py b/tests/nemo_text_processing/es/test_electronic.py --- a/tests/nemo_text_processing/es/test_electronic.py +++ b/tests/nemo_text_processing/es/test_electronic.py @@ -35,3 +35,31 @@ class TestElectronic: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_electronic.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_fraction.py b/tests/nemo_text_processing/es/test_fraction.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_fraction.py @@ -0,0 +1,51 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import pytest +from nemo_text_processing.text_normalization.normalize import Normalizer +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, parse_test_case_file + + +class TestFraction: + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_fraction.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_measure.py b/tests/nemo_text_processing/es/test_measure.py --- a/tests/nemo_text_processing/es/test_measure.py +++ b/tests/nemo_text_processing/es/test_measure.py @@ -36,3 +36,31 @@ class TestMeasure: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_measure.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_money.py b/tests/nemo_text_processing/es/test_money.py --- a/tests/nemo_text_processing/es/test_money.py +++ b/tests/nemo_text_processing/es/test_money.py @@ -23,7 +23,7 @@ class TestMoney: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,34 @@ class TestMoney: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_money.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_normalization_with_audio.py b/tests/nemo_text_processing/es/test_normalization_with_audio.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_normalization_with_audio.py @@ -0,0 +1,43 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import pytest +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, get_test_cases_multiple + + +class TestNormalizeWithAudio: + + normalizer_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + @parameterized.expand(get_test_cases_multiple('es/data_text_normalization/test_cases_normalize_with_audio.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, n_tagged=1000, punct_post_process=False) + print(expected) + print("pred") + print(pred) + assert len(set(pred).intersection(set(expected))) == len( + expected + ), f'missing: {set(expected).difference(set(pred))}' diff --git a/tests/nemo_text_processing/es/test_ordinal.py b/tests/nemo_text_processing/es/test_ordinal.py --- a/tests/nemo_text_processing/es/test_ordinal.py +++ b/tests/nemo_text_processing/es/test_ordinal.py @@ -23,7 +23,7 @@ class TestOrdinal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,33 @@ class TestOrdinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_ordinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=30, punct_post_process=False, + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh @@ -0,0 +1,84 @@ +#! /bin/sh + +PROJECT_DIR=/workspace/tests + +runtest () { + input=$1 + cd /workspace/sparrowhawk/documentation/grammars + + # read test file + while read testcase; do + IFS='~' read written spoken <<< $testcase + denorm_pred=$(echo $written | normalizer_main --config=sparrowhawk_configuration.ascii_proto 2>&1 | tail -n 1) + + # trim white space + spoken="$(echo -e "${spoken}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + denorm_pred="$(echo -e "${denorm_pred}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + + # input expected actual + assertEquals "$written" "$spoken" "$denorm_pred" + done < "$input" +} + +testTNCardinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_cardinal.txt + runtest $input +} + +testTNDate() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_date.txt + runtest $input +} + +testTNDecimal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_decimal.txt + runtest $input +} + +testTNElectronic() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_electronic.txt + runtest $input +} + +testTNFraction() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_fraction.txt + runtest $input +} + +testTNMoney() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_money.txt + runtest $input +} + +testTNOrdinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTelephone() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTime() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_time.txt + runtest $input +} + +testTNMeasure() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_measure.txt + runtest $input +} + +testTNWhitelist() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_whitelist.txt + runtest $input +} + +testTNWord() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_word.txt + runtest $input +} + +# Load shUnit2 +. $PROJECT_DIR/../shunit2/shunit2 diff --git a/tests/nemo_text_processing/es/test_telephone.py b/tests/nemo_text_processing/es/test_telephone.py --- a/tests/nemo_text_processing/es/test_telephone.py +++ b/tests/nemo_text_processing/es/test_telephone.py @@ -36,3 +36,31 @@ class TestTelephone: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_telephone.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_time.py b/tests/nemo_text_processing/es/test_time.py --- a/tests/nemo_text_processing/es/test_time.py +++ b/tests/nemo_text_processing/es/test_time.py @@ -35,3 +35,31 @@ class TestTime: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_time.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_whitelist.py b/tests/nemo_text_processing/es/test_whitelist.py --- a/tests/nemo_text_processing/es/test_whitelist.py +++ b/tests/nemo_text_processing/es/test_whitelist.py @@ -35,3 +35,30 @@ class TestWhitelist: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_whitelist.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=10, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_word.py b/tests/nemo_text_processing/es/test_word.py --- a/tests/nemo_text_processing/es/test_word.py +++ b/tests/nemo_text_processing/es/test_word.py @@ -35,3 +35,30 @@ class TestWord: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer_es = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_word.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, verbose=False) + assert pred == expected, f"input: {test_input}" + + if self.normalizer_with_audio_es: + pred_non_deterministic = self.normalizer_with_audio_es.normalize( + test_input, n_tagged=150, punct_post_process=False + ) + assert expected in pred_non_deterministic, f"input: {test_input}"
1.0
NVIDIA__NeMo-7582
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| .. |status| image:: http://www.repostatus.org/badges/latest/active.svg :target: http://www.repostatus.org/#active :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main :alt: Documentation :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE :alt: NeMo core license and license for collections in this repo .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg :target: https://badge.fury.io/py/nemo-toolkit :alt: Release version .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg :target: https://badge.fury.io/py/nemo-toolkit :alt: Python version .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads :target: https://pepy.tech/project/nemo-toolkit :alt: PyPi total downloads .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml :alt: CodeQL .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg :target: https://github.com/psf/black :alt: Code style: black .. _main-readme: **NVIDIA NeMo** =============== Introduction ------------ NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), text-to-speech synthesis (TTS), large language models (LLMs), and natural language processing (NLP). The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and training is automatically scalable to 1000s of GPUs. Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. Getting started with NeMo is simple. State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that can all be run on `Google Colab <https://colab.research.google.com>`_. For advanced users that want to train NeMo models from scratch or finetune existing NeMo models we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ which can be used to find the optimal model parallel configuration for training on a specific cluster. Also see our `introductory video <https://www.youtube.com/embed/wBgpMf_KQVw>`_ for a high level overview of NeMo. Key Features ------------ * Speech processing * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ * Jasper, QuartzNet, CitriNet, ContextNet * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer * Squeezeformer-CTC and Squeezeformer-Transducer * LSTM-Transducer (RNNT) and LSTM-CTC * Supports the following decoders/losses: * CTC * Transducer/RNNT * Hybrid Transducer/CTC * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ * Beam Search decoding * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ * Natural Language Processing * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ * Text-to-Speech Synthesis (TTS): * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 * Vocoders: HiFiGAN, UnivNet, WaveGlow * End-to-End Models: VITS * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. Requirements ------------ 1) Python 3.10 or above 2) Pytorch 1.13.1 or above 3) NVIDIA GPU for training Documentation ------------- .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main :alt: Documentation Status :scale: 100% :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable :alt: Documentation Status :scale: 100% :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ | Version | Status | Description | +=========+=============+==========================================================================================================================================+ | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ Tutorials --------- A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. Getting help with NeMo ---------------------- FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. Installation ------------ Conda ~~~~~ We recommend installing NeMo in a fresh Conda environment. .. code-block:: bash conda create --name nemo python==3.10.12 conda activate nemo Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. .. code-block:: bash conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. Pip ~~~ Use this installation mode if you want the latest released version. .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg pip install Cython pip install nemo_toolkit['all'] Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. Pip from source ~~~~~~~~~~~~~~~ Use this installation mode if you want the version from a particular GitHub branch (e.g main). .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg pip install Cython python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] From source ~~~~~~~~~~~ Use this installation mode if you are contributing to NeMo. .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg git clone https://github.com/NVIDIA/NeMo cd NeMo ./reinstall.sh If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` with ``pip install -e .`` when your PWD is the root of the NeMo repository. RNNT ~~~~ Note that RNNT requires numba to be installed from conda. .. code-block:: bash conda remove numba pip uninstall numba conda install -c conda-forge numba NeMo Megatron ~~~~~~~~~~~~~ NeMo Megatron training requires NVIDIA Apex to be installed. Install it manually if not using the NVIDIA PyTorch container. To install Apex, run .. code-block:: bash git clone https://github.com/NVIDIA/apex.git cd apex git checkout 52e18c894223800cb611682dce27d88050edf1de pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: .. code-block:: bash conda install -c nvidia cuda-nvprof=11.8 packaging is also needed: .. code-block:: bash pip install packaging With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. Transformer Engine ~~~~~~~~~~~~~~~~~~ NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. .. code-block:: bash pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. Transformer Engine requires PyTorch to be built with CUDA 11.8. Flash Attention ~~~~~~~~~~~~~~~~~~~~ Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. .. code-block:: bash pip install flash-attn pip install triton==2.0.0.dev20221202 NLP inference UI ~~~~~~~~~~~~~~~~~~~~ To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. .. code-block:: bash pip install gradio==3.34.0 NeMo Text Processing ~~~~~~~~~~~~~~~~~~~~ NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. Docker containers: ~~~~~~~~~~~~~~~~~~ We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. To use built container, please run .. code-block:: bash docker pull nvcr.io/nvidia/nemo:23.06 To build a nemo container with Dockerfile from a branch, please run .. code-block:: bash DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. .. code-block:: bash docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 Examples -------- Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. Contributing ------------ We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. Publications ------------ We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! License ------- NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. [end of README.rst] [start of examples/asr/experimental/k2/align_speech_parallel.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ # Based on examples/asr/transcribe_speech_parallel.py # ASR alignment with multi-GPU/multi-node support for large datasets # It supports both tarred and non-tarred datasets # Arguments # model: path to a nemo/PTL checkpoint file or name of a pretrained model # predict_ds: config of the dataset/dataloader # aligner_args: aligner config # output_path: path to store the predictions # model_stride: model downsampling factor, 8 for Citrinet models and 4 for Conformer models # # Results of each GPU/worker is written into a file named 'predictions_{rank}.json, and aggregated results of all workers are written into 'predictions_all.json' Example for non-tarred datasets: python align_speech_parallel.py \ model=stt_en_conformer_ctc_large \ predict_ds.manifest_filepath=/dataset/manifest_file.json \ predict_ds.batch_size=16 \ output_path=/tmp/ Example for tarred datasets: python align_speech_parallel.py \ predict_ds.is_tarred=true \ predict_ds.manifest_filepath=/tarred_dataset/tarred_audio_manifest.json \ predict_ds.tarred_audio_filepaths=/tarred_dataset/audio__OP_0..127_CL_.tar \ ... By default the trainer uses all the GPUs available and default precision is FP32. By setting the trainer config you may control these configs. For example to do the predictions with AMP on just two GPUs: python align_speech_parallel.py \ trainer.precision=16 \ trainer.gpus=2 \ ... You may control the dataloader's config by setting the predict_ds: python align_speech_parallel.py \ predict_ds.num_workers=8 \ predict_ds.min_duration=2.0 \ predict_ds.sample_rate=16000 \ model=stt_en_conformer_ctc_small \ ... You may control the aligner's config by setting the aligner_args: aligner_args.alignment_type=argmax \ aligner_args.word_output=False \ aligner_args.cpu_decoding=True \ aligner_args.decode_batch_size=8 \ aligner_args.ctc_cfg.prob_suppress_index=-1 \ aligner_args.ctc_cfg.prob_suppress_value=0.5 \ aligner_args.rnnt_cfg.predictor_window_size=10 \ aligner_args.decoder_module_cfg.intersect_pruned=true \ aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ ... """ import os from dataclasses import dataclass, is_dataclass from typing import Optional import pytorch_lightning as ptl import torch from omegaconf import MISSING, OmegaConf from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter from nemo.collections.asr.models import ASRModel from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel from nemo.core.config import TrainerConfig, hydra_runner from nemo.utils import logging from nemo.utils.get_rank import is_global_rank_zero @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() output_path: str = MISSING model_stride: int = 8 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") # there arguments will be ignored return_predictions: bool = False use_cer: bool = False def match_train_config(predict_ds, train_ds): # It copies the important configurations from the train dataset of the model # into the predict_ds to be used for prediction. It is needed to match the training configurations. if train_ds is None: return predict_ds.sample_rate = train_ds.get("sample_rate", 16000) cfg_name_list = [ "int_values", "use_start_end_token", "blank_index", "unk_index", "normalize", "parser", "eos_id", "bos_id", "pad_id", ] if is_dataclass(predict_ds): predict_ds = OmegaConf.structured(predict_ds) for cfg_name in cfg_name_list: if hasattr(train_ds, cfg_name): setattr(predict_ds, cfg_name, getattr(train_ds, cfg_name)) return predict_ds @hydra_runner(config_name="AlignmentConfig", schema=ParallelAlignmentConfig) def main(cfg: ParallelAlignmentConfig): if cfg.model.endswith(".nemo"): logging.info("Attempting to initialize from .nemo file") model = ASRModel.restore_from(restore_path=cfg.model, map_location="cpu") elif cfg.model.endswith(".ckpt"): logging.info("Attempting to initialize from .ckpt file") model = ASRModel.load_from_checkpoint(checkpoint_path=cfg.model, map_location="cpu") else: logging.info( "Attempting to initialize from a pretrained model as the model name does not have the extension of .nemo or .ckpt" ) model = ASRModel.from_pretrained(model_name=cfg.model, map_location="cpu") trainer = ptl.Trainer(**cfg.trainer) cfg.predict_ds.return_sample_id = True cfg.return_predictions = False cfg.use_cer = False cfg.predict_ds = match_train_config(predict_ds=cfg.predict_ds, train_ds=model._cfg.train_ds) data_loader = model._setup_dataloader_from_config(cfg.predict_ds) os.makedirs(cfg.output_path, exist_ok=True) # trainer.global_rank is not valid before predict() is called. Need this hack to find the correct global_rank. global_rank = trainer.node_rank * trainer.num_devices + int(os.environ.get("LOCAL_RANK", 0)) output_file = os.path.join(cfg.output_path, f"predictions_{global_rank}.json") output_ctm_dir = os.path.join(cfg.output_path, "ctm") predictor_writer = ASRCTMPredictionWriter( dataset=data_loader.dataset, output_file=output_file, output_ctm_dir=output_ctm_dir, time_per_frame=cfg.model_stride * model._cfg.preprocessor['window_stride'], ) trainer.callbacks.extend([predictor_writer]) aligner_wrapper = AlignerWrapperModel(model=model, cfg=cfg.aligner_args) trainer.predict(model=aligner_wrapper, dataloaders=data_loader, return_predictions=cfg.return_predictions) samples_num = predictor_writer.close_output_file() logging.info( f"Prediction on rank {global_rank} is done for {samples_num} samples and results are stored in {output_file}." ) if torch.distributed.is_initialized(): torch.distributed.barrier() samples_num = 0 if is_global_rank_zero(): output_file = os.path.join(cfg.output_path, f"predictions_all.json") logging.info(f"Prediction files are being aggregated in {output_file}.") with open(output_file, 'tw', encoding="utf-8") as outf: for rank in range(trainer.world_size): input_file = os.path.join(cfg.output_path, f"predictions_{rank}.json") with open(input_file, 'r', encoding="utf-8") as inpf: lines = inpf.readlines() samples_num += len(lines) outf.writelines(lines) logging.info( f"Prediction is done for {samples_num} samples in total on all workers and results are aggregated in {output_file}." ) if __name__ == '__main__': main() [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import re from abc import abstractmethod from dataclasses import dataclass, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance import numpy as np import torch from omegaconf import OmegaConf from torchmetrics import Metric from nemo.collections.asr.metrics.wer import move_dimension_to_the_front from nemo.collections.asr.parts.submodules import rnnt_beam_decoding as beam_decode from nemo.collections.asr.parts.submodules import rnnt_greedy_decoding as greedy_decode from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses from nemo.utils import logging __all__ = ['RNNTDecoding', 'RNNTWER'] class AbstractRNNTDecoding(ConfidenceMixin): """ Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy, greedy_batch (for greedy decoding). - beam, tsd, alsd (for beam search decoding). compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded tokens as well as the decoded string. Default is False in order to avoid double decoding unless required. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function with the `return_hypotheses` flag set to True. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrete intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. rnnt_timestamp_type: A str value, which represents the types of timestamps that should be calculated. Can take the following values - "char" for character/subword time stamps, "word" for word level time stamps and "all" (default), for both character level and word level time stamps. word_seperator: Str token representing the seperator between words. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of ints. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. The config may further contain the following sub-dictionaries: "greedy": max_symbols: int, describing the maximum number of target tokens to decode per timestep during greedy decoding. Setting to larger values allows longer sentences to be decoded, at the cost of increased execution time. preserve_frame_confidence: Same as above, overrides above value. confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. Set to True by default. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, at increased cost to execution time. alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. If an integer is provided, it can decode sequences of that particular maximum length. If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), where seq_len is the length of the acoustic model output (T). NOTE: If a float is provided, it can be greater than 1! By default, a float of 2.0 is used so that a target sequence can be at most twice as long as the acoustic model output length T. maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 in order to reduce expensive beam search cost later. int >= 0. maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, and affects the speed of inference since large values will perform large beam search in the next step. maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for expansion apart from the "most likely" candidate. Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally tuned on a validation set. softmax_temperature: Scales the logits of the joint prior to computing log_softmax. decoder: The Decoder/Prediction network module. joint: The Joint network module. blank_id: The id of the RNNT blank token. """ def __init__(self, decoding_cfg, decoder, joint, blank_id: int): super(AbstractRNNTDecoding, self).__init__() # Convert dataclass to config object if is_dataclass(decoding_cfg): decoding_cfg = OmegaConf.structured(decoding_cfg) self.cfg = decoding_cfg self.blank_id = blank_id self.num_extra_outputs = joint.num_extra_outputs self.big_blank_durations = self.cfg.get("big_blank_durations", None) self.durations = self.cfg.get("durations", None) self.compute_hypothesis_token_set = self.cfg.get("compute_hypothesis_token_set", False) self.compute_langs = decoding_cfg.get('compute_langs', False) self.preserve_alignments = self.cfg.get('preserve_alignments', None) self.joint_fused_batch_size = self.cfg.get('fused_batch_size', None) self.compute_timestamps = self.cfg.get('compute_timestamps', None) self.word_seperator = self.cfg.get('word_seperator', ' ') if self.durations is not None: # this means it's a TDT model. if blank_id == 0: raise ValueError("blank_id must equal len(non_blank_vocabs) for TDT models") if self.big_blank_durations is not None: raise ValueError("duration and big_blank_durations can't both be not None") if self.cfg.strategy not in ['greedy', 'greedy_batch']: raise ValueError("currently only greedy and greedy_batch inference is supported for TDT models") if self.big_blank_durations is not None: # this means it's a multi-blank model. if blank_id == 0: raise ValueError("blank_id must equal len(vocabs) for multi-blank RNN-T models") if self.cfg.strategy not in ['greedy', 'greedy_batch']: raise ValueError( "currently only greedy and greedy_batch inference is supported for multi-blank models" ) possible_strategies = ['greedy', 'greedy_batch', 'beam', 'tsd', 'alsd', 'maes'] if self.cfg.strategy not in possible_strategies: raise ValueError(f"Decoding strategy must be one of {possible_strategies}") # Update preserve alignments if self.preserve_alignments is None: if self.cfg.strategy in ['greedy', 'greedy_batch']: self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) # Update compute timestamps if self.compute_timestamps is None: if self.cfg.strategy in ['greedy', 'greedy_batch']: self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) # Test if alignments are being preserved for RNNT if self.compute_timestamps is True and self.preserve_alignments is False: raise ValueError("If `compute_timesteps` flag is set, then `preserve_alignments` flag must also be set.") # initialize confidence-related fields self._init_confidence(self.cfg.get('confidence_cfg', None)) # Confidence estimation is not implemented for these strategies if ( not self.preserve_frame_confidence and self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes'] and self.cfg.beam.get('preserve_frame_confidence', False) ): raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") if self.cfg.strategy == 'greedy': if self.big_blank_durations is None: if self.durations is None: self.decoding = greedy_decode.GreedyRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_measure_cfg=self.confidence_measure_cfg, ) else: self.decoding = greedy_decode.GreedyTDTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, durations=self.durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_measure_cfg=self.confidence_measure_cfg, ) else: self.decoding = greedy_decode.GreedyMultiblankRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, big_blank_durations=self.big_blank_durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_measure_cfg=self.confidence_measure_cfg, ) elif self.cfg.strategy == 'greedy_batch': if self.big_blank_durations is None: if self.durations is None: self.decoding = greedy_decode.GreedyBatchedRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_measure_cfg=self.confidence_measure_cfg, ) else: self.decoding = greedy_decode.GreedyBatchedTDTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, durations=self.durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_measure_cfg=self.confidence_measure_cfg, ) else: self.decoding = greedy_decode.GreedyBatchedMultiblankRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, big_blank_durations=self.big_blank_durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_measure_cfg=self.confidence_measure_cfg, ) elif self.cfg.strategy == 'beam': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='default', score_norm=self.cfg.beam.get('score_norm', True), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ) elif self.cfg.strategy == 'tsd': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='tsd', score_norm=self.cfg.beam.get('score_norm', True), tsd_max_sym_exp_per_step=self.cfg.beam.get('tsd_max_sym_exp', 10), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ) elif self.cfg.strategy == 'alsd': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='alsd', score_norm=self.cfg.beam.get('score_norm', True), alsd_max_target_len=self.cfg.beam.get('alsd_max_target_len', 2), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ) elif self.cfg.strategy == 'maes': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='maes', score_norm=self.cfg.beam.get('score_norm', True), maes_num_steps=self.cfg.beam.get('maes_num_steps', 2), maes_prefix_alpha=self.cfg.beam.get('maes_prefix_alpha', 1), maes_expansion_gamma=self.cfg.beam.get('maes_expansion_gamma', 2.3), maes_expansion_beta=self.cfg.beam.get('maes_expansion_beta', 2.0), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ngram_lm_model=self.cfg.beam.get('ngram_lm_model', None), ngram_lm_alpha=self.cfg.beam.get('ngram_lm_alpha', 0.0), hat_subtract_ilm=self.cfg.beam.get('hat_subtract_ilm', False), hat_ilm_weight=self.cfg.beam.get('hat_ilm_weight', 0.0), ) else: raise ValueError( f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" f"but was provided {self.cfg.strategy}" ) # Update the joint fused batch size or disable it entirely if needed. self.update_joint_fused_batch_size() def rnnt_decoder_predictions_tensor( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, return_hypotheses: bool = False, partial_hypotheses: Optional[List[Hypothesis]] = None, ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: """ Decode an encoder output by autoregressive decoding of the Decoder+Joint networks. Args: encoder_output: torch.Tensor of shape [B, D, T]. encoded_lengths: torch.Tensor containing lengths of the padded encoder outputs. Shape [B]. return_hypotheses: bool. If set to True it will return list of Hypothesis or NBestHypotheses Returns: If `return_best_hypothesis` is set: A tuple (hypotheses, None): hypotheses - list of Hypothesis (best hypothesis per sample). Look at rnnt_utils.Hypothesis for more information. If `return_best_hypothesis` is not set: A tuple(hypotheses, all_hypotheses) hypotheses - list of Hypothesis (best hypothesis per sample). Look at rnnt_utils.Hypothesis for more information. all_hypotheses - list of NBestHypotheses. Each NBestHypotheses further contains a sorted list of all the hypotheses of the model per sample. Look at rnnt_utils.NBestHypotheses for more information. """ # Compute hypotheses with torch.inference_mode(): hypotheses_list = self.decoding( encoder_output=encoder_output, encoded_lengths=encoded_lengths, partial_hypotheses=partial_hypotheses ) # type: [List[Hypothesis]] # extract the hypotheses hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] prediction_list = hypotheses_list if isinstance(prediction_list[0], NBestHypotheses): hypotheses = [] all_hypotheses = [] for nbest_hyp in prediction_list: # type: NBestHypotheses n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample decoded_hyps = self.decode_hypothesis(n_hyps) # type: List[str] # If computing timestamps if self.compute_timestamps is True: timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') for hyp_idx in range(len(decoded_hyps)): decoded_hyps[hyp_idx] = self.compute_rnnt_timestamps(decoded_hyps[hyp_idx], timestamp_type) hypotheses.append(decoded_hyps[0]) # best hypothesis all_hypotheses.append(decoded_hyps) if return_hypotheses: return hypotheses, all_hypotheses best_hyp_text = [h.text for h in hypotheses] all_hyp_text = [h.text for hh in all_hypotheses for h in hh] return best_hyp_text, all_hyp_text else: hypotheses = self.decode_hypothesis(prediction_list) # type: List[str] # If computing timestamps if self.compute_timestamps is True: timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') for hyp_idx in range(len(hypotheses)): hypotheses[hyp_idx] = self.compute_rnnt_timestamps(hypotheses[hyp_idx], timestamp_type) if return_hypotheses: # greedy decoding, can get high-level confidence scores if self.preserve_frame_confidence and ( self.preserve_word_confidence or self.preserve_token_confidence ): hypotheses = self.compute_confidence(hypotheses) return hypotheses, None best_hyp_text = [h.text for h in hypotheses] return best_hyp_text, None def decode_hypothesis(self, hypotheses_list: List[Hypothesis]) -> List[Union[Hypothesis, NBestHypotheses]]: """ Decode a list of hypotheses into a list of strings. Args: hypotheses_list: List of Hypothesis. Returns: A list of strings. """ for ind in range(len(hypotheses_list)): # Extract the integer encoded hypothesis prediction = hypotheses_list[ind].y_sequence if type(prediction) != list: prediction = prediction.tolist() # RNN-T sample level is already preprocessed by implicit RNNT decoding # Simply remove any blank and possibly big blank tokens if self.big_blank_durations is not None: # multi-blank RNNT num_extra_outputs = len(self.big_blank_durations) prediction = [p for p in prediction if p < self.blank_id - num_extra_outputs] elif self.durations is not None: # TDT model. prediction = [p for p in prediction if p < self.blank_id] else: # standard RNN-T prediction = [p for p in prediction if p != self.blank_id] # De-tokenize the integer tokens; if not computing timestamps if self.compute_timestamps is True: # keep the original predictions, wrap with the number of repetitions per token and alignments # this is done so that `rnnt_decoder_predictions_tensor()` can process this hypothesis # in order to compute exact time stamps. alignments = copy.deepcopy(hypotheses_list[ind].alignments) token_repetitions = [1] * len(alignments) # preserve number of repetitions per token hypothesis = (prediction, alignments, token_repetitions) else: hypothesis = self.decode_tokens_to_str(prediction) # TODO: remove # collapse leading spaces before . , ? for PC models hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) if self.compute_hypothesis_token_set: hypotheses_list[ind].tokens = self.decode_ids_to_tokens(prediction) # De-tokenize the integer tokens hypotheses_list[ind].text = hypothesis return hypotheses_list def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: """ Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. Assumes that `frame_confidence` is present in the hypotheses. Args: hypotheses_list: List of Hypothesis. Returns: A list of hypotheses with high-level confidence scores. """ if self.exclude_blank_from_confidence: for hyp in hypotheses_list: hyp.token_confidence = hyp.non_blank_frame_confidence else: for hyp in hypotheses_list: offset = 0 token_confidence = [] if len(hyp.timestep) > 0: for ts, te in zip(hyp.timestep, hyp.timestep[1:] + [len(hyp.frame_confidence)]): if ts != te: # <blank> tokens are considered to belong to the last non-blank token, if any. token_confidence.append( self._aggregate_confidence( [hyp.frame_confidence[ts][offset]] + [fc[0] for fc in hyp.frame_confidence[ts + 1 : te]] ) ) offset = 0 else: token_confidence.append(hyp.frame_confidence[ts][offset]) offset += 1 hyp.token_confidence = token_confidence if self.preserve_word_confidence: for hyp in hypotheses_list: hyp.word_confidence = self._aggregate_token_confidence(hyp) return hypotheses_list @abstractmethod def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token id list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ raise NotImplementedError() @abstractmethod def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ raise NotImplementedError() @abstractmethod def decode_tokens_to_lang(self, tokens: List[int]) -> str: """ Implemented by subclass in order to compute the most likely language ID (LID) string given the tokens. Args: tokens: List of int representing the token ids. Returns: A decoded LID string. """ raise NotImplementedError() @abstractmethod def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into language ID (LID) list. Args: tokens: List of int representing the token ids. Returns: A list of decoded LIDS. """ raise NotImplementedError() def update_joint_fused_batch_size(self): if self.joint_fused_batch_size is None: # do nothing and let the Joint itself handle setting up of the fused batch return if not hasattr(self.decoding.joint, 'set_fused_batch_size'): logging.warning( "The joint module does not have `set_fused_batch_size(int)` as a setter function.\n" "Ignoring update of joint fused batch size." ) return if not hasattr(self.decoding.joint, 'set_fuse_loss_wer'): logging.warning( "The joint module does not have `set_fuse_loss_wer(bool, RNNTLoss, RNNTWER)` " "as a setter function.\n" "Ignoring update of joint fused batch size." ) return if self.joint_fused_batch_size > 0: self.decoding.joint.set_fused_batch_size(self.joint_fused_batch_size) else: logging.info("Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.") self.decoding.joint.set_fuse_loss_wer(False) def compute_rnnt_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): assert timestamp_type in ['char', 'word', 'all'] # Unpack the temporary storage decoded_prediction, alignments, token_repetitions = hypothesis.text # Retrieve offsets char_offsets = word_offsets = None char_offsets = self._compute_offsets(hypothesis, token_repetitions, self.blank_id) # finally, set the flattened decoded predictions to text field for later text decoding hypothesis.text = decoded_prediction # Assert number of offsets and hypothesis tokens are 1:1 match. num_flattened_tokens = 0 for t in range(len(char_offsets)): # Subtract one here for the extra RNNT BLANK token emitted to designate "End of timestep" num_flattened_tokens += len(char_offsets[t]['char']) - 1 if num_flattened_tokens != len(hypothesis.text): raise ValueError( f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" " have to be of the same length, but are: " f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" f" {len(hypothesis.text)}" ) encoded_char_offsets = copy.deepcopy(char_offsets) # Correctly process the token ids to chars/subwords. for i, offsets in enumerate(char_offsets): decoded_chars = [] for char in offsets['char'][:-1]: # ignore the RNNT Blank token at end of every timestep with -1 subset decoded_chars.append(self.decode_tokens_to_str([int(char)])) char_offsets[i]["char"] = decoded_chars # detect char vs subword models lens = [] for v in char_offsets: tokens = v["char"] # each token may be either 1 unicode token or multiple unicode token # for character based models, only 1 token is used # for subword, more than one token can be used. # Computing max, then summing up total lens is a test to check for char vs subword # For char models, len(lens) == sum(lens) # but this is violated for subword models. max_len = max(len(c) for c in tokens) lens.append(max_len) # array of one or more chars implies subword based model with multiple char emitted per TxU step (via subword) if sum(lens) > len(lens): text_type = 'subword' else: # full array of ones implies character based model with 1 char emitted per TxU step text_type = 'char' # retrieve word offsets from character offsets word_offsets = None if timestamp_type in ['word', 'all']: if text_type == 'char': word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) else: # utilize the copy of char offsets with the correct integer ids for tokens # so as to avoid tokenize -> detokenize -> compare -> merge steps. word_offsets = self._get_word_offsets_subwords_sentencepiece( encoded_char_offsets, hypothesis, decode_ids_to_tokens=self.decode_ids_to_tokens, decode_tokens_to_str=self.decode_tokens_to_str, ) # attach results if len(hypothesis.timestep) > 0: timestep_info = hypothesis.timestep else: timestep_info = [] # Setup defaults hypothesis.timestep = {"timestep": timestep_info} # Add char / subword time stamps if char_offsets is not None and timestamp_type in ['char', 'all']: hypothesis.timestep['char'] = char_offsets # Add word time stamps if word_offsets is not None and timestamp_type in ['word', 'all']: hypothesis.timestep['word'] = word_offsets # Convert the flattened token indices to text hypothesis.text = self.decode_tokens_to_str(hypothesis.text) return hypothesis @staticmethod def _compute_offsets( hypothesis: Hypothesis, token_repetitions: List[int], rnnt_token: int ) -> List[Dict[str, Union[str, int]]]: """ Utility method that calculates the indidual time indices where a token starts and ends. Args: hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token emitted at every time step after rnnt collapse. token_repetitions: A list of ints representing the number of repetitions of each emitted token. rnnt_token: The integer of the rnnt blank token used during rnnt collapse. Returns: """ start_index = 0 # If the exact timestep information is available, utilize the 1st non-rnnt blank token timestep # as the start index. if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: start_index = max(0, hypothesis.timestep[0] - 1) # Construct the start and end indices brackets end_indices = np.asarray(token_repetitions).cumsum() start_indices = np.concatenate(([start_index], end_indices[:-1])) # Process the TxU dangling alignment tensor, containing pairs of (logits, label) alignment_labels = [al_logits_labels for al_logits_labels in hypothesis.text[1]] for t in range(len(alignment_labels)): for u in range(len(alignment_labels[t])): alignment_labels[t][u] = alignment_labels[t][u][1] # pick label from (logit, label) tuple # Merge the results per token into a list of dictionaries offsets = [ {"char": a, "start_offset": s, "end_offset": e} for a, s, e in zip(alignment_labels, start_indices, end_indices) ] # Filter out RNNT token (blank at [t][0] position). This is because blank can only occur at end of a # time step for RNNT, so if 0th token is blank, then that timestep is skipped. offsets = list(filter(lambda offsets: offsets["char"][0] != rnnt_token, offsets)) return offsets @staticmethod def _get_word_offsets_chars( offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of character time stamps. References: This code is a port of the Hugging Face code for word time stamp construction. Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". word_delimiter_char: Character token that represents the word delimiter. By default, " ". Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] last_state = "SPACE" word = "" start_offset = 0 end_offset = 0 for i, offset in enumerate(offsets): chars = offset["char"] for char in chars: state = "SPACE" if char == word_delimiter_char else "WORD" if state == last_state: # If we are in the same state as before, we simply repeat what we've done before end_offset = offset["end_offset"] word += char else: # Switching state if state == "SPACE": # Finishing a word word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) else: # Starting a new word start_offset = offset["start_offset"] end_offset = offset["end_offset"] word = char last_state = state if last_state == "WORD": word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) return word_offsets @staticmethod def _get_word_offsets_subwords_sentencepiece( offsets: Dict[str, Union[str, float]], hypothesis: Hypothesis, decode_ids_to_tokens: Callable[[List[int]], str], decode_tokens_to_str: Callable[[List[int]], str], ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of sub-word time stamps. **Note**: Only supports Sentencepiece based tokenizers ! Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id after rnnt collapse. decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] built_token = [] previous_token_index = 0 # For every offset token for i, offset in enumerate(offsets): # For every subword token in offset token list (ignoring the RNNT Blank token at the end) for char in offset['char'][:-1]: char = int(char) # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). token = decode_ids_to_tokens([char])[0] token_text = decode_tokens_to_str([char]) # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped # after forcing partial text conversion of the token. if token != token_text: # If there are any partially or fully built sub-word token ids, construct to text. # Note: This is "old" subword, that occurs *after* current sub-word has started. if built_token: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[previous_token_index]["start_offset"], "end_offset": offsets[i]["start_offset"], } ) # Prepare list of new sub-word ids built_token.clear() built_token.append(char) previous_token_index = i else: # If the token does not contain any sub-word start mark, then the sub-word has not completed yet # Append to current sub-word list. built_token.append(char) # Inject the start offset of the first token to word offsets # This is because we always skip the delay the injection of the first sub-word due to the loop # condition and check whether built token is ready or not. # Therefore without this forced injection, the start_offset appears as off by 1. # This should only be done when these arrays contain more than one element. if offsets and word_offsets: word_offsets[0]["start_offset"] = offsets[0]["start_offset"] # If there are any remaining tokens left, inject them all into the final word offset. # The start offset of this token is the start time of the next token to process. # The end offset of this token is the end time of the last token from offsets. # Note that built_token is a flat list; but offsets contains a nested list which # may have different dimensionality. # As such, we can't rely on the length of the list of built_token to index offsets. if built_token: # start from the previous token index as this hasn't been committed to word_offsets yet # if we still have content in built_token start_offset = offsets[previous_token_index]["start_offset"] word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": start_offset, "end_offset": offsets[-1]["end_offset"], } ) built_token.clear() return word_offsets class RNNTDecoding(AbstractRNNTDecoding): """ Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy, greedy_batch (for greedy decoding). - beam, tsd, alsd (for beam search decoding). compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded tokens as well as the decoded string. Default is False in order to avoid double decoding unless required. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function with the `return_hypotheses` flag set to True. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. The config may further contain the following sub-dictionaries: "greedy": max_symbols: int, describing the maximum number of target tokens to decode per timestep during greedy decoding. Setting to larger values allows longer sentences to be decoded, at the cost of increased execution time. preserve_frame_confidence: Same as above, overrides above value. confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. Set to True by default. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, at increased cost to execution time. alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. If an integer is provided, it can decode sequences of that particular maximum length. If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), where seq_len is the length of the acoustic model output (T). NOTE: If a float is provided, it can be greater than 1! By default, a float of 2.0 is used so that a target sequence can be at most twice as long as the acoustic model output length T. maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 in order to reduce expensive beam search cost later. int >= 0. maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, and affects the speed of inference since large values will perform large beam search in the next step. maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for expansion apart from the "most likely" candidate. Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally tuned on a validation set. softmax_temperature: Scales the logits of the joint prior to computing log_softmax. decoder: The Decoder/Prediction network module. joint: The Joint network module. vocabulary: The vocabulary (excluding the RNNT blank token) which will be used for decoding. """ def __init__( self, decoding_cfg, decoder, joint, vocabulary, ): # we need to ensure blank is the last token in the vocab for the case of RNNT and Multi-blank RNNT. blank_id = len(vocabulary) + joint.num_extra_outputs if hasattr(decoding_cfg, 'model_type') and decoding_cfg.model_type == 'tdt': blank_id = len(vocabulary) self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) super(RNNTDecoding, self).__init__( decoding_cfg=decoding_cfg, decoder=decoder, joint=joint, blank_id=blank_id, ) if isinstance(self.decoding, beam_decode.BeamRNNTInfer): self.decoding.set_decoding_type('char') def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: """ Implemented by subclass in order to aggregate token confidence to a word-level confidence. Args: hypothesis: Hypothesis Returns: A list of word-level confidence scores. """ return self._aggregate_token_confidence_chars(hypothesis.words, hypothesis.token_confidence) def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) return hypothesis def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ token_list = [self.labels_map[c] for c in tokens if c < self.blank_id - self.num_extra_outputs] return token_list def decode_tokens_to_lang(self, tokens: List[int]) -> str: """ Compute the most likely language ID (LID) string given the tokens. Args: tokens: List of int representing the token ids. Returns: A decoded LID string. """ lang = self.tokenizer.ids_to_lang(tokens) return lang def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: """ Decode a token id list into language ID (LID) list. Args: tokens: List of int representing the token ids. Returns: A list of decoded LIDS. """ lang_list = self.tokenizer.ids_to_text_and_langs(tokens) return lang_list class RNNTWER(Metric): """ This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. When doing distributed training/evaluation the result of res=WER(predictions, targets, target_lengths) calls will be all-reduced between all workers using SUM operations. Here contains two numbers res=[wer_numerator, wer_denominator]. WER=wer_numerator/wer_denominator. If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. Example: def validation_step(self, batch, batch_idx): ... wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} return self.val_outputs def on_validation_epoch_end(self): ... wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} self.val_outputs.clear() # free memory return {'val_loss': val_loss_mean, 'log': tensorboard_logs} Args: decoding: RNNTDecoding object that will perform autoregressive decoding of the RNNT model. batch_dim_index: Index of the batch dimension. use_cer: Whether to use Character Error Rate isntead of Word Error Rate. log_prediction: Whether to log a single decoded sample per call. Returns: res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenshtein's distances for all prediction - reference pairs, total number of words in all references. """ full_state_update = True def __init__( self, decoding: RNNTDecoding, batch_dim_index=0, use_cer=False, log_prediction=True, dist_sync_on_step=False ): super(RNNTWER, self).__init__(dist_sync_on_step=dist_sync_on_step) self.decoding = decoding self.batch_dim_index = batch_dim_index self.use_cer = use_cer self.log_prediction = log_prediction self.blank_id = self.decoding.blank_id self.labels_map = self.decoding.labels_map self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) def update( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, targets: torch.Tensor, target_lengths: torch.Tensor, ) -> torch.Tensor: words = 0 scores = 0 references = [] with torch.no_grad(): # prediction_cpu_tensor = tensors[0].long().cpu() targets_cpu_tensor = targets.long().cpu() targets_cpu_tensor = move_dimension_to_the_front(targets_cpu_tensor, self.batch_dim_index) tgt_lenths_cpu_tensor = target_lengths.long().cpu() # iterate over batch for ind in range(targets_cpu_tensor.shape[0]): tgt_len = tgt_lenths_cpu_tensor[ind].item() target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() reference = self.decoding.decode_tokens_to_str(target) references.append(reference) hypotheses, _ = self.decoding.rnnt_decoder_predictions_tensor(encoder_output, encoded_lengths) if self.log_prediction: logging.info(f"\n") logging.info(f"reference :{references[0]}") logging.info(f"predicted :{hypotheses[0]}") for h, r in zip(hypotheses, references): if self.use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() words += len(r_list) # Compute Levenshtein's distance scores += editdistance.eval(h_list, r_list) self.scores += torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) self.words += torch.tensor(words, device=self.words.device, dtype=self.words.dtype) # return torch.tensor([scores, words]).to(predictions.device) def compute(self): wer = self.scores.float() / self.words return wer, self.scores.detach(), self.words.detach() @dataclass class RNNTDecodingConfig: model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" strategy: str = "greedy_batch" compute_hypothesis_token_set: bool = False # preserve decoding alignments preserve_alignments: Optional[bool] = None # confidence config confidence_cfg: ConfidenceConfig = ConfidenceConfig() # RNNT Joint fused batch size fused_batch_size: Optional[int] = None # compute RNNT time stamps compute_timestamps: Optional[bool] = None # compute language IDs compute_langs: bool = False # token representing word seperator word_seperator: str = " " # type of timestamps to calculate rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() # beam decoding config beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) # can be used to change temperature for decoding temperature: float = 1.0 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/metrics/wer.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import re from abc import abstractmethod from dataclasses import dataclass, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance import jiwer import numpy as np import torch from omegaconf import DictConfig, OmegaConf from torchmetrics import Metric from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses from nemo.utils import logging, logging_mode __all__ = ['word_error_rate', 'word_error_rate_detail', 'WER', 'move_dimension_to_the_front'] def word_error_rate(hypotheses: List[str], references: List[str], use_cer=False) -> float: """ Computes Average Word Error rate between two texts represented as corresponding lists of string. Hypotheses and references must have same length. Args: hypotheses (list): list of hypotheses references(list) : list of references use_cer (bool): set True to enable cer Returns: wer (float): average word error rate """ scores = 0 words = 0 if len(hypotheses) != len(references): raise ValueError( "In word error rate calculation, hypotheses and reference" " lists must have the same number of elements. But I got:" "{0} and {1} correspondingly".format(len(hypotheses), len(references)) ) for h, r in zip(hypotheses, references): if use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() words += len(r_list) # May deprecate using editdistance in future release for here and rest of codebase # once we confirm jiwer is reliable. scores += editdistance.eval(h_list, r_list) if words != 0: wer = 1.0 * scores / words else: wer = float('inf') return wer def word_error_rate_detail( hypotheses: List[str], references: List[str], use_cer=False ) -> Tuple[float, int, float, float, float]: """ Computes Average Word Error Rate with details (insertion rate, deletion rate, substitution rate) between two texts represented as corresponding lists of string. Hypotheses and references must have same length. Args: hypotheses (list): list of hypotheses references(list) : list of references use_cer (bool): set True to enable cer Returns: wer (float): average word error rate words (int): Total number of words/charactors of given reference texts ins_rate (float): average insertion error rate del_rate (float): average deletion error rate sub_rate (float): average substitution error rate """ scores = 0 words = 0 ops_count = {'substitutions': 0, 'insertions': 0, 'deletions': 0} if len(hypotheses) != len(references): raise ValueError( "In word error rate calculation, hypotheses and reference" " lists must have the same number of elements. But I got:" "{0} and {1} correspondingly".format(len(hypotheses), len(references)) ) for h, r in zip(hypotheses, references): if use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() # To get rid of the issue that jiwer does not allow empty string if len(r_list) == 0: if len(h_list) != 0: errors = len(h_list) ops_count['insertions'] += errors else: errors = 0 else: if use_cer: measures = jiwer.cer(r, h, return_dict=True) else: measures = jiwer.compute_measures(r, h) errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] ops_count['insertions'] += measures['insertions'] ops_count['deletions'] += measures['deletions'] ops_count['substitutions'] += measures['substitutions'] scores += errors words += len(r_list) if words != 0: wer = 1.0 * scores / words ins_rate = 1.0 * ops_count['insertions'] / words del_rate = 1.0 * ops_count['deletions'] / words sub_rate = 1.0 * ops_count['substitutions'] / words else: wer, ins_rate, del_rate, sub_rate = float('inf'), float('inf'), float('inf'), float('inf') return wer, words, ins_rate, del_rate, sub_rate def word_error_rate_per_utt(hypotheses: List[str], references: List[str], use_cer=False) -> Tuple[List[float], float]: """ Computes Word Error Rate per utterance and the average WER between two texts represented as corresponding lists of string. Hypotheses and references must have same length. Args: hypotheses (list): list of hypotheses references(list) : list of references use_cer (bool): set True to enable cer Returns: wer_per_utt (List[float]): word error rate per utterance avg_wer (float): average word error rate """ scores = 0 words = 0 wer_per_utt = [] if len(hypotheses) != len(references): raise ValueError( "In word error rate calculation, hypotheses and reference" " lists must have the same number of elements. But I got:" "{0} and {1} correspondingly".format(len(hypotheses), len(references)) ) for h, r in zip(hypotheses, references): if use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() # To get rid of the issue that jiwer does not allow empty string if len(r_list) == 0: if len(h_list) != 0: errors = len(h_list) wer_per_utt.append(float('inf')) else: if use_cer: measures = jiwer.cer(r, h, return_dict=True) er = measures['cer'] else: measures = jiwer.compute_measures(r, h) er = measures['wer'] errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] wer_per_utt.append(er) scores += errors words += len(r_list) if words != 0: avg_wer = 1.0 * scores / words else: avg_wer = float('inf') return wer_per_utt, avg_wer def move_dimension_to_the_front(tensor, dim_index): all_dims = list(range(tensor.ndim)) return tensor.permute(*([dim_index] + all_dims[:dim_index] + all_dims[dim_index + 1 :])) class AbstractCTCDecoding(ConfidenceMixin): """ Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy (for greedy decoding). - beam (for DeepSpeed KenLM based decoding). compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. Can take the following values - "char" for character/subword time stamps, "word" for word level time stamps and "all" (default), for both character level and word level time stamps. word_seperator: Str token representing the seperator between words. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. The config may further contain the following sub-dictionaries: "greedy": preserve_alignments: Same as above, overrides above value. compute_timestamps: Same as above, overrides above value. preserve_frame_confidence: Same as above, overrides above value. confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. beam_alpha: float, the strength of the Language model on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. beam_beta: float, the strength of the sequence length penalty on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). If the path is invalid (file is not found at path), will raise a deferred error at the moment of calculation of beam search, so that users may update / change the decoding strategy to point to the correct file. blank_id: The id of the RNNT blank token. """ def __init__(self, decoding_cfg, blank_id: int): super().__init__() # Convert dataclas to config if is_dataclass(decoding_cfg): decoding_cfg = OmegaConf.structured(decoding_cfg) if not isinstance(decoding_cfg, DictConfig): decoding_cfg = OmegaConf.create(decoding_cfg) OmegaConf.set_struct(decoding_cfg, False) # update minimal config minimal_cfg = ['greedy'] for item in minimal_cfg: if item not in decoding_cfg: decoding_cfg[item] = OmegaConf.create({}) self.cfg = decoding_cfg self.blank_id = blank_id self.preserve_alignments = self.cfg.get('preserve_alignments', None) self.compute_timestamps = self.cfg.get('compute_timestamps', None) self.batch_dim_index = self.cfg.get('batch_dim_index', 0) self.word_seperator = self.cfg.get('word_seperator', ' ') possible_strategies = ['greedy', 'beam', 'pyctcdecode', 'flashlight'] if self.cfg.strategy not in possible_strategies: raise ValueError(f"Decoding strategy must be one of {possible_strategies}. Given {self.cfg.strategy}") # Update preserve alignments if self.preserve_alignments is None: if self.cfg.strategy in ['greedy']: self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) else: self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) # Update compute timestamps if self.compute_timestamps is None: if self.cfg.strategy in ['greedy']: self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) elif self.cfg.strategy in ['beam']: self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) # initialize confidence-related fields self._init_confidence(self.cfg.get('confidence_cfg', None)) # Confidence estimation is not implemented for strategies other than `greedy` if ( not self.preserve_frame_confidence and self.cfg.strategy != 'greedy' and self.cfg.beam.get('preserve_frame_confidence', False) ): raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") # we need timestamps to extract non-blank per-frame confidence if self.compute_timestamps is not None: self.compute_timestamps |= self.preserve_frame_confidence if self.cfg.strategy == 'greedy': self.decoding = ctc_greedy_decoding.GreedyCTCInfer( blank_id=self.blank_id, preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, preserve_frame_confidence=self.preserve_frame_confidence, confidence_measure_cfg=self.confidence_measure_cfg, ) elif self.cfg.strategy == 'beam': self.decoding = ctc_beam_decoding.BeamCTCInfer( blank_id=blank_id, beam_size=self.cfg.beam.get('beam_size', 1), search_type='default', return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), beam_beta=self.cfg.beam.get('beam_beta', 0.0), kenlm_path=self.cfg.beam.get('kenlm_path', None), ) self.decoding.override_fold_consecutive_value = False elif self.cfg.strategy == 'pyctcdecode': self.decoding = ctc_beam_decoding.BeamCTCInfer( blank_id=blank_id, beam_size=self.cfg.beam.get('beam_size', 1), search_type='pyctcdecode', return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), beam_beta=self.cfg.beam.get('beam_beta', 0.0), kenlm_path=self.cfg.beam.get('kenlm_path', None), pyctcdecode_cfg=self.cfg.beam.get('pyctcdecode_cfg', None), ) self.decoding.override_fold_consecutive_value = False elif self.cfg.strategy == 'flashlight': self.decoding = ctc_beam_decoding.BeamCTCInfer( blank_id=blank_id, beam_size=self.cfg.beam.get('beam_size', 1), search_type='flashlight', return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), beam_beta=self.cfg.beam.get('beam_beta', 0.0), kenlm_path=self.cfg.beam.get('kenlm_path', None), flashlight_cfg=self.cfg.beam.get('flashlight_cfg', None), ) self.decoding.override_fold_consecutive_value = False else: raise ValueError( f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" f"but was provided {self.cfg.strategy}" ) def ctc_decoder_predictions_tensor( self, decoder_outputs: torch.Tensor, decoder_lengths: torch.Tensor = None, fold_consecutive: bool = True, return_hypotheses: bool = False, ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: """ Decodes a sequence of labels to words Args: decoder_outputs: An integer torch.Tensor of shape [Batch, Time, {Vocabulary}] (if ``batch_index_dim == 0``) or [Time, Batch] (if ``batch_index_dim == 1``) of integer indices that correspond to the index of some character in the label set. decoder_lengths: Optional tensor of length `Batch` which contains the integer lengths of the sequence in the padded `predictions` tensor. fold_consecutive: Bool, determine whether to perform "ctc collapse", folding consecutive tokens into a single token. return_hypotheses: Bool flag whether to return just the decoding predictions of the model or a Hypothesis object that holds information such as the decoded `text`, the `alignment` of emited by the CTC Model, and the `length` of the sequence (if available). May also contain the log-probabilities of the decoder (if this method is called via transcribe()) Returns: Either a list of str which represent the CTC decoded strings per sample, or a list of Hypothesis objects containing additional information. """ if isinstance(decoder_outputs, torch.Tensor): decoder_outputs = move_dimension_to_the_front(decoder_outputs, self.batch_dim_index) if ( hasattr(self.decoding, 'override_fold_consecutive_value') and self.decoding.override_fold_consecutive_value is not None ): logging.info( f"Beam search requires that consecutive ctc tokens are not folded. \n" f"Overriding provided value of `fold_consecutive` = {fold_consecutive} to " f"{self.decoding.override_fold_consecutive_value}", mode=logging_mode.ONCE, ) fold_consecutive = self.decoding.override_fold_consecutive_value with torch.inference_mode(): # Resolve the forward step of the decoding strategy hypotheses_list = self.decoding( decoder_output=decoder_outputs, decoder_lengths=decoder_lengths ) # type: List[List[Hypothesis]] # extract the hypotheses hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] if isinstance(hypotheses_list[0], NBestHypotheses): hypotheses = [] all_hypotheses = [] for nbest_hyp in hypotheses_list: # type: NBestHypotheses n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample decoded_hyps = self.decode_hypothesis( n_hyps, fold_consecutive ) # type: List[Union[Hypothesis, NBestHypotheses]] # If computing timestamps if self.compute_timestamps is True: timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') for hyp_idx in range(len(decoded_hyps)): decoded_hyps[hyp_idx] = self.compute_ctc_timestamps(decoded_hyps[hyp_idx], timestamp_type) hypotheses.append(decoded_hyps[0]) # best hypothesis all_hypotheses.append(decoded_hyps) if return_hypotheses: return hypotheses, all_hypotheses best_hyp_text = [h.text for h in hypotheses] all_hyp_text = [h.text for hh in all_hypotheses for h in hh] return best_hyp_text, all_hyp_text else: hypotheses = self.decode_hypothesis( hypotheses_list, fold_consecutive ) # type: List[Union[Hypothesis, NBestHypotheses]] # If computing timestamps if self.compute_timestamps is True: # greedy decoding, can get high-level confidence scores if return_hypotheses and (self.preserve_word_confidence or self.preserve_token_confidence): hypotheses = self.compute_confidence(hypotheses) else: # remove unused token_repetitions from Hypothesis.text for hyp in hypotheses: hyp.text = hyp.text[:2] timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') for hyp_idx in range(len(hypotheses)): hypotheses[hyp_idx] = self.compute_ctc_timestamps(hypotheses[hyp_idx], timestamp_type) if return_hypotheses: return hypotheses, None best_hyp_text = [h.text for h in hypotheses] return best_hyp_text, None def decode_hypothesis( self, hypotheses_list: List[Hypothesis], fold_consecutive: bool ) -> List[Union[Hypothesis, NBestHypotheses]]: """ Decode a list of hypotheses into a list of strings. Args: hypotheses_list: List of Hypothesis. fold_consecutive: Whether to collapse the ctc blank tokens or not. Returns: A list of strings. """ for ind in range(len(hypotheses_list)): # Extract the integer encoded hypothesis hyp = hypotheses_list[ind] prediction = hyp.y_sequence predictions_len = hyp.length if hyp.length > 0 else None if fold_consecutive: if type(prediction) != list: prediction = prediction.numpy().tolist() if predictions_len is not None: prediction = prediction[:predictions_len] # CTC decoding procedure decoded_prediction = [] token_lengths = [] # preserve token lengths token_repetitions = [] # preserve number of repetitions per token previous = self.blank_id last_length = 0 last_repetition = 1 for pidx, p in enumerate(prediction): if (p != previous or previous == self.blank_id) and p != self.blank_id: decoded_prediction.append(p) token_lengths.append(pidx - last_length) last_length = pidx token_repetitions.append(last_repetition) last_repetition = 1 if p == previous and previous != self.blank_id: last_repetition += 1 previous = p if len(token_repetitions) > 0: token_repetitions = token_repetitions[1:] + [last_repetition] else: if predictions_len is not None: prediction = prediction[:predictions_len] decoded_prediction = prediction[prediction != self.blank_id].tolist() token_lengths = [1] * len(decoded_prediction) # preserve number of repetitions per token token_repetitions = [1] * len(decoded_prediction) # preserve number of repetitions per token # De-tokenize the integer tokens; if not computing timestamps if self.compute_timestamps is True: # keep the original predictions, wrap with the number of repetitions per token # this is done so that `ctc_decoder_predictions_tensor()` can process this hypothesis # in order to compute exact time stamps. hypothesis = (decoded_prediction, token_lengths, token_repetitions) else: hypothesis = self.decode_tokens_to_str(decoded_prediction) # TODO: remove # collapse leading spaces before . , ? for PC models hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) # Preserve this wrapped hypothesis or decoded text tokens. hypotheses_list[ind].text = hypothesis return hypotheses_list def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: """ Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. Assumes that `frame_confidence` is present in the hypotheses. Args: hypotheses_list: List of Hypothesis. Returns: A list of hypotheses with high-level confidence scores. """ for hyp in hypotheses_list: if not isinstance(hyp.text, tuple) or len(hyp.text) != 3: # the method must have been called in the wrong place raise ValueError( """Wrong format of the `text` attribute of a hypothesis.\n Expected: (decoded_prediction, token_repetitions)\n The method invocation is expected between .decode_hypothesis() and .compute_ctc_timestamps()""" ) token_repetitions = hyp.text[2] hyp.text = hyp.text[:2] token_confidence = [] if self.exclude_blank_from_confidence: non_blank_frame_confidence = hyp.non_blank_frame_confidence i = 0 for tr in token_repetitions: # token repetition can be zero j = i + tr token_confidence.append(self._aggregate_confidence(non_blank_frame_confidence[i:j])) i = j else: # <blank> tokens are considered to belong to the last non-blank token, if any. token_lengths = hyp.text[1] if len(token_lengths) > 0: ts = token_lengths[0] for tl in token_lengths[1:] + [len(hyp.frame_confidence)]: token_confidence.append(self._aggregate_confidence(hyp.frame_confidence[ts : ts + tl])) ts += tl hyp.token_confidence = token_confidence if self.preserve_word_confidence: for hyp in hypotheses_list: hyp.word_confidence = self._aggregate_token_confidence(hyp) return hypotheses_list @abstractmethod def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token id list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ raise NotImplementedError() @abstractmethod def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ raise NotImplementedError() def compute_ctc_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): """ Method to compute time stamps at char/subword, and word level given some hypothesis. Requires the input hypothesis to contain a `text` field that is the tuple. The tuple contains - the ctc collapsed integer ids, and the number of repetitions of each token. Args: hypothesis: A Hypothesis object, with a wrapped `text` field. The `text` field must contain a tuple with two values - The ctc collapsed integer ids A list of integers that represents the number of repetitions per token. timestamp_type: A str value that represents the type of time stamp calculated. Can be one of "char", "word" or "all" Returns: A Hypothesis object with a modified `timestep` value, which is now a dictionary containing the time stamp information. """ assert timestamp_type in ['char', 'word', 'all'] # Unpack the temporary storage, and set the decoded predictions decoded_prediction, token_lengths = hypothesis.text hypothesis.text = decoded_prediction # Retrieve offsets char_offsets = word_offsets = None char_offsets = self._compute_offsets(hypothesis, token_lengths, self.blank_id) # Assert number of offsets and hypothesis tokens are 1:1 match. if len(char_offsets) != len(hypothesis.text): raise ValueError( f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" " have to be of the same length, but are: " f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" f" {len(hypothesis.text)}" ) # Correctly process the token ids to chars/subwords. for i, char in enumerate(hypothesis.text): char_offsets[i]["char"] = self.decode_tokens_to_str([char]) # detect char vs subword models lens = [len(list(v["char"])) > 1 for v in char_offsets] if any(lens): text_type = 'subword' else: text_type = 'char' # retrieve word offsets from character offsets word_offsets = None if timestamp_type in ['word', 'all']: if text_type == 'char': word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) else: word_offsets = self._get_word_offsets_subwords_sentencepiece( char_offsets, hypothesis, decode_ids_to_tokens=self.decode_ids_to_tokens, decode_tokens_to_str=self.decode_tokens_to_str, ) # attach results if len(hypothesis.timestep) > 0: timestep_info = hypothesis.timestep else: timestep_info = [] # Setup defaults hypothesis.timestep = {"timestep": timestep_info} # Add char / subword time stamps if char_offsets is not None and timestamp_type in ['char', 'all']: hypothesis.timestep['char'] = char_offsets # Add word time stamps if word_offsets is not None and timestamp_type in ['word', 'all']: hypothesis.timestep['word'] = word_offsets # Convert the token indices to text hypothesis.text = self.decode_tokens_to_str(hypothesis.text) return hypothesis @staticmethod def _compute_offsets( hypothesis: Hypothesis, token_lengths: List[int], ctc_token: int ) -> List[Dict[str, Union[str, int]]]: """ Utility method that calculates the indidual time indices where a token starts and ends. Args: hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token emitted at every time step after ctc collapse. token_lengths: A list of ints representing the lengths of each emitted token. ctc_token: The integer of the ctc blank token used during ctc collapse. Returns: """ start_index = 0 # If the exact timestep information is available, utilize the 1st non-ctc blank token timestep # as the start index. if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: start_index = max(0, hypothesis.timestep[0] - 1) # Construct the start and end indices brackets end_indices = np.asarray(token_lengths).cumsum() start_indices = np.concatenate(([start_index], end_indices[:-1])) # Merge the results per token into a list of dictionaries offsets = [ {"char": t, "start_offset": s, "end_offset": e} for t, s, e in zip(hypothesis.text, start_indices, end_indices) ] # Filter out CTC token offsets = list(filter(lambda offsets: offsets["char"] != ctc_token, offsets)) return offsets @staticmethod def _get_word_offsets_chars( offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of character time stamps. References: This code is a port of the Hugging Face code for word time stamp construction. Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". word_delimiter_char: Character token that represents the word delimiter. By default, " ". Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] last_state = "SPACE" word = "" start_offset = 0 end_offset = 0 for i, offset in enumerate(offsets): char = offset["char"] state = "SPACE" if char == word_delimiter_char else "WORD" if state == last_state: # If we are in the same state as before, we simply repeat what we've done before end_offset = offset["end_offset"] word += char else: # Switching state if state == "SPACE": # Finishing a word word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) else: # Starting a new word start_offset = offset["start_offset"] end_offset = offset["end_offset"] word = char last_state = state if last_state == "WORD": word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) return word_offsets @staticmethod def _get_word_offsets_subwords_sentencepiece( offsets: Dict[str, Union[str, float]], hypothesis: Hypothesis, decode_ids_to_tokens: Callable[[List[int]], str], decode_tokens_to_str: Callable[[List[int]], str], ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of sub-word time stamps. **Note**: Only supports Sentencepiece based tokenizers ! Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id after ctc collapse. decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] built_token = [] previous_token_index = 0 # For every collapsed sub-word token for i, char in enumerate(hypothesis.text): # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). token = decode_ids_to_tokens([char])[0] token_text = decode_tokens_to_str([char]) # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped # after forcing partial text conversion of the token. if token != token_text: # If there are any partially or fully built sub-word token ids, construct to text. # Note: This is "old" subword, that occurs *after* current sub-word has started. if len(built_token) > 0: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[previous_token_index]["start_offset"], "end_offset": offsets[i]["start_offset"], } ) # Prepare list of new sub-word ids built_token.clear() built_token.append(char) previous_token_index = i else: # If the token does not contain any sub-word start mark, then the sub-word has not completed yet # Append to current sub-word list. built_token.append(char) # Inject the start offset of the first token to word offsets # This is because we always skip the delay the injection of the first sub-word due to the loop # condition and check whether built token is ready or not. # Therefore without this forced injection, the start_offset appears as off by 1. if len(word_offsets) == 0: # alaptev: sometimes word_offsets can be empty if len(built_token) > 0: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[0]["start_offset"], "end_offset": offsets[-1]["end_offset"], } ) built_token.clear() else: word_offsets[0]["start_offset"] = offsets[0]["start_offset"] # If there are any remaining tokens left, inject them all into the final word offset. # Note: The start offset of this token is the start time of the first token inside build_token. # Note: The end offset of this token is the end time of the last token inside build_token if len(built_token) > 0: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[-(len(built_token))]["start_offset"], "end_offset": offsets[-1]["end_offset"], } ) built_token.clear() return word_offsets @property def preserve_alignments(self): return self._preserve_alignments @preserve_alignments.setter def preserve_alignments(self, value): self._preserve_alignments = value if hasattr(self, 'decoding'): self.decoding.preserve_alignments = value @property def compute_timestamps(self): return self._compute_timestamps @compute_timestamps.setter def compute_timestamps(self, value): self._compute_timestamps = value if hasattr(self, 'decoding'): self.decoding.compute_timestamps = value @property def preserve_frame_confidence(self): return self._preserve_frame_confidence @preserve_frame_confidence.setter def preserve_frame_confidence(self, value): self._preserve_frame_confidence = value if hasattr(self, 'decoding'): self.decoding.preserve_frame_confidence = value class CTCDecoding(AbstractCTCDecoding): """ Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs for character based models. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy (for greedy decoding). - beam (for DeepSpeed KenLM based decoding). compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. Can take the following values - "char" for character/subword time stamps, "word" for word level time stamps and "all" (default), for both character level and word level time stamps. word_seperator: Str token representing the seperator between words. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. The config may further contain the following sub-dictionaries: "greedy": preserve_alignments: Same as above, overrides above value. compute_timestamps: Same as above, overrides above value. preserve_frame_confidence: Same as above, overrides above value. confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. beam_alpha: float, the strength of the Language model on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. beam_beta: float, the strength of the sequence length penalty on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). If the path is invalid (file is not found at path), will raise a deferred error at the moment of calculation of beam search, so that users may update / change the decoding strategy to point to the correct file. blank_id: The id of the RNNT blank token. """ def __init__( self, decoding_cfg, vocabulary, ): blank_id = len(vocabulary) self.vocabulary = vocabulary self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) super().__init__(decoding_cfg=decoding_cfg, blank_id=blank_id) # Finalize Beam Search Decoding framework if isinstance(self.decoding, ctc_beam_decoding.AbstractBeamCTCInfer): self.decoding.set_vocabulary(self.vocabulary) self.decoding.set_decoding_type('char') def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: """ Implemented by subclass in order to aggregate token confidence to a word-level confidence. Args: hypothesis: Hypothesis Returns: A list of word-level confidence scores. """ return self._aggregate_token_confidence_chars( self.decode_tokens_to_str(hypothesis.text[0]).split(), hypothesis.token_confidence ) def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) return hypothesis def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ token_list = [self.labels_map[c] for c in tokens if c != self.blank_id] return token_list class WER(Metric): """ This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. When doing distributed training/evaluation the result of ``res=WER(predictions, targets, target_lengths)`` calls will be all-reduced between all workers using SUM operations. Here ``res`` contains three numbers ``res=[wer, total_levenstein_distance, total_number_of_words]``. If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. Example: def validation_step(self, batch, batch_idx): ... wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} return self.val_outputs def on_validation_epoch_end(self): ... wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} self.val_outputs.clear() # free memory return {'val_loss': val_loss_mean, 'log': tensorboard_logs} Args: decoding: An instance of CTCDecoding. use_cer: Whether to use Character Error Rate instead of Word Error Rate. log_prediction: Whether to log a single decoded sample per call. fold_consecutive: Whether repeated consecutive characters should be folded into one when decoding. Returns: res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenstein's distances for all prediction - reference pairs, total number of words in all references. """ full_state_update: bool = True def __init__( self, decoding: CTCDecoding, use_cer=False, log_prediction=True, fold_consecutive=True, dist_sync_on_step=False, ): super().__init__(dist_sync_on_step=dist_sync_on_step) self.decoding = decoding self.use_cer = use_cer self.log_prediction = log_prediction self.fold_consecutive = fold_consecutive self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) def update( self, predictions: torch.Tensor, targets: torch.Tensor, target_lengths: torch.Tensor, predictions_lengths: torch.Tensor = None, ): """ Updates metric state. Args: predictions: an integer torch.Tensor of shape ``[Batch, Time, {Vocabulary}]`` (if ``batch_dim_index == 0``) or ``[Time, Batch]`` (if ``batch_dim_index == 1``) targets: an integer torch.Tensor of shape ``[Batch, Time]`` (if ``batch_dim_index == 0``) or ``[Time, Batch]`` (if ``batch_dim_index == 1``) target_lengths: an integer torch.Tensor of shape ``[Batch]`` predictions_lengths: an integer torch.Tensor of shape ``[Batch]`` """ words = 0 scores = 0 references = [] with torch.no_grad(): # prediction_cpu_tensor = tensors[0].long().cpu() targets_cpu_tensor = targets.long().cpu() tgt_lenths_cpu_tensor = target_lengths.long().cpu() # iterate over batch for ind in range(targets_cpu_tensor.shape[0]): tgt_len = tgt_lenths_cpu_tensor[ind].item() target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() reference = self.decoding.decode_tokens_to_str(target) references.append(reference) hypotheses, _ = self.decoding.ctc_decoder_predictions_tensor( predictions, predictions_lengths, fold_consecutive=self.fold_consecutive ) if self.log_prediction: logging.info(f"\n") logging.info(f"reference:{references[0]}") logging.info(f"predicted:{hypotheses[0]}") for h, r in zip(hypotheses, references): if self.use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() words += len(r_list) # Compute Levenstein's distance scores += editdistance.eval(h_list, r_list) self.scores = torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) self.words = torch.tensor(words, device=self.words.device, dtype=self.words.dtype) # return torch.tensor([scores, words]).to(predictions.device) def compute(self): scores = self.scores.detach().float() words = self.words.detach().float() return scores / words, scores, words @dataclass class CTCDecodingConfig: strategy: str = "greedy" # preserve decoding alignments preserve_alignments: Optional[bool] = None # compute ctc time stamps compute_timestamps: Optional[bool] = None # token representing word seperator word_seperator: str = " " # type of timestamps to calculate ctc_timestamp_type: str = "all" # can be char, word or all for both # batch dimension batch_dim_index: int = 0 # greedy decoding config greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() # beam decoding config beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) # confidence config confidence_cfg: ConfidenceConfig = ConfidenceConfig() # can be used to change temperature for decoding temperature: float = 1.0 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/asr/models/configs/aligner_config.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @dataclass class AlignerCTCConfig: prob_suppress_index: int = -1 prob_suppress_value: float = 1.0 @dataclass class AlignerRNNTConfig: predictor_window_size: int = 0 predictor_step_size: int = 1 @dataclass class AlignerWrapperModelConfig: alignment_type: str = "forced" word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Dict, List, Optional from omegaconf import MISSING import nemo.core.classes.dataset from nemo.collections.asr.metrics.wer import CTCDecodingConfig from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMelSpectrogramPreprocessorConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig from nemo.core.config import modelPT as model_cfg @dataclass class ASRDatasetConfig(nemo.core.classes.dataset.DatasetConfig): manifest_filepath: Optional[Any] = None sample_rate: int = MISSING labels: List[str] = MISSING trim_silence: bool = False # Tarred dataset support is_tarred: bool = False tarred_audio_filepaths: Optional[Any] = None tarred_shard_strategy: str = "scatter" shard_manifests: bool = False shuffle_n: int = 0 # Optional int_values: Optional[int] = None augmentor: Optional[Dict[str, Any]] = None max_duration: Optional[float] = None min_duration: Optional[float] = None max_utts: int = 0 blank_index: int = -1 unk_index: int = -1 normalize: bool = False trim: bool = True parser: Optional[str] = 'en' eos_id: Optional[int] = None bos_id: Optional[int] = None pad_id: int = 0 use_start_end_token: bool = False return_sample_id: Optional[bool] = False # bucketing params bucketing_strategy: str = "synced_randomized" bucketing_batch_size: Optional[Any] = None bucketing_weights: Optional[List[int]] = None @dataclass class EncDecCTCConfig(model_cfg.ModelConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = False labels: List[str] = MISSING # Dataset configs train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model component configs preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() encoder: ConvASREncoderConfig = ConvASREncoderConfig() decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() decoding: CTCDecodingConfig = CTCDecodingConfig() @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): model: EncDecCTCConfig = EncDecCTCConfig() @dataclass class CacheAwareStreamingConfig: chunk_size: int = 0 # the size of each chunk at each step, it can be a list of two integers to specify different chunk sizes for the first step and others shift_size: int = 0 # the size of the shift in each step, it can be a list of two integers to specify different shift sizes for the first step and others cache_drop_size: int = 0 # the number of steps to drop from the cache last_channel_cache_size: int = 0 # the size of the needed cache for last channel layers valid_out_len: int = 0 # the number of the steps in the final output which are valid (have the same value as in the offline mode) pre_encode_cache_size: int = 0 # the size of the needed cache for the pre-encoding part of the model to avoid caching inside the pre-encoding layers drop_extra_pre_encoded: int = 0 # the number of steps to get dropped after the pre-encoding layer last_channel_num: int = 0 # number of the last channel layers (like MHA layers) which need caching in the model last_time_num: int = 0 # number of the last time layers (like convolutions) which need caching in the model [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Dict, List, Optional from omegaconf import MISSING import nemo.core.classes.dataset from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMFCCPreprocessorConfig, CropOrPadSpectrogramAugmentationConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ConvASRDecoderClassificationConfig, ConvASREncoderConfig from nemo.core.config import modelPT as model_cfg @dataclass class EncDecClassificationDatasetConfig(nemo.core.classes.dataset.DatasetConfig): manifest_filepath: Optional[str] = None sample_rate: int = MISSING labels: List[str] = MISSING trim_silence: bool = False # Tarred dataset support is_tarred: bool = False tarred_audio_filepaths: Optional[str] = None tarred_shard_strategy: str = "scatter" shuffle_n: int = 0 # Optional int_values: Optional[int] = None augmentor: Optional[Dict[str, Any]] = None max_duration: Optional[float] = None min_duration: Optional[float] = None cal_labels_occurrence: Optional[bool] = False # VAD Optional vad_stream: Optional[bool] = None window_length_in_sec: float = 0.31 shift_length_in_sec: float = 0.01 normalize_audio: bool = False is_regression_task: bool = False # bucketing params bucketing_strategy: str = "synced_randomized" bucketing_batch_size: Optional[Any] = None bucketing_weights: Optional[List[int]] = None @dataclass class EncDecClassificationConfig(model_cfg.ModelConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = True kernel_size_factor: float = 1.0 labels: List[str] = MISSING timesteps: int = MISSING # Dataset configs train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=True, trim_silence=False ) validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model component configs preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( audio_length=timesteps ) encoder: ConvASREncoderConfig = ConvASREncoderConfig() decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): model: EncDecClassificationConfig = EncDecClassificationConfig() [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import asdict, dataclass from typing import Any, Dict, Optional, Tuple, Union @dataclass class DiarizerComponentConfig: """Dataclass to imitate HydraConfig dict when accessing parameters.""" def get(self, name: str, default: Optional[Any] = None): return getattr(self, name, default) def __iter__(self): for key in asdict(self): yield key def dict(self) -> Dict: return asdict(self) @dataclass class ASRDiarizerCTCDecoderParams: pretrained_language_model: Optional[str] = None # KenLM model file: .arpa model file or .bin binary file. beam_width: int = 32 alpha: float = 0.5 beta: float = 2.5 @dataclass class ASRRealigningLMParams: # Provide a KenLM language model in .arpa format. arpa_language_model: Optional[str] = None # Min number of words for the left context. min_number_of_words: int = 3 # Max number of words for the right context. max_number_of_words: int = 10 # The threshold for the difference between two log probability values from two hypotheses. logprob_diff_threshold: float = 1.2 @dataclass class ASRDiarizerParams(DiarizerComponentConfig): # if True, speech segmentation for diarization is based on word-timestamps from ASR inference. asr_based_vad: bool = False # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD. asr_based_vad_threshold: float = 1.0 # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null. asr_batch_size: Optional[int] = None # Native decoder delay. null is recommended to use the default values for each ASR model. decoder_delay_in_sec: Optional[float] = None # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. word_ts_anchor_offset: Optional[float] = None # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. word_ts_anchor_pos: str = "start" # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. fix_word_ts_with_VAD: bool = False # If True, use colored text to distinguish speakers in the output transcript. colored_text: bool = False # If True, the start and end time of each speaker turn is printed in the output transcript. print_time: bool = True # If True, the output transcript breaks the line to fix the line width (default is 90 chars) break_lines: bool = False @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" parameters: ASRDiarizerParams = ASRDiarizerParams() ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() @dataclass class VADParams(DiarizerComponentConfig): window_length_in_sec: float = 0.15 # Window length in sec for VAD context input shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech offset: float = 0.1 # Offset threshold for detecting the end of a speech pad_onset: float = 0.1 # Adding durations before each speech segment pad_offset: float = 0 # Adding durations after each speech segment min_duration_on: float = 0 # Threshold for small non_speech deletion min_duration_off: float = 0.2 # Threshold for short speech segment deletion filter_speech_first: bool = True @dataclass class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None parameters: VADParams = VADParams() @dataclass class SpeakerEmbeddingsParams(DiarizerComponentConfig): # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. save_embeddings: bool = True @dataclass class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() @dataclass class ClusteringParams(DiarizerComponentConfig): # If True, use num of speakers value provided in manifest file. oracle_num_speakers: bool = False # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. max_num_speakers: int = 8 # If the number of segments is lower than this number, enhanced speaker counting is activated. enhanced_count_thres: int = 80 # Determines the range of p-value search: 0 < p <= max_rp_threshold. max_rp_threshold: float = 0.25 # The higher the number, the more values will be examined with more time. sparse_search_volume: int = 30 # If True, take a majority vote on multiple p-values to estimate the number of speakers. maj_vote_spk_count: bool = False @dataclass class ClusteringConfig(DiarizerComponentConfig): parameters: ClusteringParams = ClusteringParams() @dataclass class MSDDParams(DiarizerComponentConfig): # If True, use speaker embedding model in checkpoint, else provided speaker embedding model in config will be used. use_speaker_model_from_ckpt: bool = True # Batch size for MSDD inference. infer_batch_size: int = 25 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. sigmoid_threshold: Tuple[float] = (0.7,) # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. seq_eval_mode: bool = False # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. split_infer: bool = True # The length of split short sequence when split_infer is True. diar_window_length: int = 50 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. overlap_infer_spk_limit: int = 5 @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" parameters: MSDDParams = MSDDParams() @dataclass class DiarizerConfig(DiarizerComponentConfig): manifest_filepath: Optional[str] = None out_dir: Optional[str] = None oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring vad: VADConfig = VADConfig() speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() clustering: ClusteringConfig = ClusteringConfig() msdd_model: MSDDConfig = MSDDConfig() asr: ASRDiarizerConfig = ASRDiarizerConfig() @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): diarizer: DiarizerConfig = DiarizerConfig() device: str = "cpu" verbose: bool = False batch_size: int = 64 num_workers: int = 1 sample_rate: int = 16000 name: str = "" @classmethod def init_config(cls, diar_model_path: str, vad_model_path: str, map_location: str, verbose: bool): return NeuralDiarizerInferenceConfig( DiarizerConfig( vad=VADConfig(model_path=vad_model_path), msdd_model=MSDDConfig(model_path=diar_model_path), ), device=map_location, verbose=verbose, ) [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig from nemo.core.config.modelPT import NemoConfig @dataclass class GraphModuleConfig: criterion_type: str = "ml" loss_type: str = "ctc" split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True backend_cfg: BackendConfig = BackendConfig() @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): graph_module_cfg: GraphModuleConfig = GraphModuleConfig() @dataclass class EncDecK2SeqModelConfig(NemoConfig): model: EncDecK2SeqConfig = EncDecK2SeqConfig() [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING from nemo.collections.asr.models.configs import classification_models_config as clf_cfg from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMFCCPreprocessorConfig, CropOrPadSpectrogramAugmentationConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ( ConvASRDecoderClassificationConfig, ConvASREncoderConfig, JasperEncoderConfig, ) from nemo.core.config import modelPT as model_cfg # fmt: off def matchboxnet_3x1x64(): config = [ JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config def matchboxnet_3x1x64_vad(): config = [ JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config # fmt: on @dataclass class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = True kernel_size_factor: float = 1.0 timesteps: int = 128 labels: List[str] = MISSING # Dataset configs train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=True, trim_silence=False ) validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model general component configs preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 ) crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( audio_length=128 ) encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() @dataclass class MatchboxNetVADModelConfig(MatchboxNetModelConfig): timesteps: int = 64 labels: List[str] = field(default_factory=lambda: ['background', 'speech']) crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = None class EncDecClassificationModelConfigBuilder(model_cfg.ModelConfigBuilder): VALID_CONFIGS = ['matchboxnet_3x1x64', 'matchboxnet_3x1x64_vad'] def __init__(self, name: str = 'matchboxnet_3x1x64', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): if name not in EncDecClassificationModelConfigBuilder.VALID_CONFIGS: raise ValueError("`name` must be one of : \n" f"{EncDecClassificationModelConfigBuilder.VALID_CONFIGS}") self.name = name if 'matchboxnet_3x1x64_vad' in name: if encoder_cfg_func is None: encoder_cfg_func = matchboxnet_3x1x64_vad model_cfg = MatchboxNetVADModelConfig( repeat=1, separable=True, encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderClassificationConfig(), ) elif 'matchboxnet_3x1x64' in name: if encoder_cfg_func is None: encoder_cfg_func = matchboxnet_3x1x64 model_cfg = MatchboxNetModelConfig( repeat=1, separable=False, spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderClassificationConfig(), ) else: raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") super(EncDecClassificationModelConfigBuilder, self).__init__(model_cfg) self.model_cfg: clf_cfg.EncDecClassificationConfig = model_cfg # enable type hinting def set_labels(self, labels: List[str]): self.model_cfg.labels = labels def set_separable(self, separable: bool): self.model_cfg.separable = separable def set_repeat(self, repeat: int): self.model_cfg.repeat = repeat def set_sample_rate(self, sample_rate: int): self.model_cfg.sample_rate = sample_rate def set_dropout(self, dropout: float = 0.0): self.model_cfg.dropout = dropout def set_timesteps(self, timesteps: int): self.model_cfg.timesteps = timesteps def set_is_regression_task(self, is_regression_task: bool): self.model_cfg.is_regression_task = is_regression_task # Note: Autocomplete for users wont work without these overrides # But practically it is not needed since python will infer at runtime # def set_train_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): # super().set_train_ds(cfg) # # def set_validation_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): # super().set_validation_ds(cfg) # # def set_test_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): # super().set_test_ds(cfg) def _finalize_cfg(self): # propagate labels self.model_cfg.train_ds.labels = self.model_cfg.labels self.model_cfg.validation_ds.labels = self.model_cfg.labels self.model_cfg.test_ds.labels = self.model_cfg.labels self.model_cfg.decoder.vocabulary = self.model_cfg.labels # propagate num classes self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) # propagate sample rate self.model_cfg.sample_rate = self.model_cfg.sample_rate self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate # propagate filters self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters # propagate timeteps if self.model_cfg.crop_or_pad_augment is not None: self.model_cfg.crop_or_pad_augment.audio_length = self.model_cfg.timesteps # propagate separable for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig layer.separable = self.model_cfg.separable # propagate repeat for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig layer.repeat = self.model_cfg.repeat # propagate dropout for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig layer.dropout = self.model_cfg.dropout def build(self) -> clf_cfg.EncDecClassificationConfig: return super().build() [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Callable, List, Optional from omegaconf import MISSING from nemo.collections.asr.models.configs import asr_models_config as ctc_cfg from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMelSpectrogramPreprocessorConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig, JasperEncoderConfig from nemo.core.config import modelPT as model_cfg # fmt: off def qn_15x5(): config = [ JasperEncoderConfig(filters=256, repeat=1, kernel=[33], stride=[2], dilation=[1], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=1, kernel=[87], stride=[1], dilation=[2], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config def jasper_10x5_dr(): config = [ JasperEncoderConfig(filters=256, repeat=1, kernel=[11], stride=[2], dilation=[1], dropout=0.2, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=896, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.4, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.4, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config # fmt: on @dataclass class JasperModelConfig(ctc_cfg.EncDecCTCConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = False labels: List[str] = MISSING # Dataset configs train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( manifest_filepath=None, shuffle=True, trim_silence=True ) validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model general component configs preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() @dataclass class QuartzNetModelConfig(JasperModelConfig): separable: bool = True class EncDecCTCModelConfigBuilder(model_cfg.ModelConfigBuilder): VALID_CONFIGS = ['quartznet_15x5', 'quartznet_15x5_zh', 'jasper_10x5dr'] def __init__(self, name: str = 'quartznet_15x5', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): if name not in EncDecCTCModelConfigBuilder.VALID_CONFIGS: raise ValueError("`name` must be one of : \n" f"{EncDecCTCModelConfigBuilder.VALID_CONFIGS}") self.name = name if 'quartznet_15x5' in name: if encoder_cfg_func is None: encoder_cfg_func = qn_15x5 model_cfg = QuartzNetModelConfig( repeat=5, separable=True, spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderConfig(), ) elif 'jasper_10x5' in name: if encoder_cfg_func is None: encoder_cfg_func = jasper_10x5_dr model_cfg = JasperModelConfig( repeat=5, separable=False, spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderConfig(), ) else: raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") super(EncDecCTCModelConfigBuilder, self).__init__(model_cfg) self.model_cfg: ctc_cfg.EncDecCTCConfig = model_cfg # enable type hinting if 'zh' in name: self.set_dataset_normalize(normalize=False) def set_labels(self, labels: List[str]): self.model_cfg.labels = labels def set_separable(self, separable: bool): self.model_cfg.separable = separable def set_repeat(self, repeat: int): self.model_cfg.repeat = repeat def set_sample_rate(self, sample_rate: int): self.model_cfg.sample_rate = sample_rate def set_dropout(self, dropout: float = 0.0): self.model_cfg.dropout = dropout def set_dataset_normalize(self, normalize: bool): self.model_cfg.train_ds.normalize = normalize self.model_cfg.validation_ds.normalize = normalize self.model_cfg.test_ds.normalize = normalize # Note: Autocomplete for users wont work without these overrides # But practically it is not needed since python will infer at runtime # def set_train_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): # super().set_train_ds(cfg) # # def set_validation_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): # super().set_validation_ds(cfg) # # def set_test_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): # super().set_test_ds(cfg) def _finalize_cfg(self): # propagate labels self.model_cfg.train_ds.labels = self.model_cfg.labels self.model_cfg.validation_ds.labels = self.model_cfg.labels self.model_cfg.test_ds.labels = self.model_cfg.labels self.model_cfg.decoder.vocabulary = self.model_cfg.labels # propagate num classes self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) # propagate sample rate self.model_cfg.sample_rate = self.model_cfg.sample_rate self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate # propagate filters self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters # propagate separable for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig layer.separable = self.model_cfg.separable # propagate repeat for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig layer.repeat = self.model_cfg.repeat # propagate dropout for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig layer.dropout = self.model_cfg.dropout def build(self) -> ctc_cfg.EncDecCTCConfig: return super().build() [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import random from abc import ABC, abstractmethod from dataclasses import dataclass from typing import Any, Dict, Optional, Tuple import torch from packaging import version from nemo.collections.asr.parts.numba.spec_augment import SpecAugmentNumba, spec_augment_launch_heuristics from nemo.collections.asr.parts.preprocessing.features import ( FilterbankFeatures, FilterbankFeaturesTA, make_seq_mask_like, ) from nemo.collections.asr.parts.submodules.spectr_augment import SpecAugment, SpecCutout from nemo.core.classes import Exportable, NeuralModule, typecheck from nemo.core.neural_types import ( AudioSignal, LengthsType, MelSpectrogramType, MFCCSpectrogramType, NeuralType, SpectrogramType, ) from nemo.core.utils import numba_utils from nemo.core.utils.numba_utils import __NUMBA_MINIMUM_VERSION__ from nemo.utils import logging try: import torchaudio import torchaudio.functional import torchaudio.transforms TORCHAUDIO_VERSION = version.parse(torchaudio.__version__) TORCHAUDIO_VERSION_MIN = version.parse('0.5') HAVE_TORCHAUDIO = True except ModuleNotFoundError: HAVE_TORCHAUDIO = False __all__ = [ 'AudioToMelSpectrogramPreprocessor', 'AudioToSpectrogram', 'SpectrogramToAudio', 'AudioToMFCCPreprocessor', 'SpectrogramAugmentation', 'MaskedPatchAugmentation', 'CropOrPadSpectrogramAugmentation', ] class AudioPreprocessor(NeuralModule, ABC): """ An interface for Neural Modules that performs audio pre-processing, transforming the wav files to features. """ def __init__(self, win_length, hop_length): super().__init__() self.win_length = win_length self.hop_length = hop_length self.torch_windows = { 'hann': torch.hann_window, 'hamming': torch.hamming_window, 'blackman': torch.blackman_window, 'bartlett': torch.bartlett_window, 'ones': torch.ones, None: torch.ones, } @typecheck() @torch.no_grad() def forward(self, input_signal, length): processed_signal, processed_length = self.get_features(input_signal, length) return processed_signal, processed_length @abstractmethod def get_features(self, input_signal, length): # Called by forward(). Subclasses should implement this. pass class AudioToMelSpectrogramPreprocessor(AudioPreprocessor, Exportable): """Featurizer module that converts wavs to mel spectrograms. Args: sample_rate (int): Sample rate of the input audio data. Defaults to 16000 window_size (float): Size of window for fft in seconds Defaults to 0.02 window_stride (float): Stride of window for fft in seconds Defaults to 0.01 n_window_size (int): Size of window for fft in samples Defaults to None. Use one of window_size or n_window_size. n_window_stride (int): Stride of window for fft in samples Defaults to None. Use one of window_stride or n_window_stride. window (str): Windowing function for fft. can be one of ['hann', 'hamming', 'blackman', 'bartlett'] Defaults to "hann" normalize (str): Can be one of ['per_feature', 'all_features']; all other options disable feature normalization. 'all_features' normalizes the entire spectrogram to be mean 0 with std 1. 'pre_features' normalizes per channel / freq instead. Defaults to "per_feature" n_fft (int): Length of FT window. If None, it uses the smallest power of 2 that is larger than n_window_size. Defaults to None preemph (float): Amount of pre emphasis to add to audio. Can be disabled by passing None. Defaults to 0.97 features (int): Number of mel spectrogram freq bins to output. Defaults to 64 lowfreq (int): Lower bound on mel basis in Hz. Defaults to 0 highfreq (int): Lower bound on mel basis in Hz. Defaults to None log (bool): Log features. Defaults to True log_zero_guard_type(str): Need to avoid taking the log of zero. There are two options: "add" or "clamp". Defaults to "add". log_zero_guard_value(float, or str): Add or clamp requires the number to add with or clamp to. log_zero_guard_value can either be a float or "tiny" or "eps". torch.finfo is used if "tiny" or "eps" is passed. Defaults to 2**-24. dither (float): Amount of white-noise dithering. Defaults to 1e-5 pad_to (int): Ensures that the output size of the time dimension is a multiple of pad_to. Defaults to 16 frame_splicing (int): Defaults to 1 exact_pad (bool): If True, sets stft center to False and adds padding, such that num_frames = audio_length // hop_length. Defaults to False. pad_value (float): The value that shorter mels are padded with. Defaults to 0 mag_power (float): The power that the linear spectrogram is raised to prior to multiplication with mel basis. Defaults to 2 for a power spec rng : Random number generator nb_augmentation_prob (float) : Probability with which narrowband augmentation would be applied to samples in the batch. Defaults to 0.0 nb_max_freq (int) : Frequency above which all frequencies will be masked for narrowband augmentation. Defaults to 4000 use_torchaudio: Whether to use the `torchaudio` implementation. mel_norm: Normalization used for mel filterbank weights. Defaults to 'slaney' (area normalization) stft_exact_pad: Deprecated argument, kept for compatibility with older checkpoints. stft_conv: Deprecated argument, kept for compatibility with older checkpoints. """ def save_to(self, save_path: str): pass @classmethod def restore_from(cls, restore_path: str): pass @property def input_types(self): """Returns definitions of module input ports. """ return { "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), "length": NeuralType( tuple('B'), LengthsType() ), # Please note that length should be in samples not seconds. } @property def output_types(self): """Returns definitions of module output ports. processed_signal: 0: AxisType(BatchTag) 1: AxisType(MelSpectrogramSignalTag) 2: AxisType(ProcessedTimeTag) processed_length: 0: AxisType(BatchTag) """ return { "processed_signal": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "processed_length": NeuralType(tuple('B'), LengthsType()), } def __init__( self, sample_rate=16000, window_size=0.02, window_stride=0.01, n_window_size=None, n_window_stride=None, window="hann", normalize="per_feature", n_fft=None, preemph=0.97, features=64, lowfreq=0, highfreq=None, log=True, log_zero_guard_type="add", log_zero_guard_value=2 ** -24, dither=1e-5, pad_to=16, frame_splicing=1, exact_pad=False, pad_value=0, mag_power=2.0, rng=None, nb_augmentation_prob=0.0, nb_max_freq=4000, use_torchaudio: bool = False, mel_norm="slaney", stft_exact_pad=False, # Deprecated arguments; kept for config compatibility stft_conv=False, # Deprecated arguments; kept for config compatibility ): super().__init__(n_window_size, n_window_stride) self._sample_rate = sample_rate if window_size and n_window_size: raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") if window_stride and n_window_stride: raise ValueError( f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." ) if window_size: n_window_size = int(window_size * self._sample_rate) if window_stride: n_window_stride = int(window_stride * self._sample_rate) # Given the long and similar argument list, point to the class and instantiate it by reference if not use_torchaudio: featurizer_class = FilterbankFeatures else: featurizer_class = FilterbankFeaturesTA self.featurizer = featurizer_class( sample_rate=self._sample_rate, n_window_size=n_window_size, n_window_stride=n_window_stride, window=window, normalize=normalize, n_fft=n_fft, preemph=preemph, nfilt=features, lowfreq=lowfreq, highfreq=highfreq, log=log, log_zero_guard_type=log_zero_guard_type, log_zero_guard_value=log_zero_guard_value, dither=dither, pad_to=pad_to, frame_splicing=frame_splicing, exact_pad=exact_pad, pad_value=pad_value, mag_power=mag_power, rng=rng, nb_augmentation_prob=nb_augmentation_prob, nb_max_freq=nb_max_freq, mel_norm=mel_norm, stft_exact_pad=stft_exact_pad, # Deprecated arguments; kept for config compatibility stft_conv=stft_conv, # Deprecated arguments; kept for config compatibility ) def input_example(self, max_batch: int = 8, max_dim: int = 32000, min_length: int = 200): batch_size = torch.randint(low=1, high=max_batch, size=[1]).item() max_length = torch.randint(low=min_length, high=max_dim, size=[1]).item() signals = torch.rand(size=[batch_size, max_length]) * 2 - 1 lengths = torch.randint(low=min_length, high=max_dim, size=[batch_size]) lengths[0] = max_length return signals, lengths def get_features(self, input_signal, length): return self.featurizer(input_signal, length) @property def filter_banks(self): return self.featurizer.filter_banks class AudioToMFCCPreprocessor(AudioPreprocessor): """Preprocessor that converts wavs to MFCCs. Uses torchaudio.transforms.MFCC. Args: sample_rate: The sample rate of the audio. Defaults to 16000. window_size: Size of window for fft in seconds. Used to calculate the win_length arg for mel spectrogram. Defaults to 0.02 window_stride: Stride of window for fft in seconds. Used to caculate the hop_length arg for mel spect. Defaults to 0.01 n_window_size: Size of window for fft in samples Defaults to None. Use one of window_size or n_window_size. n_window_stride: Stride of window for fft in samples Defaults to None. Use one of window_stride or n_window_stride. window: Windowing function for fft. can be one of ['hann', 'hamming', 'blackman', 'bartlett', 'none', 'null']. Defaults to 'hann' n_fft: Length of FT window. If None, it uses the smallest power of 2 that is larger than n_window_size. Defaults to None lowfreq (int): Lower bound on mel basis in Hz. Defaults to 0 highfreq (int): Lower bound on mel basis in Hz. Defaults to None n_mels: Number of mel filterbanks. Defaults to 64 n_mfcc: Number of coefficients to retain Defaults to 64 dct_type: Type of discrete cosine transform to use norm: Type of norm to use log: Whether to use log-mel spectrograms instead of db-scaled. Defaults to True. """ @property def input_types(self): """Returns definitions of module input ports. """ return { "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return { "processed_signal": NeuralType(('B', 'D', 'T'), MFCCSpectrogramType()), "processed_length": NeuralType(tuple('B'), LengthsType()), } def save_to(self, save_path: str): pass @classmethod def restore_from(cls, restore_path: str): pass def __init__( self, sample_rate=16000, window_size=0.02, window_stride=0.01, n_window_size=None, n_window_stride=None, window='hann', n_fft=None, lowfreq=0.0, highfreq=None, n_mels=64, n_mfcc=64, dct_type=2, norm='ortho', log=True, ): self._sample_rate = sample_rate if not HAVE_TORCHAUDIO: logging.error('Could not import torchaudio. Some features might not work.') raise ModuleNotFoundError( "torchaudio is not installed but is necessary for " "AudioToMFCCPreprocessor. We recommend you try " "building it from source for the PyTorch version you have." ) if window_size and n_window_size: raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") if window_stride and n_window_stride: raise ValueError( f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." ) # Get win_length (n_window_size) and hop_length (n_window_stride) if window_size: n_window_size = int(window_size * self._sample_rate) if window_stride: n_window_stride = int(window_stride * self._sample_rate) super().__init__(n_window_size, n_window_stride) mel_kwargs = {} mel_kwargs['f_min'] = lowfreq mel_kwargs['f_max'] = highfreq mel_kwargs['n_mels'] = n_mels mel_kwargs['n_fft'] = n_fft or 2 ** math.ceil(math.log2(n_window_size)) mel_kwargs['win_length'] = n_window_size mel_kwargs['hop_length'] = n_window_stride # Set window_fn. None defaults to torch.ones. window_fn = self.torch_windows.get(window, None) if window_fn is None: raise ValueError( f"Window argument for AudioProcessor is invalid: {window}." f"For no window function, use 'ones' or None." ) mel_kwargs['window_fn'] = window_fn # Use torchaudio's implementation of MFCCs as featurizer self.featurizer = torchaudio.transforms.MFCC( sample_rate=self._sample_rate, n_mfcc=n_mfcc, dct_type=dct_type, norm=norm, log_mels=log, melkwargs=mel_kwargs, ) def get_features(self, input_signal, length): features = self.featurizer(input_signal) seq_len = torch.ceil(length.to(torch.float32) / self.hop_length).to(dtype=torch.long) return features, seq_len class SpectrogramAugmentation(NeuralModule): """ Performs time and freq cuts in one of two ways. SpecAugment zeroes out vertical and horizontal sections as described in SpecAugment (https://arxiv.org/abs/1904.08779). Arguments for use with SpecAugment are `freq_masks`, `time_masks`, `freq_width`, and `time_width`. SpecCutout zeroes out rectangulars as described in Cutout (https://arxiv.org/abs/1708.04552). Arguments for use with Cutout are `rect_masks`, `rect_freq`, and `rect_time`. Args: freq_masks (int): how many frequency segments should be cut. Defaults to 0. time_masks (int): how many time segments should be cut Defaults to 0. freq_width (int): maximum number of frequencies to be cut in one segment. Defaults to 10. time_width (int): maximum number of time steps to be cut in one segment Defaults to 10. rect_masks (int): how many rectangular masks should be cut Defaults to 0. rect_freq (int): maximum size of cut rectangles along the frequency dimension Defaults to 5. rect_time (int): maximum size of cut rectangles along the time dimension Defaults to 25. """ @property def input_types(self): """Returns definitions of module input types """ return { "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output types """ return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} def __init__( self, freq_masks=0, time_masks=0, freq_width=10, time_width=10, rect_masks=0, rect_time=5, rect_freq=20, rng=None, mask_value=0.0, use_numba_spec_augment: bool = True, ): super().__init__() if rect_masks > 0: self.spec_cutout = SpecCutout(rect_masks=rect_masks, rect_time=rect_time, rect_freq=rect_freq, rng=rng,) # self.spec_cutout.to(self._device) else: self.spec_cutout = lambda input_spec: input_spec if freq_masks + time_masks > 0: self.spec_augment = SpecAugment( freq_masks=freq_masks, time_masks=time_masks, freq_width=freq_width, time_width=time_width, rng=rng, mask_value=mask_value, ) else: self.spec_augment = lambda input_spec, length: input_spec # Check if numba is supported, and use a Numba kernel if it is if use_numba_spec_augment and numba_utils.numba_cuda_is_supported(__NUMBA_MINIMUM_VERSION__): logging.info('Numba CUDA SpecAugment kernel is being used') self.spec_augment_numba = SpecAugmentNumba( freq_masks=freq_masks, time_masks=time_masks, freq_width=freq_width, time_width=time_width, rng=rng, mask_value=mask_value, ) else: self.spec_augment_numba = None @typecheck() def forward(self, input_spec, length): augmented_spec = self.spec_cutout(input_spec=input_spec) # To run the Numba kernel, correct numba version is required as well as # tensor must be on GPU and length must be provided if self.spec_augment_numba is not None and spec_augment_launch_heuristics(augmented_spec, length): augmented_spec = self.spec_augment_numba(input_spec=augmented_spec, length=length) else: augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) return augmented_spec class MaskedPatchAugmentation(NeuralModule): """ Zeroes out fixed size time patches of the spectrogram. All samples in batch are guaranteed to have the same amount of masked time steps. Optionally also performs frequency masking in the same way as SpecAugment. Args: patch_size (int): up to how many time steps does one patch consist of. Defaults to 48. mask_patches (float): how many patches should be masked in each sample. if >= 1., interpreted as number of patches (after converting to int) if <1., interpreted as fraction of total tokens to be masked (number of patches is rounded up) Defaults to 10. freq_masks (int): how many frequency segments should be cut. Defaults to 0. freq_width (int): maximum number of frequencies to be cut in a segment. Defaults to 0. """ @property def input_types(self): """Returns definitions of module input types """ return { "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output types """ return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} def __init__( self, patch_size: int = 48, mask_patches: float = 10.0, freq_masks: int = 0, freq_width: int = 0, ): super().__init__() self.patch_size = patch_size if mask_patches >= 1: self.mask_patches = int(mask_patches) elif mask_patches >= 0: self._mask_fraction = mask_patches self.mask_patches = None else: raise ValueError('mask_patches cannot be negative') if freq_masks > 0: self.spec_augment = SpecAugment(freq_masks=freq_masks, time_masks=0, freq_width=freq_width, time_width=0,) else: self.spec_augment = None @typecheck() def forward(self, input_spec, length): augmented_spec = input_spec min_len = torch.min(length) if self.mask_patches is None: # masking specified as fraction len_fraction = int(min_len * self._mask_fraction) mask_patches = len_fraction // self.patch_size + int(len_fraction % self.patch_size != 0) else: mask_patches = self.mask_patches if min_len < self.patch_size * mask_patches: mask_patches = min_len // self.patch_size for idx in range(input_spec.shape[0]): cur_len = length[idx] patches = range(cur_len // self.patch_size) masked_patches = random.sample(patches, mask_patches) for mp in masked_patches: augmented_spec[idx, :, mp * self.patch_size : (mp + 1) * self.patch_size] = 0.0 if self.spec_augment is not None: augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) return augmented_spec class CropOrPadSpectrogramAugmentation(NeuralModule): """ Pad or Crop the incoming Spectrogram to a certain shape. Args: audio_length (int): the final number of timesteps that is required. The signal will be either padded or cropped temporally to this size. """ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length @typecheck() @torch.no_grad() def forward(self, input_signal, length): image = input_signal num_images = image.shape[0] audio_length = self.audio_length image_len = image.shape[-1] # Crop long signal if image_len > audio_length: # randomly slice cutout_images = [] offset = torch.randint(low=0, high=image_len - audio_length + 1, size=[num_images]) for idx, offset in enumerate(offset): cutout_images.append(image[idx : idx + 1, :, offset : offset + audio_length]) image = torch.cat(cutout_images, dim=0) del cutout_images else: # symmetrically pad short signal with zeros pad_left = (audio_length - image_len) // 2 pad_right = (audio_length - image_len) // 2 if (audio_length - image_len) % 2 == 1: pad_right += 1 image = torch.nn.functional.pad(image, [pad_left, pad_right], mode="constant", value=0) # Replace dynamic length sequences with static number of timesteps length = (length * 0) + audio_length return image, length @property def input_types(self): """Returns definitions of module output ports. """ return { "input_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return { "processed_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), "processed_length": NeuralType(tuple('B'), LengthsType()), } def save_to(self, save_path: str): pass @classmethod def restore_from(cls, restore_path: str): pass class AudioToSpectrogram(NeuralModule): """Transform a batch of input multi-channel signals into a batch of STFT-based spectrograms. Args: fft_length: length of FFT hop_length: length of hops/shifts of the sliding window power: exponent for magnitude spectrogram. Default `None` will return a complex-valued spectrogram """ def __init__(self, fft_length: int, hop_length: int, power: Optional[float] = None): if not HAVE_TORCHAUDIO: logging.error('Could not import torchaudio. Some features might not work.') raise ModuleNotFoundError( "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" ) super().__init__() # For now, assume FFT length is divisible by two if fft_length % 2 != 0: raise ValueError(f'fft_length = {fft_length} must be divisible by 2') self.stft = torchaudio.transforms.Spectrogram( n_fft=fft_length, hop_length=hop_length, power=power, pad_mode='constant' ) # number of subbands self.F = fft_length // 2 + 1 @property def num_subbands(self) -> int: return self.F @property def input_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "input": NeuralType(('B', 'C', 'T'), AudioSignal()), "input_length": NeuralType(('B',), LengthsType(), optional=True), } @property def output_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "output": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), "output_length": NeuralType(('B',), LengthsType()), } @typecheck() def forward( self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None ) -> Tuple[torch.Tensor, torch.Tensor]: """Convert a batch of C-channel input signals into a batch of complex-valued spectrograms. Args: input: Time-domain input signal with C channels, shape (B, C, T) input_length: Length of valid entries along the time dimension, shape (B,) Returns: Output spectrogram with F subbands and N time frames, shape (B, C, F, N) and output length with shape (B,). """ B, T = input.size(0), input.size(-1) input = input.view(B, -1, T) # STFT output (B, C, F, N) with torch.cuda.amp.autocast(enabled=False): output = self.stft(input.float()) if input_length is not None: # Mask padded frames output_length = self.get_output_length(input_length=input_length) length_mask: torch.Tensor = make_seq_mask_like( lengths=output_length, like=output, time_dim=-1, valid_ones=False ) output = output.masked_fill(length_mask, 0.0) else: # Assume all frames are valid for all examples in the batch output_length = output.size(-1) * torch.ones(B, device=output.device).long() return output, output_length def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: """Get length of valid frames for the output. Args: input_length: number of valid samples, shape (B,) Returns: Number of valid frames, shape (B,) """ output_length = input_length.div(self.stft.hop_length, rounding_mode='floor').add(1).long() return output_length class SpectrogramToAudio(NeuralModule): """Transform a batch of input multi-channel spectrograms into a batch of time-domain multi-channel signals. Args: fft_length: length of FFT hop_length: length of hops/shifts of the sliding window power: exponent for magnitude spectrogram. Default `None` will return a complex-valued spectrogram """ def __init__(self, fft_length: int, hop_length: int): if not HAVE_TORCHAUDIO: logging.error('Could not import torchaudio. Some features might not work.') raise ModuleNotFoundError( "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" ) super().__init__() # For now, assume FFT length is divisible by two if fft_length % 2 != 0: raise ValueError(f'fft_length = {fft_length} must be divisible by 2') self.istft = torchaudio.transforms.InverseSpectrogram( n_fft=fft_length, hop_length=hop_length, pad_mode='constant' ) self.F = fft_length // 2 + 1 @property def num_subbands(self) -> int: return self.F @property def input_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "input": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), "input_length": NeuralType(('B',), LengthsType(), optional=True), } @property def output_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "output": NeuralType(('B', 'C', 'T'), AudioSignal()), "output_length": NeuralType(('B',), LengthsType()), } @typecheck() def forward(self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None) -> torch.Tensor: """Convert input complex-valued spectrogram to a time-domain signal. Multi-channel IO is supported. Args: input: Input spectrogram for C channels, shape (B, C, F, N) input_length: Length of valid entries along the time dimension, shape (B,) Returns: Time-domain signal with T time-domain samples and C channels, (B, C, T) and output length with shape (B,). """ B, F, N = input.size(0), input.size(-2), input.size(-1) assert F == self.F, f'Number of subbands F={F} not matching self.F={self.F}' input = input.view(B, -1, F, N) # iSTFT output (B, C, T) with torch.cuda.amp.autocast(enabled=False): output = self.istft(input.cfloat()) if input_length is not None: # Mask padded samples output_length = self.get_output_length(input_length=input_length) length_mask: torch.Tensor = make_seq_mask_like( lengths=output_length, like=output, time_dim=-1, valid_ones=False ) output = output.masked_fill(length_mask, 0.0) else: # Assume all frames are valid for all examples in the batch output_length = output.size(-1) * torch.ones(B, device=output.device).long() return output, output_length def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: """Get length of valid samples for the output. Args: input_length: number of valid frames, shape (B,) Returns: Number of valid samples, shape (B,) """ output_length = input_length.sub(1).mul(self.istft.hop_length).long() return output_length @dataclass class AudioToMelSpectrogramPreprocessorConfig: _target_: str = "nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor" sample_rate: int = 16000 window_size: float = 0.02 window_stride: float = 0.01 n_window_size: Optional[int] = None n_window_stride: Optional[int] = None window: str = "hann" normalize: str = "per_feature" n_fft: Optional[int] = None preemph: float = 0.97 features: int = 64 lowfreq: int = 0 highfreq: Optional[int] = None log: bool = True log_zero_guard_type: str = "add" log_zero_guard_value: float = 2 ** -24 dither: float = 1e-5 pad_to: int = 16 frame_splicing: int = 1 exact_pad: bool = False pad_value: int = 0 mag_power: float = 2.0 rng: Optional[str] = None nb_augmentation_prob: float = 0.0 nb_max_freq: int = 4000 use_torchaudio: bool = False mel_norm: str = "slaney" stft_exact_pad: bool = False # Deprecated argument, kept for compatibility with older checkpoints. stft_conv: bool = False # Deprecated argument, kept for compatibility with older checkpoints. @dataclass class AudioToMFCCPreprocessorConfig: _target_: str = 'nemo.collections.asr.modules.AudioToMFCCPreprocessor' sample_rate: int = 16000 window_size: float = 0.02 window_stride: float = 0.01 n_window_size: Optional[int] = None n_window_stride: Optional[int] = None window: str = 'hann' n_fft: Optional[int] = None lowfreq: Optional[float] = 0.0 highfreq: Optional[float] = None n_mels: int = 64 n_mfcc: int = 64 dct_type: int = 2 norm: str = 'ortho' log: bool = True @dataclass class SpectrogramAugmentationConfig: _target_: str = "nemo.collections.asr.modules.SpectrogramAugmentation" freq_masks: int = 0 time_masks: int = 0 freq_width: int = 0 time_width: Optional[Any] = 0 rect_masks: int = 0 rect_time: int = 0 rect_freq: int = 0 mask_value: float = 0 rng: Optional[Any] = None # random.Random() type use_numba_spec_augment: bool = True @dataclass class CropOrPadSpectrogramAugmentationConfig: audio_length: int _target_: str = "nemo.collections.asr.modules.CropOrPadSpectrogramAugmentation" @dataclass class MaskedPatchAugmentationConfig: patch_size: int = 48 mask_patches: float = 10.0 freq_masks: int = 0 freq_width: int = 0 _target_: str = "nemo.collections.asr.modules.MaskedPatchAugmentation" [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/parts/k2/classes.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from abc import ABC from dataclasses import dataclass from typing import Any, Optional, Tuple import torch from omegaconf import DictConfig from nemo.utils import logging @dataclass class GraphIntersectDenseConfig: """Graph dense intersection config. """ search_beam: float = 20.0 output_beam: float = 10.0 min_active_states: int = 30 max_active_states: int = 10000 @dataclass class GraphModuleConfig: """Config for graph modules. Typically used with graph losses and decoders. """ topo_type: str = "default" topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 class ASRK2Mixin(ABC): """k2 Mixin class that simplifies the construction of various models with k2-based losses. It does the following: - Sets up the graph loss and decoder (methods _init_k2 and update_k2_modules). - Registers external graphs, if needed. - Augments forward(...) with optional graph decoding to get accurate predictions. """ def _init_k2(self): """ k2-related initialization implementation. This method is expected to run after the __init__ which sets self._cfg self._cfg is expected to have the attribute graph_module_cfg """ if not hasattr(self, "_cfg"): raise ValueError("self._cfg must be set before calling _init_k2().") if not hasattr(self._cfg, "graph_module_cfg") or self._cfg.graph_module_cfg is None: raise ValueError("self._cfg.graph_module_cfg must be set and cannot be None.") self.graph_module_cfg = self._cfg.graph_module_cfg # register token_lm for MAPLoss criterion_type = self.graph_module_cfg.get("criterion_type", "ml") self.use_graph_lm = criterion_type == "map" if self.use_graph_lm: token_lm_path = self.graph_module_cfg.backend_cfg.get("token_lm", None) if token_lm_path is None: raise ValueError( f"graph_module_cfg.backend_cfg.token_lm is empty. It must be set for criterion_type == `{criterion_type}`" ) token_lm_path = self.register_artifact('graph_module_cfg.backend_cfg.token_lm', token_lm_path) self.graph_module_cfg.backend_cfg["token_lm"] = token_lm_path self.update_k2_modules(self.graph_module_cfg) def update_k2_modules(self, input_cfg: DictConfig): """ Helper function to initialize or update k2 loss and transcribe_decoder. Args: input_cfg: DictConfig to take new parameters from. Schema is expected as in nemo.collections.asr.models.configs.k2_sequence_models_config.GraphModuleConfig """ del self.loss if hasattr(self, "transcribe_decoder"): del self.transcribe_decoder if hasattr(self, "joint"): # RNNT num_classes = self.joint.num_classes_with_blank - 1 else: # CTC, MMI, ... num_classes = self.decoder.num_classes_with_blank - 1 remove_consecutive = input_cfg.backend_cfg.get("topo_with_self_loops", True) and input_cfg.backend_cfg.get( "topo_type", "default" ) not in ["forced_blank", "identity",] self._wer.remove_consecutive = remove_consecutive from nemo.collections.asr.losses.lattice_losses import LatticeLoss self.loss = LatticeLoss( num_classes=num_classes, reduction=self._cfg.get("ctc_reduction", "mean_batch"), backend="k2", criterion_type=input_cfg.get("criterion_type", "ml"), loss_type=input_cfg.get("loss_type", "ctc"), split_batch_size=input_cfg.get("split_batch_size", 0), graph_module_cfg=input_cfg.backend_cfg, ) criterion_type = self.loss.criterion_type self.use_graph_lm = criterion_type == "map" transcribe_training = input_cfg.get("transcribe_training", False) if transcribe_training and criterion_type == "ml": logging.warning( f"""You do not need to use transcribe_training=`{transcribe_training}` with criterion_type=`{criterion_type}`. transcribe_training will be set to False.""" ) transcribe_training = False self.transcribe_training = transcribe_training if self.use_graph_lm: from nemo.collections.asr.modules.graph_decoder import ViterbiDecoderWithGraph self.transcribe_decoder = ViterbiDecoderWithGraph( num_classes=num_classes, backend="k2", dec_type="token_lm", return_type="1best", return_ilabels=True, output_aligned=True, split_batch_size=input_cfg.get("split_batch_size", 0), graph_module_cfg=input_cfg.backend_cfg, ) def _forward_k2_post_processing( self, log_probs: torch.Tensor, encoded_length: torch.Tensor, greedy_predictions: torch.Tensor ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """ k2-related post-processing parf of .forward() Args: log_probs: The log probabilities tensor of shape [B, T, D]. encoded_length: The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. greedy_predictions: The greedy token predictions of the model of shape [B, T] Returns: A tuple of 3 elements - 1) The log probabilities tensor of shape [B, T, D]. 2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 3) The greedy token predictions of the model of shape [B, T] (via argmax) """ # greedy_predictions from .forward() are incorrect for criterion_type=`map` # getting correct greedy_predictions, if needed if self.use_graph_lm and (not self.training or self.transcribe_training): greedy_predictions, encoded_length, _ = self.transcribe_decoder.forward( log_probs=log_probs, log_probs_length=encoded_length ) return log_probs, encoded_length, greedy_predictions [end of nemo/collections/asr/parts/k2/classes.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math from dataclasses import dataclass from typing import Any, Optional import torch from torch import nn as nn from nemo.collections.asr.parts.submodules import multi_head_attention as mha from nemo.collections.common.parts import adapter_modules from nemo.core.classes.mixins import adapter_mixin_strategies class MHAResidualAddAdapterStrategy(adapter_mixin_strategies.ResidualAddAdapterStrategy): """ An implementation of residual addition of an adapter module with its input for the MHA Adapters. """ def forward(self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin'): """ A basic strategy, comprising of a residual connection over the input, after forward pass by the underlying adapter. Additional work is done to pack and unpack the dictionary of inputs and outputs. Note: The `value` tensor is added to the output of the attention adapter as the residual connection. Args: input: A dictionary of multiple input arguments for the adapter module. `query`, `key`, `value`: Original output tensor of the module, or the output of the previous adapter (if more than one adapters are enabled). `mask`: Attention mask. `pos_emb`: Optional positional embedding for relative encoding. adapter: The adapter module that is currently required to perform the forward pass. module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, therefore the strategy can access all other adapters in this module via `module.adapter_layer`. Returns: The result tensor, after one of the active adapters has finished its forward passes. """ out = self.compute_output(input, adapter, module=module) # If not in training mode, or probability of stochastic depth is 0, skip step. p = self.stochastic_depth if not module.training or p == 0.0: pass else: out = self.apply_stochastic_depth(out, input['value'], adapter, module=module) # Return the residual connection output = input + adapter(input) result = input['value'] + out # If l2_lambda is activated, register the loss value self.compute_auxiliary_losses(result, input['value'], adapter, module=module) return result def compute_output( self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin' ) -> torch.Tensor: """ Compute the output of a single adapter to some input. Args: input: Original output tensor of the module, or the output of the previous adapter (if more than one adapters are enabled). adapter: The adapter module that is currently required to perform the forward pass. module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, therefore the strategy can access all other adapters in this module via `module.adapter_layer`. Returns: The result tensor, after one of the active adapters has finished its forward passes. """ if isinstance(input, (list, tuple)): out = adapter(*input) elif isinstance(input, dict): out = adapter(**input) else: out = adapter(input) return out @dataclass class MHAResidualAddAdapterStrategyConfig(adapter_mixin_strategies.ResidualAddAdapterStrategyConfig): _target_: str = "{0}.{1}".format( MHAResidualAddAdapterStrategy.__module__, MHAResidualAddAdapterStrategy.__name__ ) # mandatory field class MultiHeadAttentionAdapter(mha.MultiHeadAttention, adapter_modules.AdapterModuleUtil): """Multi-Head Attention layer of Transformer. Args: n_head (int): number of heads n_feat (int): size of the features dropout_rate (float): dropout rate proj_dim (int, optional): Optional integer value for projection before computing attention. If None, then there is no projection (equivalent to proj_dim = n_feat). If > 0, then will project the n_feat to proj_dim before calculating attention. If <0, then will equal n_head, so that each head has a projected dimension of 1. adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, n_head: int, n_feat: int, dropout_rate: float, proj_dim: Optional[int] = None, adapter_strategy: MHAResidualAddAdapterStrategy = None, ): super().__init__(n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, max_cache_len=0) self.pre_norm = nn.LayerNorm(n_feat) # Set the projection dim to number of heads automatically if proj_dim is not None and proj_dim < 1: proj_dim = n_head self.proj_dim = proj_dim # Recompute weights for projection dim if self.proj_dim is not None: if self.proj_dim % n_head != 0: raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") self.d_k = self.proj_dim // n_head self.s_d_k = math.sqrt(self.d_k) self.linear_q = nn.Linear(n_feat, self.proj_dim) self.linear_k = nn.Linear(n_feat, self.proj_dim) self.linear_v = nn.Linear(n_feat, self.proj_dim) self.linear_out = nn.Linear(self.proj_dim, n_feat) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) # reset parameters for Q to be identity operation self.reset_parameters() def forward(self, query, key, value, mask, pos_emb=None, cache=None): """Compute 'Scaled Dot Product Attention'. Args: query (torch.Tensor): (batch, time1, size) key (torch.Tensor): (batch, time2, size) value(torch.Tensor): (batch, time2, size) mask (torch.Tensor): (batch, time1, time2) cache (torch.Tensor) : (batch, time_cache, size) returns: output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention cache (torch.Tensor) : (batch, time_cache_next, size) """ # Need to perform duplicate computations as at this point the tensors have been # separated by the adapter forward query = self.pre_norm(query) key = self.pre_norm(key) value = self.pre_norm(value) return super().forward(query, key, value, mask, pos_emb, cache=cache) def reset_parameters(self): with torch.no_grad(): nn.init.zeros_(self.linear_out.weight) nn.init.zeros_(self.linear_out.bias) def get_default_strategy_config(self) -> 'dataclass': return MHAResidualAddAdapterStrategyConfig() @dataclass class MultiHeadAttentionAdapterConfig: n_head: int n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): """Multi-Head Attention layer of Transformer-XL with support of relative positional encoding. Paper: https://arxiv.org/abs/1901.02860 Args: n_head (int): number of heads n_feat (int): size of the features dropout_rate (float): dropout rate proj_dim (int, optional): Optional integer value for projection before computing attention. If None, then there is no projection (equivalent to proj_dim = n_feat). If > 0, then will project the n_feat to proj_dim before calculating attention. If <0, then will equal n_head, so that each head has a projected dimension of 1. adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, n_head: int, n_feat: int, dropout_rate: float, proj_dim: Optional[int] = None, adapter_strategy: MHAResidualAddAdapterStrategyConfig = None, ): super().__init__( n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, pos_bias_u=None, pos_bias_v=None, max_cache_len=0 ) self.pre_norm = nn.LayerNorm(n_feat) # Set the projection dim to number of heads automatically if proj_dim is not None and proj_dim < 1: proj_dim = n_head self.proj_dim = proj_dim # Recompute weights for projection dim if self.proj_dim is not None: if self.proj_dim % n_head != 0: raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") self.d_k = self.proj_dim // n_head self.s_d_k = math.sqrt(self.d_k) self.linear_q = nn.Linear(n_feat, self.proj_dim) self.linear_k = nn.Linear(n_feat, self.proj_dim) self.linear_v = nn.Linear(n_feat, self.proj_dim) self.linear_out = nn.Linear(self.proj_dim, n_feat) self.linear_pos = nn.Linear(n_feat, self.proj_dim, bias=False) self.pos_bias_u = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) self.pos_bias_v = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) # reset parameters for Q to be identity operation self.reset_parameters() def forward(self, query, key, value, mask, pos_emb, cache=None): """Compute 'Scaled Dot Product Attention' with rel. positional encoding. Args: query (torch.Tensor): (batch, time1, size) key (torch.Tensor): (batch, time2, size) value(torch.Tensor): (batch, time2, size) mask (torch.Tensor): (batch, time1, time2) pos_emb (torch.Tensor) : (batch, time1, size) cache (torch.Tensor) : (batch, time_cache, size) Returns: output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention cache_next (torch.Tensor) : (batch, time_cache_next, size) """ # Need to perform duplicate computations as at this point the tensors have been # separated by the adapter forward query = self.pre_norm(query) key = self.pre_norm(key) value = self.pre_norm(value) return super().forward(query, key, value, mask, pos_emb, cache=cache) def reset_parameters(self): with torch.no_grad(): nn.init.zeros_(self.linear_out.weight) nn.init.zeros_(self.linear_out.bias) # NOTE: This exact procedure apparently highly important. # Above operation is safe to do as self.linear_out.weight *= 0.0 (similar for bias) # However: # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 # For some reason at init sometimes it will cause the value of the tensor to become NaN # All operations to compute matrix_ac and matrix_bd will then fail. nn.init.zeros_(self.pos_bias_u) nn.init.zeros_(self.pos_bias_v) def get_default_strategy_config(self) -> 'dataclass': return MHAResidualAddAdapterStrategyConfig() @dataclass class RelPositionMultiHeadAttentionAdapterConfig: n_head: int n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) class PositionalEncodingAdapter(mha.PositionalEncoding, adapter_modules.AdapterModuleUtil): """ Absolute positional embedding adapter. .. note:: Absolute positional embedding value is added to the input tensor *without residual connection* ! Therefore, the input is changed, if you only require the positional embedding, drop the returned `x` ! Args: d_model (int): The input dimension of x. max_len (int): The max sequence length. xscale (float): The input scaling factor. Defaults to 1.0. adapter_strategy (AbstractAdapterStrategy): By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. NOTE: Since this is a positional encoding, it will not add a residual ! """ def __init__( self, d_model: int, max_len: int = 5000, xscale=1.0, adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, ): super().__init__( d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, ) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) def get_default_strategy_config(self) -> 'dataclass': return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() @dataclass class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): """ Relative positional encoding for TransformerXL's layers See : Appendix B in https://arxiv.org/abs/1901.02860 .. note:: Relative positional embedding value is **not** added to the input tensor ! Therefore, the input should be updated changed, if you only require the positional embedding, drop the returned `x` ! Args: d_model (int): embedding dim max_len (int): maximum input length xscale (bool): whether to scale the input by sqrt(d_model) adapter_strategy: By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, d_model: int, max_len: int = 5000, xscale=1.0, adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, ): super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) def get_default_strategy_config(self) -> 'dataclass': return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() @dataclass class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import os from dataclasses import dataclass from typing import List, Optional, Tuple, Union import torch from nemo.collections.asr.parts.utils import rnnt_utils from nemo.collections.common.tokenizers.tokenizer_spec import TokenizerSpec from nemo.core.classes import Typing, typecheck from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType from nemo.utils import logging DEFAULT_TOKEN_OFFSET = 100 def pack_hypotheses( hypotheses: List[rnnt_utils.NBestHypotheses], logitlen: torch.Tensor, ) -> List[rnnt_utils.NBestHypotheses]: if logitlen is not None: if hasattr(logitlen, 'cpu'): logitlen_cpu = logitlen.to('cpu') else: logitlen_cpu = logitlen for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.NBestHypotheses for candidate_idx, cand in enumerate(hyp.n_best_hypotheses): cand.y_sequence = torch.tensor(cand.y_sequence, dtype=torch.long) if logitlen is not None: cand.length = logitlen_cpu[idx] if cand.dec_state is not None: cand.dec_state = _states_to_device(cand.dec_state) return hypotheses def _states_to_device(dec_state, device='cpu'): if torch.is_tensor(dec_state): dec_state = dec_state.to(device) elif isinstance(dec_state, (list, tuple)): dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) return dec_state class AbstractBeamCTCInfer(Typing): """A beam CTC decoder. Provides a common abstraction for sample level beam decoding. Args: blank_id: int, index of the blank token. Can be 0 or len(vocabulary). beam_size: int, size of the beam used in the underlying beam search engine. """ @property def input_types(self): """Returns definitions of module input ports. """ return { "decoder_output": NeuralType(('B', 'T', 'D'), LogprobsType()), "decoder_lengths": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return {"predictions": [NeuralType(elements_type=HypothesisType())]} def __init__(self, blank_id: int, beam_size: int): self.blank_id = blank_id if beam_size < 1: raise ValueError("Beam search size cannot be less than 1!") self.beam_size = beam_size # Variables set by corresponding setter methods self.vocab = None self.decoding_type = None self.tokenizer = None # Utility maps for vocabulary self.vocab_index_map = None self.index_vocab_map = None # Internal variable, used to prevent double reduction of consecutive tokens (ctc collapse) self.override_fold_consecutive_value = None def set_vocabulary(self, vocab: List[str]): """ Set the vocabulary of the decoding framework. Args: vocab: List of str. Each token corresponds to its location in the vocabulary emitted by the model. Note that this vocabulary must NOT contain the "BLANK" token. """ self.vocab = vocab self.vocab_index_map = {v: i for i, v in enumerate(vocab)} self.index_vocab_map = {i: v for i, v in enumerate(vocab)} def set_decoding_type(self, decoding_type: str): """ Sets the decoding type of the framework. Can support either char or subword models. Args: decoding_type: Str corresponding to decoding type. Only supports "char" and "subword". """ decoding_type = decoding_type.lower() supported_types = ['char', 'subword'] if decoding_type not in supported_types: raise ValueError( f"Unsupported decoding type. Supported types = {supported_types}.\n" f"Given = {decoding_type}" ) self.decoding_type = decoding_type def set_tokenizer(self, tokenizer: TokenizerSpec): """ Set the tokenizer of the decoding framework. Args: tokenizer: NeMo tokenizer object, which inherits from TokenizerSpec. """ self.tokenizer = tokenizer @typecheck() def forward( self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-repressively. Args: decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). decoder_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ raise NotImplementedError() def __call__(self, *args, **kwargs): return self.forward(*args, **kwargs) class BeamCTCInfer(AbstractBeamCTCInfer): """A greedy CTC decoder. Provides a common abstraction for sample level and batch level greedy decoding. Args: blank_index: int index of the blank token. Can be 0 or len(vocabulary). preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. """ def __init__( self, blank_id: int, beam_size: int, search_type: str = "default", return_best_hypothesis: bool = True, preserve_alignments: bool = False, compute_timestamps: bool = False, beam_alpha: float = 1.0, beam_beta: float = 0.0, kenlm_path: str = None, flashlight_cfg: Optional['FlashlightConfig'] = None, pyctcdecode_cfg: Optional['PyCTCDecodeConfig'] = None, ): super().__init__(blank_id=blank_id, beam_size=beam_size) self.search_type = search_type self.return_best_hypothesis = return_best_hypothesis self.preserve_alignments = preserve_alignments self.compute_timestamps = compute_timestamps if self.compute_timestamps: raise ValueError(f"Currently this flag is not supported for beam search algorithms.") self.vocab = None # This must be set by specific method by user before calling forward() ! if search_type == "default" or search_type == "nemo": self.search_algorithm = self.default_beam_search elif search_type == "pyctcdecode": self.search_algorithm = self._pyctcdecode_beam_search elif search_type == "flashlight": self.search_algorithm = self.flashlight_beam_search else: raise NotImplementedError( f"The search type ({search_type}) supplied is not supported!\n" f"Please use one of : (default, nemo, pyctcdecode)" ) # Log the beam search algorithm logging.info(f"Beam search algorithm: {search_type}") self.beam_alpha = beam_alpha self.beam_beta = beam_beta # Default beam search args self.kenlm_path = kenlm_path # PyCTCDecode params if pyctcdecode_cfg is None: pyctcdecode_cfg = PyCTCDecodeConfig() self.pyctcdecode_cfg = pyctcdecode_cfg # type: PyCTCDecodeConfig if flashlight_cfg is None: flashlight_cfg = FlashlightConfig() self.flashlight_cfg = flashlight_cfg # Default beam search scorer functions self.default_beam_scorer = None self.pyctcdecode_beam_scorer = None self.flashlight_beam_scorer = None self.token_offset = 0 @typecheck() def forward( self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-repressively. Args: decoder_output: A tensor of size (batch, timesteps, features). decoder_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ if self.vocab is None: raise RuntimeError("Please set the vocabulary with `set_vocabulary()` before calling this function.") if self.decoding_type is None: raise ValueError("Please set the decoding type with `set_decoding_type()` before calling this function.") with torch.no_grad(), torch.inference_mode(): # Process each sequence independently prediction_tensor = decoder_output if prediction_tensor.ndim != 3: raise ValueError( f"`decoder_output` must be a tensor of shape [B, T, V] (log probs, float). " f"Provided shape = {prediction_tensor.shape}" ) # determine type of input - logprobs or labels out_len = decoder_lengths if decoder_lengths is not None else None hypotheses = self.search_algorithm(prediction_tensor, out_len) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, decoder_lengths) # Pack the result if self.return_best_hypothesis and isinstance(packed_result[0], rnnt_utils.NBestHypotheses): packed_result = [res.n_best_hypotheses[0] for res in packed_result] # type: Hypothesis return (packed_result,) @torch.no_grad() def default_beam_search( self, x: torch.Tensor, out_len: torch.Tensor ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: """ Open Seq2Seq Beam Search Algorithm (DeepSpeed) Args: x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, and V is the vocabulary size. The tensor contains log-probabilities. out_len: Tensor of shape [B], contains lengths of each sequence in the batch. Returns: A list of NBestHypotheses objects, one for each sequence in the batch. """ if self.compute_timestamps: raise ValueError( f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" ) if self.default_beam_scorer is None: # Check for filepath if self.kenlm_path is None or not os.path.exists(self.kenlm_path): raise FileNotFoundError( f"KenLM binary file not found at : {self.kenlm_path}. " f"Please set a valid path in the decoding config." ) # perform token offset for subword models if self.decoding_type == 'subword': vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] else: # char models vocab = self.vocab # Must import at runtime to avoid circular dependency due to module level import. from nemo.collections.asr.modules.beam_search_decoder import BeamSearchDecoderWithLM self.default_beam_scorer = BeamSearchDecoderWithLM( vocab=vocab, lm_path=self.kenlm_path, beam_width=self.beam_size, alpha=self.beam_alpha, beta=self.beam_beta, num_cpus=max(1, os.cpu_count()), input_tensor=False, ) x = x.to('cpu') with typecheck.disable_checks(): data = [x[sample_id, : out_len[sample_id], :].softmax(dim=-1) for sample_id in range(len(x))] beams_batch = self.default_beam_scorer.forward(log_probs=data, log_probs_length=None) # For each sample in the batch nbest_hypotheses = [] for beams_idx, beams in enumerate(beams_batch): # For each beam candidate / hypothesis in each sample hypotheses = [] for candidate_idx, candidate in enumerate(beams): hypothesis = rnnt_utils.Hypothesis( score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None ) # For subword encoding, NeMo will double encode the subword (multiple tokens) into a # singular unicode id. In doing so, we preserve the semantic of the unicode token, and # compress the size of the final KenLM ARPA / Binary file. # In order to do double encoding, we shift the subword by some token offset. # This step is ignored for character based models. if self.decoding_type == 'subword': pred_token_ids = [ord(c) - self.token_offset for c in candidate[1]] else: # Char models pred_token_ids = [self.vocab_index_map[c] for c in candidate[1]] # We preserve the token ids and the score for this hypothesis hypothesis.y_sequence = pred_token_ids hypothesis.score = candidate[0] # If alignment must be preserved, we preserve a view of the output logprobs. # Note this view is shared amongst all beams within the sample, be sure to clone it if you # require specific processing for each sample in the beam. # This is done to preserve memory. if self.preserve_alignments: hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] hypotheses.append(hypothesis) # Wrap the result in NBestHypothesis. hypotheses = rnnt_utils.NBestHypotheses(hypotheses) nbest_hypotheses.append(hypotheses) return nbest_hypotheses @torch.no_grad() def _pyctcdecode_beam_search( self, x: torch.Tensor, out_len: torch.Tensor ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: """ PyCTCDecode Beam Search Algorithm. Should support Char and Subword models. Args: x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, and V is the vocabulary size. The tensor contains log-probabilities. out_len: Tensor of shape [B], contains lengths of each sequence in the batch. Returns: A list of NBestHypotheses objects, one for each sequence in the batch. """ if self.compute_timestamps: raise ValueError( f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" ) try: import pyctcdecode except (ImportError, ModuleNotFoundError): raise ImportError( f"Could not load `pyctcdecode` library. Please install it from pip using :\n" f"pip install --upgrade pyctcdecode" ) if self.pyctcdecode_beam_scorer is None: self.pyctcdecode_beam_scorer = pyctcdecode.build_ctcdecoder( labels=self.vocab, kenlm_model_path=self.kenlm_path, alpha=self.beam_alpha, beta=self.beam_beta ) # type: pyctcdecode.BeamSearchDecoderCTC x = x.to('cpu').numpy() with typecheck.disable_checks(): beams_batch = [] for sample_id in range(len(x)): logprobs = x[sample_id, : out_len[sample_id], :] result = self.pyctcdecode_beam_scorer.decode_beams( logprobs, beam_width=self.beam_size, beam_prune_logp=self.pyctcdecode_cfg.beam_prune_logp, token_min_logp=self.pyctcdecode_cfg.token_min_logp, prune_history=self.pyctcdecode_cfg.prune_history, hotwords=self.pyctcdecode_cfg.hotwords, hotword_weight=self.pyctcdecode_cfg.hotword_weight, lm_start_state=None, ) # Output format: text, last_lm_state, text_frames, logit_score, lm_score beams_batch.append(result) nbest_hypotheses = [] for beams_idx, beams in enumerate(beams_batch): hypotheses = [] for candidate_idx, candidate in enumerate(beams): # Candidate = (text, last_lm_state, text_frames, logit_score, lm_score) hypothesis = rnnt_utils.Hypothesis( score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None ) # TODO: Requires token ids to be returned rather than text. if self.decoding_type == 'subword': if self.tokenizer is None: raise ValueError("Tokenizer must be provided for subword decoding. Use set_tokenizer().") pred_token_ids = self.tokenizer.text_to_ids(candidate[0]) else: if self.vocab is None: raise ValueError("Vocab must be provided for character decoding. Use set_vocab().") chars = list(candidate[0]) pred_token_ids = [self.vocab_index_map[c] for c in chars] hypothesis.y_sequence = pred_token_ids hypothesis.text = candidate[0] # text hypothesis.score = candidate[4] # score # Inject word level timestamps hypothesis.timestep = candidate[2] # text_frames if self.preserve_alignments: hypothesis.alignments = torch.from_numpy(x[beams_idx][: out_len[beams_idx]]) hypotheses.append(hypothesis) hypotheses = rnnt_utils.NBestHypotheses(hypotheses) nbest_hypotheses.append(hypotheses) return nbest_hypotheses @torch.no_grad() def flashlight_beam_search( self, x: torch.Tensor, out_len: torch.Tensor ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: """ Flashlight Beam Search Algorithm. Should support Char and Subword models. Args: x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, and V is the vocabulary size. The tensor contains log-probabilities. out_len: Tensor of shape [B], contains lengths of each sequence in the batch. Returns: A list of NBestHypotheses objects, one for each sequence in the batch. """ if self.compute_timestamps: raise ValueError( f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" ) if self.flashlight_beam_scorer is None: # Check for filepath if self.kenlm_path is None or not os.path.exists(self.kenlm_path): raise FileNotFoundError( f"KenLM binary file not found at : {self.kenlm_path}. " f"Please set a valid path in the decoding config." ) # perform token offset for subword models # if self.decoding_type == 'subword': # vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] # else: # # char models # vocab = self.vocab # Must import at runtime to avoid circular dependency due to module level import. from nemo.collections.asr.modules.flashlight_decoder import FlashLightKenLMBeamSearchDecoder self.flashlight_beam_scorer = FlashLightKenLMBeamSearchDecoder( lm_path=self.kenlm_path, vocabulary=self.vocab, tokenizer=self.tokenizer, lexicon_path=self.flashlight_cfg.lexicon_path, boost_path=self.flashlight_cfg.boost_path, beam_size=self.beam_size, beam_size_token=self.flashlight_cfg.beam_size_token, beam_threshold=self.flashlight_cfg.beam_threshold, lm_weight=self.beam_alpha, word_score=self.beam_beta, unk_weight=self.flashlight_cfg.unk_weight, sil_weight=self.flashlight_cfg.sil_weight, ) x = x.to('cpu') with typecheck.disable_checks(): beams_batch = self.flashlight_beam_scorer.forward(log_probs=x) # For each sample in the batch nbest_hypotheses = [] for beams_idx, beams in enumerate(beams_batch): # For each beam candidate / hypothesis in each sample hypotheses = [] for candidate_idx, candidate in enumerate(beams): hypothesis = rnnt_utils.Hypothesis( score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None ) # We preserve the token ids and the score for this hypothesis hypothesis.y_sequence = candidate['tokens'].tolist() hypothesis.score = candidate['score'] # If alignment must be preserved, we preserve a view of the output logprobs. # Note this view is shared amongst all beams within the sample, be sure to clone it if you # require specific processing for each sample in the beam. # This is done to preserve memory. if self.preserve_alignments: hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] hypotheses.append(hypothesis) # Wrap the result in NBestHypothesis. hypotheses = rnnt_utils.NBestHypotheses(hypotheses) nbest_hypotheses.append(hypotheses) return nbest_hypotheses def set_decoding_type(self, decoding_type: str): super().set_decoding_type(decoding_type) # Please check train_kenlm.py in scripts/asr_language_modeling/ to find out why we need # TOKEN_OFFSET for BPE-based models if self.decoding_type == 'subword': self.token_offset = DEFAULT_TOKEN_OFFSET @dataclass class PyCTCDecodeConfig: # These arguments cannot be imported from pyctcdecode (optional dependency) # Therefore we copy the values explicitly # Taken from pyctcdecode.constant beam_prune_logp: float = -10.0 token_min_logp: float = -5.0 prune_history: bool = False hotwords: Optional[List[str]] = None hotword_weight: float = 10.0 @dataclass class FlashlightConfig: lexicon_path: Optional[str] = None boost_path: Optional[str] = None beam_size_token: int = 16 beam_threshold: float = 20.0 unk_weight: float = -math.inf sil_weight: float = 0.0 @dataclass class BeamCTCInferConfig: beam_size: int search_type: str = 'default' preserve_alignments: bool = False compute_timestamps: bool = False return_best_hypothesis: bool = True beam_alpha: float = 1.0 beam_beta: float = 0.0 kenlm_path: Optional[str] = None flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import List, Optional import torch from omegaconf import DictConfig, OmegaConf from nemo.collections.asr.parts.utils import rnnt_utils from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMeasureConfig, ConfidenceMeasureMixin from nemo.core.classes import Typing, typecheck from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType from nemo.utils import logging def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: if logitlen is not None: if hasattr(logitlen, 'cpu'): logitlen_cpu = logitlen.to('cpu') else: logitlen_cpu = logitlen for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) if logitlen is not None: hyp.length = logitlen_cpu[idx] if hyp.dec_state is not None: hyp.dec_state = _states_to_device(hyp.dec_state) return hypotheses def _states_to_device(dec_state, device='cpu'): if torch.is_tensor(dec_state): dec_state = dec_state.to(device) elif isinstance(dec_state, (list, tuple)): dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) return dec_state class GreedyCTCInfer(Typing, ConfidenceMeasureMixin): """A greedy CTC decoder. Provides a common abstraction for sample level and batch level greedy decoding. Args: blank_index: int index of the blank token. Can be 0 or len(vocabulary). preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ @property def input_types(self): """Returns definitions of module input ports. """ # Input can be of dimention - # ('B', 'T', 'D') [Log probs] or ('B', 'T') [Labels] return { "decoder_output": NeuralType(None, LogprobsType()), "decoder_lengths": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return {"predictions": [NeuralType(elements_type=HypothesisType())]} def __init__( self, blank_id: int, preserve_alignments: bool = False, compute_timestamps: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__() self.blank_id = blank_id self.preserve_alignments = preserve_alignments # we need timestamps to extract non-blank per-frame confidence self.compute_timestamps = compute_timestamps | preserve_frame_confidence self.preserve_frame_confidence = preserve_frame_confidence # set confidence calculation measure self._init_confidence_measure(confidence_measure_cfg) @typecheck() def forward( self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-repressively. Args: decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). decoder_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ with torch.inference_mode(): hypotheses = [] # Process each sequence independently prediction_cpu_tensor = decoder_output.cpu() if prediction_cpu_tensor.ndim < 2 or prediction_cpu_tensor.ndim > 3: raise ValueError( f"`decoder_output` must be a tensor of shape [B, T] (labels, int) or " f"[B, T, V] (log probs, float). Provided shape = {prediction_cpu_tensor.shape}" ) # determine type of input - logprobs or labels if prediction_cpu_tensor.ndim == 2: # labels greedy_decode = self._greedy_decode_labels else: greedy_decode = self._greedy_decode_logprobs for ind in range(prediction_cpu_tensor.shape[0]): out_len = decoder_lengths[ind] if decoder_lengths is not None else None hypothesis = greedy_decode(prediction_cpu_tensor[ind], out_len) hypotheses.append(hypothesis) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, decoder_lengths) return (packed_result,) @torch.no_grad() def _greedy_decode_logprobs(self, x: torch.Tensor, out_len: torch.Tensor): # x: [T, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) prediction = x.detach().cpu() if out_len is not None: prediction = prediction[:out_len] prediction_logprobs, prediction_labels = prediction.max(dim=-1) non_blank_ids = prediction_labels != self.blank_id hypothesis.y_sequence = prediction_labels.numpy().tolist() hypothesis.score = (prediction_logprobs[non_blank_ids]).sum() if self.preserve_alignments: # Preserve the logprobs, as well as labels after argmax hypothesis.alignments = (prediction.clone(), prediction_labels.clone()) if self.compute_timestamps: hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() if self.preserve_frame_confidence: hypothesis.frame_confidence = self._get_confidence(prediction) return hypothesis @torch.no_grad() def _greedy_decode_labels(self, x: torch.Tensor, out_len: torch.Tensor): # x: [T] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) prediction_labels = x.detach().cpu() if out_len is not None: prediction_labels = prediction_labels[:out_len] non_blank_ids = prediction_labels != self.blank_id hypothesis.y_sequence = prediction_labels.numpy().tolist() hypothesis.score = -1.0 if self.preserve_alignments: raise ValueError("Requested for alignments, but predictions provided were labels, not log probabilities.") if self.compute_timestamps: hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() if self.preserve_frame_confidence: raise ValueError( "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." ) return hypothesis def __call__(self, *args, **kwargs): return self.forward(*args, **kwargs) @dataclass class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.confidence_measure_cfg = OmegaConf.structured( self.confidence_measure_cfg if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.confidence_measure_cfg) ) if self.confidence_method_cfg != "DEPRECATED": logging.warning( "`confidence_method_cfg` is deprecated and will be removed in the future. " "Please use `confidence_measure_cfg` instead." ) # TODO (alaptev): delete the following two lines sometime in the future logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") # OmegaConf.structured ensures that post_init check is always executed self.confidence_measure_cfg = OmegaConf.structured( self.confidence_method_cfg if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.confidence_method_cfg) ) [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # Copyright 2017 Johns Hopkins University (Shinji Watanabe) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import List, Optional, Tuple, Union import numpy as np import torch from omegaconf import DictConfig, OmegaConf from nemo.collections.asr.modules import rnnt_abstract from nemo.collections.asr.parts.utils import rnnt_utils from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMeasureConfig, ConfidenceMeasureMixin from nemo.collections.common.parts.rnn import label_collate from nemo.core.classes import Typing, typecheck from nemo.core.neural_types import AcousticEncodedRepresentation, ElementType, HypothesisType, LengthsType, NeuralType from nemo.utils import logging def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: if hasattr(logitlen, 'cpu'): logitlen_cpu = logitlen.to('cpu') else: logitlen_cpu = logitlen for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) hyp.length = logitlen_cpu[idx] if hyp.dec_state is not None: hyp.dec_state = _states_to_device(hyp.dec_state) return hypotheses def _states_to_device(dec_state, device='cpu'): if torch.is_tensor(dec_state): dec_state = dec_state.to(device) elif isinstance(dec_state, (list, tuple)): dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) return dec_state class _GreedyRNNTInfer(Typing, ConfidenceMeasureMixin): """A greedy transducer decoder. Provides a common abstraction for sample level and batch level greedy decoding. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Can be 0 or len(vocabulary). max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ @property def input_types(self): """Returns definitions of module input ports. """ return { "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), "encoded_lengths": NeuralType(tuple('B'), LengthsType()), "partial_hypotheses": [NeuralType(elements_type=HypothesisType(), optional=True)], # must always be last } @property def output_types(self): """Returns definitions of module output ports. """ return {"predictions": [NeuralType(elements_type=HypothesisType())]} def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__() self.decoder = decoder_model self.joint = joint_model self._blank_index = blank_index self._SOS = blank_index # Start of single index self.max_symbols = max_symbols_per_step self.preserve_alignments = preserve_alignments self.preserve_frame_confidence = preserve_frame_confidence # set confidence calculation measure self._init_confidence_measure(confidence_measure_cfg) def __call__(self, *args, **kwargs): return self.forward(*args, **kwargs) @torch.no_grad() def _pred_step( self, label: Union[torch.Tensor, int], hidden: Optional[torch.Tensor], add_sos: bool = False, batch_size: Optional[int] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: """ Common prediction step based on the AbstractRNNTDecoder implementation. Args: label: (int/torch.Tensor): Label or "Start-of-Signal" token. hidden: (Optional torch.Tensor): RNN State vector add_sos (bool): Whether to add a zero vector at the begging as "start of sentence" token. batch_size: Batch size of the output tensor. Returns: g: (B, U, H) if add_sos is false, else (B, U + 1, H) hid: (h, c) where h is the final sequence hidden state and c is the final cell state: h (tensor), shape (L, B, H) c (tensor), shape (L, B, H) """ if isinstance(label, torch.Tensor): # label: [batch, 1] if label.dtype != torch.long: label = label.long() else: # Label is an integer if label == self._SOS: return self.decoder.predict(None, hidden, add_sos=add_sos, batch_size=batch_size) label = label_collate([[label]]) # output: [B, 1, K] return self.decoder.predict(label, hidden, add_sos=add_sos, batch_size=batch_size) def _joint_step(self, enc, pred, log_normalize: Optional[bool] = None): """ Common joint step based on AbstractRNNTJoint implementation. Args: enc: Output of the Encoder model. A torch.Tensor of shape [B, 1, H1] pred: Output of the Decoder model. A torch.Tensor of shape [B, 1, H2] log_normalize: Whether to log normalize or not. None will log normalize only for CPU. Returns: logits of shape (B, T=1, U=1, V + 1) """ with torch.no_grad(): logits = self.joint.joint(enc, pred) if log_normalize is None: if not logits.is_cuda: # Use log softmax only if on CPU logits = logits.log_softmax(dim=len(logits.shape) - 1) else: if log_normalize: logits = logits.log_softmax(dim=len(logits.shape) - 1) return logits class GreedyRNNTInfer(_GreedyRNNTInfer): """A greedy transducer decoder. Sequence level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Can be 0 or len(vocabulary). max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_measure_cfg=confidence_measure_cfg, ) @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) self.decoder.eval() self.joint.eval() hypotheses = [] # Process each sequence independently with self.decoder.as_frozen(), self.joint.as_frozen(): for batch_idx in range(encoder_output.size(0)): inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] logitlen = encoded_lengths[batch_idx] partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) hypotheses.append(hypothesis) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, encoded_lengths) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) @torch.no_grad() def _greedy_decode( self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None ): # x: [T, 1, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) if partial_hypotheses is not None: hypothesis.last_token = partial_hypotheses.last_token hypothesis.y_sequence = ( partial_hypotheses.y_sequence.cpu().tolist() if isinstance(partial_hypotheses.y_sequence, torch.Tensor) else partial_hypotheses.y_sequence ) if partial_hypotheses.dec_state is not None: hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) if self.preserve_alignments: # Alignments is a 2-dimensional dangling list representing T x U hypothesis.alignments = [[]] if self.preserve_frame_confidence: hypothesis.frame_confidence = [[]] # For timestep t in X_t for time_idx in range(out_len): # Extract encoder embedding at timestep t # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] f = x.narrow(dim=0, start=time_idx, length=1) # Setup exit flags and counter not_blank = True symbols_added = 0 # While blank is not predicted, or we dont run out of max symbols per timestep while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # In the first timestep, we initialize the network with RNNT Blank # In later timesteps, we provide previous predicted label as input. if hypothesis.last_token is None and hypothesis.dec_state is None: last_label = self._SOS else: last_label = label_collate([[hypothesis.last_token]]) # Perform prediction network and joint network steps. g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 0, 0, 0, : ] del g # torch.max(0) op doesnt exist for FP 16. if logp.dtype != torch.float32: logp = logp.float() # get index k, of max prob v, k = logp.max(0) k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. if self.preserve_alignments: # insert logprobs into last timestep hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) if self.preserve_frame_confidence: # insert confidence into last timestep hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) del logp # If blank token is predicted, exit inner loop, move onto next timestep t if k == self._blank_index: not_blank = False else: # Append token to label set, update RNN state. hypothesis.y_sequence.append(k) hypothesis.score += float(v) hypothesis.timestep.append(time_idx) hypothesis.dec_state = hidden_prime hypothesis.last_token = k # Increment token counter. symbols_added += 1 if self.preserve_alignments: # convert Ti-th logits into a torch array hypothesis.alignments.append([]) # blank buffer for next timestep if self.preserve_frame_confidence: hypothesis.frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of Alignments if self.preserve_alignments: if len(hypothesis.alignments[-1]) == 0: del hypothesis.alignments[-1] # Remove trailing empty list of per-frame confidence if self.preserve_frame_confidence: if len(hypothesis.frame_confidence[-1]) == 0: del hypothesis.frame_confidence[-1] # Unpack the hidden states hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) return hypothesis class GreedyBatchedRNNTInfer(_GreedyRNNTInfer): """A batch level greedy transducer decoder. Batch level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Can be 0 or len(vocabulary). max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_measure_cfg=confidence_measure_cfg, ) # Depending on availability of `blank_as_pad` support # switch between more efficient batch decoding technique if self.decoder.blank_as_pad: self._greedy_decode = self._greedy_decode_blank_as_pad else: self._greedy_decode = self._greedy_decode_masked @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) logitlen = encoded_lengths self.decoder.eval() self.joint.eval() with self.decoder.as_frozen(), self.joint.as_frozen(): inseq = encoder_output # [B, T, D] hypotheses = self._greedy_decode( inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses ) # Pack the hypotheses results packed_result = pack_hypotheses(hypotheses, logitlen) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) def _greedy_decode_blank_as_pad( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") with torch.inference_mode(): # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize list of Hypothesis batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a dangling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] # If confidence scores need to be preserved, register a dangling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() for time_idx in range(max_out_len): f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) all_blanks = torch.all(blank_mask) del k_is_blank # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx, is_blank in enumerate(blank_mask): # we only want to update non-blanks, unless we are at the last step in the loop where # all elements produced blanks, otherwise there will be duplicate predictions # saved in alignments if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx, is_blank in enumerate(blank_mask): if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if all_blanks: not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.clone().view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses def _greedy_decode_masked( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize state batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] else: alignments = None # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) last_label_without_blank = last_label.clone() # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() with torch.inference_mode(): for time_idx in range(max_out_len): f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Set a dummy label for the blank value # This value will be overwritten by "blank" again the last label update below # This is done as vocabulary of prediction network does not contain "blank" token of RNNT last_label_without_blank_mask = last_label == self._blank_index last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label last_label_without_blank[~last_label_without_blank_mask] = last_label[ ~last_label_without_blank_mask ] # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) all_blanks = torch.all(blank_mask) # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx, is_blank in enumerate(blank_mask): # we only want to update non-blanks, unless we are at the last step in the loop where # all elements produced blanks, otherwise there will be duplicate predictions # saved in alignments if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx, is_blank in enumerate(blank_mask): if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses class ExportedModelGreedyBatchedRNNTInfer: def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = None): self.encoder_model_path = encoder_model self.decoder_joint_model_path = decoder_joint_model self.max_symbols_per_step = max_symbols_per_step # Will be populated at runtime self._blank_index = None def __call__(self, audio_signal: torch.Tensor, length: torch.Tensor): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ with torch.no_grad(): # Apply optional preprocessing encoder_output, encoded_lengths = self.run_encoder(audio_signal=audio_signal, length=length) if torch.is_tensor(encoder_output): encoder_output = encoder_output.transpose(1, 2) else: encoder_output = encoder_output.transpose([0, 2, 1]) # (B, T, D) logitlen = encoded_lengths inseq = encoder_output # [B, T, D] hypotheses, timestamps = self._greedy_decode(inseq, logitlen) # Pack the hypotheses results packed_result = [rnnt_utils.Hypothesis(score=-1.0, y_sequence=[]) for _ in range(len(hypotheses))] for i in range(len(packed_result)): packed_result[i].y_sequence = torch.tensor(hypotheses[i], dtype=torch.long) packed_result[i].length = timestamps[i] del hypotheses return packed_result def _greedy_decode(self, x, out_len): # x: [B, T, D] # out_len: [B] # Initialize state batchsize = x.shape[0] hidden = self._get_initial_states(batchsize) target_lengths = torch.ones(batchsize, dtype=torch.int32) # Output string buffer label = [[] for _ in range(batchsize)] timesteps = [[] for _ in range(batchsize)] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long).numpy() if torch.is_tensor(x): last_label = torch.from_numpy(last_label).to(self.device) # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool).numpy() # Get max sequence length max_out_len = out_len.max() for time_idx in range(max_out_len): f = x[:, time_idx : time_idx + 1, :] # [B, 1, D] if torch.is_tensor(f): f = f.transpose(1, 2) else: f = f.transpose([0, 2, 1]) # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask *= False # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols_per_step is None or symbols_added < self.max_symbols_per_step): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0: g = torch.tensor([self._blank_index] * batchsize, dtype=torch.int32).view(-1, 1) else: if torch.is_tensor(last_label): g = last_label.type(torch.int32) else: g = last_label.astype(np.int32) # Batched joint step - Output = [B, V + 1] joint_out, hidden_prime = self.run_decoder_joint(f, g, target_lengths, *hidden) logp, pred_lengths = joint_out logp = logp[:, 0, 0, :] # Get index k, of max prob for batch if torch.is_tensor(logp): v, k = logp.max(1) else: k = np.argmax(logp, axis=1).astype(np.int32) # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask |= k_is_blank del k_is_blank del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past if torch.is_tensor(blank_mask): blank_indices = (blank_mask == 1).nonzero(as_tuple=False) else: blank_indices = blank_mask.astype(np.int32).nonzero() if type(blank_indices) in (list, tuple): blank_indices = blank_indices[0] # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states for state_id in range(len(hidden)): hidden_prime[state_id][:, blank_indices, :] = hidden[state_id][:, blank_indices, :] elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states for state_id in range(len(hidden_prime)): hidden_prime[state_id][:, blank_indices, :] *= 0.0 # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration if torch.is_tensor(k): last_label = k.clone().reshape(-1, 1) else: last_label = k.copy().reshape(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: label[kidx].append(ki) timesteps[kidx].append(time_idx) symbols_added += 1 return label, timesteps def _setup_blank_index(self): raise NotImplementedError() def run_encoder(self, audio_signal, length): raise NotImplementedError() def run_decoder_joint(self, enc_logits, targets, target_length, *states): raise NotImplementedError() def _get_initial_states(self, batchsize): raise NotImplementedError() class ONNXGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = 10): super().__init__( encoder_model=encoder_model, decoder_joint_model=decoder_joint_model, max_symbols_per_step=max_symbols_per_step, ) try: import onnx import onnxruntime except (ModuleNotFoundError, ImportError): raise ImportError(f"`onnx` or `onnxruntime` could not be imported, please install the libraries.\n") if torch.cuda.is_available(): # Try to use onnxruntime-gpu providers = ['TensorrtExecutionProvider', 'CUDAExecutionProvider'] else: # Fall back to CPU and onnxruntime-cpu providers = ['CPUExecutionProvider'] onnx_session_opt = onnxruntime.SessionOptions() onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL onnx_model = onnx.load(self.encoder_model_path) onnx.checker.check_model(onnx_model, full_check=True) self.encoder_model = onnx_model self.encoder = onnxruntime.InferenceSession( onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt ) onnx_model = onnx.load(self.decoder_joint_model_path) onnx.checker.check_model(onnx_model, full_check=True) self.decoder_joint_model = onnx_model self.decoder_joint = onnxruntime.InferenceSession( onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt ) logging.info("Successfully loaded encoder, decoder and joint onnx models !") # Will be populated at runtime self._blank_index = None self.max_symbols_per_step = max_symbols_per_step self._setup_encoder_input_output_keys() self._setup_decoder_joint_input_output_keys() self._setup_blank_index() def _setup_encoder_input_output_keys(self): self.encoder_inputs = list(self.encoder_model.graph.input) self.encoder_outputs = list(self.encoder_model.graph.output) def _setup_decoder_joint_input_output_keys(self): self.decoder_joint_inputs = list(self.decoder_joint_model.graph.input) self.decoder_joint_outputs = list(self.decoder_joint_model.graph.output) def _setup_blank_index(self): # ASSUME: Single input with no time length information dynamic_dim = 257 shapes = self.encoder_inputs[0].type.tensor_type.shape.dim ip_shape = [] for shape in shapes: if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: ip_shape.append(dynamic_dim) # replace dynamic axes with constant else: ip_shape.append(int(shape.dim_value)) enc_logits, encoded_length = self.run_encoder( audio_signal=torch.randn(*ip_shape), length=torch.randint(0, 1, size=(dynamic_dim,)) ) # prepare states states = self._get_initial_states(batchsize=dynamic_dim) # run decoder 1 step joint_out, states = self.run_decoder_joint(enc_logits, None, None, *states) log_probs, lengths = joint_out self._blank_index = log_probs.shape[-1] - 1 # last token of vocab size is blank token logging.info( f"Enc-Dec-Joint step was evaluated, blank token id = {self._blank_index}; vocab size = {log_probs.shape[-1]}" ) def run_encoder(self, audio_signal, length): if hasattr(audio_signal, 'cpu'): audio_signal = audio_signal.cpu().numpy() if hasattr(length, 'cpu'): length = length.cpu().numpy() ip = { self.encoder_inputs[0].name: audio_signal, self.encoder_inputs[1].name: length, } enc_out = self.encoder.run(None, ip) enc_out, encoded_length = enc_out # ASSUME: single output return enc_out, encoded_length def run_decoder_joint(self, enc_logits, targets, target_length, *states): # ASSUME: Decoder is RNN Transducer if targets is None: targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32) target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32) if hasattr(targets, 'cpu'): targets = targets.cpu().numpy() if hasattr(target_length, 'cpu'): target_length = target_length.cpu().numpy() ip = { self.decoder_joint_inputs[0].name: enc_logits, self.decoder_joint_inputs[1].name: targets, self.decoder_joint_inputs[2].name: target_length, } num_states = 0 if states is not None and len(states) > 0: num_states = len(states) for idx, state in enumerate(states): if hasattr(state, 'cpu'): state = state.cpu().numpy() ip[self.decoder_joint_inputs[len(ip)].name] = state dec_out = self.decoder_joint.run(None, ip) # unpack dec output if num_states > 0: new_states = dec_out[-num_states:] dec_out = dec_out[:-num_states] else: new_states = None return dec_out, new_states def _get_initial_states(self, batchsize): # ASSUME: LSTM STATES of shape (layers, batchsize, dim) input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] num_states = len(input_state_nodes) if num_states == 0: return input_states = [] for state_id in range(num_states): node = input_state_nodes[state_id] ip_shape = [] for shape_idx, shape in enumerate(node.type.tensor_type.shape.dim): if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: ip_shape.append(batchsize) # replace dynamic axes with constant else: ip_shape.append(int(shape.dim_value)) input_states.append(torch.zeros(*ip_shape)) return input_states class TorchscriptGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): def __init__( self, encoder_model: str, decoder_joint_model: str, cfg: DictConfig, device: str, max_symbols_per_step: Optional[int] = 10, ): super().__init__( encoder_model=encoder_model, decoder_joint_model=decoder_joint_model, max_symbols_per_step=max_symbols_per_step, ) self.cfg = cfg self.device = device self.encoder = torch.jit.load(self.encoder_model_path, map_location=self.device) self.decoder_joint = torch.jit.load(self.decoder_joint_model_path, map_location=self.device) logging.info("Successfully loaded encoder, decoder and joint torchscript models !") # Will be populated at runtime self._blank_index = None self.max_symbols_per_step = max_symbols_per_step self._setup_encoder_input_keys() self._setup_decoder_joint_input_keys() self._setup_blank_index() def _setup_encoder_input_keys(self): arguments = self.encoder.forward.schema.arguments[1:] self.encoder_inputs = [arg for arg in arguments] def _setup_decoder_joint_input_keys(self): arguments = self.decoder_joint.forward.schema.arguments[1:] self.decoder_joint_inputs = [arg for arg in arguments] def _setup_blank_index(self): self._blank_index = len(self.cfg.joint.vocabulary) logging.info(f"Blank token id = {self._blank_index}; vocab size = {len(self.cfg.joint.vocabulary) + 1}") def run_encoder(self, audio_signal, length): enc_out = self.encoder(audio_signal, length) enc_out, encoded_length = enc_out # ASSUME: single output return enc_out, encoded_length def run_decoder_joint(self, enc_logits, targets, target_length, *states): # ASSUME: Decoder is RNN Transducer if targets is None: targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32, device=enc_logits.device) target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32, device=enc_logits.device) num_states = 0 if states is not None and len(states) > 0: num_states = len(states) dec_out = self.decoder_joint(enc_logits, targets, target_length, *states) # unpack dec output if num_states > 0: new_states = dec_out[-num_states:] dec_out = dec_out[:-num_states] else: new_states = None return dec_out, new_states def _get_initial_states(self, batchsize): # ASSUME: LSTM STATES of shape (layers, batchsize, dim) input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] num_states = len(input_state_nodes) if num_states == 0: return input_states = [] for state_id in range(num_states): # Hardcode shape size for LSTM (1 is for num layers in LSTM, which is flattened for export) ip_shape = [1, batchsize, self.cfg.model_defaults.pred_hidden] input_states.append(torch.zeros(*ip_shape, device=self.device)) return input_states class GreedyMultiblankRNNTInfer(GreedyRNNTInfer): """A greedy transducer decoder for multi-blank RNN-T. Sequence level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. big_blank_durations: a list containing durations for big blanks the model supports. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, big_blank_durations: list, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_measure_cfg=confidence_measure_cfg, ) self.big_blank_durations = big_blank_durations self._SOS = blank_index - len(big_blank_durations) @torch.no_grad() def _greedy_decode( self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None ): # x: [T, 1, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) if partial_hypotheses is not None: hypothesis.last_token = partial_hypotheses.last_token hypothesis.y_sequence = ( partial_hypotheses.y_sequence.cpu().tolist() if isinstance(partial_hypotheses.y_sequence, torch.Tensor) else partial_hypotheses.y_sequence ) if partial_hypotheses.dec_state is not None: hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) if self.preserve_alignments: # Alignments is a 2-dimensional dangling list representing T x U hypothesis.alignments = [[]] if self.preserve_frame_confidence: hypothesis.frame_confidence = [[]] # if this variable is > 1, it means the last emission was a big-blank and we need to skip frames. big_blank_duration = 1 # For timestep t in X_t for time_idx in range(out_len): if big_blank_duration > 1: # skip frames until big_blank_duration == 1. big_blank_duration -= 1 continue # Extract encoder embedding at timestep t # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] f = x.narrow(dim=0, start=time_idx, length=1) # Setup exit flags and counter not_blank = True symbols_added = 0 # While blank is not predicted, or we dont run out of max symbols per timestep while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # In the first timestep, we initialize the network with RNNT Blank # In later timesteps, we provide previous predicted label as input. if hypothesis.last_token is None and hypothesis.dec_state is None: last_label = self._SOS else: last_label = label_collate([[hypothesis.last_token]]) # Perform prediction network and joint network steps. g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 0, 0, 0, : ] del g # torch.max(0) op doesnt exist for FP 16. if logp.dtype != torch.float32: logp = logp.float() # get index k, of max prob v, k = logp.max(0) k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. # Note, we have non-blanks in the vocab first, followed by big blanks, and standard blank at last. # here we check if it's a big blank and if yes, set the duration variable. if k >= self._blank_index - len(self.big_blank_durations) and k < self._blank_index: big_blank_duration = self.big_blank_durations[self._blank_index - k - 1] if self.preserve_alignments: # insert logprobs into last timestep hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) if self.preserve_frame_confidence: # insert confidence into last timestep hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) del logp # If any type of blank token is predicted, exit inner loop, move onto next timestep t if k >= self._blank_index - len(self.big_blank_durations): not_blank = False else: # Append token to label set, update RNN state. hypothesis.y_sequence.append(k) hypothesis.score += float(v) hypothesis.timestep.append(time_idx) hypothesis.dec_state = hidden_prime hypothesis.last_token = k # Increment token counter. symbols_added += 1 if self.preserve_alignments: # convert Ti-th logits into a torch array hypothesis.alignments.append([]) # blank buffer for next timestep if self.preserve_frame_confidence: hypothesis.frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of Alignments if self.preserve_alignments: if len(hypothesis.alignments[-1]) == 0: del hypothesis.alignments[-1] # Remove trailing empty list of per-frame confidence if self.preserve_frame_confidence: if len(hypothesis.frame_confidence[-1]) == 0: del hypothesis.frame_confidence[-1] # Unpack the hidden states hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) return hypothesis class GreedyBatchedMultiblankRNNTInfer(GreedyBatchedRNNTInfer): """A batch level greedy transducer decoder. Batch level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. big_blank_durations: a list containing durations for big blanks the model supports. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, big_blank_durations: List[int], max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_measure_cfg=confidence_measure_cfg, ) self.big_blank_durations = big_blank_durations # Depending on availability of `blank_as_pad` support # switch between more efficient batch decoding technique if self.decoder.blank_as_pad: self._greedy_decode = self._greedy_decode_blank_as_pad else: self._greedy_decode = self._greedy_decode_masked self._SOS = blank_index - len(big_blank_durations) def _greedy_decode_blank_as_pad( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") with torch.inference_mode(): # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize list of Hypothesis batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._SOS, dtype=torch.long, device=device) # this mask is true for if the emission is *any type* of blank. blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() # We have a mask for each big blank. A mask is "true" means: the previous emission is exactly the big-blank # with the corresponding duration, or has larger duration. E.g., for big_blank_mask for duration 2, it will # be set true if the previous emission was a big blank with duration 4, or 3 or 2; but false if prevoius # emission was a standard blank (with duration = 1). big_blank_masks = [torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device)] * len( self.big_blank_durations ) # if this variable > 1, it means the previous emission was big-blank and we need to skip frames. big_blank_duration = 1 for time_idx in range(max_out_len): if big_blank_duration > 1: # skip frames until big_blank_duration == 1 big_blank_duration -= 1 continue f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset all blank masks blank_mask.mul_(False) for i in range(len(big_blank_masks)): big_blank_masks[i].mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len for i in range(len(big_blank_masks)): big_blank_masks[i] = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1 + num-big-blanks] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k >= self._blank_index - len(self.big_blank_durations) blank_mask.bitwise_or_(k_is_blank) for i in range(len(big_blank_masks)): # using <= since as we mentioned before, the mask doesn't store exact matches. # instead, it is True when the predicted blank's duration is >= the duration that the # mask corresponds to. k_is_big_blank = k <= self._blank_index - 1 - i # need to do a bitwise_and since it could also be a non-blank. k_is_big_blank.bitwise_and_(k_is_blank) big_blank_masks[i].bitwise_or_(k_is_big_blank) del k_is_blank # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.clone().view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 for i in range(len(big_blank_masks) + 1): # The task here is find the shortest blank duration of all batches. # so we start from the shortest blank duration and go up, # and stop once we found the duration whose corresponding mask isn't all True. if i == len(big_blank_masks) or not big_blank_masks[i].all(): big_blank_duration = self.big_blank_durations[i - 1] if i > 0 else 1 break # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses def _greedy_decode_masked( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") if self.big_blank_durations != [1] * len(self.big_blank_durations): raise NotImplementedError( "Efficient frame-skipping version for multi-blank masked decoding is not supported." ) # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize state batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] else: hyp.alignments = None # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) last_label_without_blank = last_label.clone() # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() with torch.inference_mode(): for time_idx in range(max_out_len): f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Set a dummy label for the blank value # This value will be overwritten by "blank" again the last label update below # This is done as vocabulary of prediction network does not contain "blank" token of RNNT last_label_without_blank_mask = last_label >= self._blank_index last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label last_label_without_blank[~last_label_without_blank_mask] = last_label[ ~last_label_without_blank_mask ] # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1 + num-big-blanks] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses @dataclass class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.confidence_measure_cfg = OmegaConf.structured( self.confidence_measure_cfg if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.confidence_measure_cfg) ) if self.confidence_method_cfg != "DEPRECATED": logging.warning( "`confidence_method_cfg` is deprecated and will be removed in the future. " "Please use `confidence_measure_cfg` instead." ) # TODO (alaptev): delete the following two lines sometime in the future logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") # OmegaConf.structured ensures that post_init check is always executed self.confidence_measure_cfg = OmegaConf.structured( self.confidence_method_cfg if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.confidence_method_cfg) ) self.confidence_method_cfg = "DEPRECATED" @dataclass class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.confidence_measure_cfg = OmegaConf.structured( self.confidence_measure_cfg if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.confidence_measure_cfg) ) if self.confidence_method_cfg != "DEPRECATED": logging.warning( "`confidence_method_cfg` is deprecated and will be removed in the future. " "Please use `confidence_measure_cfg` instead." ) # TODO (alaptev): delete the following two lines sometime in the future logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") # OmegaConf.structured ensures that post_init check is always executed self.confidence_measure_cfg = OmegaConf.structured( self.confidence_method_cfg if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.confidence_method_cfg) ) self.confidence_method_cfg = "DEPRECATED" class GreedyTDTInfer(_GreedyRNNTInfer): """A greedy TDT decoder. Sequence level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. durations: a list containing durations for TDT. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, durations: list, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_measure_cfg=confidence_measure_cfg, ) self.durations = durations @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) self.decoder.eval() self.joint.eval() hypotheses = [] # Process each sequence independently with self.decoder.as_frozen(), self.joint.as_frozen(): for batch_idx in range(encoder_output.size(0)): inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] logitlen = encoded_lengths[batch_idx] partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) hypotheses.append(hypothesis) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, encoded_lengths) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) @torch.no_grad() def _greedy_decode( self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None ): # x: [T, 1, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) if partial_hypotheses is not None: hypothesis.last_token = partial_hypotheses.last_token hypothesis.y_sequence = ( partial_hypotheses.y_sequence.cpu().tolist() if isinstance(partial_hypotheses.y_sequence, torch.Tensor) else partial_hypotheses.y_sequence ) if partial_hypotheses.dec_state is not None: hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) if self.preserve_alignments: # Alignments is a 2-dimensional dangling list representing T x U hypothesis.alignments = [[]] if self.preserve_frame_confidence: hypothesis.frame_confidence = [[]] time_idx = 0 while time_idx < out_len: # Extract encoder embedding at timestep t # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] f = x.narrow(dim=0, start=time_idx, length=1) # Setup exit flags and counter not_blank = True symbols_added = 0 need_loop = True # While blank is not predicted, or we dont run out of max symbols per timestep while need_loop and (self.max_symbols is None or symbols_added < self.max_symbols): # In the first timestep, we initialize the network with RNNT Blank # In later timesteps, we provide previous predicted label as input. if hypothesis.last_token is None and hypothesis.dec_state is None: last_label = self._SOS else: last_label = label_collate([[hypothesis.last_token]]) # Perform prediction network and joint network steps. g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) # If preserving per-frame confidence, log_normalize must be true logits = self._joint_step(f, g, log_normalize=False) logp = logits[0, 0, 0, : -len(self.durations)] if self.preserve_frame_confidence: logp = torch.log_softmax(logp, -1) duration_logp = torch.log_softmax(logits[0, 0, 0, -len(self.durations) :], dim=-1) del g # torch.max(0) op doesnt exist for FP 16. if logp.dtype != torch.float32: logp = logp.float() # get index k, of max prob v, k = logp.max(0) k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. d_v, d_k = duration_logp.max(0) d_k = d_k.item() skip = self.durations[d_k] if self.preserve_alignments: # insert logprobs into last timestep hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) if self.preserve_frame_confidence: # insert confidence into last timestep hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) del logp # If blank token is predicted, exit inner loop, move onto next timestep t if k == self._blank_index: not_blank = False else: # Append token to label set, update RNN state. hypothesis.y_sequence.append(k) hypothesis.score += float(v) hypothesis.timestep.append(time_idx) hypothesis.dec_state = hidden_prime hypothesis.last_token = k # Increment token counter. symbols_added += 1 time_idx += skip need_loop = skip == 0 # this rarely happens, but we manually increment the `skip` number # if blank is emitted and duration=0 is predicted. This prevents possible # infinite loops. if skip == 0: skip = 1 if self.preserve_alignments: # convert Ti-th logits into a torch array hypothesis.alignments.append([]) # blank buffer for next timestep if self.preserve_frame_confidence: hypothesis.frame_confidence.append([]) # blank buffer for next timestep if symbols_added == self.max_symbols: time_idx += 1 # Remove trailing empty list of Alignments if self.preserve_alignments: if len(hypothesis.alignments[-1]) == 0: del hypothesis.alignments[-1] # Remove trailing empty list of per-frame confidence if self.preserve_frame_confidence: if len(hypothesis.frame_confidence[-1]) == 0: del hypothesis.frame_confidence[-1] # Unpack the hidden states hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) return hypothesis class GreedyBatchedTDTInfer(_GreedyRNNTInfer): """A batch level greedy TDT decoder. Batch level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. durations: a list containing durations. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, durations: List[int], max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_measure_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_measure_cfg=confidence_measure_cfg, ) self.durations = durations # Depending on availability of `blank_as_pad` support # switch between more efficient batch decoding technique if self.decoder.blank_as_pad: self._greedy_decode = self._greedy_decode_blank_as_pad else: self._greedy_decode = self._greedy_decode_masked @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) logitlen = encoded_lengths self.decoder.eval() self.joint.eval() with self.decoder.as_frozen(), self.joint.as_frozen(): inseq = encoder_output # [B, T, D] hypotheses = self._greedy_decode( inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses ) # Pack the hypotheses results packed_result = pack_hypotheses(hypotheses, logitlen) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) def _greedy_decode_blank_as_pad( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") with torch.inference_mode(): # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize list of Hypothesis batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() # skip means the number of frames the next decoding step should "jump" to. When skip == 1 # it means the next decoding step will just use the next input frame. skip = 1 for time_idx in range(max_out_len): if skip > 1: # if skip > 1 at the current step, we decrement it and skip the current frame. skip -= 1 continue f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # need_to_stay is a boolean indicates whether the next decoding step should remain in the same frame. need_to_stay = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while need_to_stay and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1 + num-big-blanks] # Note: log_normalize must not be True here since the joiner output is contanetation of both token logits and duration logits, # and they need to be normalized independently. joined = self._joint_step(f, g, log_normalize=None) logp = joined[:, 0, 0, : -len(self.durations)] duration_logp = joined[:, 0, 0, -len(self.durations) :] if logp.dtype != torch.float32: logp = logp.float() duration_logp = duration_logp.float() # get the max for both token and duration predictions. v, k = logp.max(1) dv, dk = duration_logp.max(1) # here we set the skip value to be the minimum of all predicted durations, hense the "torch.min(dk)" call there. # Please refer to Section 5.2 of our paper https://arxiv.org/pdf/2304.06795.pdf for explanation of this. skip = self.durations[int(torch.min(dk))] # this is a special case: if all batches emit blanks, we require that skip be at least 1 # so we don't loop forever at the current frame. if blank_mask.all(): if skip == 0: skip = 1 need_to_stay = skip == 0 del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) del k_is_blank del logp, duration_logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if not blank_mask.all(): # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.clone().view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses def _greedy_decode_masked( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): raise NotImplementedError("masked greedy-batched decode is not supported for TDT models.") [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math from abc import ABC, abstractmethod from dataclasses import dataclass from functools import partial from typing import List, Optional import torch from omegaconf import DictConfig, OmegaConf from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis from nemo.utils import logging class ConfidenceMeasureConstants: NAMES = ("max_prob", "entropy") ENTROPY_TYPES = ("gibbs", "tsallis", "renyi") ENTROPY_NORMS = ("lin", "exp") @classmethod def print(cls): return ( cls.__name__ + ": " + str({"NAMES": cls.NAMES, "ENTROPY_TYPES": cls.ENTROPY_TYPES, "ENTROPY_NORMS": cls.ENTROPY_NORMS}) ) class ConfidenceConstants: AGGREGATIONS = ("mean", "min", "max", "prod") @classmethod def print(cls): return cls.__name__ + ": " + str({"AGGREGATIONS": cls.AGGREGATIONS}) @dataclass class ConfidenceMeasureConfig: """A Config which contains the measure name and settings to compute per-frame confidence scores. Args: name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ name: str = "entropy" entropy_type: str = "tsallis" alpha: float = 0.33 entropy_norm: str = "exp" temperature: str = "DEPRECATED" def __post_init__(self): if self.temperature != "DEPRECATED": logging.warning( "`temperature` is deprecated and will be removed in the future. Please use `alpha` instead." ) # TODO (alaptev): delete the following two lines sometime in the future logging.warning("Re-writing `alpha` with the value of `temperature`.") # self.temperature has type str self.alpha = float(self.temperature) self.temperature = "DEPRECATED" if self.name not in ConfidenceMeasureConstants.NAMES: raise ValueError( f"`name` must be one of the following: " f"{'`' + '`, `'.join(ConfidenceMeasureConstants.NAMES) + '`'}. Provided: `{self.name}`" ) if self.entropy_type not in ConfidenceMeasureConstants.ENTROPY_TYPES: raise ValueError( f"`entropy_type` must be one of the following: " f"{'`' + '`, `'.join(ConfidenceMeasureConstants.ENTROPY_TYPES) + '`'}. Provided: `{self.entropy_type}`" ) if self.alpha <= 0.0: raise ValueError(f"`alpha` must be > 0. Provided: {self.alpha}") if self.entropy_norm not in ConfidenceMeasureConstants.ENTROPY_NORMS: raise ValueError( f"`entropy_norm` must be one of the following: " f"{'`' + '`, `'.join(ConfidenceMeasureConstants.ENTROPY_NORMS) + '`'}. Provided: `{self.entropy_norm}`" ) @dataclass class ConfidenceConfig: """A config which contains the following key-value pairs related to confidence scores. Args: preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame confidence scores. name: The measure name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ preserve_frame_confidence: bool = False preserve_token_confidence: bool = False preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() method_cfg: str = "DEPRECATED" def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.measure_cfg = OmegaConf.structured( self.measure_cfg if isinstance(self.measure_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.measure_cfg) ) if self.method_cfg != "DEPRECATED": logging.warning( "`method_cfg` is deprecated and will be removed in the future. Please use `measure_cfg` instead." ) # TODO (alaptev): delete the following two lines sometime in the future logging.warning("Re-writing `measure_cfg` with the value of `method_cfg`.") # OmegaConf.structured ensures that post_init check is always executed self.measure_cfg = OmegaConf.structured( self.method_cfg if isinstance(self.method_cfg, ConfidenceMeasureConfig) else ConfidenceMeasureConfig(**self.method_cfg) ) self.method_cfg = "DEPRECATED" if self.aggregation not in ConfidenceConstants.AGGREGATIONS: raise ValueError( f"`aggregation` has to be one of the following: " f"{'`' + '`, `'.join(ConfidenceMeasureConstants.AGGREGATIONS) + '`'}. Provided: `{self.aggregation}`" ) def get_confidence_measure_bank(): """Generate a dictionary with confidence measure functionals. Supported confidence measures: max_prob: normalized maximum probability entropy_gibbs_lin: Gibbs entropy with linear normalization entropy_gibbs_exp: Gibbs entropy with exponential normalization entropy_tsallis_lin: Tsallis entropy with linear normalization entropy_tsallis_exp: Tsallis entropy with exponential normalization entropy_renyi_lin: Rรฉnyi entropy with linear normalization entropy_renyi_exp: Rรฉnyi entropy with exponential normalization Returns: dictionary with lambda functions. """ # helper functions # Gibbs entropy is implemented without alpha neg_entropy_gibbs = lambda x: (x.exp() * x).sum(-1) neg_entropy_alpha = lambda x, t: (x * t).exp().sum(-1) neg_entropy_alpha_gibbs = lambda x, t: ((x * t).exp() * x).sum(-1) # too big for a lambda def entropy_tsallis_exp(x, v, t): exp_neg_max_ent = math.exp((1 - math.pow(v, 1 - t)) / (1 - t)) return (((1 - neg_entropy_alpha(x, t)) / (1 - t)).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) def entropy_gibbs_exp(x, v, t): exp_neg_max_ent = math.pow(v, -t * math.pow(v, 1 - t)) return ((neg_entropy_alpha_gibbs(x, t) * t).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) # use Gibbs entropies for Tsallis and Rรฉnyi with t == 1.0 entropy_gibbs_lin_baseline = lambda x, v: 1 + neg_entropy_gibbs(x) / math.log(v) entropy_gibbs_exp_baseline = lambda x, v: (neg_entropy_gibbs(x).exp() * v - 1) / (v - 1) # fill the measure bank confidence_measure_bank = {} # Maximum probability measure is implemented without alpha confidence_measure_bank["max_prob"] = ( lambda x, v, t: (x.max(dim=-1)[0].exp() * v - 1) / (v - 1) if t == 1.0 else ((x.max(dim=-1)[0] * t).exp() * math.pow(v, t) - 1) / (math.pow(v, t) - 1) ) confidence_measure_bank["entropy_gibbs_lin"] = ( lambda x, v, t: entropy_gibbs_lin_baseline(x, v) if t == 1.0 else 1 + neg_entropy_alpha_gibbs(x, t) / math.log(v) / math.pow(v, 1 - t) ) confidence_measure_bank["entropy_gibbs_exp"] = ( lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_gibbs_exp(x, v, t) ) confidence_measure_bank["entropy_tsallis_lin"] = ( lambda x, v, t: entropy_gibbs_lin_baseline(x, v) if t == 1.0 else 1 + (1 - neg_entropy_alpha(x, t)) / (math.pow(v, 1 - t) - 1) ) confidence_measure_bank["entropy_tsallis_exp"] = ( lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_tsallis_exp(x, v, t) ) confidence_measure_bank["entropy_renyi_lin"] = ( lambda x, v, t: entropy_gibbs_lin_baseline(x, v) if t == 1.0 else 1 + neg_entropy_alpha(x, t).log2() / (t - 1) / math.log(v, 2) ) confidence_measure_bank["entropy_renyi_exp"] = ( lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else (neg_entropy_alpha(x, t).pow(1 / (t - 1)) * v - 1) / (v - 1) ) return confidence_measure_bank def get_confidence_aggregation_bank(): """Generate a dictionary with confidence aggregation functions. Supported confidence measures: min: minimum max: maximum mean: arithmetic mean prod: product Returns: dictionary with functions. """ confidence_aggregation_bank = {"mean": lambda x: sum(x) / len(x), "min": min, "max": max} # python 3.7 and earlier do not have math.prod if hasattr(math, "prod"): confidence_aggregation_bank["prod"] = math.prod else: import operator from functools import reduce confidence_aggregation_bank["prod"] = lambda x: reduce(operator.mul, x, 1) return confidence_aggregation_bank class ConfidenceMeasureMixin(ABC): """Confidence Measure Mixin class. It initializes per-frame confidence measure. """ def _init_confidence_measure(self, confidence_measure_cfg: Optional[DictConfig] = None): """Initialize per-frame confidence measure from config. """ # OmegaConf.structured ensures that post_init check is always executed confidence_measure_cfg = OmegaConf.structured( ConfidenceMeasureConfig() if confidence_measure_cfg is None else ConfidenceMeasureConfig(**confidence_measure_cfg) ) # set confidence calculation measure # we suppose that self.blank_id == len(vocabulary) self.num_tokens = (self.blank_id if hasattr(self, "blank_id") else self._blank_index) + 1 self.alpha = confidence_measure_cfg.alpha # init confidence measure bank self.confidence_measure_bank = get_confidence_measure_bank() measure = None # construct measure_name measure_name = "" if confidence_measure_cfg.name == "max_prob": measure_name = "max_prob" elif confidence_measure_cfg.name == "entropy": measure_name = '_'.join( [confidence_measure_cfg.name, confidence_measure_cfg.entropy_type, confidence_measure_cfg.entropy_norm] ) else: raise ValueError(f"Unsupported `confidence_measure_cfg.name`: `{confidence_measure_cfg.name}`") if measure_name not in self.confidence_measure_bank: raise ValueError(f"Unsupported measure setup: `{measure_name}`") measure = partial(self.confidence_measure_bank[measure_name], v=self.num_tokens, t=self.alpha) self._get_confidence = lambda x: measure(torch.nan_to_num(x)).tolist() class ConfidenceMixin(ABC): """Confidence Mixin class. It is responsible for confidence estimation method initialization and high-level confidence score calculation. """ def _init_confidence(self, confidence_cfg: Optional[DictConfig] = None): """Initialize confidence-related fields and confidence aggregation function from config. """ # OmegaConf.structured ensures that post_init check is always executed confidence_cfg = OmegaConf.structured( ConfidenceConfig() if confidence_cfg is None else ConfidenceConfig(**confidence_cfg) ) self.confidence_measure_cfg = confidence_cfg.measure_cfg # extract the config self.preserve_word_confidence = confidence_cfg.get('preserve_word_confidence', False) # set preserve_frame_confidence and preserve_token_confidence to True # if preserve_word_confidence is True self.preserve_token_confidence = ( confidence_cfg.get('preserve_token_confidence', False) | self.preserve_word_confidence ) # set preserve_frame_confidence to True if preserve_token_confidence is True self.preserve_frame_confidence = ( confidence_cfg.get('preserve_frame_confidence', False) | self.preserve_token_confidence ) self.exclude_blank_from_confidence = confidence_cfg.get('exclude_blank', True) self.word_confidence_aggregation = confidence_cfg.get('aggregation', "min") # define aggregation functions self.confidence_aggregation_bank = get_confidence_aggregation_bank() self._aggregate_confidence = self.confidence_aggregation_bank[self.word_confidence_aggregation] # Update preserve frame confidence if self.preserve_frame_confidence is False: if self.cfg.strategy in ['greedy', 'greedy_batch']: self.preserve_frame_confidence = self.cfg.greedy.get('preserve_frame_confidence', False) # OmegaConf.structured ensures that post_init check is always executed confidence_measure_cfg = OmegaConf.structured(self.cfg.greedy).get('confidence_measure_cfg', None) self.confidence_measure_cfg = ( OmegaConf.structured(ConfidenceMeasureConfig()) if confidence_measure_cfg is None else OmegaConf.structured(ConfidenceMeasureConfig(**confidence_measure_cfg)) ) @abstractmethod def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: """Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. Assumes that `frame_confidence` is present in the hypotheses. Args: hypotheses_list: List of Hypothesis. Returns: A list of hypotheses with high-level confidence scores. """ raise NotImplementedError() @abstractmethod def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: """Implemented by subclass in order to aggregate token confidence to a word-level confidence. Args: hypothesis: Hypothesis Returns: A list of word-level confidence scores. """ raise NotImplementedError() def _aggregate_token_confidence_chars(self, words: List[str], token_confidence: List[float]) -> List[float]: """Implementation of token confidence aggregation for character-based models. Args: words: List of words of a hypothesis. token_confidence: List of token-level confidence scores of a hypothesis. Returns: A list of word-level confidence scores. """ word_confidence = [] i = 0 for word in words: word_len = len(word) word_confidence.append(self._aggregate_confidence(token_confidence[i : i + word_len])) # we assume that there is exactly one space token between words and exclude it from word confidence i += word_len + 1 return word_confidence def _aggregate_token_confidence_subwords_sentencepiece( self, words: List[str], token_confidence: List[float], token_ids: List[int] ) -> List[float]: """Implementation of token confidence aggregation for subword-based models. **Note**: Only supports Sentencepiece based tokenizers ! Args: words: List of words of a hypothesis. token_confidence: List of token-level confidence scores of a hypothesis. token_ids: List of token ids of a hypothesis. Returns: A list of word-level confidence scores. """ word_confidence = [] # run only if there are final words if len(words) > 0: j = 0 prev_unk = False prev_underline = False for i, token_id in enumerate(token_ids): token = self.decode_ids_to_tokens([int(token_id)])[0] token_text = self.decode_tokens_to_str([int(token_id)]) # treat `<unk>` as a separate word regardless of the next token # to match the result of `tokenizer.ids_to_text` if (token != token_text or prev_unk) and i > j: # do not add confidence for `โ–` if the current token starts with `โ–` # to match the result of `tokenizer.ids_to_text` if not prev_underline: word_confidence.append(self._aggregate_confidence(token_confidence[j:i])) j = i prev_unk = token == '<unk>' prev_underline = token == 'โ–' if not prev_underline: word_confidence.append(self._aggregate_confidence(token_confidence[j : len(token_ids)])) if len(words) != len(word_confidence): raise RuntimeError( f"""Something went wrong with word-level confidence aggregation.\n Please check these values for debugging:\n len(words): {len(words)},\n len(word_confidence): {len(word_confidence)},\n recognized text: `{' '.join(words)}`""" ) return word_confidence [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of nemo/collections/common/parts/adapter_modules.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass, is_dataclass from typing import Any, Optional from hydra.utils import instantiate from omegaconf import OmegaConf from torch import nn as nn from nemo.collections.common.parts.utils import activation_registry from nemo.core.classes.mixins import access_mixins, adapter_mixin_strategies class AdapterModuleUtil(access_mixins.AccessMixin): """ Base class of Adapter Modules, providing common functionality to all Adapter Modules. """ def setup_adapter_strategy(self, adapter_strategy: Optional[adapter_mixin_strategies.AbstractAdapterStrategy]): """ Setup adapter strategy of this class, enabling dynamic change in the way the adapter output is merged with the input. When called successfully, will assign the variable `adapter_strategy` to the module. Args: adapter_strategy: Can be a None or an implementation of AbstractAdapterStrategy. """ # set default adapter strategy if adapter_strategy is None: adapter_strategy = self.get_default_strategy_config() if is_dataclass(adapter_strategy): adapter_strategy = OmegaConf.structured(adapter_strategy) OmegaConf.set_struct(adapter_strategy, False) # The config must have the `_target_` field pointing to the actual adapter strategy class # which will load that strategy dynamically to this module. if isinstance(adapter_strategy, dict) or OmegaConf.is_config(adapter_strategy): self.adapter_strategy = instantiate(adapter_strategy) elif isinstance(adapter_strategy, adapter_mixin_strategies.AbstractAdapterStrategy): self.adapter_strategy = adapter_strategy else: raise AttributeError(f'`adapter_strategy` provided is invalid : {adapter_strategy}') def get_default_strategy_config(self) -> 'dataclass': """ Returns a default adapter module strategy. """ return adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() def adapter_unfreeze(self,): """ Sets the requires grad for all parameters in the adapter to True. This method should be overridden for any custom unfreeze behavior that is required. For example, if not all params of the adapter should be unfrozen. """ for param in self.parameters(): param.requires_grad_(True) class LinearAdapter(nn.Module, AdapterModuleUtil): """ Simple Linear Feedforward Adapter module with LayerNorm and singe hidden layer with activation function. Note: The adapter explicitly initializes its final layer with all zeros in order to avoid affecting the original model when all adapters are disabled. Args: in_features: Input dimension of the module. Note that for adapters, input_dim == output_dim. dim: Hidden dimension of the feed forward network. activation: Str name for an activation function. norm_position: Str, can be `pre` or `post`. Defaults to `pre`. Determines whether the normalization will occur in the first layer or the last layer. Certain architectures may prefer one over the other. dropout: float value, whether to perform dropout on the output of the last layer of the adapter. adapter_strategy: By default, ResidualAddAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, in_features: int, dim: int, activation: str = 'swish', norm_position: str = 'pre', dropout: float = 0.0, adapter_strategy: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig = None, ): super().__init__() activation = activation_registry[activation]() # If the activation can be executed in place, do so. if hasattr(activation, 'inplace'): activation.inplace = True assert norm_position in ['pre', 'post'] self.norm_position = norm_position if norm_position == 'pre': self.module = nn.Sequential( nn.LayerNorm(in_features), nn.Linear(in_features, dim, bias=False), activation, nn.Linear(dim, in_features, bias=False), ) elif norm_position == 'post': self.module = nn.Sequential( nn.Linear(in_features, dim, bias=False), activation, nn.Linear(dim, in_features, bias=False), nn.LayerNorm(in_features), ) if dropout > 0.0: self.dropout = nn.Dropout(dropout) else: self.dropout = None # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) # reset parameters self.reset_parameters() def reset_parameters(self): # Final layer initializations must be 0 if self.norm_position == 'pre': self.module[-1].weight.data *= 0 elif self.norm_position == 'post': self.module[-1].weight.data *= 0 self.module[-1].bias.data *= 0 def forward(self, x): x = self.module(x) # Add dropout if available if self.dropout is not None: x = self.dropout(x) return x @dataclass class LinearAdapterConfig: in_features: int dim: int activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import re from typing import List import ipadic import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer class EnJaProcessor: """ Tokenizer, Detokenizer and Normalizer utilities for Japanese & English Args: lang_id: One of ['en', 'ja']. """ def __init__(self, lang_id: str): self.lang_id = lang_id self.moses_tokenizer = MosesTokenizer(lang=lang_id) self.moses_detokenizer = MosesDetokenizer(lang=lang_id) self.normalizer = MosesPunctNormalizer( lang=lang_id, pre_replace_unicode_punct=True, post_remove_control_chars=True ) def detokenize(self, tokens: List[str]) -> str: """ Detokenizes a list of tokens Args: tokens: list of strings as tokens Returns: detokenized Japanese or English string """ return self.moses_detokenizer.detokenize(tokens) def tokenize(self, text) -> str: """ Tokenizes text using Moses. Returns a string of tokens. """ tokens = self.moses_tokenizer.tokenize(text) return ' '.join(tokens) def normalize(self, text) -> str: # Normalization doesn't handle Japanese periods correctly; # 'ใ€‚'becomes '.'. if self.lang_id == 'en': return self.normalizer.normalize(text) else: return text class JaMecabProcessor: """ Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English """ def __init__(self): self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: RE_WS_IN_FW = re.compile( r'([\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])\s+(?=[\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])' ) detokenize = lambda s: spacing(RE_WS_IN_FW.sub(r'\1', s)).strip() return detokenize(' '.join(text)) def tokenize(self, text) -> str: """ Tokenizes text using Moses. Returns a string of tokens. """ return self.mecab_tokenizer.parse(text).strip() def normalize(self, text) -> str: return text [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING from nemo.collections.nlp.data.machine_translation.machine_translation_dataset import TranslationDataConfig from nemo.collections.nlp.models.enc_dec_nlp_model import EncDecNLPModelConfig from nemo.collections.nlp.modules.common.token_classifier import TokenClassifierConfig from nemo.collections.nlp.modules.common.tokenizer_utils import TokenizerConfig from nemo.collections.nlp.modules.common.transformer.transformer import ( NeMoTransformerConfig, NeMoTransformerEncoderConfig, ) from nemo.collections.nlp.modules.common.transformer.transformer_bottleneck import ( NeMoTransformerBottleneckDecoderConfig, NeMoTransformerBottleneckEncoderConfig, ) from nemo.core.config.modelPT import OptimConfig, SchedConfig @dataclass class MTSchedConfig(SchedConfig): name: str = 'InverseSquareRootAnnealing' warmup_ratio: Optional[float] = None last_epoch: int = -1 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). @dataclass class MTOptimConfig(OptimConfig): name: str = 'adam' lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 sched: Optional[MTSchedConfig] = MTSchedConfig() @dataclass class MTEncDecModelConfig(EncDecNLPModelConfig): # machine translation configurations num_val_examples: int = 3 num_test_examples: int = 3 max_generation_delta: int = 10 label_smoothing: Optional[float] = 0.0 beam_size: int = 4 len_pen: float = 0.0 src_language: Any = 'en' # Any = str or List[str] tgt_language: Any = 'en' # Any = str or List[str] find_unused_parameters: Optional[bool] = True shared_tokenizer: Optional[bool] = True multilingual: Optional[bool] = False preproc_out_dir: Optional[str] = None validate_input_ids: Optional[bool] = True shared_embeddings: bool = False # network architecture configuration encoder_tokenizer: Any = MISSING encoder: Any = MISSING decoder_tokenizer: Any = MISSING decoder: Any = MISSING head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) # dataset configurations train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( src_file_name=MISSING, tgt_file_name=MISSING, tokens_in_batch=512, clean=True, shuffle=True, cache_ids=False, use_cache=False, ) validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( src_file_name=MISSING, tgt_file_name=MISSING, tokens_in_batch=512, clean=False, shuffle=False, cache_ids=False, use_cache=False, ) test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( src_file_name=MISSING, tgt_file_name=MISSING, tokens_in_batch=512, clean=False, shuffle=False, cache_ids=False, use_cache=False, ) optim: Optional[OptimConfig] = MTOptimConfig() @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( library='nemo', model_name=None, pretrained=False, hidden_size=512, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, ) decoder: NeMoTransformerConfig = NeMoTransformerConfig( library='nemo', model_name=None, pretrained=False, hidden_size=512, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, ) @dataclass class MTBottleneckModelConfig(AAYNBaseConfig): model_type: str = 'nll' min_logv: float = -6 latent_size: int = -1 # -1 will take value of encoder hidden non_recon_warmup_batches: int = 200000 recon_per_token: bool = True log_timing: bool = True encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( library='nemo', model_name=None, pretrained=False, hidden_size=512, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, arch='seq2seq', hidden_steps=32, hidden_blocks=1, hidden_init_method='params', ) decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( library='nemo', model_name=None, pretrained=False, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, arch='seq2seq', ) [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict from nemo.collections.common.parts.adapter_modules import LinearAdapterConfig from nemo.collections.nlp.data.token_classification.punctuation_capitalization_dataset import ( PunctuationCapitalizationEvalDataConfig, PunctuationCapitalizationTrainDataConfig, legacy_data_config_to_new_data_config, ) from nemo.core.config import TrainerConfig from nemo.core.config.modelPT import NemoConfig from nemo.utils.exp_manager import ExpManagerConfig @dataclass class FreezeConfig: is_enabled: bool = False """Freeze audio encoder weight and add Conformer Layers on top of it""" d_model: Optional[int] = 256 """`d_model` parameter of ``ConformerLayer``""" d_ff: Optional[int] = 1024 """``d_ff`` parameter of ``ConformerLayer``""" num_layers: Optional[int] = 8 """``num_layers`` number of ``ConformerLayer`` modules to add on top of audio encoder""" @dataclass class AdapterConfig: config: Optional[LinearAdapterConfig] = None """Linear adapter config see ``collections.common.parts.LinearAdapterConfig``""" enable: bool = False """Use adapters for audio encoder""" @dataclass class FusionConfig: num_layers: Optional[int] = 4 """"Number of layers to use in fusion""" num_attention_heads: Optional[int] = 4 """Number of attention heads to use in fusion""" inner_size: Optional[int] = 2048 """Fusion inner size""" @dataclass class AudioEncoderConfig: pretrained_model: str = MISSING """A configuration for restoring pretrained audio encoder""" freeze: Optional[FreezeConfig] = None adapter: Optional[AdapterConfig] = None fusion: Optional[FusionConfig] = None @dataclass class TokenizerConfig: """A structure and default values of source text tokenizer.""" vocab_file: Optional[str] = None """A path to vocabulary file which is used in ``'word'``, ``'char'``, and HuggingFace tokenizers""" tokenizer_name: str = MISSING """A name of the tokenizer used for tokenization of source sequences. Possible options are ``'sentencepiece'``, ``'word'``, ``'char'``, HuggingFace tokenizers (e.g. ``'bert-base-uncased'``). For more options see function ``nemo.collections.nlp.modules.common.get_tokenizer``. The tokenizer must have properties ``cls_id``, ``pad_id``, ``sep_id``, ``unk_id``.""" special_tokens: Optional[Dict[str, str]] = None """A dictionary with special tokens passed to constructors of ``'char'``, ``'word'``, ``'sentencepiece'``, and various HuggingFace tokenizers.""" tokenizer_model: Optional[str] = None """A path to a tokenizer model required for ``'sentencepiece'`` tokenizer.""" @dataclass class LanguageModelConfig: """ A structure and default values of language model configuration of punctuation and capitalization model. BERT like HuggingFace models are supported. Provide a valid ``pretrained_model_name`` and, optionally, you may reinitialize model via ``config_file`` or ``config``. Alternatively you can initialize the language model using ``lm_checkpoint``. This config is a part of :class:`PunctuationCapitalizationModelConfig` config. """ pretrained_model_name: str = MISSING """A mandatory parameter containing name of HuggingFace pretrained model. For example, ``'bert-base-uncased'``.""" config_file: Optional[str] = None """A path to a file with HuggingFace model config which is used to reinitialize language model.""" config: Optional[Dict] = None """A HuggingFace config which is used to reinitialize language model.""" lm_checkpoint: Optional[str] = None """A path to a ``torch`` checkpoint of a language model.""" @dataclass class HeadConfig: """ A structure and default values of configuration of capitalization or punctuation model head. This config defines a multilayer perceptron which is applied to output of a language model. Number of units in the hidden layer is equal to the dimension of the language model. This config is a part of :class:`PunctuationCapitalizationModelConfig` config. """ num_fc_layers: int = 1 """A number of hidden layers in a multilayer perceptron.""" fc_dropout: float = 0.1 """A dropout used in an MLP.""" activation: str = 'relu' """An activation used in hidden layers.""" use_transformer_init: bool = True """Whether to initialize the weights of the classifier head with the approach that was used for language model initialization.""" @dataclass class ClassLabelsConfig: """ A structure and default values of a mandatory part of config which contains names of files which are saved in .nemo checkpoint. These files can also be used for passing label vocabulary to the model. For using them as label vocabularies you will need to provide path these files in parameter ``model.common_dataset_parameters.label_vocab_dir``. Each line in labels files contains 1 label. The values are sorted, ``<line number>==<label id>``, starting from ``0``. A label with ``0`` id must contain neutral label which must be equal to ``model.common_dataset_parameters.pad_label``. This config is a part of :class:`~CommonDatasetParametersConfig`. """ punct_labels_file: str = MISSING """A name of punctuation labels file.""" capit_labels_file: str = MISSING """A name of capitalization labels file.""" @dataclass class CommonDatasetParametersConfig: """ A structure and default values of common dataset parameters config which includes label and loss mask information. If you omit parameters ``punct_label_ids``, ``capit_label_ids``, ``label_vocab_dir``, then labels will be inferred from a training dataset or loaded from a checkpoint. Parameters ``ignore_extra_tokens`` and ``ignore_start_end`` are responsible for forming loss mask. A loss mask defines on which tokens loss is computed. This parameter is a part of config :class:`~PunctuationCapitalizationModelConfig`. """ pad_label: str = MISSING """A mandatory parameter which should contain label used for punctuation and capitalization label padding. It also serves as a neutral label for both punctuation and capitalization. If any of ``punct_label_ids``, ``capit_label_ids`` parameters is provided, then ``pad_label`` must have ``0`` id in them. In addition, if ``label_vocab_dir`` is provided, then ``pad_label`` must be on the first lines in files ``class_labels.punct_labels_file`` and ``class_labels.capit_labels_file``.""" ignore_extra_tokens: bool = False """Whether to compute loss on not first tokens in words. If this parameter is ``True``, then loss mask is ``False`` for all tokens in a word except the first.""" ignore_start_end: bool = True """If ``False``, then loss is computed on [CLS] and [SEP] tokens.""" punct_label_ids: Optional[Dict[str, int]] = None """A dictionary with punctuation label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this parameter and pass label ids through ``class_labels.punct_labels_file`` or let the model to infer label ids from dataset or load them from checkpoint.""" capit_label_ids: Optional[Dict[str, int]] = None """A dictionary with capitalization label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this parameter and pass label ids through ``class_labels.capit_labels_file`` or let model to infer label ids from dataset or load them from checkpoint.""" label_vocab_dir: Optional[str] = None """A path to directory which contains class labels files. See :class:`ClassLabelsConfig`. If this parameter is provided, then labels will be loaded from files which are located in ``label_vocab_dir`` and have names specified in ``model.class_labels`` configuration section. A label specified in ``pad_label`` has to be on the first lines of ``model.class_labels`` files.""" @dataclass class PunctuationCapitalizationModelConfig: """ A configuration of :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model. See an example of model config in `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ This config is a part of :class:`~PunctuationCapitalizationConfig`. """ class_labels: ClassLabelsConfig = ClassLabelsConfig() """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None """A configuration for creating training dataset and data loader.""" validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating validation datasets and data loaders.""" test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" punct_head: HeadConfig = HeadConfig() """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" capit_head: HeadConfig = HeadConfig() """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" tokenizer: Any = TokenizerConfig() """A configuration for source text tokenizer.""" language_model: LanguageModelConfig = LanguageModelConfig() """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For description see `Optimizers <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/core/core.html#optimizers>`_ section in documentation and `primer <https://github.com/NVIDIA/NeMo/blob/main/tutorials/00_NeMo_Primer.ipynb>_ tutorial.""" @dataclass class PunctuationCapitalizationLexicalAudioModelConfig(PunctuationCapitalizationModelConfig): """ A configuration of :class:`~nemo.collections.nlp.models.token_classification.punctuation_lexical_audio_capitalization_model.PunctuationCapitalizationLexicalAudioModel` model. See an example of model config in `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_lexical_audio_config.yaml>`_ Audio encoder can be frozen during training with ``freeze_audio_encoder`` parameter. Adapter can be added to audio encoder with ``use_adapters`` and ``adapter_config`` parameters. More conformer layers can be added on top of pretrained audio encoder with ``frozen_conf_d_model``, ``frozen_conf_d_ff`` and ``frozen_conf_num_layers`` parameters. """ train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None """A configuration for creating training dataset and data loader.""" validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating validation datasets and data loaders.""" test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" audio_encoder: Optional[AudioEncoderConfig] = None restore_lexical_encoder_from: Optional[str] = None """"Path to .nemo checkpoint to load weights from""" # add more comments use_weighted_loss: Optional[bool] = False """If set to ``True`` CrossEntropyLoss will be weighted""" @dataclass class PunctuationCapitalizationConfig(NemoConfig): """ A config for punctuation model training and testing. See an example of full config in `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ """ pretrained_model: Optional[str] = None """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options by calling method :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. """ name: Optional[str] = 'Punctuation_and_Capitalization' """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" do_training: bool = True """Whether to perform training of the model.""" do_testing: bool = False """Whether ot perform testing of the model.""" model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" trainer: Optional[TrainerConfig] = TrainerConfig() """Contains ``Trainer`` Lightning class constructor parameters.""" exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() def is_legacy_model_config(model_cfg: DictConfig) -> bool: """ Test if model config is old style config. Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support tarred datasets. Args: model_cfg: model configuration Returns: whether ``model_config`` is legacy """ return 'common_dataset_parameters' not in model_cfg def legacy_model_config_to_new_model_config(model_cfg: DictConfig) -> DictConfig: """ Transform old style config into :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig`. Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support tarred datasets. Args: model_cfg: old style config Returns: model config which follows dataclass :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig` """ train_ds = model_cfg.get('train_ds') validation_ds = model_cfg.get('validation_ds') test_ds = model_cfg.get('test_ds') dataset = model_cfg.dataset punct_head_config = model_cfg.get('punct_head', {}) capit_head_config = model_cfg.get('capit_head', {}) omega_conf = OmegaConf.structured( PunctuationCapitalizationModelConfig( class_labels=model_cfg.class_labels, common_dataset_parameters=CommonDatasetParametersConfig( pad_label=dataset.pad_label, ignore_extra_tokens=dataset.get( 'ignore_extra_tokens', CommonDatasetParametersConfig.ignore_extra_tokens ), ignore_start_end=dataset.get('ignore_start_end', CommonDatasetParametersConfig.ignore_start_end), punct_label_ids=model_cfg.punct_label_ids, capit_label_ids=model_cfg.capit_label_ids, ), train_ds=None if train_ds is None else legacy_data_config_to_new_data_config(train_ds, dataset, train=True), validation_ds=None if validation_ds is None else legacy_data_config_to_new_data_config(validation_ds, dataset, train=False), test_ds=None if test_ds is None else legacy_data_config_to_new_data_config(test_ds, dataset, train=False), punct_head=HeadConfig( num_fc_layers=punct_head_config.get('punct_num_fc_layers', HeadConfig.num_fc_layers), fc_dropout=punct_head_config.get('fc_dropout', HeadConfig.fc_dropout), activation=punct_head_config.get('activation', HeadConfig.activation), use_transformer_init=punct_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), ), capit_head=HeadConfig( num_fc_layers=capit_head_config.get('capit_num_fc_layers', HeadConfig.num_fc_layers), fc_dropout=capit_head_config.get('fc_dropout', HeadConfig.fc_dropout), activation=capit_head_config.get('activation', HeadConfig.activation), use_transformer_init=capit_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), ), tokenizer=model_cfg.tokenizer, language_model=model_cfg.language_model, optim=model_cfg.optim, ) ) with open_dict(omega_conf): retain_during_legacy_conversion = model_cfg.get('retain_during_legacy_conversion', {}) for key in retain_during_legacy_conversion.keys(): omega_conf[key] = retain_during_legacy_conversion[key] return omega_conf [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Transformer based language model.""" from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( MegatronRetrievalTransformerEncoderModule, ) from nemo.collections.nlp.modules.common.megatron.utils import ( ApexGuardDefaults, init_method_normal, scaled_init_method_normal, ) try: from apex.transformer.enums import AttnMaskType, ModelType HAVE_APEX = True except (ImportError, ModuleNotFoundError): HAVE_APEX = False # fake missing classes with None attributes AttnMaskType = ApexGuardDefaults() ModelType = ApexGuardDefaults() try: from megatron.core import ModelParallelConfig HAVE_MEGATRON_CORE = True except (ImportError, ModuleNotFoundError): ModelParallelConfig = ApexGuardDefaults HAVE_MEGATRON_CORE = False __all__ = [] AVAILABLE_ENCODERS = ["transformer", "perceiver", "retro"] def get_encoder_model( config: ModelParallelConfig, arch, hidden_size, ffn_hidden_size, num_layers, num_attention_heads, apply_query_key_layer_scaling=False, kv_channels=None, init_method=None, scaled_init_method=None, encoder_attn_mask_type=AttnMaskType.padding, pre_process=True, post_process=True, init_method_std=0.02, megatron_amp_O2=False, hidden_dropout=0.1, attention_dropout=0.1, ffn_dropout=0.0, precision=16, fp32_residual_connection=False, activations_checkpoint_method=None, activations_checkpoint_num_layers=1, activations_checkpoint_granularity=None, layernorm_epsilon=1e-5, bias_activation_fusion=True, bias_dropout_add_fusion=True, masked_softmax_fusion=True, persist_layer_norm=False, openai_gelu=False, activation="gelu", onnx_safe=False, bias=True, normalization="layernorm", headscale=False, transformer_block_type="pre_ln", hidden_steps=32, parent_model_type=ModelType.encoder_or_decoder, layer_type=None, chunk_size=64, num_self_attention_per_cross_attention=1, layer_number_offset=0, # this is use only for attention norm_factor scaling megatron_legacy=False, normalize_attention_scores=True, sequence_parallel=False, num_moe_experts=1, moe_frequency=1, moe_dropout=0.0, turn_off_rop=False, # turn off the RoP positional embedding version=1, # model version position_embedding_type='learned_absolute', use_flash_attention=False, ): """Build language model and return along with the key to save.""" if kv_channels is None: assert ( hidden_size % num_attention_heads == 0 ), 'hidden_size must be divisible by num_attention_heads if kv_channels is None' kv_channels = hidden_size // num_attention_heads if init_method is None: init_method = init_method_normal(init_method_std) if scaled_init_method is None: scaled_init_method = scaled_init_method_normal(init_method_std, num_layers) if arch == "transformer": # Language encoder. encoder = MegatronTransformerEncoderModule( config=config, init_method=init_method, output_layer_init_method=scaled_init_method, hidden_size=hidden_size, num_layers=num_layers, num_attention_heads=num_attention_heads, apply_query_key_layer_scaling=apply_query_key_layer_scaling, kv_channels=kv_channels, ffn_hidden_size=ffn_hidden_size, encoder_attn_mask_type=encoder_attn_mask_type, pre_process=pre_process, post_process=post_process, megatron_amp_O2=megatron_amp_O2, hidden_dropout=hidden_dropout, attention_dropout=attention_dropout, ffn_dropout=ffn_dropout, precision=precision, fp32_residual_connection=fp32_residual_connection, activations_checkpoint_method=activations_checkpoint_method, activations_checkpoint_num_layers=activations_checkpoint_num_layers, activations_checkpoint_granularity=activations_checkpoint_granularity, layernorm_epsilon=layernorm_epsilon, bias_activation_fusion=bias_activation_fusion, bias_dropout_add_fusion=bias_dropout_add_fusion, masked_softmax_fusion=masked_softmax_fusion, persist_layer_norm=persist_layer_norm, openai_gelu=openai_gelu, onnx_safe=onnx_safe, activation=activation, bias=bias, normalization=normalization, transformer_block_type=transformer_block_type, headscale=headscale, parent_model_type=parent_model_type, megatron_legacy=megatron_legacy, normalize_attention_scores=normalize_attention_scores, num_moe_experts=num_moe_experts, moe_frequency=moe_frequency, moe_dropout=moe_dropout, position_embedding_type=position_embedding_type, use_flash_attention=use_flash_attention, ) elif arch == "retro": encoder = MegatronRetrievalTransformerEncoderModule( config=config, init_method=init_method, output_layer_init_method=scaled_init_method, hidden_size=hidden_size, num_layers=num_layers, num_attention_heads=num_attention_heads, apply_query_key_layer_scaling=apply_query_key_layer_scaling, kv_channels=kv_channels, layer_type=layer_type, ffn_hidden_size=ffn_hidden_size, pre_process=pre_process, post_process=post_process, megatron_amp_O2=megatron_amp_O2, hidden_dropout=hidden_dropout, attention_dropout=attention_dropout, precision=precision, fp32_residual_connection=fp32_residual_connection, activations_checkpoint_method=activations_checkpoint_method, activations_checkpoint_num_layers=activations_checkpoint_num_layers, activations_checkpoint_granularity=activations_checkpoint_granularity, layernorm_epsilon=layernorm_epsilon, bias_activation_fusion=bias_activation_fusion, bias_dropout_add_fusion=bias_dropout_add_fusion, masked_softmax_fusion=masked_softmax_fusion, persist_layer_norm=persist_layer_norm, openai_gelu=openai_gelu, onnx_safe=onnx_safe, activation=activation, bias=bias, normalization=normalization, transformer_block_type=transformer_block_type, parent_model_type=parent_model_type, chunk_size=chunk_size, layer_number_offset=layer_number_offset, megatron_legacy=megatron_legacy, normalize_attention_scores=normalize_attention_scores, turn_off_rop=turn_off_rop, version=version, ) elif arch == "perceiver": encoder = MegatronPerceiverEncoderModule( config=config, init_method=init_method, output_layer_init_method=scaled_init_method, hidden_size=hidden_size, num_layers=num_layers, num_attention_heads=num_attention_heads, apply_query_key_layer_scaling=apply_query_key_layer_scaling, kv_channels=kv_channels, ffn_hidden_size=ffn_hidden_size, encoder_attn_mask_type=encoder_attn_mask_type, pre_process=pre_process, post_process=post_process, megatron_amp_O2=megatron_amp_O2, hidden_dropout=hidden_dropout, attention_dropout=attention_dropout, ffn_dropout=ffn_dropout, precision=precision, fp32_residual_connection=fp32_residual_connection, activations_checkpoint_method=activations_checkpoint_method, activations_checkpoint_num_layers=activations_checkpoint_num_layers, activations_checkpoint_granularity=activations_checkpoint_granularity, layernorm_epsilon=layernorm_epsilon, bias_activation_fusion=bias_activation_fusion, bias_dropout_add_fusion=bias_dropout_add_fusion, masked_softmax_fusion=masked_softmax_fusion, persist_layer_norm=persist_layer_norm, openai_gelu=openai_gelu, onnx_safe=onnx_safe, activation=activation, bias=bias, normalization=normalization, transformer_block_type=transformer_block_type, headscale=headscale, parent_model_type=parent_model_type, hidden_steps=hidden_steps, num_self_attention_per_cross_attention=num_self_attention_per_cross_attention, megatron_legacy=megatron_legacy, normalize_attention_scores=normalize_attention_scores, ) else: raise ValueError(f"Unknown encoder arch = {arch}. Available encoder arch = {AVAILABLE_ENCODERS}") return encoder [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/tts/models/fastpitch.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import contextlib from dataclasses import dataclass from pathlib import Path from typing import List, Optional import torch from hydra.utils import instantiate from omegaconf import DictConfig, OmegaConf, open_dict from pytorch_lightning import Trainer from pytorch_lightning.loggers import TensorBoardLogger from nemo.collections.common.parts.preprocessing import parsers from nemo.collections.tts.losses.aligner_loss import BinLoss, ForwardSumLoss from nemo.collections.tts.losses.fastpitchloss import DurationLoss, EnergyLoss, MelLoss, PitchLoss from nemo.collections.tts.models.base import SpectrogramGenerator from nemo.collections.tts.modules.fastpitch import FastPitchModule from nemo.collections.tts.parts.mixins import FastPitchAdapterModelMixin from nemo.collections.tts.parts.utils.callbacks import LoggingCallback from nemo.collections.tts.parts.utils.helpers import ( batch_from_ragged, g2p_backward_compatible_support, plot_alignment_to_numpy, plot_spectrogram_to_numpy, process_batch, sample_tts_input, ) from nemo.core.classes import Exportable from nemo.core.classes.common import PretrainedModelInfo, typecheck from nemo.core.neural_types.elements import ( Index, LengthsType, MelSpectrogramType, ProbsType, RegressionValuesType, TokenDurationType, TokenIndex, TokenLogDurationType, ) from nemo.core.neural_types.neural_type import NeuralType from nemo.utils import logging, model_utils @dataclass class G2PConfig: _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" phoneme_probability: float = 0.5 @dataclass class TextTokenizer: _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" punct: bool = True stresses: bool = True chars: bool = True apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True g2p: G2PConfig = G2PConfig() @dataclass class TextTokenizerConfig: text_tokenizer: TextTokenizer = TextTokenizer() class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" def __init__(self, cfg: DictConfig, trainer: Trainer = None): # Convert to Hydra 1.0 compatible DictConfig cfg = model_utils.convert_model_config_to_dict_config(cfg) cfg = model_utils.maybe_update_config_version(cfg) # Setup normalizer self.normalizer = None self.text_normalizer_call = None self.text_normalizer_call_kwargs = {} self._setup_normalizer(cfg) self.learn_alignment = cfg.get("learn_alignment", False) # Setup vocabulary (=tokenizer) and input_fft_kwargs (supported only with self.learn_alignment=True) input_fft_kwargs = {} if self.learn_alignment: self.vocab = None self.ds_class = cfg.train_ds.dataset._target_ self.ds_class_name = self.ds_class.split(".")[-1] if not self.ds_class in [ "nemo.collections.tts.data.dataset.TTSDataset", "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset", "nemo.collections.tts.torch.data.TTSDataset", ]: raise ValueError(f"Unknown dataset class: {self.ds_class}.") self._setup_tokenizer(cfg) assert self.vocab is not None input_fft_kwargs["n_embed"] = len(self.vocab.tokens) input_fft_kwargs["padding_idx"] = self.vocab.pad self._parser = None self._tb_logger = None super().__init__(cfg=cfg, trainer=trainer) self.bin_loss_warmup_epochs = cfg.get("bin_loss_warmup_epochs", 100) self.log_images = cfg.get("log_images", False) self.log_train_images = False default_prosody_loss_scale = 0.1 if self.learn_alignment else 1.0 dur_loss_scale = cfg.get("dur_loss_scale", default_prosody_loss_scale) pitch_loss_scale = cfg.get("pitch_loss_scale", default_prosody_loss_scale) energy_loss_scale = cfg.get("energy_loss_scale", default_prosody_loss_scale) self.mel_loss_fn = MelLoss() self.pitch_loss_fn = PitchLoss(loss_scale=pitch_loss_scale) self.duration_loss_fn = DurationLoss(loss_scale=dur_loss_scale) self.energy_loss_fn = EnergyLoss(loss_scale=energy_loss_scale) self.aligner = None if self.learn_alignment: aligner_loss_scale = cfg.get("aligner_loss_scale", 1.0) self.aligner = instantiate(self._cfg.alignment_module) self.forward_sum_loss_fn = ForwardSumLoss(loss_scale=aligner_loss_scale) self.bin_loss_fn = BinLoss(loss_scale=aligner_loss_scale) self.preprocessor = instantiate(self._cfg.preprocessor) input_fft = instantiate(self._cfg.input_fft, **input_fft_kwargs) output_fft = instantiate(self._cfg.output_fft) duration_predictor = instantiate(self._cfg.duration_predictor) pitch_predictor = instantiate(self._cfg.pitch_predictor) speaker_encoder = instantiate(self._cfg.get("speaker_encoder", None)) energy_embedding_kernel_size = cfg.get("energy_embedding_kernel_size", 0) energy_predictor = instantiate(self._cfg.get("energy_predictor", None)) # [TODO] may remove if we change the pre-trained config # cfg: condition_types = [ "add" ] n_speakers = cfg.get("n_speakers", 0) speaker_emb_condition_prosody = cfg.get("speaker_emb_condition_prosody", False) speaker_emb_condition_decoder = cfg.get("speaker_emb_condition_decoder", False) speaker_emb_condition_aligner = cfg.get("speaker_emb_condition_aligner", False) min_token_duration = cfg.get("min_token_duration", 0) use_log_energy = cfg.get("use_log_energy", True) if n_speakers > 1 and "add" not in input_fft.cond_input.condition_types: input_fft.cond_input.condition_types.append("add") if speaker_emb_condition_prosody: duration_predictor.cond_input.condition_types.append("add") pitch_predictor.cond_input.condition_types.append("add") if speaker_emb_condition_decoder: output_fft.cond_input.condition_types.append("add") if speaker_emb_condition_aligner and self.aligner is not None: self.aligner.cond_input.condition_types.append("add") self.fastpitch = FastPitchModule( input_fft, output_fft, duration_predictor, pitch_predictor, energy_predictor, self.aligner, speaker_encoder, n_speakers, cfg.symbols_embedding_dim, cfg.pitch_embedding_kernel_size, energy_embedding_kernel_size, cfg.n_mel_channels, min_token_duration, cfg.max_token_duration, use_log_energy, ) self._input_types = self._output_types = None self.export_config = { "emb_range": (0, self.fastpitch.encoder.word_emb.num_embeddings), "enable_volume": False, "enable_ragged_batches": False, } if self.fastpitch.speaker_emb is not None: self.export_config["num_speakers"] = cfg.n_speakers self.log_config = cfg.get("log_config", None) # Adapter modules setup (from FastPitchAdapterModelMixin) self.setup_adapters() def _get_default_text_tokenizer_conf(self): text_tokenizer: TextTokenizerConfig = TextTokenizerConfig() return OmegaConf.create(OmegaConf.to_yaml(text_tokenizer)) def _setup_normalizer(self, cfg): if "text_normalizer" in cfg: normalizer_kwargs = {} if "whitelist" in cfg.text_normalizer: normalizer_kwargs["whitelist"] = self.register_artifact( 'text_normalizer.whitelist', cfg.text_normalizer.whitelist ) try: import nemo_text_processing self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) except Exception as e: logging.error(e) raise ImportError( "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" ) self.text_normalizer_call = self.normalizer.normalize if "text_normalizer_call_kwargs" in cfg: self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs def _setup_tokenizer(self, cfg): text_tokenizer_kwargs = {} if "g2p" in cfg.text_tokenizer: # for backward compatibility if ( self._is_model_being_restored() and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") ): cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( cfg.text_tokenizer.g2p["_target_"] ) g2p_kwargs = {} if "phoneme_dict" in cfg.text_tokenizer.g2p: g2p_kwargs["phoneme_dict"] = self.register_artifact( 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, ) if "heteronyms" in cfg.text_tokenizer.g2p: g2p_kwargs["heteronyms"] = self.register_artifact( 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, ) # for backward compatability text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) # TODO @xueyang: rename the instance of tokenizer because vocab is misleading. self.vocab = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) @property def tb_logger(self): if self._tb_logger is None: if self.logger is None and self.logger.experiment is None: return None tb_logger = self.logger.experiment for logger in self.trainer.loggers: if isinstance(logger, TensorBoardLogger): tb_logger = logger.experiment break self._tb_logger = tb_logger return self._tb_logger @property def parser(self): if self._parser is not None: return self._parser if self.learn_alignment: self._parser = self.vocab.encode else: self._parser = parsers.make_parser( labels=self._cfg.labels, name='en', unk_id=-1, blank_id=-1, do_normalize=True, abbreviation_version="fastpitch", make_table=False, ) return self._parser def parse(self, str_input: str, normalize=True) -> torch.tensor: if self.training: logging.warning("parse() is meant to be called in eval mode.") if normalize and self.text_normalizer_call is not None: str_input = self.text_normalizer_call(str_input, **self.text_normalizer_call_kwargs) if self.learn_alignment: eval_phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): eval_phon_mode = self.vocab.set_phone_prob(prob=1.0) # Disable mixed g2p representation if necessary with eval_phon_mode: tokens = self.parser(str_input) else: tokens = self.parser(str_input) x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device) return x @typecheck( input_types={ "text": NeuralType(('B', 'T_text'), TokenIndex()), "durs": NeuralType(('B', 'T_text'), TokenDurationType()), "pitch": NeuralType(('B', 'T_audio'), RegressionValuesType()), "energy": NeuralType(('B', 'T_audio'), RegressionValuesType(), optional=True), "speaker": NeuralType(('B'), Index(), optional=True), "pace": NeuralType(optional=True), "spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), "attn_prior": NeuralType(('B', 'T_spec', 'T_text'), ProbsType(), optional=True), "mel_lens": NeuralType(('B'), LengthsType(), optional=True), "input_lens": NeuralType(('B'), LengthsType(), optional=True), # reference_* data is used for multi-speaker FastPitch training "reference_spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), "reference_spec_lens": NeuralType(('B'), LengthsType(), optional=True), } ) def forward( self, *, text, durs=None, pitch=None, energy=None, speaker=None, pace=1.0, spec=None, attn_prior=None, mel_lens=None, input_lens=None, reference_spec=None, reference_spec_lens=None, ): return self.fastpitch( text=text, durs=durs, pitch=pitch, energy=energy, speaker=speaker, pace=pace, spec=spec, attn_prior=attn_prior, mel_lens=mel_lens, input_lens=input_lens, reference_spec=reference_spec, reference_spec_lens=reference_spec_lens, ) @typecheck(output_types={"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType())}) def generate_spectrogram( self, tokens: 'torch.tensor', speaker: Optional[int] = None, pace: float = 1.0, reference_spec: Optional['torch.tensor'] = None, reference_spec_lens: Optional['torch.tensor'] = None, ) -> torch.tensor: if self.training: logging.warning("generate_spectrogram() is meant to be called in eval mode.") if isinstance(speaker, int): speaker = torch.tensor([speaker]).to(self.device) spect, *_ = self( text=tokens, durs=None, pitch=None, speaker=speaker, pace=pace, reference_spec=reference_spec, reference_spec_lens=reference_spec_lens, ) return spect def training_step(self, batch, batch_idx): attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( None, None, None, None, None, None, ) if self.learn_alignment: if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": batch_dict = batch else: batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) audio = batch_dict.get("audio") audio_lens = batch_dict.get("audio_lens") text = batch_dict.get("text") text_lens = batch_dict.get("text_lens") attn_prior = batch_dict.get("align_prior_matrix", None) pitch = batch_dict.get("pitch", None) energy = batch_dict.get("energy", None) speaker = batch_dict.get("speaker_id", None) reference_audio = batch_dict.get("reference_audio", None) reference_audio_len = batch_dict.get("reference_audio_lens", None) else: audio, audio_lens, text, text_lens, durs, pitch, speaker = batch mels, spec_len = self.preprocessor(input_signal=audio, length=audio_lens) reference_spec, reference_spec_len = None, None if reference_audio is not None: reference_spec, reference_spec_len = self.preprocessor( input_signal=reference_audio, length=reference_audio_len ) ( mels_pred, _, _, log_durs_pred, pitch_pred, attn_soft, attn_logprob, attn_hard, attn_hard_dur, pitch, energy_pred, energy_tgt, ) = self( text=text, durs=durs, pitch=pitch, energy=energy, speaker=speaker, pace=1.0, spec=mels if self.learn_alignment else None, reference_spec=reference_spec, reference_spec_lens=reference_spec_len, attn_prior=attn_prior, mel_lens=spec_len, input_lens=text_lens, ) if durs is None: durs = attn_hard_dur mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) loss = mel_loss + dur_loss if self.learn_alignment: ctc_loss = self.forward_sum_loss_fn(attn_logprob=attn_logprob, in_lens=text_lens, out_lens=spec_len) bin_loss_weight = min(self.current_epoch / self.bin_loss_warmup_epochs, 1.0) * 1.0 bin_loss = self.bin_loss_fn(hard_attention=attn_hard, soft_attention=attn_soft) * bin_loss_weight loss += ctc_loss + bin_loss pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) loss += pitch_loss + energy_loss self.log("t_loss", loss) self.log("t_mel_loss", mel_loss) self.log("t_dur_loss", dur_loss) self.log("t_pitch_loss", pitch_loss) if energy_tgt is not None: self.log("t_energy_loss", energy_loss) if self.learn_alignment: self.log("t_ctc_loss", ctc_loss) self.log("t_bin_loss", bin_loss) # Log images to tensorboard if self.log_images and self.log_train_images and isinstance(self.logger, TensorBoardLogger): self.log_train_images = False self.tb_logger.add_image( "train_mel_target", plot_spectrogram_to_numpy(mels[0].data.cpu().float().numpy()), self.global_step, dataformats="HWC", ) spec_predict = mels_pred[0].data.cpu().float().numpy() self.tb_logger.add_image( "train_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", ) if self.learn_alignment: attn = attn_hard[0].data.cpu().float().numpy().squeeze() self.tb_logger.add_image( "train_attn", plot_alignment_to_numpy(attn.T), self.global_step, dataformats="HWC", ) soft_attn = attn_soft[0].data.cpu().float().numpy().squeeze() self.tb_logger.add_image( "train_soft_attn", plot_alignment_to_numpy(soft_attn.T), self.global_step, dataformats="HWC", ) return loss def validation_step(self, batch, batch_idx): attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( None, None, None, None, None, None, ) if self.learn_alignment: if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": batch_dict = batch else: batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) audio = batch_dict.get("audio") audio_lens = batch_dict.get("audio_lens") text = batch_dict.get("text") text_lens = batch_dict.get("text_lens") attn_prior = batch_dict.get("align_prior_matrix", None) pitch = batch_dict.get("pitch", None) energy = batch_dict.get("energy", None) speaker = batch_dict.get("speaker_id", None) reference_audio = batch_dict.get("reference_audio", None) reference_audio_len = batch_dict.get("reference_audio_lens", None) else: audio, audio_lens, text, text_lens, durs, pitch, speaker = batch mels, mel_lens = self.preprocessor(input_signal=audio, length=audio_lens) reference_spec, reference_spec_len = None, None if reference_audio is not None: reference_spec, reference_spec_len = self.preprocessor( input_signal=reference_audio, length=reference_audio_len ) # Calculate val loss on ground truth durations to better align L2 loss in time (mels_pred, _, _, log_durs_pred, pitch_pred, _, _, _, attn_hard_dur, pitch, energy_pred, energy_tgt,) = self( text=text, durs=durs, pitch=pitch, energy=energy, speaker=speaker, pace=1.0, spec=mels if self.learn_alignment else None, reference_spec=reference_spec, reference_spec_lens=reference_spec_len, attn_prior=attn_prior, mel_lens=mel_lens, input_lens=text_lens, ) if durs is None: durs = attn_hard_dur mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) loss = mel_loss + dur_loss + pitch_loss + energy_loss val_outputs = { "val_loss": loss, "mel_loss": mel_loss, "dur_loss": dur_loss, "pitch_loss": pitch_loss, "energy_loss": energy_loss if energy_tgt is not None else None, "mel_target": mels if batch_idx == 0 else None, "mel_pred": mels_pred if batch_idx == 0 else None, } self.validation_step_outputs.append(val_outputs) return val_outputs def on_validation_epoch_end(self): collect = lambda key: torch.stack([x[key] for x in self.validation_step_outputs]).mean() val_loss = collect("val_loss") mel_loss = collect("mel_loss") dur_loss = collect("dur_loss") pitch_loss = collect("pitch_loss") self.log("val_loss", val_loss, sync_dist=True) self.log("val_mel_loss", mel_loss, sync_dist=True) self.log("val_dur_loss", dur_loss, sync_dist=True) self.log("val_pitch_loss", pitch_loss, sync_dist=True) if self.validation_step_outputs[0]["energy_loss"] is not None: energy_loss = collect("energy_loss") self.log("val_energy_loss", energy_loss, sync_dist=True) _, _, _, _, _, spec_target, spec_predict = self.validation_step_outputs[0].values() if self.log_images and isinstance(self.logger, TensorBoardLogger): self.tb_logger.add_image( "val_mel_target", plot_spectrogram_to_numpy(spec_target[0].data.cpu().float().numpy()), self.global_step, dataformats="HWC", ) spec_predict = spec_predict[0].data.cpu().float().numpy() self.tb_logger.add_image( "val_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", ) self.log_train_images = True self.validation_step_outputs.clear() # free memory) def _setup_train_dataloader(self, cfg): phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): phon_mode = self.vocab.set_phone_prob(self.vocab.phoneme_probability) with phon_mode: dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) sampler = dataset.get_sampler(cfg.dataloader_params.batch_size) return torch.utils.data.DataLoader( dataset, collate_fn=dataset.collate_fn, sampler=sampler, **cfg.dataloader_params ) def _setup_test_dataloader(self, cfg): phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): phon_mode = self.vocab.set_phone_prob(0.0) with phon_mode: dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): raise ValueError(f"No dataset for {name}") if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): raise ValueError(f"No dataloader_params for {name}") if shuffle_should_be: if 'shuffle' not in cfg.dataloader_params: logging.warning( f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " "config. Manually setting to True" ) with open_dict(cfg.dataloader_params): cfg.dataloader_params.shuffle = True elif not cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") elif cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") if self.ds_class == "nemo.collections.tts.data.dataset.TTSDataset": phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): phon_mode = self.vocab.set_phone_prob(prob=None if name == "val" else self.vocab.phoneme_probability) with phon_mode: dataset = instantiate( cfg.dataset, text_normalizer=self.normalizer, text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, text_tokenizer=self.vocab, ) else: dataset = instantiate(cfg.dataset) return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) def setup_training_data(self, cfg): if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": self._train_dl = self._setup_train_dataloader(cfg) else: self._train_dl = self.__setup_dataloader_from_config(cfg) def setup_validation_data(self, cfg): if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": self._validation_dl = self._setup_test_dataloader(cfg) else: self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="val") def setup_test_data(self, cfg): """Omitted.""" pass def configure_callbacks(self): if not self.log_config: return [] sample_ds_class = self.log_config.dataset._target_ if sample_ds_class != "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": raise ValueError(f"Logging callback only supported for TextToSpeechDataset, got {sample_ds_class}") data_loader = self._setup_test_dataloader(self.log_config) generators = instantiate(self.log_config.generators) log_dir = Path(self.log_config.log_dir) if self.log_config.log_dir else None log_callback = LoggingCallback( generators=generators, data_loader=data_loader, log_epochs=self.log_config.log_epochs, epoch_frequency=self.log_config.epoch_frequency, output_dir=log_dir, loggers=self.trainer.loggers, log_tensorboard=self.log_config.log_tensorboard, log_wandb=self.log_config.log_wandb, ) return [log_callback] @classmethod def list_available_models(cls) -> 'List[PretrainedModelInfo]': """ This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. Returns: List of available pre-trained models. """ list_of_models = [] # en-US, single speaker, 22050Hz, LJSpeech (ARPABET). model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/1.8.1/files/tts_en_fastpitch_align.nemo", description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is ARPABET-based.", class_=cls, ) list_of_models.append(model) # en-US, single speaker, 22050Hz, LJSpeech (IPA). model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch_ipa", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/IPA_1.13.0/files/tts_en_fastpitch_align_ipa.nemo", description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is IPA-based.", class_=cls, ) list_of_models.append(model) # en-US, multi-speaker, 44100Hz, HiFiTTS. model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch_multispeaker", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_multispeaker_fastpitchhifigan/versions/1.10.0/files/tts_en_fastpitch_multispeaker.nemo", description="This model is trained on HiFITTS sampled at 44100Hz with and can be used to generate male and female English voices with an American accent.", class_=cls, ) list_of_models.append(model) # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 21.02 model = PretrainedModelInfo( pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2102", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2102.nemo", description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 21.02 Dataset sampled at 22050Hz and can be used to generate male German voices.", class_=cls, ) list_of_models.append(model) # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 22.10 model = PretrainedModelInfo( pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2210", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2210.nemo", description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 22.10 Dataset sampled at 22050Hz and can be used to generate male German voices.", class_=cls, ) list_of_models.append(model) # de-DE, multi-speaker, 5 speakers, 44100 Hz, HUI-Audio-Corpus-German Clean. model = PretrainedModelInfo( pretrained_model_name="tts_de_fastpitch_multispeaker_5", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitch_multispeaker_5/versions/1.11.0/files/tts_de_fastpitch_multispeaker_5.nemo", description="This model is trained on 5 speakers in HUI-Audio-Corpus-German clean subset sampled at 44100Hz with and can be used to generate male and female German voices.", class_=cls, ) list_of_models.append(model) # es, 174 speakers, 44100Hz, OpenSLR (IPA) model = PretrainedModelInfo( pretrained_model_name="tts_es_fastpitch_multispeaker", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_es_multispeaker_fastpitchhifigan/versions/1.15.0/files/tts_es_fastpitch_multispeaker.nemo", description="This model is trained on 174 speakers in 6 crowdsourced Latin American Spanish OpenSLR datasets sampled at 44100Hz and can be used to generate male and female Spanish voices with Latin American accents.", class_=cls, ) list_of_models.append(model) # zh, single female speaker, 22050Hz, SFSpeech Bilingual Chinese/English dataset, improved model using richer # dict and jieba word segmenter for polyphone disambiguation. model = PretrainedModelInfo( pretrained_model_name="tts_zh_fastpitch_sfspeech", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_zh_fastpitch_hifigan_sfspeech/versions/1.15.0/files/tts_zh_fastpitch_sfspeech.nemo", description="This model is trained on a single female speaker in SFSpeech Bilingual Chinese/English dataset" " sampled at 22050Hz and can be used to generate female Mandarin Chinese voices. It is improved" " using richer dict and jieba word segmenter for polyphone disambiguation.", class_=cls, ) list_of_models.append(model) # en, multi speaker, LibriTTS, 16000 Hz # stft 25ms 10ms matching ASR params # for use during Enhlish ASR training/adaptation model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch_for_asr_finetuning", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch_spectrogram_enhancer_for_asr_finetuning/versions/1.20.0/files/tts_en_fastpitch_for_asr_finetuning.nemo", description="This model is trained on LibriSpeech, train-960 subset." " STFT parameters follow those commonly used in ASR: 25 ms window, 10 ms hop." " This model is supposed to be used with its companion SpetrogramEnhancer for " " ASR fine-tuning. Usage for regular TTS tasks is not advised.", class_=cls, ) list_of_models.append(model) return list_of_models # Methods for model exportability def _prepare_for_export(self, **kwargs): super()._prepare_for_export(**kwargs) tensor_shape = ('T') if self.export_config["enable_ragged_batches"] else ('B', 'T') # Define input_types and output_types as required by export() self._input_types = { "text": NeuralType(tensor_shape, TokenIndex()), "pitch": NeuralType(tensor_shape, RegressionValuesType()), "pace": NeuralType(tensor_shape), "volume": NeuralType(tensor_shape, optional=True), "batch_lengths": NeuralType(('B'), optional=True), "speaker": NeuralType(('B'), Index(), optional=True), } self._output_types = { "spect": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "num_frames": NeuralType(('B'), TokenDurationType()), "durs_predicted": NeuralType(('B', 'T'), TokenDurationType()), "log_durs_predicted": NeuralType(('B', 'T'), TokenLogDurationType()), "pitch_predicted": NeuralType(('B', 'T'), RegressionValuesType()), } if self.export_config["enable_volume"]: self._output_types["volume_aligned"] = NeuralType(('B', 'T'), RegressionValuesType()) def _export_teardown(self): self._input_types = self._output_types = None @property def disabled_deployment_input_names(self): """Implement this method to return a set of input names disabled for export""" disabled_inputs = set() if self.fastpitch.speaker_emb is None: disabled_inputs.add("speaker") if not self.export_config["enable_ragged_batches"]: disabled_inputs.add("batch_lengths") if not self.export_config["enable_volume"]: disabled_inputs.add("volume") return disabled_inputs @property def input_types(self): return self._input_types @property def output_types(self): return self._output_types def input_example(self, max_batch=1, max_dim=44): """ Generates input examples for tracing etc. Returns: A tuple of input examples. """ par = next(self.fastpitch.parameters()) inputs = sample_tts_input(self.export_config, par.device, max_batch=max_batch, max_dim=max_dim) if 'enable_ragged_batches' not in self.export_config: inputs.pop('batch_lengths', None) return (inputs,) def forward_for_export(self, text, pitch, pace, volume=None, batch_lengths=None, speaker=None): if self.export_config["enable_ragged_batches"]: text, pitch, pace, volume_tensor, lens = batch_from_ragged( text, pitch, pace, batch_lengths, padding_idx=self.fastpitch.encoder.padding_idx, volume=volume ) if volume is not None: volume = volume_tensor return self.fastpitch.infer(text=text, pitch=pitch, pace=pace, volume=volume, speaker=speaker) def interpolate_speaker( self, original_speaker_1, original_speaker_2, weight_speaker_1, weight_speaker_2, new_speaker_id ): """ This method performs speaker interpolation between two original speakers the model is trained on. Inputs: original_speaker_1: Integer speaker ID of first existing speaker in the model original_speaker_2: Integer speaker ID of second existing speaker in the model weight_speaker_1: Floating point weight associated in to first speaker during weight combination weight_speaker_2: Floating point weight associated in to second speaker during weight combination new_speaker_id: Integer speaker ID of new interpolated speaker in the model """ if self.fastpitch.speaker_emb is None: raise Exception( "Current FastPitch model is not a multi-speaker FastPitch model. Speaker interpolation can only \ be performed with a multi-speaker model" ) n_speakers = self.fastpitch.speaker_emb.weight.data.size()[0] if original_speaker_1 >= n_speakers or original_speaker_2 >= n_speakers or new_speaker_id >= n_speakers: raise Exception( f"Parameters original_speaker_1, original_speaker_2, new_speaker_id should be less than the total \ total number of speakers FastPitch was trained on (n_speakers = {n_speakers})." ) speaker_emb_1 = ( self.fastpitch.speaker_emb(torch.tensor(original_speaker_1, dtype=torch.int32).cuda()).clone().detach() ) speaker_emb_2 = ( self.fastpitch.speaker_emb(torch.tensor(original_speaker_2, dtype=torch.int32).cuda()).clone().detach() ) new_speaker_emb = weight_speaker_1 * speaker_emb_1 + weight_speaker_2 * speaker_emb_2 self.fastpitch.speaker_emb.weight.data[new_speaker_id] = new_speaker_emb [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/tts/models/tacotron2.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import contextlib from dataclasses import dataclass from typing import Any, Dict, List, Optional import torch from hydra.utils import instantiate from omegaconf import MISSING, DictConfig, OmegaConf, open_dict from omegaconf.errors import ConfigAttributeError from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger from torch import nn from nemo.collections.common.parts.preprocessing import parsers from nemo.collections.tts.losses.tacotron2loss import Tacotron2Loss from nemo.collections.tts.models.base import SpectrogramGenerator from nemo.collections.tts.parts.utils.helpers import ( g2p_backward_compatible_support, get_mask_from_lengths, tacotron2_log_to_tb_func, tacotron2_log_to_wandb_func, ) from nemo.core.classes.common import PretrainedModelInfo, typecheck from nemo.core.neural_types.elements import ( AudioSignal, EmbeddedTextType, LengthsType, LogitsType, MelSpectrogramType, SequenceToSequenceAlignmentType, ) from nemo.core.neural_types.neural_type import NeuralType from nemo.utils import logging, model_utils @dataclass class Preprocessor: _target_: str = MISSING pad_value: float = MISSING @dataclass class Tacotron2Config: preprocessor: Preprocessor = Preprocessor() encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING labels: List = MISSING train_ds: Optional[Dict[Any, Any]] = None validation_ds: Optional[Dict[Any, Any]] = None class Tacotron2Model(SpectrogramGenerator): """Tacotron 2 Model that is used to generate mel spectrograms from text""" def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None): # Convert to Hydra 1.0 compatible DictConfig cfg = model_utils.convert_model_config_to_dict_config(cfg) cfg = model_utils.maybe_update_config_version(cfg) # setup normalizer self.normalizer = None self.text_normalizer_call = None self.text_normalizer_call_kwargs = {} self._setup_normalizer(cfg) # setup tokenizer self.tokenizer = None if hasattr(cfg, 'text_tokenizer'): self._setup_tokenizer(cfg) self.num_tokens = len(self.tokenizer.tokens) self.tokenizer_pad = self.tokenizer.pad self.tokenizer_unk = self.tokenizer.oov # assert self.tokenizer is not None else: self.num_tokens = len(cfg.labels) + 3 super().__init__(cfg=cfg, trainer=trainer) schema = OmegaConf.structured(Tacotron2Config) # ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes if isinstance(cfg, dict): cfg = OmegaConf.create(cfg) elif not isinstance(cfg, DictConfig): raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") # Ensure passed cfg is compliant with schema try: OmegaConf.merge(cfg, schema) self.pad_value = cfg.preprocessor.pad_value except ConfigAttributeError: self.pad_value = cfg.preprocessor.params.pad_value logging.warning( "Your config is using an old NeMo yaml configuration. Please ensure that the yaml matches the " "current version in the main branch for future compatibility." ) self._parser = None self.audio_to_melspec_precessor = instantiate(cfg.preprocessor) self.text_embedding = nn.Embedding(self.num_tokens, 512) self.encoder = instantiate(self._cfg.encoder) self.decoder = instantiate(self._cfg.decoder) self.postnet = instantiate(self._cfg.postnet) self.loss = Tacotron2Loss() self.calculate_loss = True @property def parser(self): if self._parser is not None: return self._parser ds_class_name = self._cfg.train_ds.dataset._target_.split(".")[-1] if ds_class_name == "TTSDataset": self._parser = None elif hasattr(self._cfg, "labels"): self._parser = parsers.make_parser( labels=self._cfg.labels, name='en', unk_id=-1, blank_id=-1, do_normalize=True, abbreviation_version="fastpitch", make_table=False, ) else: raise ValueError("Wanted to setup parser, but model does not have necessary paramaters") return self._parser def parse(self, text: str, normalize=True) -> torch.Tensor: if self.training: logging.warning("parse() is meant to be called in eval mode.") if normalize and self.text_normalizer_call is not None: text = self.text_normalizer_call(text, **self.text_normalizer_call_kwargs) eval_phon_mode = contextlib.nullcontext() if hasattr(self.tokenizer, "set_phone_prob"): eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0) with eval_phon_mode: if self.tokenizer is not None: tokens = self.tokenizer.encode(text) else: tokens = self.parser(text) # Old parser doesn't add bos and eos ids, so maunally add it tokens = [len(self._cfg.labels)] + tokens + [len(self._cfg.labels) + 1] tokens_tensor = torch.tensor(tokens).unsqueeze_(0).to(self.device) return tokens_tensor @property def input_types(self): if self.training: return { "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), "token_len": NeuralType(('B'), LengthsType()), "audio": NeuralType(('B', 'T'), AudioSignal()), "audio_len": NeuralType(('B'), LengthsType()), } else: return { "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), "token_len": NeuralType(('B'), LengthsType()), "audio": NeuralType(('B', 'T'), AudioSignal(), optional=True), "audio_len": NeuralType(('B'), LengthsType(), optional=True), } @property def output_types(self): if not self.calculate_loss and not self.training: return { "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "gate_pred": NeuralType(('B', 'T'), LogitsType()), "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), "pred_length": NeuralType(('B'), LengthsType()), } return { "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "gate_pred": NeuralType(('B', 'T'), LogitsType()), "spec_target": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "spec_target_len": NeuralType(('B'), LengthsType()), "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), } @typecheck() def forward(self, *, tokens, token_len, audio=None, audio_len=None): if audio is not None and audio_len is not None: spec_target, spec_target_len = self.audio_to_melspec_precessor(audio, audio_len) else: if self.training or self.calculate_loss: raise ValueError( f"'audio' and 'audio_len' can not be None when either 'self.training' or 'self.calculate_loss' is True." ) token_embedding = self.text_embedding(tokens).transpose(1, 2) encoder_embedding = self.encoder(token_embedding=token_embedding, token_len=token_len) if self.training: spec_pred_dec, gate_pred, alignments = self.decoder( memory=encoder_embedding, decoder_inputs=spec_target, memory_lengths=token_len ) else: spec_pred_dec, gate_pred, alignments, pred_length = self.decoder( memory=encoder_embedding, memory_lengths=token_len ) spec_pred_postnet = self.postnet(mel_spec=spec_pred_dec) if not self.calculate_loss and not self.training: return spec_pred_dec, spec_pred_postnet, gate_pred, alignments, pred_length return spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments @typecheck( input_types={"tokens": NeuralType(('B', 'T'), EmbeddedTextType())}, output_types={"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType())}, ) def generate_spectrogram(self, *, tokens): self.eval() self.calculate_loss = False token_len = torch.tensor([len(i) for i in tokens]).to(self.device) tensors = self(tokens=tokens, token_len=token_len) spectrogram_pred = tensors[1] if spectrogram_pred.shape[0] > 1: # Silence all frames past the predicted end mask = ~get_mask_from_lengths(tensors[-1]) mask = mask.expand(spectrogram_pred.shape[1], mask.size(0), mask.size(1)) mask = mask.permute(1, 0, 2) spectrogram_pred.data.masked_fill_(mask, self.pad_value) return spectrogram_pred def training_step(self, batch, batch_idx): audio, audio_len, tokens, token_len = batch spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, _ = self.forward( audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len ) loss, _ = self.loss( spec_pred_dec=spec_pred_dec, spec_pred_postnet=spec_pred_postnet, gate_pred=gate_pred, spec_target=spec_target, spec_target_len=spec_target_len, pad_value=self.pad_value, ) output = { 'loss': loss, 'progress_bar': {'training_loss': loss}, 'log': {'loss': loss}, } return output def validation_step(self, batch, batch_idx): audio, audio_len, tokens, token_len = batch spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments = self.forward( audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len ) loss, gate_target = self.loss( spec_pred_dec=spec_pred_dec, spec_pred_postnet=spec_pred_postnet, gate_pred=gate_pred, spec_target=spec_target, spec_target_len=spec_target_len, pad_value=self.pad_value, ) loss = { "val_loss": loss, "mel_target": spec_target, "mel_postnet": spec_pred_postnet, "gate": gate_pred, "gate_target": gate_target, "alignments": alignments, } self.validation_step_outputs.append(loss) return loss def on_validation_epoch_end(self): if self.logger is not None and self.logger.experiment is not None: logger = self.logger.experiment for logger in self.trainer.loggers: if isinstance(logger, TensorBoardLogger): logger = logger.experiment break if isinstance(logger, TensorBoardLogger): tacotron2_log_to_tb_func( logger, self.validation_step_outputs[0].values(), self.global_step, tag="val", log_images=True, add_audio=False, ) elif isinstance(logger, WandbLogger): tacotron2_log_to_wandb_func( logger, self.validation_step_outputs[0].values(), self.global_step, tag="val", log_images=True, add_audio=False, ) avg_loss = torch.stack( [x['val_loss'] for x in self.validation_step_outputs] ).mean() # This reduces across batches, not workers! self.log('val_loss', avg_loss) self.validation_step_outputs.clear() # free memory def _setup_normalizer(self, cfg): if "text_normalizer" in cfg: normalizer_kwargs = {} if "whitelist" in cfg.text_normalizer: normalizer_kwargs["whitelist"] = self.register_artifact( 'text_normalizer.whitelist', cfg.text_normalizer.whitelist ) try: import nemo_text_processing self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) except Exception as e: logging.error(e) raise ImportError( "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" ) self.text_normalizer_call = self.normalizer.normalize if "text_normalizer_call_kwargs" in cfg: self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs def _setup_tokenizer(self, cfg): text_tokenizer_kwargs = {} if "g2p" in cfg.text_tokenizer and cfg.text_tokenizer.g2p is not None: # for backward compatibility if ( self._is_model_being_restored() and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") ): cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( cfg.text_tokenizer.g2p["_target_"] ) g2p_kwargs = {} if "phoneme_dict" in cfg.text_tokenizer.g2p: g2p_kwargs["phoneme_dict"] = self.register_artifact( 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, ) if "heteronyms" in cfg.text_tokenizer.g2p: g2p_kwargs["heteronyms"] = self.register_artifact( 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, ) text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) self.tokenizer = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): raise ValueError(f"No dataset for {name}") if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): raise ValueError(f"No dataloder_params for {name}") if shuffle_should_be: if 'shuffle' not in cfg.dataloader_params: logging.warning( f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " "config. Manually setting to True" ) with open_dict(cfg.dataloader_params): cfg.dataloader_params.shuffle = True elif not cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") elif not shuffle_should_be and cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") dataset = instantiate( cfg.dataset, text_normalizer=self.normalizer, text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, text_tokenizer=self.tokenizer, ) return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) def setup_training_data(self, cfg): self._train_dl = self.__setup_dataloader_from_config(cfg) def setup_validation_data(self, cfg): self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="validation") @classmethod def list_available_models(cls) -> 'List[PretrainedModelInfo]': """ This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. Returns: List of available pre-trained models. """ list_of_models = [] model = PretrainedModelInfo( pretrained_model_name="tts_en_tacotron2", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_tacotron2/versions/1.10.0/files/tts_en_tacotron2.nemo", description="This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.", class_=cls, aliases=["Tacotron2-22050Hz"], ) list_of_models.append(model) return list_of_models [end of nemo/collections/tts/models/tacotron2.py] [start of nemo/core/config/modelPT.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf import MISSING from nemo.core import config from nemo.core.classes.dataset import DatasetConfig from nemo.utils import exp_manager @dataclass class SchedConfig: name: str = MISSING min_lr: float = 0.0 last_epoch: int = -1 @dataclass class OptimConfig: name: str = MISSING sched: Optional[SchedConfig] = None @dataclass class ModelConfig: """ Model component inside ModelPT """ # ... train_ds: Optional[DatasetConfig] = None validation_ds: Optional[DatasetConfig] = None test_ds: Optional[DatasetConfig] = None optim: Optional[OptimConfig] = None @dataclass class HydraConfig: run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) @dataclass class NemoConfig: name: str = MISSING model: ModelConfig = MISSING trainer: config.TrainerConfig = config.TrainerConfig( strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' ) exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() hydra: HydraConfig = HydraConfig() class ModelConfigBuilder: def __init__(self, model_cfg: ModelConfig): """ Base class for any Model Config Builder. A Model Config Builder is a utility class that accepts a ModelConfig dataclass, and via a set of utility methods (that are implemented by the subclassed ModelConfigBuilder), builds a finalized ModelConfig that can be supplied to a NemoModel dataclass as the `model` component. Subclasses *must* implement the private method `_finalize_cfg`. Inside this method, they must update `self.model_cfg` with all interdependent config options that need to be set (either updated by user explicitly or with their default value). The updated model config must then be preserved in `self.model_cfg`. Example: # Create the config builder config_builder = <subclass>ModelConfigBuilder() # Update the components of the config that are modifiable config_builder.set_X(X) config_builder.set_Y(Y) # Create a "finalized" config dataclass that will contain all the updates # that were specified by the builder model_config = config_builder.build() # Use model config as is (or further update values), then create a new Model model = nemo.<domain>.models.<ModelName>Model(cfg=model_config, trainer=Trainer()) Supported build methods: - set_train_ds: All model configs can accept a subclass of `DatasetConfig` as their training config. Subclasses can override this method to enable auto-complete by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. - set_validation_ds: All model configs can accept a subclass of `DatasetConfig` as their validation config. Subclasses can override this method to enable auto-complete by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. - set_test_ds: All model configs can accept a subclass of `DatasetConfig` as their test config. Subclasses can override this method to enable auto-complete by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. - set_optim: A build method that supports changes to the Optimizer (and optionally, the Scheduler) used for training the model. The function accepts two inputs - `cfg`: A subclass of `OptimizerParams` - any OptimizerParams subclass can be used, in order to select an appropriate Optimizer. Examples: AdamParams. `sched_cfg`: A subclass of `SchedulerParams` - any SchedulerParams subclass can be used, in order to select an appropriate Scheduler. Examples: CosineAnnealingParams. Note that this argument is optional. - build(): The method which should return a "finalized" ModelConfig dataclass. Subclasses *should* always override this method, and update the signature of this method with the return type of the Dataclass, so that it enables autocomplete for the user. Example: def build(self) -> EncDecCTCConfig: return super().build() Any additional build methods must be added by subclasses of ModelConfigBuilder. Args: model_cfg: """ self.model_cfg = model_cfg self.train_ds_cfg = None self.validation_ds_cfg = None self.test_ds_cfg = None self.optim_cfg = None def set_train_ds(self, cfg: Optional[DatasetConfig] = None): self.model_cfg.train_ds = cfg def set_validation_ds(self, cfg: Optional[DatasetConfig] = None): self.model_cfg.validation_ds = cfg def set_test_ds(self, cfg: Optional[DatasetConfig] = None): self.model_cfg.test_ds = cfg def set_optim(self, cfg: config.OptimizerParams, sched_cfg: Optional[config.SchedulerParams] = None): @dataclass class WrappedOptimConfig(OptimConfig, cfg.__class__): pass # Setup optim optim_name = cfg.__class__.__name__.replace("Params", "").lower() wrapped_cfg = WrappedOptimConfig(name=optim_name, sched=None, **vars(cfg)) if sched_cfg is not None: @dataclass class WrappedSchedConfig(SchedConfig, sched_cfg.__class__): pass # Setup scheduler sched_name = sched_cfg.__class__.__name__.replace("Params", "") wrapped_sched_cfg = WrappedSchedConfig(name=sched_name, **vars(sched_cfg)) wrapped_cfg.sched = wrapped_sched_cfg self.model_cfg.optim = wrapped_cfg def _finalize_cfg(self): raise NotImplementedError() def build(self) -> ModelConfig: # validate config self._finalize_cfg() return self.model_cfg [end of nemo/core/config/modelPT.py] [start of nemo/utils/exp_manager.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import glob import os import subprocess import sys import time import warnings from dataclasses import dataclass from datetime import timedelta from pathlib import Path from shutil import copy, move from typing import Any, Dict, List, Optional, Tuple, Union import pytorch_lightning import torch from hydra.core.hydra_config import HydraConfig from hydra.utils import get_original_cwd from omegaconf import DictConfig, OmegaConf, open_dict from pytorch_lightning.callbacks import Callback, ModelCheckpoint from pytorch_lightning.callbacks.early_stopping import EarlyStopping from pytorch_lightning.callbacks.timer import Interval, Timer from pytorch_lightning.loggers import MLFlowLogger, TensorBoardLogger, WandbLogger from pytorch_lightning.loops import _TrainingEpochLoop from pytorch_lightning.strategies.ddp import DDPStrategy from nemo.collections.common.callbacks import EMA from nemo.constants import NEMO_ENV_VARNAME_TESTING, NEMO_ENV_VARNAME_VERSION from nemo.utils import logging, timers from nemo.utils.app_state import AppState from nemo.utils.callbacks import NeMoModelCheckpoint, PreemptionCallback from nemo.utils.env_var_parsing import get_envbool from nemo.utils.exceptions import NeMoBaseException from nemo.utils.get_rank import is_global_rank_zero from nemo.utils.lightning_logger_patch import add_filehandlers_to_pl_logger from nemo.utils.loggers import ClearMLLogger, ClearMLParams, DLLogger, DLLoggerParams, MLFlowParams from nemo.utils.model_utils import uninject_model_parallel_rank class NotFoundError(NeMoBaseException): """ Raised when a file or folder is not found""" class LoggerMisconfigurationError(NeMoBaseException): """ Raised when a mismatch between trainer.logger and exp_manager occurs""" def __init__(self, message): message = ( message + " You can disable lighning's trainer from creating a logger by passing logger=False to its constructor." ) super().__init__(message) class CheckpointMisconfigurationError(NeMoBaseException): """ Raised when a mismatch between trainer.callbacks and exp_manager occurs""" @dataclass class EarlyStoppingParams: monitor: str = "val_loss" # The metric that early stopping should consider. mode: str = "min" # inform early stopping whether to look for increase or decrease in monitor. min_delta: float = 0.001 # smallest change to consider as improvement. patience: int = 10 # how many (continuous) validation cycles to wait with no improvement and stopping training. verbose: bool = True strict: bool = True check_finite: bool = True stopping_threshold: Optional[float] = None divergence_threshold: Optional[float] = None check_on_train_epoch_end: Optional[bool] = None log_rank_zero_only: bool = False @dataclass class CallbackParams: filepath: Optional[str] = None # Deprecated dirpath: Optional[str] = None # If None, exp_manager will attempt to handle the filepath filename: Optional[str] = None # If None, exp_manager will attempt to handle the filepath monitor: Optional[str] = "val_loss" verbose: Optional[bool] = True save_last: Optional[bool] = True save_top_k: Optional[int] = 3 save_weights_only: Optional[bool] = False mode: Optional[str] = "min" auto_insert_metric_name: bool = True every_n_epochs: Optional[int] = 1 every_n_train_steps: Optional[int] = None train_time_interval: Optional[str] = None prefix: Optional[str] = None # If None, exp_manager will attempt to handle the filepath postfix: str = ".nemo" save_best_model: bool = False always_save_nemo: bool = False save_nemo_on_train_end: Optional[bool] = True # Whether to automatically save .nemo file durin on_train_end hook model_parallel_size: Optional[int] = None # tensor parallel size * pipeline parallel size save_on_train_epoch_end: Optional[bool] = False # Save after training, not after validation @dataclass class StepTimingParams: reduction: Optional[str] = "mean" # if True torch.cuda.synchronize() is called on start/stop sync_cuda: Optional[bool] = False # if positive, defines the size of a sliding window for computing mean buffer_size: Optional[int] = 1 @dataclass class EMAParams: enable: Optional[bool] = False decay: Optional[float] = 0.999 cpu_offload: Optional[bool] = False validate_original_weights: Optional[bool] = False every_n_steps: int = 1 @dataclass class ExpManagerConfig: """Experiment Manager config for validation of passed arguments. """ # Log dir creation parameters explicit_log_dir: Optional[str] = None exp_dir: Optional[str] = None name: Optional[str] = None version: Optional[str] = None use_datetime_version: Optional[bool] = True resume_if_exists: Optional[bool] = False resume_past_end: Optional[bool] = False resume_ignore_no_checkpoint: Optional[bool] = False resume_from_checkpoint: Optional[str] = None # Logging parameters create_tensorboard_logger: Optional[bool] = True summary_writer_kwargs: Optional[Dict[Any, Any]] = None create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() create_dllogger_logger: Optional[bool] = False dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() create_clearml_logger: Optional[bool] = False clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() create_early_stopping_callback: Optional[bool] = False early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True ema: Optional[EMAParams] = EMAParams() # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization seconds_to_sleep: float = 5 class TimingCallback(Callback): """ Logs execution time of train/val/test steps """ def __init__(self, timer_kwargs={}): self.timer = timers.NamedTimer(**timer_kwargs) def _on_batch_start(self, name): # reset only if we do not return mean of a sliding window if self.timer.buffer_size <= 0: self.timer.reset(name) self.timer.start(name) def _on_batch_end(self, name, pl_module): self.timer.stop(name) # Set the `batch_size=1` as WAR for `dataloader_iter`, which is not used for any metric pl_module.log( name + ' in s', self.timer[name], on_step=True, on_epoch=False, batch_size=1, prog_bar=(name == "train_step_timing"), ) def on_train_batch_start(self, trainer, pl_module, batch, batch_idx): self._on_batch_start("train_step_timing") def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): self._on_batch_end("train_step_timing", pl_module) def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): self._on_batch_start("validation_step_timing") def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): self._on_batch_end("validation_step_timing", pl_module) def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): self._on_batch_start("test_step_timing") def on_test_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): self._on_batch_end("test_step_timing", pl_module) def on_before_backward(self, trainer, pl_module, loss): self._on_batch_start("train_backward_timing") def on_after_backward(self, trainer, pl_module): self._on_batch_end("train_backward_timing", pl_module) def exp_manager(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None) -> Optional[Path]: """ exp_manager is a helper function used to manage folders for experiments. It follows the pytorch lightning paradigm of exp_dir/model_or_experiment_name/version. If the lightning trainer has a logger, exp_manager will get exp_dir, name, and version from the logger. Otherwise it will use the exp_dir and name arguments to create the logging directory. exp_manager also allows for explicit folder creation via explicit_log_dir. The version can be a datetime string or an integer. Datestime version can be disabled if use_datetime_version is set to False. It optionally creates TensorBoardLogger, WandBLogger, DLLogger, MLFlowLogger, ClearMLLogger, ModelCheckpoint objects from pytorch lightning. It copies sys.argv, and git information if available to the logging directory. It creates a log file for each process to log their output into. exp_manager additionally has a resume feature (resume_if_exists) which can be used to continuing training from the constructed log_dir. When you need to continue the training repeatedly (like on a cluster which you need multiple consecutive jobs), you need to avoid creating the version folders. Therefore from v1.0.0, when resume_if_exists is set to True, creating the version folders is ignored. Args: trainer (pytorch_lightning.Trainer): The lightning trainer. cfg (DictConfig, dict): Can have the following keys: - explicit_log_dir (str, Path): Can be used to override exp_dir/name/version folder creation. Defaults to None, which will use exp_dir, name, and version to construct the logging directory. - exp_dir (str, Path): The base directory to create the logging directory. Defaults to None, which logs to ./nemo_experiments. - name (str): The name of the experiment. Defaults to None which turns into "default" via name = name or "default". - version (str): The version of the experiment. Defaults to None which uses either a datetime string or lightning's TensorboardLogger system of using version_{int}. - use_datetime_version (bool): Whether to use a datetime string for version. Defaults to True. - resume_if_exists (bool): Whether this experiment is resuming from a previous run. If True, it sets trainer._checkpoint_connector._ckpt_path so that the trainer should auto-resume. exp_manager will move files under log_dir to log_dir/run_{int}. Defaults to False. From v1.0.0, when resume_if_exists is True, we would not create version folders to make it easier to find the log folder for next runs. - resume_past_end (bool): exp_manager errors out if resume_if_exists is True and a checkpoint matching ``*end.ckpt`` indicating a previous training run fully completed. This behaviour can be disabled, in which case the ``*end.ckpt`` will be loaded by setting resume_past_end to True. Defaults to False. - resume_ignore_no_checkpoint (bool): exp_manager errors out if resume_if_exists is True and no checkpoint could be found. This behaviour can be disabled, in which case exp_manager will print a message and continue without restoring, by setting resume_ignore_no_checkpoint to True. Defaults to False. - resume_from_checkpoint (str): Can be used to specify a path to a specific checkpoint file to load from. This will override any checkpoint found when resume_if_exists is True. Defaults to None. - create_tensorboard_logger (bool): Whether to create a tensorboard logger and attach it to the pytorch lightning trainer. Defaults to True. - summary_writer_kwargs (dict): A dictionary of kwargs that can be passed to lightning's TensorboardLogger class. Note that log_dir is passed by exp_manager and cannot exist in this dict. Defaults to None. - create_wandb_logger (bool): Whether to create a Weights and Baises logger and attach it to the pytorch lightning trainer. Defaults to False. - wandb_logger_kwargs (dict): A dictionary of kwargs that can be passed to lightning's WandBLogger class. Note that name and project are required parameters if create_wandb_logger is True. Defaults to None. - create_mlflow_logger (bool): Whether to create an MLFlow logger and attach it to the pytorch lightning training. Defaults to False - mlflow_logger_kwargs (dict): optional parameters for the MLFlow logger - create_dllogger_logger (bool): Whether to create an DLLogger logger and attach it to the pytorch lightning training. Defaults to False - dllogger_logger_kwargs (dict): optional parameters for the DLLogger logger - create_clearml_logger (bool): Whether to create an ClearML logger and attach it to the pytorch lightning training. Defaults to False - clearml_logger_kwargs (dict): optional parameters for the ClearML logger - create_checkpoint_callback (bool): Whether to create a ModelCheckpoint callback and attach it to the pytorch lightning trainer. The ModelCheckpoint saves the top 3 models with the best "val_loss", the most recent checkpoint under ``*last.ckpt``, and the final checkpoint after training completes under ``*end.ckpt``. Defaults to True. - create_early_stopping_callback (bool): Flag to decide if early stopping should be used to stop training. Default is False. See EarlyStoppingParams dataclass above. - create_preemption_callback (bool): Flag to decide whether to enable preemption callback to save checkpoints and exit training immediately upon preemption. Default is True. - files_to_copy (list): A list of files to copy to the experiment logging directory. Defaults to None which copies no files. - log_local_rank_0_only (bool): Whether to only create log files for local rank 0. Defaults to False. Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. - log_global_rank_0_only (bool): Whether to only create log files for global rank 0. Defaults to False. Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. - max_time (str): The maximum wall clock time *per run*. This is intended to be used on clusters where you want a checkpoint to be saved after this specified time and be able to resume from that checkpoint. Defaults to None. - seconds_to_sleep (float): seconds to sleep non rank 0 processes for. Used to give enough time for rank 0 to initialize returns: log_dir (Path): The final logging directory where logging files are saved. Usually the concatenation of exp_dir, name, and version. """ # Add rank information to logger # Note: trainer.global_rank and trainer.is_global_zero are not set until trainer.fit, so have to hack around it local_rank = int(os.environ.get("LOCAL_RANK", 0)) global_rank = trainer.node_rank * trainer.num_devices + local_rank logging.rank = global_rank if cfg is None: logging.error("exp_manager did not receive a cfg argument. It will be disabled.") return if trainer.fast_dev_run: logging.info("Trainer was called with fast_dev_run. exp_manager will return without any functionality.") return # Ensure passed cfg is compliant with ExpManagerConfig schema = OmegaConf.structured(ExpManagerConfig) if isinstance(cfg, dict): cfg = OmegaConf.create(cfg) elif not isinstance(cfg, DictConfig): raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True)) cfg = OmegaConf.merge(schema, cfg) error_checks(trainer, cfg) # Ensures that trainer options are compliant with NeMo and exp_manager arguments log_dir, exp_dir, name, version = get_log_dir( trainer=trainer, exp_dir=cfg.exp_dir, name=cfg.name, version=cfg.version, explicit_log_dir=cfg.explicit_log_dir, use_datetime_version=cfg.use_datetime_version, resume_if_exists=cfg.resume_if_exists, ) check_resume( trainer, log_dir, cfg.resume_if_exists, cfg.resume_past_end, cfg.resume_ignore_no_checkpoint, cfg.checkpoint_callback_params.dirpath, cfg.resume_from_checkpoint, ) checkpoint_name = name # If name returned from get_log_dir is "", use cfg.name for checkpointing if checkpoint_name is None or checkpoint_name == '': checkpoint_name = cfg.name or "default" # Set mlflow name if it's not set, before the main name is erased if cfg.create_mlflow_logger and (not cfg.mlflow_logger_kwargs.get("experiment_name", None)): cfg.mlflow_logger_kwargs.experiment_name = cfg.name logging.warning( 'mlflow logger specified but no experiment name set. Using the same as Tensorboard: %s', cfg.mlflow_logger_kwargs.experiment_name, ) cfg.name = name # Used for configure_loggers so that the log_dir is properly set even if name is "" cfg.version = version # update app_state with log_dir, exp_dir, etc app_state = AppState() app_state.log_dir = log_dir app_state.exp_dir = exp_dir app_state.name = name app_state.version = version app_state.checkpoint_name = checkpoint_name app_state.create_checkpoint_callback = cfg.create_checkpoint_callback app_state.checkpoint_callback_params = cfg.checkpoint_callback_params # Create the logging directory if it does not exist os.makedirs(log_dir, exist_ok=True) # Cannot limit creation to global zero as all ranks write to own log file logging.info(f'Experiments will be logged at {log_dir}') trainer._default_root_dir = log_dir if cfg.log_local_rank_0_only is True and cfg.log_global_rank_0_only is True: raise ValueError( f"Cannot set both log_local_rank_0_only and log_global_rank_0_only to True. Please set either one or neither." ) # This is set if the env var NEMO_TESTING is set to True. nemo_testing = get_envbool(NEMO_ENV_VARNAME_TESTING, False) # Handle logging to file log_file = log_dir / f'nemo_log_globalrank-{global_rank}_localrank-{local_rank}.txt' if cfg.log_local_rank_0_only is True and not nemo_testing: if local_rank == 0: logging.add_file_handler(log_file) elif cfg.log_global_rank_0_only is True and not nemo_testing: if global_rank == 0: logging.add_file_handler(log_file) else: # Logs on all ranks. logging.add_file_handler(log_file) # For some reason, LearningRateLogger requires trainer to have a logger. Safer to create logger on all ranks # not just global rank 0. if ( cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger or cfg.create_dllogger_logger or cfg.create_clearml_logger ): configure_loggers( trainer, exp_dir, log_dir, cfg.name, cfg.version, cfg.checkpoint_callback_params, cfg.create_tensorboard_logger, cfg.summary_writer_kwargs, cfg.create_wandb_logger, cfg.wandb_logger_kwargs, cfg.create_mlflow_logger, cfg.mlflow_logger_kwargs, cfg.create_dllogger_logger, cfg.dllogger_logger_kwargs, cfg.create_clearml_logger, cfg.clearml_logger_kwargs, ) # add loggers timing callbacks if cfg.log_step_timing: timing_callback = TimingCallback(timer_kwargs=cfg.step_timing_kwargs or {}) trainer.callbacks.insert(0, timing_callback) if cfg.ema.enable: ema_callback = EMA( decay=cfg.ema.decay, validate_original_weights=cfg.ema.validate_original_weights, cpu_offload=cfg.ema.cpu_offload, every_n_steps=cfg.ema.every_n_steps, ) trainer.callbacks.append(ema_callback) if cfg.create_early_stopping_callback: early_stop_callback = EarlyStopping(**cfg.early_stopping_callback_params) trainer.callbacks.append(early_stop_callback) if cfg.create_checkpoint_callback: configure_checkpointing( trainer, log_dir, checkpoint_name, cfg.resume_if_exists, cfg.checkpoint_callback_params, cfg.create_preemption_callback, ) if cfg.disable_validation_on_resume: # extend training loop to skip initial validation when resuming from checkpoint configure_no_restart_validation_training_loop(trainer) # Setup a stateless timer for use on clusters. if cfg.max_time_per_run is not None: found_ptl_timer = False for idx, callback in enumerate(trainer.callbacks): if isinstance(callback, Timer): # NOTE: PTL does not expose a `trainer.max_time`. By the time we are in this function, PTL has already setup a timer if the user specifies `trainer.max_time` so best we can do is replace that. # Working: If only `trainer.max_time` is set - it behaves as a normal PTL timer. If only `exp_manager.max_time_per_run` is set - it behaves as a StateLessTimer. If both are set, it also behaves as a StateLessTimer. logging.warning( f'Found a PTL Timer callback, replacing with a StatelessTimer callback. This will happen if you set trainer.max_time as well as exp_manager.max_time_per_run.' ) trainer.callbacks[idx] = StatelessTimer(cfg.max_time_per_run) found_ptl_timer = True break if not found_ptl_timer: trainer.max_time = cfg.max_time_per_run trainer.callbacks.append(StatelessTimer(cfg.max_time_per_run)) if is_global_rank_zero(): # Move files_to_copy to folder and add git information if present if cfg.files_to_copy: for _file in cfg.files_to_copy: copy(Path(_file), log_dir) # Create files for cmd args and git info with open(log_dir / 'cmd-args.log', 'w', encoding='utf-8') as _file: _file.write(" ".join(sys.argv)) # Try to get git hash git_repo, git_hash = get_git_hash() if git_repo: with open(log_dir / 'git-info.log', 'w', encoding='utf-8') as _file: _file.write(f'commit hash: {git_hash}') _file.write(get_git_diff()) # Add err_file logging to global_rank zero logging.add_err_file_handler(log_dir / 'nemo_error_log.txt') # Add lightning file logging to global_rank zero add_filehandlers_to_pl_logger(log_dir / 'lightning_logs.txt', log_dir / 'nemo_error_log.txt') elif trainer.num_nodes * trainer.num_devices > 1: # sleep other ranks so rank 0 can finish # doing the initialization such as moving files time.sleep(cfg.seconds_to_sleep) return log_dir def error_checks(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None): """ Checks that the passed trainer is compliant with NeMo and exp_manager's passed configuration. Checks that: - Throws error when hydra has changed the working directory. This causes issues with lightning's DDP - Throws error when trainer has loggers defined but create_tensorboard_logger or create_wandB_logger or create_mlflow_logger or create_dllogger_logger is True - Prints error messages when 1) run on multi-node and not Slurm, and 2) run on multi-gpu without DDP """ if HydraConfig.initialized() and get_original_cwd() != os.getcwd(): raise ValueError( "Hydra changed the working directory. This interferes with ExpManger's functionality. Please pass " "hydra.run.dir=. to your python script." ) if trainer.logger is not None and ( cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger ): raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger, and either " f"create_tensorboard_logger: {cfg.create_tensorboard_logger} or create_wandb_logger: " f"{cfg.create_wandb_logger} or create_mlflow_logger: {cfg.create_mlflow_logger}" f"or create_dllogger_logger: {cfg.create_mlflow_logger} was set to True. " "These can only be used if trainer does not already have a logger." ) if trainer.num_nodes > 1 and not check_slurm(trainer): logging.error( "You are running multi-node training without SLURM handling the processes." " Please note that this is not tested in NeMo and could result in errors." ) if trainer.num_devices > 1 and not isinstance(trainer.strategy, DDPStrategy): logging.error( "You are running multi-gpu without ddp.Please note that this is not tested in NeMo and could result in " "errors." ) def check_resume( trainer: 'pytorch_lightning.Trainer', log_dir: str, resume_if_exists: bool = False, resume_past_end: bool = False, resume_ignore_no_checkpoint: bool = False, dirpath: str = None, resume_from_checkpoint: str = None, ): """Checks that resume=True was used correctly with the arguments pass to exp_manager. Sets trainer._checkpoint_connector._ckpt_path as necessary. Returns: log_dir (Path): The log_dir exp_dir (str): The base exp_dir without name nor version name (str): The name of the experiment version (str): The version of the experiment Raises: NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. ValueError: If resume is True, and there were more than 1 checkpoint could found. """ if not log_dir: raise ValueError(f"Resuming requires the log_dir {log_dir} to be passed to exp_manager") checkpoint = None if resume_from_checkpoint: checkpoint = resume_from_checkpoint if resume_if_exists: # Use <log_dir>/checkpoints/ unless `dirpath` is set checkpoint_dir = Path(dirpath) if dirpath else Path(Path(log_dir) / "checkpoints") # when using distributed checkpointing, checkpoint_dir is a directory of directories # we check for this here dist_checkpoints = [d for d in list(checkpoint_dir.glob("*")) if d.is_dir()] end_dist_checkpoints = [d for d in dist_checkpoints if d.match("*end")] last_dist_checkpoints = [d for d in dist_checkpoints if d.match("*last")] end_checkpoints = end_dist_checkpoints if end_dist_checkpoints else list(checkpoint_dir.rglob("*end.ckpt")) last_checkpoints = last_dist_checkpoints if last_dist_checkpoints else list(checkpoint_dir.rglob("*last.ckpt")) if not checkpoint_dir.exists() or (not len(end_checkpoints) > 0 and not len(last_checkpoints) > 0): if resume_ignore_no_checkpoint: warn = f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. " if checkpoint is None: warn += "Training from scratch." elif checkpoint == resume_from_checkpoint: warn += f"Training from {resume_from_checkpoint}." logging.warning(warn) else: raise NotFoundError( f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. Cannot resume." ) elif len(end_checkpoints) > 0: if resume_past_end: if len(end_checkpoints) > 1: if 'mp_rank' in str(end_checkpoints[0]): checkpoint = end_checkpoints[0] else: raise ValueError(f"Multiple checkpoints {end_checkpoints} that matches *end.ckpt.") else: raise ValueError( f"Found {end_checkpoints[0]} indicating that the last training run has already completed." ) elif len(last_checkpoints) > 1: if 'mp_rank' in str(last_checkpoints[0]) or 'tp_rank' in str(last_checkpoints[0]): checkpoint = last_checkpoints[0] checkpoint = uninject_model_parallel_rank(checkpoint) else: raise ValueError(f"Multiple checkpoints {last_checkpoints} that matches *last.ckpt.") else: checkpoint = last_checkpoints[0] # PTL 2.0 supports ckpt_path instead of resume_from_checkpoint as the trainer flag if checkpoint is not None: trainer.ckpt_path = str(checkpoint) logging.info(f'Resuming training from checkpoint: {trainer.ckpt_path}') if is_global_rank_zero(): # Check to see if any files exist that need to be moved files_to_move = [] if Path(log_dir).exists(): for child in Path(log_dir).iterdir(): if child.is_file(): files_to_move.append(child) if len(files_to_move) > 0: # Move old files to a new folder other_run_dirs = Path(log_dir).glob("run_*") run_count = 0 for fold in other_run_dirs: if fold.is_dir(): run_count += 1 new_run_dir = Path(Path(log_dir) / f"run_{run_count}") new_run_dir.mkdir() for _file in files_to_move: move(str(_file), str(new_run_dir)) def check_explicit_log_dir( trainer: 'pytorch_lightning.Trainer', explicit_log_dir: Union[Path, str], exp_dir: str, name: str, version: str ) -> Tuple[Path, str, str, str]: """ Checks that the passed arguments are compatible with explicit_log_dir. Returns: log_dir (Path): the log_dir exp_dir (str): the base exp_dir without name nor version name (str): The name of the experiment version (str): The version of the experiment Raise: LoggerMisconfigurationError """ if trainer.logger is not None: raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger and explicit_log_dir: " f"{explicit_log_dir} was pass to exp_manager. Please remove the logger from the lightning trainer." ) # Checking only (explicit_log_dir) vs (exp_dir and version). # The `name` will be used as the actual name of checkpoint/archive. if exp_dir or version: logging.error( f"exp_manager received explicit_log_dir: {explicit_log_dir} and at least one of exp_dir: {exp_dir}, " f"or version: {version}. Please note that exp_dir, name, and version will be ignored." ) if is_global_rank_zero() and Path(explicit_log_dir).exists(): logging.warning(f"Exp_manager is logging to {explicit_log_dir}, but it already exists.") return Path(explicit_log_dir), str(explicit_log_dir), "", "" def get_log_dir( trainer: 'pytorch_lightning.Trainer', exp_dir: str = None, name: str = None, version: str = None, explicit_log_dir: str = None, use_datetime_version: bool = True, resume_if_exists: bool = False, ) -> Tuple[Path, str, str, str]: """ Obtains the log_dir used for exp_manager. Returns: log_dir (Path): the log_dir exp_dir (str): the base exp_dir without name nor version name (str): The name of the experiment version (str): The version of the experiment explicit_log_dir (str): The explicit path to the log folder. Defaults to False. use_datetime_version (bool): Uses date and time as the version of the log folder. Defaults to True. resume_if_exists (bool): if resume_if_exists of the exp_manager's config is enabled or not. When enabled, the version folders would not get created. Raise: LoggerMisconfigurationError: If trainer is incompatible with arguments NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. ValueError: If resume is True, and there were more than 1 checkpoint could found. """ if explicit_log_dir: # If explicit log_dir was passed, short circuit return check_explicit_log_dir(trainer, explicit_log_dir, exp_dir, name, version) # Default exp_dir to ./nemo_experiments if None was passed _exp_dir = exp_dir if exp_dir is None: _exp_dir = str(Path.cwd() / 'nemo_experiments') # If the user has already defined a logger for the trainer, use the logger defaults for logging directory if trainer.logger is not None: if trainer.logger.save_dir: if exp_dir: raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger, the logger's " f"save_dir was not None, and exp_dir ({exp_dir}) was not None. If trainer.logger.save_dir " "exists, exp_manager will use trainer.logger.save_dir as the logging directory and exp_dir " "must be None." ) _exp_dir = trainer.logger.save_dir if name: raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger, and name: " f"{name} was also passed to exp_manager. If the trainer contains a " "logger, exp_manager will use trainer.logger.name, and name passed to exp_manager must be None." ) name = trainer.logger.name version = f"version_{trainer.logger.version}" # Use user-defined exp_dir, project_name, exp_name, and versioning options else: name = name or "default" version = version or os.environ.get(NEMO_ENV_VARNAME_VERSION, None) if not version: if resume_if_exists: logging.warning( "No version folders would be created under the log folder as 'resume_if_exists' is enabled." ) version = None elif is_global_rank_zero(): if use_datetime_version: version = time.strftime('%Y-%m-%d_%H-%M-%S') else: tensorboard_logger = TensorBoardLogger(save_dir=Path(_exp_dir), name=name, version=version) version = f"version_{tensorboard_logger.version}" os.environ[NEMO_ENV_VARNAME_VERSION] = "" if version is None else version log_dir = Path(_exp_dir) / Path(str(name)) / Path("" if version is None else str(version)) return log_dir, str(_exp_dir), name, version def get_git_hash(): """ Helper function that tries to get the commit hash if running inside a git folder returns: Bool: Whether the git subprocess ran without error str: git subprocess output or error message """ try: return ( True, subprocess.check_output(['git', 'rev-parse', 'HEAD'], stderr=subprocess.STDOUT).decode(), ) except subprocess.CalledProcessError as err: return False, "{}\n".format(err.output.decode("utf-8")) def get_git_diff(): """ Helper function that tries to get the git diff if running inside a git folder returns: Bool: Whether the git subprocess ran without error str: git subprocess output or error message """ try: return subprocess.check_output(['git', 'diff'], stderr=subprocess.STDOUT).decode() except subprocess.CalledProcessError as err: return "{}\n".format(err.output.decode("utf-8")) def configure_loggers( trainer: 'pytorch_lightning.Trainer', exp_dir: [Path, str], log_dir: [Path, str], name: str, version: str, checkpoint_callback_params: dict, create_tensorboard_logger: bool, summary_writer_kwargs: dict, create_wandb_logger: bool, wandb_kwargs: dict, create_mlflow_logger: bool, mlflow_kwargs: dict, create_dllogger_logger: bool, dllogger_kwargs: dict, create_clearml_logger: bool, clearml_kwargs: dict, ): """ Creates TensorboardLogger and/or WandBLogger / MLFlowLogger / DLlogger / ClearMLLogger and attach them to trainer. Raises ValueError if summary_writer_kwargs or wandb_kwargs are misconfigured. """ # Potentially create tensorboard logger and/or WandBLogger / MLFlowLogger / DLLogger logger_list = [] if create_tensorboard_logger: if summary_writer_kwargs is None: summary_writer_kwargs = {} elif "log_dir" in summary_writer_kwargs: raise ValueError( "You cannot pass `log_dir` as part of `summary_writer_kwargs`. `log_dir` is handled by lightning's " "TensorBoardLogger logger." ) tensorboard_logger = TensorBoardLogger(save_dir=exp_dir, name=name, version=version, **summary_writer_kwargs) logger_list.append(tensorboard_logger) logging.info("TensorboardLogger has been set up") if create_wandb_logger: if wandb_kwargs is None: wandb_kwargs = {} if "name" not in wandb_kwargs and "project" not in wandb_kwargs: raise ValueError("name and project are required for wandb_logger") # Update the wandb save_dir if wandb_kwargs.get('save_dir', None) is None: wandb_kwargs['save_dir'] = exp_dir os.makedirs(wandb_kwargs['save_dir'], exist_ok=True) wandb_logger = WandbLogger(version=version, **wandb_kwargs) logger_list.append(wandb_logger) logging.info("WandBLogger has been set up") if create_mlflow_logger: mlflow_logger = MLFlowLogger(run_name=version, **mlflow_kwargs) logger_list.append(mlflow_logger) logging.info("MLFlowLogger has been set up") if create_dllogger_logger: dllogger_logger = DLLogger(**dllogger_kwargs) logger_list.append(dllogger_logger) logging.info("DLLogger has been set up") if create_clearml_logger: clearml_logger = ClearMLLogger( clearml_cfg=clearml_kwargs, log_dir=log_dir, prefix=name, save_best_model=checkpoint_callback_params.save_best_model, ) logger_list.append(clearml_logger) logging.info("ClearMLLogger has been set up") trainer._logger_connector.configure_logger(logger_list) def configure_checkpointing( trainer: 'pytorch_lightning.Trainer', log_dir: Path, name: str, resume: bool, params: 'DictConfig', create_preemption_callback: bool, ): """ Adds ModelCheckpoint to trainer. Raises CheckpointMisconfigurationError if trainer already has a ModelCheckpoint callback """ for callback in trainer.callbacks: if isinstance(callback, ModelCheckpoint): raise CheckpointMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a ModelCheckpoint " "and create_checkpoint_callback was set to True. Please either set create_checkpoint_callback " "to False, or remove ModelCheckpoint from the lightning trainer" ) # Create the callback and attach it to trainer if "filepath" in params: if params.filepath is not None: logging.warning("filepath is deprecated. Please switch to dirpath and filename instead") if params.dirpath is None: params.dirpath = Path(params.filepath).parent if params.filename is None: params.filename = Path(params.filepath).name with open_dict(params): del params["filepath"] if params.dirpath is None: params.dirpath = Path(log_dir / 'checkpoints') if params.filename is None: params.filename = f'{name}--{{{params.monitor}:.4f}}-{{epoch}}' if params.prefix is None: params.prefix = name NeMoModelCheckpoint.CHECKPOINT_NAME_LAST = params.filename + '-last' logging.debug(params.dirpath) logging.debug(params.filename) logging.debug(params.prefix) if "val" in params.monitor: if ( trainer.max_epochs is not None and trainer.max_epochs != -1 and trainer.max_epochs < trainer.check_val_every_n_epoch ): logging.error( "The checkpoint callback was told to monitor a validation value but trainer.max_epochs(" f"{trainer.max_epochs}) was less than trainer.check_val_every_n_epoch({trainer.check_val_every_n_epoch}" f"). It is very likely this run will fail with ModelCheckpoint(monitor='{params.monitor}') not found " "in the returned metrics. Please ensure that validation is run within trainer.max_epochs." ) elif trainer.max_steps is not None and trainer.max_steps != -1: logging.warning( "The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to " f"{trainer.max_steps}. Please ensure that max_steps will run for at least " f"{trainer.check_val_every_n_epoch} epochs to ensure that checkpointing will not error out." ) checkpoint_callback = NeMoModelCheckpoint(n_resume=resume, **params) checkpoint_callback.last_model_path = trainer.ckpt_path or "" if 'mp_rank' in checkpoint_callback.last_model_path or 'tp_rank' in checkpoint_callback.last_model_path: checkpoint_callback.last_model_path = uninject_model_parallel_rank(checkpoint_callback.last_model_path) trainer.callbacks.append(checkpoint_callback) if create_preemption_callback: # Check if cuda is avialable as preemption is supported only on GPUs if torch.cuda.is_available(): ## By default PreemptionCallback handles SIGTERM. To handle other signals pass the signal in the call as below: ## PreemptionCallback(checkpoint_callback, signal.SIGCHLD) preemption_callback = PreemptionCallback(checkpoint_callback) trainer.callbacks.append(preemption_callback) else: logging.info("Preemption is supported only on GPUs, disabling preemption") def check_slurm(trainer): try: return trainer.accelerator_connector.is_slurm_managing_tasks except AttributeError: return False class StatelessTimer(Timer): """Extension of PTL timers to be per run.""" def __init__(self, duration: timedelta = None, interval: str = Interval.step, verbose: bool = True,) -> None: super().__init__(duration, interval, verbose) # Override PTL Timer's state dict to not store elapsed time information so that we can restore and continue training. def state_dict(self) -> Dict[str, Any]: return {} def load_state_dict(self, state_dict: Dict[str, Any]) -> None: return def configure_no_restart_validation_training_loop(trainer: pytorch_lightning.Trainer) -> None: if type(trainer.fit_loop.epoch_loop) != _TrainingEpochLoop: warnings.warn("Detected custom epoch loop. Skipping no validation on restart support.", UserWarning) return ## Pass trainer object to avoid trainer getting overwritten as None loop = SkipResumeTrainingValidationLoop(trainer, trainer.min_steps, trainer.max_steps) trainer.fit_loop.epoch_loop = loop class SkipResumeTrainingValidationLoop(_TrainingEpochLoop): """ Extend the PTL Epoch loop to skip validating when resuming. This happens when resuming a checkpoint that has already run validation, but loading restores the training state before validation has run. """ def _should_check_val_fx(self) -> bool: if self.restarting and self.global_step % self.trainer.val_check_batch == 0: return False return super()._should_check_val_fx() def clean_exp_ckpt(exp_log_dir: Union[str, Path], remove_ckpt: bool = True, remove_nemo: bool = False): """ Helper method that removes Pytorch Lightning .ckpt files or NeMo .nemo files from the checkpoint directory Args: exp_log_dir: str path to the root directory of the current experiment. remove_ckpt: bool, whether to remove all *.ckpt files in the checkpoints directory. remove_nemo: bool, whether to remove all *.nemo files in the checkpoints directory. """ exp_log_dir = str(exp_log_dir) if remove_ckpt: logging.info("Deleting *.ckpt files ...") ckpt_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.ckpt")) for filepath in ckpt_files: os.remove(filepath) logging.info(f"Deleted file : {filepath}") if remove_nemo: logging.info("Deleting *.nemo files ...") nemo_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.nemo")) for filepath in nemo_files: os.remove(filepath) logging.info(f"Deleted file : {filepath}") [end of nemo/utils/exp_manager.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # """ # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in # fusion with beam search decoders on top of a trained ASR model with CTC decoder. To evaluate a model with # Transducer (RNN-T) decoder use another script 'scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py'. # NeMo's beam search decoders are capable of using the KenLM's N-gram models # to find the best candidates. This script supports both character level and BPE level # encodings and models which is detected automatically from the type of the model. # You may train the LM model with 'scripts/asr_language_modeling/ngram_lm/train_kenlm.py'. # Config Help To discover all arguments of the script, please run : python eval_beamsearch_ngram.py --help python eval_beamsearch_ngram.py --cfg job # USAGE python eval_beamsearch_ngram.py nemo_model_file=<path to the .nemo file of the model> \ input_manifest=<path to the evaluation JSON manifest file> \ kenlm_model_file=<path to the binary KenLM model> \ beam_width=[<list of the beam widths, separated with commas>] \ beam_alpha=[<list of the beam alphas, separated with commas>] \ beam_beta=[<list of the beam betas, separated with commas>] \ preds_output_folder=<optional folder to store the predictions> \ probs_cache_file=null \ decoding_mode=beamsearch_ngram ... # Grid Search for Hyper parameters For grid search, you can provide a list of arguments as follows - beam_width=[4,8,16,....] \ beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ beam_beta=[-1.0,-0.5,0.0,...,1.0] \ # You may find more info on how to use this script at: # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html """ import contextlib import json import os import pickle from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import List, Optional import editdistance import numpy as np import torch from omegaconf import MISSING, OmegaConf from sklearn.model_selection import ParameterGrid from tqdm.auto import tqdm import nemo.collections.asr as nemo_asr from nemo.collections.asr.models import EncDecHybridRNNTCTCModel from nemo.collections.asr.parts.submodules import ctc_beam_decoding from nemo.collections.asr.parts.utils.transcribe_utils import PunctuationCapitalization, TextProcessingConfig from nemo.core.config import hydra_runner from nemo.utils import logging # fmt: off @dataclass class EvalBeamSearchNGramConfig: """ Evaluate an ASR model with beam search decoding and n-gram KenLM language model. """ # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) nemo_model_file: str = MISSING # File paths input_manifest: str = MISSING # The manifest file of the evaluation set kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model # Parameters for inference acoustic_batch_size: int = 16 # The batch size to calculate log probabilities beam_batch_size: int = 128 # The batch size to be used for beam search decoding device: str = "cuda" # The device to load the model onto to calculate log probabilities use_amp: bool = False # Whether to use AMP if available to calculate log probabilities # Beam Search hyperparameters # The decoding scheme to be used for evaluation. # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] decoding_mode: str = "beamsearch_ngram" beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, ) # fmt: on def beam_search_eval( model: nemo_asr.models.ASRModel, cfg: EvalBeamSearchNGramConfig, all_probs: List[torch.Tensor], target_transcripts: List[str], preds_output_file: str = None, lm_path: str = None, beam_alpha: float = 1.0, beam_beta: float = 0.0, beam_width: int = 128, beam_batch_size: int = 128, progress_bar: bool = True, punctuation_capitalization: PunctuationCapitalization = None, ): level = logging.getEffectiveLevel() logging.setLevel(logging.CRITICAL) # Reset config model.change_decoding_strategy(None) # Override the beam search config with current search candidate configuration cfg.decoding.beam_size = beam_width cfg.decoding.beam_alpha = beam_alpha cfg.decoding.beam_beta = beam_beta cfg.decoding.return_best_hypothesis = False cfg.decoding.kenlm_path = cfg.kenlm_model_file # Update model's decoding strategy config model.cfg.decoding.strategy = cfg.decoding_strategy model.cfg.decoding.beam = cfg.decoding # Update model's decoding strategy if isinstance(model, EncDecHybridRNNTCTCModel): model.change_decoding_strategy(model.cfg.decoding, decoder_type='ctc') decoding = model.ctc_decoding else: model.change_decoding_strategy(model.cfg.decoding) decoding = model.decoding logging.setLevel(level) wer_dist_first = cer_dist_first = 0 wer_dist_best = cer_dist_best = 0 words_count = 0 chars_count = 0 sample_idx = 0 if preds_output_file: out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') if progress_bar: it = tqdm( range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=f"Beam search decoding with width={beam_width}, alpha={beam_alpha}, beta={beam_beta}", ncols=120, ) else: it = range(int(np.ceil(len(all_probs) / beam_batch_size))) for batch_idx in it: # disabling type checking probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] probs_lens = torch.tensor([prob.shape[0] for prob in probs_batch]) with torch.no_grad(): packed_batch = torch.zeros(len(probs_batch), max(probs_lens), probs_batch[0].shape[-1], device='cpu') for prob_index in range(len(probs_batch)): packed_batch[prob_index, : probs_lens[prob_index], :] = torch.tensor( probs_batch[prob_index], device=packed_batch.device, dtype=packed_batch.dtype ) _, beams_batch = decoding.ctc_decoder_predictions_tensor( packed_batch, decoder_lengths=probs_lens, return_hypotheses=True, ) for beams_idx, beams in enumerate(beams_batch): target = target_transcripts[sample_idx + beams_idx] target_split_w = target.split() target_split_c = list(target) words_count += len(target_split_w) chars_count += len(target_split_c) wer_dist_min = cer_dist_min = 10000 for candidate_idx, candidate in enumerate(beams): # type: (int, ctc_beam_decoding.rnnt_utils.Hypothesis) pred_text = candidate.text if cfg.text_processing.do_lowercase: pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] if cfg.text_processing.rm_punctuation: pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] if cfg.text_processing.separate_punctuation: pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] pred_split_w = pred_text.split() wer_dist = editdistance.eval(target_split_w, pred_split_w) pred_split_c = list(pred_text) cer_dist = editdistance.eval(target_split_c, pred_split_c) wer_dist_min = min(wer_dist_min, wer_dist) cer_dist_min = min(cer_dist_min, cer_dist) if candidate_idx == 0: # first candidate wer_dist_first += wer_dist cer_dist_first += cer_dist score = candidate.score if preds_output_file: out_file.write('{}\t{}\n'.format(pred_text, score)) wer_dist_best += wer_dist_min cer_dist_best += cer_dist_min sample_idx += len(probs_batch) if preds_output_file: out_file.close() logging.info(f"Stored the predictions of beam search decoding at '{preds_output_file}'.") if lm_path: logging.info( 'WER/CER with beam search decoding and N-gram model = {:.2%}/{:.2%}'.format( wer_dist_first / words_count, cer_dist_first / chars_count ) ) else: logging.info( 'WER/CER with beam search decoding = {:.2%}/{:.2%}'.format( wer_dist_first / words_count, cer_dist_first / chars_count ) ) logging.info( 'Oracle WER/CER in candidates with perfect LM= {:.2%}/{:.2%}'.format( wer_dist_best / words_count, cer_dist_best / chars_count ) ) logging.info(f"=================================================================================") return wer_dist_first / words_count, cer_dist_first / chars_count @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) def main(cfg: EvalBeamSearchNGramConfig): logging.warning("This file will be renamed to eval_beamsearch_ngram_ctc.py in the future NeMo (1.21) release.") if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig valid_decoding_modes = ["greedy", "beamsearch", "beamsearch_ngram"] if cfg.decoding_mode not in valid_decoding_modes: raise ValueError( f"Given decoding_mode={cfg.decoding_mode} is invalid. Available options are :\n" f"{valid_decoding_modes}" ) if cfg.nemo_model_file.endswith('.nemo'): asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) else: logging.warning( "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." ) asr_model = nemo_asr.models.ASRModel.from_pretrained( cfg.nemo_model_file, map_location=torch.device(cfg.device) ) target_transcripts = [] manifest_dir = Path(cfg.input_manifest).parent with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: audio_file_paths = [] for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): data = json.loads(line) audio_file = Path(data['audio_filepath']) if not audio_file.is_file() and not audio_file.is_absolute(): audio_file = manifest_dir / audio_file target_transcripts.append(data['text']) audio_file_paths.append(str(audio_file.absolute())) punctuation_capitalization = PunctuationCapitalization(cfg.text_processing.punctuation_marks) if cfg.text_processing.do_lowercase: target_transcripts = punctuation_capitalization.do_lowercase(target_transcripts) if cfg.text_processing.rm_punctuation: target_transcripts = punctuation_capitalization.rm_punctuation(target_transcripts) if cfg.text_processing.separate_punctuation: target_transcripts = punctuation_capitalization.separate_punctuation(target_transcripts) if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") with open(cfg.probs_cache_file, 'rb') as probs_file: all_probs = pickle.load(probs_file) if len(all_probs) != len(audio_file_paths): raise ValueError( f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " f"match the manifest file. You may need to delete the probabilities cached file." ) else: @contextlib.contextmanager def default_autocast(): yield if cfg.use_amp: if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): logging.info("AMP is enabled!\n") autocast = torch.cuda.amp.autocast else: autocast = default_autocast else: autocast = default_autocast with autocast(): with torch.no_grad(): if isinstance(asr_model, EncDecHybridRNNTCTCModel): asr_model.cur_decoder = 'ctc' all_logits = asr_model.transcribe(audio_file_paths, batch_size=cfg.acoustic_batch_size, logprobs=True) all_probs = all_logits if cfg.probs_cache_file: logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") with open(cfg.probs_cache_file, 'wb') as f_dump: pickle.dump(all_probs, f_dump) wer_dist_greedy = 0 cer_dist_greedy = 0 words_count = 0 chars_count = 0 for batch_idx, probs in enumerate(all_probs): preds = np.argmax(probs, axis=1) preds_tensor = torch.tensor(preds, device='cpu').unsqueeze(0) if isinstance(asr_model, EncDecHybridRNNTCTCModel): pred_text = asr_model.ctc_decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] else: pred_text = asr_model._wer.decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] if cfg.text_processing.do_lowercase: pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] if cfg.text_processing.rm_punctuation: pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] if cfg.text_processing.separate_punctuation: pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] pred_split_w = pred_text.split() target_split_w = target_transcripts[batch_idx].split() pred_split_c = list(pred_text) target_split_c = list(target_transcripts[batch_idx]) wer_dist = editdistance.eval(target_split_w, pred_split_w) cer_dist = editdistance.eval(target_split_c, pred_split_c) wer_dist_greedy += wer_dist cer_dist_greedy += cer_dist words_count += len(target_split_w) chars_count += len(target_split_c) logging.info('Greedy WER/CER = {:.2%}/{:.2%}'.format(wer_dist_greedy / words_count, cer_dist_greedy / chars_count)) asr_model = asr_model.to('cpu') if cfg.decoding_mode == "beamsearch_ngram": if not os.path.exists(cfg.kenlm_model_file): raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") lm_path = cfg.kenlm_model_file else: lm_path = None # 'greedy' decoding_mode would skip the beam search decoding if cfg.decoding_mode in ["beamsearch_ngram", "beamsearch"]: if cfg.beam_width is None or cfg.beam_alpha is None or cfg.beam_beta is None: raise ValueError("beam_width, beam_alpha and beam_beta are needed to perform beam search decoding.") params = {'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'beam_beta': cfg.beam_beta} hp_grid = ParameterGrid(params) hp_grid = list(hp_grid) best_wer_beam_size, best_cer_beam_size = None, None best_wer_alpha, best_cer_alpha = None, None best_wer_beta, best_cer_beta = None, None best_wer, best_cer = 1e6, 1e6 logging.info(f"==============================Starting the beam search decoding===============================") logging.info(f"Grid search size: {len(hp_grid)}") logging.info(f"It may take some time...") logging.info(f"==============================================================================================") if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): os.mkdir(cfg.preds_output_folder) for hp in hp_grid: if cfg.preds_output_folder: preds_output_file = os.path.join( cfg.preds_output_folder, f"preds_out_width{hp['beam_width']}_alpha{hp['beam_alpha']}_beta{hp['beam_beta']}.tsv", ) else: preds_output_file = None candidate_wer, candidate_cer = beam_search_eval( asr_model, cfg, all_probs=all_probs, target_transcripts=target_transcripts, preds_output_file=preds_output_file, lm_path=lm_path, beam_width=hp["beam_width"], beam_alpha=hp["beam_alpha"], beam_beta=hp["beam_beta"], beam_batch_size=cfg.beam_batch_size, progress_bar=True, punctuation_capitalization=punctuation_capitalization, ) if candidate_cer < best_cer: best_cer_beam_size = hp["beam_width"] best_cer_alpha = hp["beam_alpha"] best_cer_beta = hp["beam_beta"] best_cer = candidate_cer if candidate_wer < best_wer: best_wer_beam_size = hp["beam_width"] best_wer_alpha = hp["beam_alpha"] best_wer_beta = hp["beam_beta"] best_wer = candidate_wer logging.info( f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' f'Beam alpha = {best_wer_alpha}, Beam beta = {best_wer_beta}' ) logging.info( f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' f'Beam alpha = {best_cer_alpha}, Beam beta = {best_cer_beta}' ) logging.info(f"=================================================================================") if __name__ == '__main__': main() [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # """ # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in # fusion with beam search decoders on top of a trained ASR Transducer model. NeMo's beam search decoders are capable of using the # KenLM's N-gram models to find the best candidates. This script supports both character level and BPE level # encodings and models which is detected automatically from the type of the model. # You may train the LM model with 'scripts/ngram_lm/train_kenlm.py'. # Config Help To discover all arguments of the script, please run : python eval_beamsearch_ngram.py --help python eval_beamsearch_ngram.py --cfg job # USAGE python eval_beamsearch_ngram_transducer.py nemo_model_file=<path to the .nemo file of the model> \ input_manifest=<path to the evaluation JSON manifest file \ kenlm_model_file=<path to the binary KenLM model> \ beam_width=[<list of the beam widths, separated with commas>] \ beam_alpha=[<list of the beam alphas, separated with commas>] \ preds_output_folder=<optional folder to store the predictions> \ probs_cache_file=null \ decoding_strategy=<greedy_batch or maes decoding> maes_prefix_alpha=[<list of the maes prefix alphas, separated with commas>] \ maes_expansion_gamma=[<list of the maes expansion gammas, separated with commas>] \ hat_subtract_ilm=<in case of HAT model: subtract internal LM or not> \ hat_ilm_weight=[<in case of HAT model: list of the HAT internal LM weights, separated with commas>] \ ... # Grid Search for Hyper parameters For grid search, you can provide a list of arguments as follows - beam_width=[4,8,16,....] \ beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ # You may find more info on how to use this script at: # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html """ import contextlib import json import os import pickle import tempfile from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import List, Optional import editdistance import numpy as np import torch from omegaconf import MISSING, OmegaConf from sklearn.model_selection import ParameterGrid from tqdm.auto import tqdm import nemo.collections.asr as nemo_asr from nemo.collections.asr.parts.submodules import rnnt_beam_decoding from nemo.core.config import hydra_runner from nemo.utils import logging # fmt: off @dataclass class EvalBeamSearchNGramConfig: """ Evaluate an ASR model with beam search decoding and n-gram KenLM language model. """ # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) nemo_model_file: str = MISSING # File paths input_manifest: str = MISSING # The manifest file of the evaluation set kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model # Parameters for inference acoustic_batch_size: int = 128 # The batch size to calculate log probabilities beam_batch_size: int = 128 # The batch size to be used for beam search decoding device: str = "cuda" # The device to load the model onto to calculate log probabilities use_amp: bool = False # Whether to use AMP if available to calculate log probabilities num_workers: int = 1 # Number of workers for DataLoader # The decoding scheme to be used for evaluation decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] # Beam Search hyperparameters beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding # HAT related parameters (only for internal lm subtraction) hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) # fmt: on def decoding_step( model: nemo_asr.models.ASRModel, cfg: EvalBeamSearchNGramConfig, all_probs: List[torch.Tensor], target_transcripts: List[str], preds_output_file: str = None, beam_batch_size: int = 128, progress_bar: bool = True, ): level = logging.getEffectiveLevel() logging.setLevel(logging.CRITICAL) # Reset config model.change_decoding_strategy(None) cfg.decoding.hat_ilm_weight = cfg.decoding.hat_ilm_weight * cfg.hat_subtract_ilm # Override the beam search config with current search candidate configuration cfg.decoding.return_best_hypothesis = False cfg.decoding.ngram_lm_model = cfg.kenlm_model_file cfg.decoding.hat_subtract_ilm = cfg.hat_subtract_ilm # Update model's decoding strategy config model.cfg.decoding.strategy = cfg.decoding_strategy model.cfg.decoding.beam = cfg.decoding # Update model's decoding strategy model.change_decoding_strategy(model.cfg.decoding) logging.setLevel(level) wer_dist_first = cer_dist_first = 0 wer_dist_best = cer_dist_best = 0 words_count = 0 chars_count = 0 sample_idx = 0 if preds_output_file: out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') if progress_bar: if cfg.decoding_strategy == "greedy_batch": description = "Greedy_batch decoding.." else: description = f"{cfg.decoding_strategy} decoding with bw={cfg.decoding.beam_size}, ba={cfg.decoding.ngram_lm_alpha}, ma={cfg.decoding.maes_prefix_alpha}, mg={cfg.decoding.maes_expansion_gamma}, hat_ilmw={cfg.decoding.hat_ilm_weight}" it = tqdm(range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=description, ncols=120) else: it = range(int(np.ceil(len(all_probs) / beam_batch_size))) for batch_idx in it: # disabling type checking probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] probs_lens = torch.tensor([prob.shape[-1] for prob in probs_batch]) with torch.no_grad(): packed_batch = torch.zeros(len(probs_batch), probs_batch[0].shape[0], max(probs_lens), device='cpu') for prob_index in range(len(probs_batch)): packed_batch[prob_index, :, : probs_lens[prob_index]] = torch.tensor( probs_batch[prob_index].unsqueeze(0), device=packed_batch.device, dtype=packed_batch.dtype ) best_hyp_batch, beams_batch = model.decoding.rnnt_decoder_predictions_tensor( packed_batch, probs_lens, return_hypotheses=True, ) if cfg.decoding_strategy == "greedy_batch": beams_batch = [[x] for x in best_hyp_batch] for beams_idx, beams in enumerate(beams_batch): target = target_transcripts[sample_idx + beams_idx] target_split_w = target.split() target_split_c = list(target) words_count += len(target_split_w) chars_count += len(target_split_c) wer_dist_min = cer_dist_min = 10000 for candidate_idx, candidate in enumerate(beams): # type: (int, rnnt_beam_decoding.rnnt_utils.Hypothesis) pred_text = candidate.text pred_split_w = pred_text.split() wer_dist = editdistance.eval(target_split_w, pred_split_w) pred_split_c = list(pred_text) cer_dist = editdistance.eval(target_split_c, pred_split_c) wer_dist_min = min(wer_dist_min, wer_dist) cer_dist_min = min(cer_dist_min, cer_dist) if candidate_idx == 0: # first candidate wer_dist_first += wer_dist cer_dist_first += cer_dist score = candidate.score if preds_output_file: out_file.write('{}\t{}\n'.format(pred_text, score)) wer_dist_best += wer_dist_min cer_dist_best += cer_dist_min sample_idx += len(probs_batch) if cfg.decoding_strategy == "greedy_batch": return wer_dist_first / words_count, cer_dist_first / chars_count if preds_output_file: out_file.close() logging.info(f"Stored the predictions of {cfg.decoding_strategy} decoding at '{preds_output_file}'.") if cfg.decoding.ngram_lm_model: logging.info( f"WER/CER with {cfg.decoding_strategy} decoding and N-gram model = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" ) else: logging.info( f"WER/CER with {cfg.decoding_strategy} decoding = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" ) logging.info( f"Oracle WER/CER in candidates with perfect LM= {wer_dist_best / words_count:.2%}/{cer_dist_best / chars_count:.2%}" ) logging.info(f"=================================================================================") return wer_dist_first / words_count, cer_dist_first / chars_count @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) def main(cfg: EvalBeamSearchNGramConfig): if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig valid_decoding_strategis = ["greedy_batch", "beam", "tsd", "alsd", "maes"] if cfg.decoding_strategy not in valid_decoding_strategis: raise ValueError( f"Given decoding_strategy={cfg.decoding_strategy} is invalid. Available options are :\n" f"{valid_decoding_strategis}" ) if cfg.nemo_model_file.endswith('.nemo'): asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) else: logging.warning( "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." ) asr_model = nemo_asr.models.ASRModel.from_pretrained( cfg.nemo_model_file, map_location=torch.device(cfg.device) ) if cfg.kenlm_model_file: if not os.path.exists(cfg.kenlm_model_file): raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") if cfg.decoding_strategy != "maes": raise ValueError(f"Decoding with kenlm model is supported only for maes decoding algorithm.") lm_path = cfg.kenlm_model_file else: lm_path = None cfg.beam_alpha = [0.0] if cfg.hat_subtract_ilm: assert lm_path, "kenlm must be set for hat internal lm subtraction" if cfg.decoding_strategy != "maes": cfg.maes_prefix_alpha, cfg.maes_expansion_gamma, cfg.hat_ilm_weight = [0], [0], [0] target_transcripts = [] manifest_dir = Path(cfg.input_manifest).parent with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: audio_file_paths = [] for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): data = json.loads(line) audio_file = Path(data['audio_filepath']) if not audio_file.is_file() and not audio_file.is_absolute(): audio_file = manifest_dir / audio_file target_transcripts.append(data['text']) audio_file_paths.append(str(audio_file.absolute())) if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") with open(cfg.probs_cache_file, 'rb') as probs_file: all_probs = pickle.load(probs_file) if len(all_probs) != len(audio_file_paths): raise ValueError( f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " f"match the manifest file. You may need to delete the probabilities cached file." ) else: @contextlib.contextmanager def default_autocast(): yield if cfg.use_amp: if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): logging.info("AMP is enabled!\n") autocast = torch.cuda.amp.autocast else: autocast = default_autocast else: autocast = default_autocast # manual calculation of encoder_embeddings with autocast(): with torch.no_grad(): asr_model.eval() asr_model.encoder.freeze() device = next(asr_model.parameters()).device all_probs = [] with tempfile.TemporaryDirectory() as tmpdir: with open(os.path.join(tmpdir, 'manifest.json'), 'w', encoding='utf-8') as fp: for audio_file in audio_file_paths: entry = {'audio_filepath': audio_file, 'duration': 100000, 'text': ''} fp.write(json.dumps(entry) + '\n') config = { 'paths2audio_files': audio_file_paths, 'batch_size': cfg.acoustic_batch_size, 'temp_dir': tmpdir, 'num_workers': cfg.num_workers, 'channel_selector': None, 'augmentor': None, } temporary_datalayer = asr_model._setup_transcribe_dataloader(config) for test_batch in tqdm(temporary_datalayer, desc="Transcribing", disable=True): encoded, encoded_len = asr_model.forward( input_signal=test_batch[0].to(device), input_signal_length=test_batch[1].to(device) ) # dump encoder embeddings per file for idx in range(encoded.shape[0]): encoded_no_pad = encoded[idx, :, : encoded_len[idx]] all_probs.append(encoded_no_pad) if cfg.probs_cache_file: logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") with open(cfg.probs_cache_file, 'wb') as f_dump: pickle.dump(all_probs, f_dump) if cfg.decoding_strategy == "greedy_batch": asr_model = asr_model.to('cpu') candidate_wer, candidate_cer = decoding_step( asr_model, cfg, all_probs=all_probs, target_transcripts=target_transcripts, beam_batch_size=cfg.beam_batch_size, progress_bar=True, ) logging.info(f"Greedy batch WER/CER = {candidate_wer:.2%}/{candidate_cer:.2%}") asr_model = asr_model.to('cpu') # 'greedy_batch' decoding_strategy would skip the beam search decoding if cfg.decoding_strategy in ["beam", "tsd", "alsd", "maes"]: if cfg.beam_width is None or cfg.beam_alpha is None: raise ValueError("beam_width and beam_alpha are needed to perform beam search decoding.") params = { 'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'maes_prefix_alpha': cfg.maes_prefix_alpha, 'maes_expansion_gamma': cfg.maes_expansion_gamma, 'hat_ilm_weight': cfg.hat_ilm_weight, } hp_grid = ParameterGrid(params) hp_grid = list(hp_grid) best_wer_beam_size, best_cer_beam_size = None, None best_wer_alpha, best_cer_alpha = None, None best_wer, best_cer = 1e6, 1e6 logging.info( f"==============================Starting the {cfg.decoding_strategy} decoding===============================" ) logging.info(f"Grid search size: {len(hp_grid)}") logging.info(f"It may take some time...") logging.info(f"==============================================================================================") if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): os.mkdir(cfg.preds_output_folder) for hp in hp_grid: if cfg.preds_output_folder: results_file = f"preds_out_{cfg.decoding_strategy}_bw{hp['beam_width']}" if cfg.decoding_strategy == "maes": results_file = f"{results_file}_ma{hp['maes_prefix_alpha']}_mg{hp['maes_expansion_gamma']}" if cfg.kenlm_model_file: results_file = f"{results_file}_ba{hp['beam_alpha']}" if cfg.hat_subtract_ilm: results_file = f"{results_file}_hat_ilmw{hp['hat_ilm_weight']}" preds_output_file = os.path.join(cfg.preds_output_folder, f"{results_file}.tsv") else: preds_output_file = None cfg.decoding.beam_size = hp["beam_width"] cfg.decoding.ngram_lm_alpha = hp["beam_alpha"] cfg.decoding.maes_prefix_alpha = hp["maes_prefix_alpha"] cfg.decoding.maes_expansion_gamma = hp["maes_expansion_gamma"] cfg.decoding.hat_ilm_weight = hp["hat_ilm_weight"] candidate_wer, candidate_cer = decoding_step( asr_model, cfg, all_probs=all_probs, target_transcripts=target_transcripts, preds_output_file=preds_output_file, beam_batch_size=cfg.beam_batch_size, progress_bar=True, ) if candidate_cer < best_cer: best_cer_beam_size = hp["beam_width"] best_cer_alpha = hp["beam_alpha"] best_cer_ma = hp["maes_prefix_alpha"] best_cer_mg = hp["maes_expansion_gamma"] best_cer_hat_ilm_weight = hp["hat_ilm_weight"] best_cer = candidate_cer if candidate_wer < best_wer: best_wer_beam_size = hp["beam_width"] best_wer_alpha = hp["beam_alpha"] best_wer_ma = hp["maes_prefix_alpha"] best_wer_ga = hp["maes_expansion_gamma"] best_wer_hat_ilm_weight = hp["hat_ilm_weight"] best_wer = candidate_wer wer_hat_parameter = "" if cfg.hat_subtract_ilm: wer_hat_parameter = f"HAT ilm weight = {best_wer_hat_ilm_weight}, " logging.info( f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' f'Beam alpha = {best_wer_alpha}, {wer_hat_parameter}' f'maes_prefix_alpha = {best_wer_ma}, maes_expansion_gamma = {best_wer_ga} ' ) cer_hat_parameter = "" if cfg.hat_subtract_ilm: cer_hat_parameter = f"HAT ilm weight = {best_cer_hat_ilm_weight}" logging.info( f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' f'Beam alpha = {best_cer_alpha}, {cer_hat_parameter} ' f'maes_prefix_alpha = {best_cer_ma}, maes_expansion_gamma = {best_cer_mg}' ) logging.info(f"=================================================================================") if __name__ == '__main__': main() [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of scripts/confidence_ensembles/build_ensemble.py] # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ This script provides a functionality to create confidence-based ensembles from a collection of pretrained models. For more details see the paper https://arxiv.org/abs/2306.15824 or tutorial in tutorials/asr/Confidence_Ensembles.ipynb You would typically use this script by providing a yaml config file or overriding default options from command line. Usage examples: 1. Building an ensemble of two monolingual models with default settings (no confidence tuning). python build_ensemble.py --config-path=. --config-name=ensemble_config.yaml ensemble.0.model=stt_it_conformer_ctc_large ensemble.0.training_manifest=<path to the Italian data of 100+ utterances (no transcription required)> ensemble.1.model=stt_es_conformer_ctc_large ensemble.1.training_manifest=<path to the Spanish data of 100+ utterances (no transcription required)> output_path=<path to the desired location of the .nemo checkpoint> You can have more than 2 models and can control transcription settings (e.g., batch size) with ``transcription.<any argument of examples/asr/transcribe_speech.py>`` parameters. 2. If you want to get improved results, you can enable tuning of the confidence and logistic regression (LR) parameters. E.g. python build_ensemble.py <all arguments like in the previous example> ensemble.0.dev_manifest=<path to the dev data that's required for tuning> ... # IMPORTANT: see the note below if you use > 2 models! ensemble.N.dev_manifest=<path to the dev data that's required for tuning> tune_confidence=True # to allow confidence tuning. LR is tuned by default As with any tuning, it is recommended to have reasonably large validation set for each model, otherwise you might overfit to the validation data. Note that if you add additional models (> 2) you will need to modify ensemble_config.yaml or create a new one with added models in there. While it's theoretically possible to fully override such parameters from commandline, hydra is very unfriendly for such use-cases, so it's strongly recommended to be creating new configs. 3. If you want to precisely control tuning grid search, you can do that with python build_ensemble.py <all arguments as in the previous examples> tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set You can check the dataclasses in this file for the full list of supported arguments and their default values. """ import atexit # using default logging to be able to silence unnecessary messages from nemo import logging import os import random import sys import tempfile from copy import deepcopy from dataclasses import dataclass from pathlib import Path from typing import Dict, List, Optional, Tuple import joblib import numpy as np import pytorch_lightning as pl from omegaconf import MISSING, DictConfig, OmegaConf from sklearn.linear_model import LogisticRegression from sklearn.metrics import confusion_matrix from sklearn.pipeline import Pipeline, make_pipeline from sklearn.preprocessing import StandardScaler from tqdm import tqdm from nemo.collections.asr.models.confidence_ensemble import ( ConfidenceEnsembleModel, ConfidenceSpec, compute_confidence, get_filtered_logprobs, ) from nemo.collections.asr.parts.utils.asr_confidence_utils import ( ConfidenceConfig, ConfidenceMeasureConfig, get_confidence_aggregation_bank, get_confidence_measure_bank, ) from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis from nemo.core.config import hydra_runner LOG = logging.getLogger(__file__) # adding Python path. If not found, asking user to get the file try: sys.path.append(str(Path(__file__).parents[2] / "examples" / "asr")) import transcribe_speech except ImportError: # if users run script normally from nemo repo, this shouldn't be triggered as # we modify the path above. But if they downloaded the build_ensemble.py as # an isolated script, we'd ask them to also download corresponding version # of the transcribe_speech.py print( "Current script depends on 'examples/asr/transcribe_speech.py', but can't find it. " "If it's not present, download it from the NeMo github manually and put inside this folder." ) @dataclass class EnsembleConfig: # .nemo path or pretrained name model: str = MISSING # path to the training data manifest (non-tarred) training_manifest: str = MISSING # specify to limit the number of training samples # 100 is most likely enough, but setting higher default just in case max_training_samples: int = 1000 # specify to provide dev data manifest for HP tuning dev_manifest: Optional[str] = None @dataclass class TuneConfidenceConfig: # important parameter, so should always be tuned exclude_blank: Tuple[bool] = (True, False) # prod is pretty much always worse, so not including by default aggregation: Tuple[str] = ("mean", "min", "max") # not including max prob, as there is always an entropy-based metric # that's better but otherwise including everything confidence_type: Tuple[str] = ( "entropy_renyi_exp", "entropy_renyi_lin", "entropy_tsallis_exp", "entropy_tsallis_lin", "entropy_gibbs_lin", "entropy_gibbs_exp", ) # TODO: currently it's not possible to efficiently tune temperature, as we always # apply log-softmax in the decoder, so to try different values it will be required # to rerun the decoding, which is very slow. To support this for one-off experiments # it's possible to modify the code of CTC decoder / Transducer joint to # remove log-softmax and then apply it directly in this script with the temperature # # Alternatively, one can run this script multiple times with different values of # temperature and pick the best performing ensemble. Note that this will increase # tuning time by the number of temperature values tried. On the other hand, # the above approach is a lot more efficient and will only slightly increase # the total tuning runtime. # very important to tune for max prob, but for entropy metrics 1.0 is almost always best # temperature: Tuple[float] = (1.0,) # not that important, but can sometimes make a small difference alpha: Tuple[float] = (0.25, 0.33, 0.5, 1.0) def get_grid_size(self) -> int: """Returns the total number of points in the search space.""" if "max_prob" in self.confidence_type: return ( len(self.exclude_blank) * len(self.aggregation) * ((len(self.confidence_type) - 1) * len(self.alpha) + 1) ) return len(self.exclude_blank) * len(self.aggregation) * len(self.confidence_type) * len(self.alpha) @dataclass class TuneLogisticRegressionConfig: # will have log-uniform grid over this range with that many points # note that a value of 10000.0 (not regularization) is always added C_num_points: int = 10 C_min: float = 0.0001 C_max: float = 10.0 # not too important multi_class: Tuple[str] = ("ovr", "multinomial") # should try to include weights directly if the data is too imbalanced class_weight: Tuple = (None, "balanced") # increase if getting many warnings that algorithm didn't converge max_iter: int = 1000 @dataclass class BuildEnsembleConfig: # where to save the resulting ensemble model output_path: str = MISSING # each model specification ensemble: List[EnsembleConfig] = MISSING random_seed: int = 0 # for reproducibility # default confidence, can override confidence: ConfidenceConfig = ConfidenceConfig( # we keep frame confidences and apply aggregation manually to get full-utterance confidence preserve_frame_confidence=True, exclude_blank=True, aggregation="mean", measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() # set to True to tune the confidence. # requires dev manifests to be specified for each model tune_confidence: bool = False # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() def __post_init__(self): """Checking that if any dev data is provided, all are provided. Will also auto-set tune_logistic_regression to False if no dev data is available. If tune_confidence is set to True (user choice) and no dev data is provided, will raise an error. """ num_dev_data = 0 for ensemble_cfg in self.ensemble: num_dev_data += ensemble_cfg.dev_manifest is not None if num_dev_data == 0: if self.tune_confidence: raise ValueError("tune_confidence is set to True, but no dev data is provided") LOG.info("Setting tune_logistic_regression = False since no dev data is provided") self.tune_logistic_regression = False return if num_dev_data < len(self.ensemble): raise ValueError( "Some ensemble configs specify dev data, but some don't. Either all have to specify it or none!" ) def calculate_score(features: np.ndarray, labels: np.ndarray, pipe: Pipeline) -> Tuple[float, np.ndarray]: """Score is always calculated as mean of the per-class scores. This is done to account for possible class imbalances. Args: features: numpy array of features of shape [N x D], where N is the number of objects (typically a total number of utterances in all datasets) and D is the total number of confidence scores used to train the model (typically = number of models). labels: numpy array of shape [N] contatining ground-truth model indices. pipe: classification pipeline (currently, standardization + logistic regression). Returns: tuple: score value in [0, 1] and full classification confusion matrix. """ predictions = pipe.predict(features) conf_m = confusion_matrix(labels, predictions) score = np.diag(conf_m).sum() / conf_m.sum() return score, conf_m def train_model_selection( training_features: np.ndarray, training_labels: np.ndarray, dev_features: Optional[np.ndarray] = None, dev_labels: Optional[np.ndarray] = None, tune_lr: bool = False, tune_lr_cfg: Optional[TuneLogisticRegressionConfig] = None, verbose: bool = False, ) -> Tuple[Pipeline, float]: """Trains model selection block with an (optional) tuning of the parameters. Returns a pipeline consisting of feature standardization and logistic regression. If tune_lr is set to True, dev features/labels will be used to tune the hyperparameters of the logistic regression with the grid search that's defined via ``tune_lr_cfg``. If no tuning is requested, uses the following parameters:: best_pipe = make_pipeline( StandardScaler(), LogisticRegression( multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced", ), ) Args: training_features: numpy array of features of shape [N x D], where N is the number of objects (typically a total number of utterances in all training datasets) and D is the total number of confidence scores used to train the model (typically = number of models). training_labels: numpy array of shape [N] contatining ground-truth model indices. dev_features: same as training, but for the validation subset. dev_labels: same as training, but for the validation subset. tune_lr: controls whether tuning of LR hyperparameters is performed. If set to True, it's required to also provide dev features/labels. tune_lr_cfg: specifies what values of LR hyperparameters to try. verbose: if True, will output final training/dev scores. Returns: tuple: trained model selection pipeline, best score (or -1 if no tuning was done). """ if not tune_lr: # default parameters: C=10000.0 disables regularization best_pipe = make_pipeline( StandardScaler(), LogisticRegression(multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced"), ) max_score = -1 else: C_pms = np.append( np.exp(np.linspace(np.log(tune_lr_cfg.C_min), np.log(tune_lr_cfg.C_max), tune_lr_cfg.C_num_points)), 10000.0, ) max_score = 0 best_pipe = None for class_weight in tune_lr_cfg.class_weight: for multi_class in tune_lr_cfg.multi_class: for C in C_pms: pipe = make_pipeline( StandardScaler(), LogisticRegression( multi_class=multi_class, C=C, max_iter=tune_lr_cfg.max_iter, class_weight=class_weight ), ) pipe.fit(training_features, training_labels) score, confusion = calculate_score(dev_features, dev_labels, pipe) if score > max_score: max_score = score best_pipe = pipe best_pipe.fit(training_features, training_labels) if verbose: accuracy, confusion = calculate_score(training_features, training_labels, best_pipe) LOG.info("Training fit accuracy: %.4f", accuracy * 100.0) LOG.info("Training confusion matrix:\n%s", str(confusion)) if dev_features is not None and verbose: accuracy, confusion = calculate_score(dev_features, dev_labels, best_pipe) LOG.info("Dev fit accuracy: %.4f", accuracy * 100.0) LOG.info("Dev confusion matrix:\n%s", str(confusion)) return best_pipe, max_score def subsample_manifest(manifest_file: str, max_samples: int) -> str: """Will save a subsampled version of the manifest to the same folder. Have to save to the same folder to support relative paths. Args: manifest_file: path to the manifest file that needs subsampling. max_samples: how many samples to retain. Will randomly select that many lines from the manifest. Returns: str: the path to the subsampled manifest file. """ with open(manifest_file, "rt", encoding="utf-8") as fin: lines = fin.readlines() if max_samples < len(lines): lines = random.sample(lines, max_samples) output_file = manifest_file + "-subsampled" with open(output_file, "wt", encoding="utf-8") as fout: fout.write("".join(lines)) return output_file def cleanup_subsampled_manifests(subsampled_manifests: List[str]): """Removes all generated subsamples manifests.""" for manifest in subsampled_manifests: os.remove(manifest) def compute_all_confidences( hypothesis: Hypothesis, tune_confidence_cfg: TuneConfidenceConfig ) -> Dict[ConfidenceSpec, float]: """Computes a set of confidence scores from a given hypothesis. Works with the output of both CTC and Transducer decoding. Args: hypothesis: generated hypothesis as returned from the transcribe method of the ASR model. tune_confidence_cfg: config specifying what confidence scores to compute. Returns: dict: dictionary with confidenct spec -> confidence score mapping. """ conf_values = {} for exclude_blank in tune_confidence_cfg.exclude_blank: filtered_logprobs = get_filtered_logprobs(hypothesis, exclude_blank) vocab_size = filtered_logprobs.shape[1] for aggregation in tune_confidence_cfg.aggregation: aggr_func = get_confidence_aggregation_bank()[aggregation] for conf_type in tune_confidence_cfg.confidence_type: conf_func = get_confidence_measure_bank()[conf_type] if conf_type == "max_prob": # skipping alpha in this case conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=1.0)).cpu().item() conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, 1.0)] = conf_value else: for alpha in tune_confidence_cfg.alpha: conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=alpha)).cpu().item() conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, alpha)] = conf_value return conf_values def find_best_confidence( train_confidences: List[List[Dict[ConfidenceSpec, float]]], train_labels: List[int], dev_confidences: List[List[Dict[ConfidenceSpec, float]]], dev_labels: List[int], tune_lr: bool, tune_lr_config: TuneConfidenceConfig, ) -> Tuple[ConfidenceConfig, Pipeline]: """Finds the best confidence configuration for model selection. Will loop over all values in the confidence dictionary and fit the LR model (optionally tuning its HPs). The best performing confidence (on the dev set) will be used for the final LR model. Args: train_confidences: this is an object of type ``List[List[Dict[ConfidenceSpec, float]]]``. The shape of this object is [M, N, S], where M: number of models N: number of utterances in all training sets S: number of confidence scores to try This argument will be used to construct np.array objects for each of the confidence scores with the shape [M, N] train_labels: ground-truth labels of the correct model for each data points. This is a list of size [N] dev_confidences: same as training, but for the validation subset. dev_labels: same as training, but for the validation subset. tune_lr: controls whether tuning of LR hyperparameters is performed. tune_lr_cfg: specifies what values of LR hyperparameters to try. Returns: tuple: best confidence config, best model selection pipeline """ max_score = 0 best_pipe = None best_conf_spec = None LOG.info("Evaluation all confidences. Total grid size: %d", len(train_confidences[0][0].keys())) for conf_spec in tqdm(train_confidences[0][0].keys()): cur_train_confidences = [] for model_confs in train_confidences: cur_train_confidences.append([]) for model_conf in model_confs: cur_train_confidences[-1].append(model_conf[conf_spec]) cur_dev_confidences = [] for model_confs in dev_confidences: cur_dev_confidences.append([]) for model_conf in model_confs: cur_dev_confidences[-1].append(model_conf[conf_spec]) # transposing with zip(*list) training_features = np.array(list(zip(*cur_train_confidences))) training_labels = np.array(train_labels) dev_features = np.array(list(zip(*cur_dev_confidences))) dev_labels = np.array(dev_labels) pipe, score = train_model_selection( training_features, training_labels, dev_features, dev_labels, tune_lr, tune_lr_config, ) if max_score < score: max_score = score best_pipe = pipe best_conf_spec = conf_spec LOG.info("Found better parameters: %s. New score: %.4f", str(conf_spec), max_score) return best_conf_spec.to_confidence_config(), best_pipe @hydra_runner(config_name="BuildEnsembleConfig", schema=BuildEnsembleConfig) def main(cfg: BuildEnsembleConfig): # silencing all messages from nemo/ptl to avoid dumping tons of configs to the stdout logging.getLogger('pytorch_lightning').setLevel(logging.CRITICAL) logging.getLogger('nemo_logger').setLevel(logging.CRITICAL) LOG.info(f'Build ensemble config:\n{OmegaConf.to_yaml(cfg)}') # to ensure post init is called cfg = BuildEnsembleConfig(**cfg) pl.seed_everything(cfg.random_seed) cfg.transcription.random_seed = None # seed is already applied cfg.transcription.return_transcriptions = True cfg.transcription.preserve_alignment = True cfg.transcription.ctc_decoding.temperature = cfg.temperature cfg.transcription.rnnt_decoding.temperature = cfg.temperature # this ensures that generated output is after log-softmax for consistency with CTC train_confidences = [] dev_confidences = [] train_labels = [] dev_labels = [] # registering clean-up function that will hold on to this list and # should clean up even if there is partial error in some of the transcribe # calls subsampled_manifests = [] atexit.register(cleanup_subsampled_manifests, subsampled_manifests) # note that we loop over the same config. # This is intentional, as we need to run all models on all datasets # this loop will do the following things: # 1. Goes through each model X each training dataset # 2. Computes predictions by directly calling transcribe_speech.main # 3. Converts transcription to the confidence score(s) as specified in the config # 4. If dev sets are provided, computes the same for them # 5. Creates a list of ground-truth model indices by mapping each model # to its own training dataset as specified in the config. # 6. After the loop, we either run tuning over all confidence scores or # directly use a single score to fit logistic regression and save the # final ensemble model. for model_idx, model_cfg in enumerate(cfg.ensemble): train_model_confidences = [] dev_model_confidences = [] for data_idx, data_cfg in enumerate(cfg.ensemble): if model_idx == 0: # generating subsampled manifests only one time subsampled_manifests.append( subsample_manifest(data_cfg.training_manifest, data_cfg.max_training_samples) ) subsampled_manifest = subsampled_manifests[data_idx] if model_cfg.model.endswith(".nemo"): cfg.transcription.model_path = model_cfg.model else: # assuming pretrained model cfg.transcription.pretrained_name = model_cfg.model cfg.transcription.dataset_manifest = subsampled_manifest # training with tempfile.NamedTemporaryFile() as output_file: cfg.transcription.output_filename = output_file.name LOG.info("Transcribing training dataset %d with model %d", data_idx, model_idx) transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) LOG.info("Generating confidence scores") # TODO: parallelize this loop? for transcription in tqdm(transcriptions): if cfg.tune_confidence: train_model_confidences.append( compute_all_confidences(transcription, cfg.tune_confidence_config) ) else: train_model_confidences.append(compute_confidence(transcription, cfg.confidence)) if model_idx == 0: # labels are the same for all models train_labels.append(data_idx) # optional dev if data_cfg.dev_manifest is not None: cfg.transcription.dataset_manifest = data_cfg.dev_manifest with tempfile.NamedTemporaryFile() as output_file: cfg.transcription.output_filename = output_file.name LOG.info("Transcribing dev dataset %d with model %d", data_idx, model_idx) transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) LOG.info("Generating confidence scores") for transcription in tqdm(transcriptions): if cfg.tune_confidence: dev_model_confidences.append( compute_all_confidences(transcription, cfg.tune_confidence_config) ) else: dev_model_confidences.append(compute_confidence(transcription, cfg.confidence)) if model_idx == 0: # labels are the same for all models dev_labels.append(data_idx) train_confidences.append(train_model_confidences) if dev_model_confidences: dev_confidences.append(dev_model_confidences) if cfg.tune_confidence: best_confidence, model_selection_block = find_best_confidence( train_confidences, train_labels, dev_confidences, dev_labels, cfg.tune_logistic_regression, cfg.tune_logistic_regression_config, ) else: best_confidence = cfg.confidence # transposing with zip(*list) training_features = np.array(list(zip(*train_confidences))) training_labels = np.array(train_labels) if dev_confidences: dev_features = np.array(list(zip(*dev_confidences))) dev_labels = np.array(dev_labels) else: dev_features = None dev_labels = None model_selection_block, _ = train_model_selection( training_features, training_labels, dev_features, dev_labels, cfg.tune_logistic_regression, cfg.tune_logistic_regression_config, verbose=True, ) with tempfile.TemporaryDirectory() as tmpdir: model_selection_block_path = os.path.join(tmpdir, 'model_selection_block.pkl') joblib.dump(model_selection_block, model_selection_block_path) # creating ensemble checkpoint ensemble_model = ConfidenceEnsembleModel( cfg=DictConfig( { 'model_selection_block': model_selection_block_path, 'confidence': best_confidence, 'temperature': cfg.temperature, 'load_models': [model_cfg.model for model_cfg in cfg.ensemble], } ), trainer=None, ) ensemble_model.save_to(cfg.output_path) if __name__ == '__main__': main() [end of scripts/confidence_ensembles/build_ensemble.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os from dataclasses import dataclass, is_dataclass from pathlib import Path from typing import Optional import pytorch_lightning as pl import torch from omegaconf import MISSING, OmegaConf from sklearn.model_selection import ParameterGrid from nemo.collections.asr.metrics.rnnt_wer import RNNTDecodingConfig from nemo.collections.asr.metrics.wer import CTCDecodingConfig from nemo.collections.asr.models import ASRModel, EncDecRNNTModel from nemo.collections.asr.parts.utils.asr_confidence_benchmarking_utils import ( apply_confidence_parameters, run_confidence_benchmark, ) from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig from nemo.core.config import hydra_runner from nemo.utils import logging """ Get confidence metrics and curve plots for a given model, dataset, and confidence parameters. # Arguments model_path: Path to .nemo ASR checkpoint pretrained_name: Name of pretrained ASR model (from NGC registry) dataset_manifest: Path to dataset JSON manifest file (in NeMo format) output_dir: Output directory to store a report and curve plot directories batch_size: batch size during inference num_workers: number of workers during inference cuda: Optional int to enable or disable execution of model on certain CUDA device amp: Bool to decide if Automatic Mixed Precision should be used during inference audio_type: Str filetype of the audio. Supported = wav, flac, mp3 target_level: Word- or token-level confidence. Supported = word, token, auto (for computing both word and token) confidence_cfg: Config with confidence parameters grid_params: Dictionary with lists of parameters to iteratively benchmark on # Usage ASR model can be specified by either "model_path" or "pretrained_name". Data for transcription are defined with "dataset_manifest". Results are returned as a benchmark report and curve plots. python benchmark_asr_confidence.py \ model_path=null \ pretrained_name=null \ dataset_manifest="" \ output_dir="" \ batch_size=64 \ num_workers=8 \ cuda=0 \ amp=True \ target_level="word" \ confidence_cfg.exclude_blank=False \ 'grid_params="{\"aggregation\": [\"min\", \"prod\"], \"alpha\": [0.33, 0.5]}"' """ def get_experiment_params(cfg): """Get experiment parameters from a confidence config and generate the experiment name. Returns: List of experiment parameters. String with the experiment name. """ blank = "no_blank" if cfg.exclude_blank else "blank" aggregation = cfg.aggregation method_name = cfg.measure_cfg.name alpha = cfg.measure_cfg.alpha if method_name == "entropy": entropy_type = cfg.measure_cfg.entropy_type entropy_norm = cfg.measure_cfg.entropy_norm experiment_param_list = [ aggregation, str(cfg.exclude_blank), method_name, entropy_type, entropy_norm, str(alpha), ] experiment_str = "-".join([aggregation, blank, method_name, entropy_type, entropy_norm, str(alpha)]) else: experiment_param_list = [aggregation, str(cfg.exclude_blank), method_name, "-", "-", str(alpha)] experiment_str = "-".join([aggregation, blank, method_name, str(alpha)]) return experiment_param_list, experiment_str @dataclass class ConfidenceBenchmarkingConfig: # Required configs model_path: Optional[str] = None # Path to a .nemo file pretrained_name: Optional[str] = None # Name of a pretrained model dataset_manifest: str = MISSING output_dir: str = MISSING # General configs batch_size: int = 32 num_workers: int = 4 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA # device anyway, and do inference on CPU only if CUDA device is not found. # If `cuda` is a negative number, inference will be on CPU only. cuda: Optional[int] = None amp: bool = False audio_type: str = "wav" # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) def main(cfg: ConfidenceBenchmarkingConfig): torch.set_grad_enabled(False) logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) if cfg.model_path is None and cfg.pretrained_name is None: raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None!") # setup GPU if cfg.cuda is None: if torch.cuda.is_available(): device = [0] # use 0th CUDA device accelerator = 'gpu' else: device = 1 accelerator = 'cpu' else: device = [cfg.cuda] accelerator = 'gpu' map_location = torch.device('cuda:{}'.format(device[0]) if accelerator == 'gpu' else 'cpu') # setup model if cfg.model_path is not None: # restore model from .nemo file path model_cfg = ASRModel.restore_from(restore_path=cfg.model_path, return_config=True) classpath = model_cfg.target # original class path imported_class = model_utils.import_class_by_path(classpath) # type: ASRModel logging.info(f"Restoring model : {imported_class.__name__}") asr_model = imported_class.restore_from( restore_path=cfg.model_path, map_location=map_location ) # type: ASRModel else: # restore model by name asr_model = ASRModel.from_pretrained( model_name=cfg.pretrained_name, map_location=map_location ) # type: ASRModel trainer = pl.Trainer(devices=device, accelerator=accelerator) asr_model.set_trainer(trainer) asr_model = asr_model.eval() # Check if ctc or rnnt model is_rnnt = isinstance(asr_model, EncDecRNNTModel) # Check that the model has the `change_decoding_strategy` method if not hasattr(asr_model, 'change_decoding_strategy'): raise RuntimeError("The asr_model you are using must have the `change_decoding_strategy` method.") # get filenames and reference texts from manifest filepaths = [] reference_texts = [] if os.stat(cfg.dataset_manifest).st_size == 0: logging.error(f"The input dataset_manifest {cfg.dataset_manifest} is empty. Exiting!") return None manifest_dir = Path(cfg.dataset_manifest).parent with open(cfg.dataset_manifest, 'r') as f: for line in f: item = json.loads(line) audio_file = Path(item['audio_filepath']) if not audio_file.is_file() and not audio_file.is_absolute(): audio_file = manifest_dir / audio_file filepaths.append(str(audio_file.absolute())) reference_texts.append(item['text']) # setup AMP (optional) autocast = None if cfg.amp and torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): logging.info("AMP enabled!\n") autocast = torch.cuda.amp.autocast # do grid-based benchmarking if grid_params is provided, otherwise a regular one work_dir = Path(cfg.output_dir) os.makedirs(work_dir, exist_ok=True) report_legend = ( ",".join( [ "model_type", "aggregation", "blank", "method_name", "entropy_type", "entropy_norm", "alpha", "target_level", "auc_roc", "auc_pr", "auc_nt", "nce", "ece", "auc_yc", "std_yc", "max_yc", ] ) + "\n" ) model_typename = "RNNT" if is_rnnt else "CTC" report_file = work_dir / Path("report.csv") if cfg.grid_params: asr_model.change_decoding_strategy( RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) if is_rnnt else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) ) params = json.loads(cfg.grid_params) hp_grid = ParameterGrid(params) hp_grid = list(hp_grid) logging.info(f"==============================Running a benchmarking with grid search=========================") logging.info(f"Grid search size: {len(hp_grid)}") logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directories near the file") logging.info(f"==============================================================================================") with open(report_file, "tw", encoding="utf-8") as f: f.write(report_legend) f.flush() for i, hp in enumerate(hp_grid): logging.info(f"Run # {i + 1}, grid: `{hp}`") asr_model.change_decoding_strategy(apply_confidence_parameters(asr_model.cfg.decoding, hp)) param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) plot_dir = work_dir / Path(experiment_name) results = run_confidence_benchmark( asr_model, cfg.target_level, filepaths, reference_texts, cfg.batch_size, cfg.num_workers, plot_dir, autocast, ) for level, result in results.items(): f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") f.flush() else: asr_model.change_decoding_strategy( RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) if is_rnnt else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) ) param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) plot_dir = work_dir / Path(experiment_name) logging.info(f"==============================Running a single benchmarking===================================") logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directory `{plot_dir}`") with open(report_file, "tw", encoding="utf-8") as f: f.write(report_legend) f.flush() results = run_confidence_benchmark( asr_model, cfg.batch_size, cfg.num_workers, cfg.target_level, filepaths, reference_texts, plot_dir, autocast, ) for level, result in results.items(): f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") logging.info(f"===========================================Done===============================================") if __name__ == '__main__': main() [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ # This script converts an existing audio dataset with a manifest to # a tarred and sharded audio dataset that can be read by the # TarredAudioToTextDataLayer. # Please make sure your audio_filepath DOES NOT CONTAIN '-sub'! # Because we will use it to handle files which have duplicate filenames but with different offsets # (see function create_shard for details) # Bucketing can help to improve the training speed. You may use --buckets_num to specify the number of buckets. # It creates multiple tarred datasets, one per bucket, based on the audio durations. # The range of [min_duration, max_duration) is split into equal sized buckets. # Recommend to use --sort_in_shards to speedup the training by reducing the paddings in the batches # More info on how to use bucketing feature: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/datasets.html # If valid NVIDIA DALI version is installed, will also generate the corresponding DALI index files that need to be # supplied to the config in order to utilize webdataset for efficient large dataset handling. # NOTE: DALI + Webdataset is NOT compatible with Bucketing support ! # Usage: 1) Creating a new tarfile dataset python convert_to_tarred_audio_dataset.py \ --manifest_path=<path to the manifest file> \ --target_dir=<path to output directory> \ --num_shards=<number of tarfiles that will contain the audio> \ --max_duration=<float representing maximum duration of audio samples> \ --min_duration=<float representing minimum duration of audio samples> \ --shuffle --shuffle_seed=1 \ --sort_in_shards \ --workers=-1 2) Concatenating more tarfiles to a pre-existing tarred dataset python convert_to_tarred_audio_dataset.py \ --manifest_path=<path to the tarred manifest file> \ --metadata_path=<path to the metadata.yaml (or metadata_version_{X}.yaml) file> \ --target_dir=<path to output directory where the original tarfiles are contained> \ --max_duration=<float representing maximum duration of audio samples> \ --min_duration=<float representing minimum duration of audio samples> \ --shuffle --shuffle_seed=1 \ --sort_in_shards \ --workers=-1 \ --concat_manifest_paths \ <space separated paths to 1 or more manifest files to concatenate into the original tarred dataset> 3) Writing an empty metadata file python convert_to_tarred_audio_dataset.py \ --target_dir=<path to output directory> \ # any other optional argument --num_shards=8 \ --max_duration=16.7 \ --min_duration=0.01 \ --shuffle \ --workers=-1 \ --sort_in_shards \ --shuffle_seed=1 \ --write_metadata """ import argparse import copy import json import os import random import tarfile from collections import defaultdict from dataclasses import dataclass, field from datetime import datetime from typing import Any, List, Optional from joblib import Parallel, delayed from omegaconf import DictConfig, OmegaConf, open_dict try: import create_dali_tarred_dataset_index as dali_index DALI_INDEX_SCRIPT_AVAILABLE = True except (ImportError, ModuleNotFoundError, FileNotFoundError): DALI_INDEX_SCRIPT_AVAILABLE = False parser = argparse.ArgumentParser( description="Convert an existing ASR dataset to tarballs compatible with TarredAudioToTextDataLayer." ) parser.add_argument( "--manifest_path", default=None, type=str, required=False, help="Path to the existing dataset's manifest." ) parser.add_argument( '--concat_manifest_paths', nargs='+', default=None, type=str, required=False, help="Path to the additional dataset's manifests that will be concatenated with base dataset.", ) # Optional arguments parser.add_argument( "--target_dir", default='./tarred', type=str, help="Target directory for resulting tarballs and manifest. Defaults to `./tarred`. Creates the path if necessary.", ) parser.add_argument( "--metadata_path", required=False, default=None, type=str, help="Path to metadata file for the dataset.", ) parser.add_argument( "--num_shards", default=-1, type=int, help="Number of shards (tarballs) to create. Used for partitioning data among workers.", ) parser.add_argument( '--max_duration', default=None, required=True, type=float, help='Maximum duration of audio clip in the dataset. By default, it is None and is required to be set.', ) parser.add_argument( '--min_duration', default=None, type=float, help='Minimum duration of audio clip in the dataset. By default, it is None and will not filter files.', ) parser.add_argument( "--shuffle", action='store_true', help="Whether or not to randomly shuffle the samples in the manifest before tarring/sharding.", ) parser.add_argument( "--keep_files_together", action='store_true', help="Whether or not to keep entries from the same file (but different offsets) together when sorting before tarring/sharding.", ) parser.add_argument( "--sort_in_shards", action='store_true', help="Whether or not to sort samples inside the shards based on their duration.", ) parser.add_argument( "--buckets_num", type=int, default=1, help="Number of buckets to create based on duration.", ) parser.add_argument("--shuffle_seed", type=int, default=None, help="Random seed for use if shuffling is enabled.") parser.add_argument( '--write_metadata', action='store_true', help=( "Flag to write a blank metadata with the current call config. " "Note that the metadata will not contain the number of shards, " "and it must be filled out by the user." ), ) parser.add_argument( "--no_shard_manifests", action='store_true', help="Do not write sharded manifests along with the aggregated manifest.", ) parser.add_argument('--workers', type=int, default=1, help='Number of worker processes') args = parser.parse_args() @dataclass class ASRTarredDatasetConfig: num_shards: int = -1 shuffle: bool = False max_duration: Optional[float] = None min_duration: Optional[float] = None shuffle_seed: Optional[int] = None sort_in_shards: bool = True shard_manifests: bool = True keep_files_together: bool = False @dataclass class ASRTarredDatasetMetadata: created_datetime: Optional[str] = None version: int = 0 num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): self.created_datetime = self.get_current_datetime() def get_current_datetime(self): return datetime.now().strftime("%m-%d-%Y %H-%M-%S") @classmethod def from_config(cls, config: DictConfig): obj = cls() obj.__dict__.update(**config) return obj @classmethod def from_file(cls, filepath: str): config = OmegaConf.load(filepath) return ASRTarredDatasetMetadata.from_config(config=config) class ASRTarredDatasetBuilder: """ Helper class that constructs a tarred dataset from scratch, or concatenates tarred datasets together and constructs manifests for them. """ def __init__(self): self.config = None def configure(self, config: ASRTarredDatasetConfig): """ Sets the config generated from command line overrides. Args: config: ASRTarredDatasetConfig dataclass object. """ self.config = config # type: ASRTarredDatasetConfig if self.config.num_shards < 0: raise ValueError("`num_shards` must be > 0. Please fill in the metadata information correctly.") def create_new_dataset(self, manifest_path: str, target_dir: str = "./tarred/", num_workers: int = 0): """ Creates a new tarred dataset from a given manifest file. Args: manifest_path: Path to the original ASR manifest. target_dir: Output directory. num_workers: Integer denoting number of parallel worker processes which will write tarfiles. Defaults to 1 - which denotes sequential worker process. Output: Writes tarfiles, along with the tarred dataset compatible manifest file. Also preserves a record of the metadata used to construct this tarred dataset. """ if self.config is None: raise ValueError("Config has not been set. Please call `configure(config: ASRTarredDatasetConfig)`") if manifest_path is None: raise FileNotFoundError("Manifest filepath cannot be None !") config = self.config # type: ASRTarredDatasetConfig if not os.path.exists(target_dir): os.makedirs(target_dir) # Read the existing manifest entries, total_duration, filtered_entries, filtered_duration = self._read_manifest(manifest_path, config) if len(filtered_entries) > 0: print(f"Filtered {len(filtered_entries)} files which amounts to {filtered_duration} seconds of audio.") print( f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." ) if len(entries) == 0: print("No tarred dataset was created as there were 0 valid samples after filtering!") return if config.shuffle: random.seed(config.shuffle_seed) print("Shuffling...") if config.keep_files_together: filename_entries = defaultdict(list) for ent in entries: filename_entries[ent["audio_filepath"]].append(ent) filenames = list(filename_entries.keys()) random.shuffle(filenames) shuffled_entries = [] for filename in filenames: shuffled_entries += filename_entries[filename] entries = shuffled_entries else: random.shuffle(entries) # Create shards and updated manifest entries print(f"Number of samples added : {len(entries)}") print(f"Remainder: {len(entries) % config.num_shards}") start_indices = [] end_indices = [] # Build indices for i in range(config.num_shards): start_idx = (len(entries) // config.num_shards) * i end_idx = start_idx + (len(entries) // config.num_shards) print(f"Shard {i} has entries {start_idx} ~ {end_idx}") files = set() for ent_id in range(start_idx, end_idx): files.add(entries[ent_id]["audio_filepath"]) print(f"Shard {i} contains {len(files)} files") if i == config.num_shards - 1: # We discard in order to have the same number of entries per shard. print(f"Have {len(entries) - end_idx} entries left over that will be discarded.") start_indices.append(start_idx) end_indices.append(end_idx) manifest_folder, _ = os.path.split(manifest_path) with Parallel(n_jobs=num_workers, verbose=config.num_shards) as parallel: # Call parallel tarfile construction new_entries_list = parallel( delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, i, manifest_folder) for i, (start_idx, end_idx) in enumerate(zip(start_indices, end_indices)) ) if config.shard_manifests: sharded_manifests_dir = target_dir + '/sharded_manifests' if not os.path.exists(sharded_manifests_dir): os.makedirs(sharded_manifests_dir) for manifest in new_entries_list: shard_id = manifest[0]['shard_id'] new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: for entry in manifest: json.dump(entry, m2) m2.write('\n') # Flatten the list of list of entries to a list of entries new_entries = [sample for manifest in new_entries_list for sample in manifest] del new_entries_list print("Total number of entries in manifest :", len(new_entries)) # Write manifest new_manifest_path = os.path.join(target_dir, 'tarred_audio_manifest.json') with open(new_manifest_path, 'w', encoding='utf-8') as m2: for entry in new_entries: json.dump(entry, m2) m2.write('\n') # Write metadata (default metadata for new datasets) new_metadata_path = os.path.join(target_dir, 'metadata.yaml') metadata = ASRTarredDatasetMetadata() # Update metadata metadata.dataset_config = config metadata.num_samples_per_shard = len(new_entries) // config.num_shards # Write metadata metadata_yaml = OmegaConf.structured(metadata) OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) def create_concatenated_dataset( self, base_manifest_path: str, manifest_paths: List[str], metadata: ASRTarredDatasetMetadata, target_dir: str = "./tarred_concatenated/", num_workers: int = 1, ): """ Creates new tarfiles in order to create a concatenated dataset, whose manifest contains the data for both the original dataset as well as the new data submitted in manifest paths. Args: base_manifest_path: Path to the manifest file which contains the information for the original tarred dataset (with flattened paths). manifest_paths: List of one or more paths to manifest files that will be concatenated with above base tarred dataset. metadata: ASRTarredDatasetMetadata dataclass instance with overrides from command line. target_dir: Output directory Output: Writes tarfiles which with indices mapping to a "concatenated" tarred dataset, along with the tarred dataset compatible manifest file which includes information about all the datasets that comprise the concatenated dataset. Also preserves a record of the metadata used to construct this tarred dataset. """ if not os.path.exists(target_dir): os.makedirs(target_dir) if base_manifest_path is None: raise FileNotFoundError("Base manifest filepath cannot be None !") if manifest_paths is None or len(manifest_paths) == 0: raise FileNotFoundError("List of additional manifest filepaths cannot be None !") config = ASRTarredDatasetConfig(**(metadata.dataset_config)) # Read the existing manifest (no filtering here) base_entries, _, _, _ = self._read_manifest(base_manifest_path, config) print(f"Read base manifest containing {len(base_entries)} samples.") # Precompute number of samples per shard if metadata.num_samples_per_shard is None: num_samples_per_shard = len(base_entries) // config.num_shards else: num_samples_per_shard = metadata.num_samples_per_shard print("Number of samples per shard :", num_samples_per_shard) # Compute min and max duration and update config (if no metadata passed) print(f"Selected max duration : {config.max_duration}") print(f"Selected min duration : {config.min_duration}") entries = [] for new_manifest_idx in range(len(manifest_paths)): new_entries, total_duration, filtered_new_entries, filtered_duration = self._read_manifest( manifest_paths[new_manifest_idx], config ) if len(filtered_new_entries) > 0: print( f"Filtered {len(filtered_new_entries)} files which amounts to {filtered_duration:0.2f}" f" seconds of audio from manifest {manifest_paths[new_manifest_idx]}." ) print( f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." ) entries.extend(new_entries) if len(entries) == 0: print("No tarred dataset was created as there were 0 valid samples after filtering!") return if config.shuffle: random.seed(config.shuffle_seed) print("Shuffling...") random.shuffle(entries) # Drop last section of samples that cannot be added onto a chunk drop_count = len(entries) % num_samples_per_shard total_new_entries = len(entries) entries = entries[:-drop_count] print( f"Dropping {drop_count} samples from total new samples {total_new_entries} since they cannot " f"be added into a uniformly sized chunk." ) # Create shards and updated manifest entries num_added_shards = len(entries) // num_samples_per_shard print(f"Number of samples in base dataset : {len(base_entries)}") print(f"Number of samples in additional datasets : {len(entries)}") print(f"Number of added shards : {num_added_shards}") print(f"Remainder: {len(entries) % num_samples_per_shard}") start_indices = [] end_indices = [] shard_indices = [] for i in range(num_added_shards): start_idx = (len(entries) // num_added_shards) * i end_idx = start_idx + (len(entries) // num_added_shards) shard_idx = i + config.num_shards print(f"Shard {shard_idx} has entries {start_idx + len(base_entries)} ~ {end_idx + len(base_entries)}") start_indices.append(start_idx) end_indices.append(end_idx) shard_indices.append(shard_idx) manifest_folder, _ = os.path.split(base_manifest_path) with Parallel(n_jobs=num_workers, verbose=num_added_shards) as parallel: # Call parallel tarfile construction new_entries_list = parallel( delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, shard_idx, manifest_folder) for i, (start_idx, end_idx, shard_idx) in enumerate(zip(start_indices, end_indices, shard_indices)) ) if config.shard_manifests: sharded_manifests_dir = target_dir + '/sharded_manifests' if not os.path.exists(sharded_manifests_dir): os.makedirs(sharded_manifests_dir) for manifest in new_entries_list: shard_id = manifest[0]['shard_id'] new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: for entry in manifest: json.dump(entry, m2) m2.write('\n') # Flatten the list of list of entries to a list of entries new_entries = [sample for manifest in new_entries_list for sample in manifest] del new_entries_list # Write manifest if metadata is None: new_version = 1 # start with `1`, where `0` indicates the base manifest + dataset else: new_version = metadata.version + 1 print("Total number of entries in manifest :", len(base_entries) + len(new_entries)) new_manifest_path = os.path.join(target_dir, f'tarred_audio_manifest_version_{new_version}.json') with open(new_manifest_path, 'w', encoding='utf-8') as m2: # First write all the entries of base manifest for entry in base_entries: json.dump(entry, m2) m2.write('\n') # Finally write the new entries for entry in new_entries: json.dump(entry, m2) m2.write('\n') # Preserve historical metadata base_metadata = metadata # Write metadata (updated metadata for concatenated datasets) new_metadata_path = os.path.join(target_dir, f'metadata_version_{new_version}.yaml') metadata = ASRTarredDatasetMetadata() # Update config config.num_shards = config.num_shards + num_added_shards # Update metadata metadata.version = new_version metadata.dataset_config = config metadata.num_samples_per_shard = num_samples_per_shard metadata.is_concatenated_manifest = True metadata.created_datetime = metadata.get_current_datetime() # Attach history current_metadata = OmegaConf.structured(base_metadata.history) metadata.history = current_metadata # Write metadata metadata_yaml = OmegaConf.structured(metadata) OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) def _read_manifest(self, manifest_path: str, config: ASRTarredDatasetConfig): """Read and filters data from the manifest""" # Read the existing manifest entries = [] total_duration = 0.0 filtered_entries = [] filtered_duration = 0.0 with open(manifest_path, 'r', encoding='utf-8') as m: for line in m: entry = json.loads(line) if (config.max_duration is None or entry['duration'] < config.max_duration) and ( config.min_duration is None or entry['duration'] >= config.min_duration ): entries.append(entry) total_duration += entry["duration"] else: filtered_entries.append(entry) filtered_duration += entry['duration'] return entries, total_duration, filtered_entries, filtered_duration def _create_shard(self, entries, target_dir, shard_id, manifest_folder): """Creates a tarball containing the audio files from `entries`. """ if self.config.sort_in_shards: entries.sort(key=lambda x: x["duration"], reverse=False) new_entries = [] tar = tarfile.open(os.path.join(target_dir, f'audio_{shard_id}.tar'), mode='w', dereference=True) count = dict() for entry in entries: # We squash the filename since we do not preserve directory structure of audio files in the tarball. if os.path.exists(entry["audio_filepath"]): audio_filepath = entry["audio_filepath"] else: audio_filepath = os.path.join(manifest_folder, entry["audio_filepath"]) if not os.path.exists(audio_filepath): raise FileNotFoundError(f"Could not find {entry['audio_filepath']}!") base, ext = os.path.splitext(audio_filepath) base = base.replace('/', '_') # Need the following replacement as long as WebDataset splits on first period base = base.replace('.', '_') squashed_filename = f'{base}{ext}' if squashed_filename not in count: tar.add(audio_filepath, arcname=squashed_filename) to_write = squashed_filename count[squashed_filename] = 1 else: to_write = base + "-sub" + str(count[squashed_filename]) + ext count[squashed_filename] += 1 new_entry = { 'audio_filepath': to_write, 'duration': entry['duration'], 'shard_id': shard_id, # Keep shard ID for recordkeeping } if 'label' in entry: new_entry['label'] = entry['label'] if 'text' in entry: new_entry['text'] = entry['text'] if 'offset' in entry: new_entry['offset'] = entry['offset'] if 'lang' in entry: new_entry['lang'] = entry['lang'] new_entries.append(new_entry) tar.close() return new_entries @classmethod def setup_history(cls, base_metadata: ASRTarredDatasetMetadata, history: List[Any]): if 'history' in base_metadata.keys(): for history_val in base_metadata.history: cls.setup_history(history_val, history) if base_metadata is not None: metadata_copy = copy.deepcopy(base_metadata) with open_dict(metadata_copy): metadata_copy.pop('history', None) history.append(metadata_copy) def main(): if args.buckets_num > 1: bucket_length = (args.max_duration - args.min_duration) / float(args.buckets_num) for i in range(args.buckets_num): min_duration = args.min_duration + i * bucket_length max_duration = min_duration + bucket_length if i == args.buckets_num - 1: # add a small number to cover the samples with exactly duration of max_duration in the last bucket. max_duration += 1e-5 target_dir = os.path.join(args.target_dir, f"bucket{i+1}") print(f"Creating bucket {i+1} with min_duration={min_duration} and max_duration={max_duration} ...") print(f"Results are being saved at: {target_dir}.") create_tar_datasets(min_duration=min_duration, max_duration=max_duration, target_dir=target_dir) print(f"Bucket {i+1} is created.") else: create_tar_datasets(min_duration=args.min_duration, max_duration=args.max_duration, target_dir=args.target_dir) def create_tar_datasets(min_duration: float, max_duration: float, target_dir: str): builder = ASRTarredDatasetBuilder() shard_manifests = False if args.no_shard_manifests else True if args.write_metadata: metadata = ASRTarredDatasetMetadata() dataset_cfg = ASRTarredDatasetConfig( num_shards=args.num_shards, shuffle=args.shuffle, max_duration=max_duration, min_duration=min_duration, shuffle_seed=args.shuffle_seed, sort_in_shards=args.sort_in_shards, shard_manifests=shard_manifests, keep_files_together=args.keep_files_together, ) metadata.dataset_config = dataset_cfg output_path = os.path.join(target_dir, 'default_metadata.yaml') OmegaConf.save(metadata, output_path, resolve=True) print(f"Default metadata written to {output_path}") exit(0) if args.concat_manifest_paths is None or len(args.concat_manifest_paths) == 0: print("Creating new tarred dataset ...") # Create a tarred dataset from scratch config = ASRTarredDatasetConfig( num_shards=args.num_shards, shuffle=args.shuffle, max_duration=max_duration, min_duration=min_duration, shuffle_seed=args.shuffle_seed, sort_in_shards=args.sort_in_shards, shard_manifests=shard_manifests, keep_files_together=args.keep_files_together, ) builder.configure(config) builder.create_new_dataset(manifest_path=args.manifest_path, target_dir=target_dir, num_workers=args.workers) else: if args.buckets_num > 1: raise ValueError("Concatenation feature does not support buckets_num > 1.") print("Concatenating multiple tarred datasets ...") # Implicitly update config from base details if args.metadata_path is not None: metadata = ASRTarredDatasetMetadata.from_file(args.metadata_path) else: raise ValueError("`metadata` yaml file path must be provided!") # Preserve history history = [] builder.setup_history(OmegaConf.structured(metadata), history) metadata.history = history # Add command line overrides (everything other than num_shards) metadata.dataset_config.max_duration = max_duration metadata.dataset_config.min_duration = min_duration metadata.dataset_config.shuffle = args.shuffle metadata.dataset_config.shuffle_seed = args.shuffle_seed metadata.dataset_config.sort_in_shards = args.sort_in_shards metadata.dataset_config.shard_manifests = shard_manifests builder.configure(metadata.dataset_config) # Concatenate a tarred dataset onto a previous one builder.create_concatenated_dataset( base_manifest_path=args.manifest_path, manifest_paths=args.concat_manifest_paths, metadata=metadata, target_dir=target_dir, num_workers=args.workers, ) if DALI_INDEX_SCRIPT_AVAILABLE and dali_index.INDEX_CREATOR_AVAILABLE: print("Constructing DALI Tarfile Index - ", target_dir) index_config = dali_index.DALITarredIndexConfig(tar_dir=target_dir, workers=args.workers) dali_index.main(index_config) if __name__ == "__main__": main() [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of tools/nemo_forced_aligner/align.py] # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import math import os from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import List, Optional import torch from omegaconf import OmegaConf from utils.data_prep import ( add_t_start_end_to_utt_obj, get_batch_starts_ends, get_batch_variables, get_manifest_lines_batch, is_entry_in_all_lines, is_entry_in_any_lines, ) from utils.make_ass_files import make_ass_files from utils.make_ctm_files import make_ctm_files from utils.make_output_manifest import write_manifest_out_line from utils.viterbi_decoding import viterbi_decoding from nemo.collections.asr.models.ctc_models import EncDecCTCModel from nemo.collections.asr.models.hybrid_rnnt_ctc_models import EncDecHybridRNNTCTCModel from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchASR from nemo.collections.asr.parts.utils.transcribe_utils import setup_model from nemo.core.config import hydra_runner from nemo.utils import logging """ Align the utterances in manifest_filepath. Results are saved in ctm files in output_dir. Arguments: pretrained_name: string specifying the name of a CTC NeMo ASR model which will be automatically downloaded from NGC and used for generating the log-probs which we will use to do alignment. Note: NFA can only use CTC models (not Transducer models) at the moment. model_path: string specifying the local filepath to a CTC NeMo ASR model which will be used to generate the log-probs which we will use to do alignment. Note: NFA can only use CTC models (not Transducer models) at the moment. Note: if a model_path is provided, it will override the pretrained_name. manifest_filepath: filepath to the manifest of the data you want to align, containing 'audio_filepath' and 'text' fields. output_dir: the folder where output CTM files and new JSON manifest will be saved. align_using_pred_text: if True, will transcribe the audio using the specified model and then use that transcription as the reference text for the forced alignment. transcribe_device: None, or a string specifying the device that will be used for generating log-probs (i.e. "transcribing"). The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available (otherwise will set it to 'cpu'). viterbi_device: None, or string specifying the device that will be used for doing Viterbi decoding. The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available (otherwise will set it to 'cpu'). batch_size: int specifying batch size that will be used for generating log-probs and doing Viterbi decoding. use_local_attention: boolean flag specifying whether to try to use local attention for the ASR Model (will only work if the ASR Model is a Conformer model). If local attention is used, we will set the local attention context size to [64,64]. additional_segment_grouping_separator: an optional string used to separate the text into smaller segments. If this is not specified, then the whole text will be treated as a single segment. remove_blank_tokens_from_ctm: a boolean denoting whether to remove <blank> tokens from token-level output CTMs. audio_filepath_parts_in_utt_id: int specifying how many of the 'parts' of the audio_filepath we will use (starting from the final part of the audio_filepath) to determine the utt_id that will be used in the CTM files. Note also that any spaces that are present in the audio_filepath will be replaced with dashes, so as not to change the number of space-separated elements in the CTM files. e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 1 => utt_id will be "e1" e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 2 => utt_id will be "d_e1" e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 3 => utt_id will be "c_d_e1" use_buffered_infer: False, if set True, using streaming to do get the logits for alignment This flag is useful when aligning large audio file. However, currently the chunk streaming inference does not support batch inference, which means even you set batch_size > 1, it will only infer one by one instead of doing the whole batch inference together. chunk_len_in_secs: float chunk length in seconds total_buffer_in_secs: float Length of buffer (chunk + left and right padding) in seconds chunk_batch_size: int batch size for buffered chunk inference, which will cut one audio into segments and do inference on chunk_batch_size segments at a time simulate_cache_aware_streaming: False, if set True, using cache aware streaming to do get the logits for alignment save_output_file_formats: List of strings specifying what type of output files to save (default: ["ctm", "ass"]) ctm_file_config: CTMFileConfig to specify the configuration of the output CTM files ass_file_config: ASSFileConfig to specify the configuration of the output ASS files """ @dataclass class CTMFileConfig: remove_blank_tokens: bool = False # minimum duration (in seconds) for timestamps in the CTM.If any line in the CTM has a # duration lower than this, it will be enlarged from the middle outwards until it # meets the minimum_timestamp_duration, or reaches the beginning or end of the audio file. # Note that this may cause timestamps to overlap. minimum_timestamp_duration: float = 0 @dataclass class ASSFileConfig: fontsize: int = 20 vertical_alignment: str = "center" # if resegment_text_to_fill_space is True, the ASS files will use new segments # such that each segment will not take up more than (approximately) max_lines_per_segment # when the ASS file is applied to a video resegment_text_to_fill_space: bool = False max_lines_per_segment: int = 2 text_already_spoken_rgb: List[int] = field(default_factory=lambda: [49, 46, 61]) # dark gray text_being_spoken_rgb: List[int] = field(default_factory=lambda: [57, 171, 9]) # dark green text_not_yet_spoken_rgb: List[int] = field(default_factory=lambda: [194, 193, 199]) # light gray @dataclass class AlignmentConfig: # Required configs pretrained_name: Optional[str] = None model_path: Optional[str] = None manifest_filepath: Optional[str] = None output_dir: Optional[str] = None # General configs align_using_pred_text: bool = False transcribe_device: Optional[str] = None viterbi_device: Optional[str] = None batch_size: int = 1 use_local_attention: bool = True additional_segment_grouping_separator: Optional[str] = None audio_filepath_parts_in_utt_id: int = 1 # Buffered chunked streaming configs use_buffered_chunked_streaming: bool = False chunk_len_in_secs: float = 1.6 total_buffer_in_secs: float = 4.0 chunk_batch_size: int = 32 # Cache aware streaming configs simulate_cache_aware_streaming: Optional[bool] = False # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) ctm_file_config: CTMFileConfig = CTMFileConfig() ass_file_config: ASSFileConfig = ASSFileConfig() @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) def main(cfg: AlignmentConfig): logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) # Validate config if cfg.model_path is None and cfg.pretrained_name is None: raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None") if cfg.model_path is not None and cfg.pretrained_name is not None: raise ValueError("One of cfg.model_path and cfg.pretrained_name must be None") if cfg.manifest_filepath is None: raise ValueError("cfg.manifest_filepath must be specified") if cfg.output_dir is None: raise ValueError("cfg.output_dir must be specified") if cfg.batch_size < 1: raise ValueError("cfg.batch_size cannot be zero or a negative number") if cfg.additional_segment_grouping_separator == "" or cfg.additional_segment_grouping_separator == " ": raise ValueError("cfg.additional_grouping_separator cannot be empty string or space character") if cfg.ctm_file_config.minimum_timestamp_duration < 0: raise ValueError("cfg.minimum_timestamp_duration cannot be a negative number") if cfg.ass_file_config.vertical_alignment not in ["top", "center", "bottom"]: raise ValueError("cfg.ass_file_config.vertical_alignment must be one of 'top', 'center' or 'bottom'") for rgb_list in [ cfg.ass_file_config.text_already_spoken_rgb, cfg.ass_file_config.text_already_spoken_rgb, cfg.ass_file_config.text_already_spoken_rgb, ]: if len(rgb_list) != 3: raise ValueError( "cfg.ass_file_config.text_already_spoken_rgb," " cfg.ass_file_config.text_being_spoken_rgb," " and cfg.ass_file_config.text_already_spoken_rgb all need to contain" " exactly 3 elements." ) # Validate manifest contents if not is_entry_in_all_lines(cfg.manifest_filepath, "audio_filepath"): raise RuntimeError( "At least one line in cfg.manifest_filepath does not contain an 'audio_filepath' entry. " "All lines must contain an 'audio_filepath' entry." ) if cfg.align_using_pred_text: if is_entry_in_any_lines(cfg.manifest_filepath, "pred_text"): raise RuntimeError( "Cannot specify cfg.align_using_pred_text=True when the manifest at cfg.manifest_filepath " "contains 'pred_text' entries. This is because the audio will be transcribed and may produce " "a different 'pred_text'. This may cause confusion." ) else: if not is_entry_in_all_lines(cfg.manifest_filepath, "text"): raise RuntimeError( "At least one line in cfg.manifest_filepath does not contain a 'text' entry. " "NFA requires all lines to contain a 'text' entry when cfg.align_using_pred_text=False." ) # init devices if cfg.transcribe_device is None: transcribe_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") else: transcribe_device = torch.device(cfg.transcribe_device) logging.info(f"Device to be used for transcription step (`transcribe_device`) is {transcribe_device}") if cfg.viterbi_device is None: viterbi_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") else: viterbi_device = torch.device(cfg.viterbi_device) logging.info(f"Device to be used for viterbi step (`viterbi_device`) is {viterbi_device}") if transcribe_device.type == 'cuda' or viterbi_device.type == 'cuda': logging.warning( 'One or both of transcribe_device and viterbi_device are GPUs. If you run into OOM errors ' 'it may help to change both devices to be the CPU.' ) # load model model, _ = setup_model(cfg, transcribe_device) model.eval() if isinstance(model, EncDecHybridRNNTCTCModel): model.change_decoding_strategy(decoder_type="ctc") if cfg.use_local_attention: logging.info( "Flag use_local_attention is set to True => will try to use local attention for model if it allows it" ) model.change_attention_model(self_attention_model="rel_pos_local_attn", att_context_size=[64, 64]) if not (isinstance(model, EncDecCTCModel) or isinstance(model, EncDecHybridRNNTCTCModel)): raise NotImplementedError( f"Model is not an instance of NeMo EncDecCTCModel or ENCDecHybridRNNTCTCModel." " Currently only instances of these models are supported" ) if cfg.ctm_file_config.minimum_timestamp_duration > 0: logging.warning( f"cfg.ctm_file_config.minimum_timestamp_duration has been set to {cfg.ctm_file_config.minimum_timestamp_duration} seconds. " "This may cause the alignments for some tokens/words/additional segments to be overlapping." ) buffered_chunk_params = {} if cfg.use_buffered_chunked_streaming: model_cfg = copy.deepcopy(model._cfg) OmegaConf.set_struct(model_cfg.preprocessor, False) # some changes for streaming scenario model_cfg.preprocessor.dither = 0.0 model_cfg.preprocessor.pad_to = 0 if model_cfg.preprocessor.normalize != "per_feature": logging.error( "Only EncDecCTCModelBPE models trained with per_feature normalization are supported currently" ) # Disable config overwriting OmegaConf.set_struct(model_cfg.preprocessor, True) feature_stride = model_cfg.preprocessor['window_stride'] model_stride_in_secs = feature_stride * cfg.model_downsample_factor total_buffer = cfg.total_buffer_in_secs chunk_len = float(cfg.chunk_len_in_secs) tokens_per_chunk = math.ceil(chunk_len / model_stride_in_secs) mid_delay = math.ceil((chunk_len + (total_buffer - chunk_len) / 2) / model_stride_in_secs) logging.info(f"tokens_per_chunk is {tokens_per_chunk}, mid_delay is {mid_delay}") model = FrameBatchASR( asr_model=model, frame_len=chunk_len, total_buffer=cfg.total_buffer_in_secs, batch_size=cfg.chunk_batch_size, ) buffered_chunk_params = { "delay": mid_delay, "model_stride_in_secs": model_stride_in_secs, "tokens_per_chunk": tokens_per_chunk, } # get start and end line IDs of batches starts, ends = get_batch_starts_ends(cfg.manifest_filepath, cfg.batch_size) # init output_timestep_duration = None and we will calculate and update it during the first batch output_timestep_duration = None # init f_manifest_out os.makedirs(cfg.output_dir, exist_ok=True) tgt_manifest_name = str(Path(cfg.manifest_filepath).stem) + "_with_output_file_paths.json" tgt_manifest_filepath = str(Path(cfg.output_dir) / tgt_manifest_name) f_manifest_out = open(tgt_manifest_filepath, 'w') # get alignment and save in CTM batch-by-batch for start, end in zip(starts, ends): manifest_lines_batch = get_manifest_lines_batch(cfg.manifest_filepath, start, end) (log_probs_batch, y_batch, T_batch, U_batch, utt_obj_batch, output_timestep_duration,) = get_batch_variables( manifest_lines_batch, model, cfg.additional_segment_grouping_separator, cfg.align_using_pred_text, cfg.audio_filepath_parts_in_utt_id, output_timestep_duration, cfg.simulate_cache_aware_streaming, cfg.use_buffered_chunked_streaming, buffered_chunk_params, ) alignments_batch = viterbi_decoding(log_probs_batch, y_batch, T_batch, U_batch, viterbi_device) for utt_obj, alignment_utt in zip(utt_obj_batch, alignments_batch): utt_obj = add_t_start_end_to_utt_obj(utt_obj, alignment_utt, output_timestep_duration) if "ctm" in cfg.save_output_file_formats: utt_obj = make_ctm_files(utt_obj, cfg.output_dir, cfg.ctm_file_config,) if "ass" in cfg.save_output_file_formats: utt_obj = make_ass_files(utt_obj, cfg.output_dir, cfg.ass_file_config) write_manifest_out_line( f_manifest_out, utt_obj, ) f_manifest_out.close() return None if __name__ == "__main__": main() [end of tools/nemo_forced_aligner/align.py] </code> I need you to solve this issue by regenerating the full files in the code base that you would like to change. You can change as many files as you like. Please respond with a list of files and their revised contents in the following format. <example> [start of /src/this_file.py] import os def euclidean(a, b): if b == 0: return a return euclidean(b, a % b) [end of /src/this_file.py] [start of /src/another_file.py] def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) x, y = x0, y0 sx = -1 if x0 > x1 else 1 sy = -1 if y0 > y1 else 1 if dx > dy: err = dx / 2.0 while x != x1: points.append((x, y)) err -= dy if err < 0: y += sy err += dx x += sx else: err = dy / 2.0 while y != y1: points.append((x err -= dx if err < 0: x += sx err += dy y += sy points.append((x, y)) return points [end of /src/another_file.py] </example>
NVIDIA/NeMo
8a892b86186dbdf61803d75570cb5c58471e9dda
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-09-30T01:26:50Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,9 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,9 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): @@ -2217,7 +2219,9 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -181,7 +181,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() + measure_cfg: ConfidenceMeasureConfig = field(default_factory=lambda: ConfidenceMeasureConfig()) method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -110,7 +110,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
NVIDIA__NeMo-7616
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| .. |status| image:: http://www.repostatus.org/badges/latest/active.svg :target: http://www.repostatus.org/#active :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main :alt: Documentation :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE :alt: NeMo core license and license for collections in this repo .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg :target: https://badge.fury.io/py/nemo-toolkit :alt: Release version .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg :target: https://badge.fury.io/py/nemo-toolkit :alt: Python version .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads :target: https://pepy.tech/project/nemo-toolkit :alt: PyPi total downloads .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml :alt: CodeQL .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg :target: https://github.com/psf/black :alt: Code style: black .. _main-readme: **NVIDIA NeMo** =============== Introduction ------------ NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), text-to-speech synthesis (TTS), large language models (LLMs), and natural language processing (NLP). The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and training is automatically scalable to 1000s of GPUs. Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. Getting started with NeMo is simple. State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that can all be run on `Google Colab <https://colab.research.google.com>`_. For advanced users that want to train NeMo models from scratch or finetune existing NeMo models we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ which can be used to find the optimal model parallel configuration for training on a specific cluster. Also see the two introductory videos below for a high level overview of NeMo. * Developing State-Of-The-Art Conversational AI Models in Three Lines of Code. * NVIDIA NeMo: Toolkit for Conversational AI at PyData Yerevan 2022. |three_lines| |pydata| .. |pydata| image:: https://img.youtube.com/vi/J-P6Sczmas8/maxres3.jpg :target: https://www.youtube.com/embed/J-P6Sczmas8?mute=0&start=14&autoplay=0 :width: 600 :alt: Develop Conversational AI Models in 3 Lines .. |three_lines| image:: https://img.youtube.com/vi/wBgpMf_KQVw/maxresdefault.jpg :target: https://www.youtube.com/embed/wBgpMf_KQVw?mute=0&start=0&autoplay=0 :width: 600 :alt: Introduction at PyData@Yerevan 2022 Key Features ------------ * Speech processing * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ * Jasper, QuartzNet, CitriNet, ContextNet * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer * Squeezeformer-CTC and Squeezeformer-Transducer * LSTM-Transducer (RNNT) and LSTM-CTC * Supports the following decoders/losses: * CTC * Transducer/RNNT * Hybrid Transducer/CTC * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ * Beam Search decoding * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ * Natural Language Processing * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ * Text-to-Speech Synthesis (TTS): * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 * Vocoders: HiFiGAN, UnivNet, WaveGlow * End-to-End Models: VITS * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. Requirements ------------ 1) Python 3.10 or above 2) Pytorch 1.13.1 or above 3) NVIDIA GPU for training Documentation ------------- .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main :alt: Documentation Status :scale: 100% :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable :alt: Documentation Status :scale: 100% :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ | Version | Status | Description | +=========+=============+==========================================================================================================================================+ | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ Tutorials --------- A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. Getting help with NeMo ---------------------- FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. Installation ------------ Conda ~~~~~ We recommend installing NeMo in a fresh Conda environment. .. code-block:: bash conda create --name nemo python==3.10.12 conda activate nemo Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. .. code-block:: bash conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. Pip ~~~ Use this installation mode if you want the latest released version. .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg pip install Cython pip install nemo_toolkit['all'] Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. Pip from source ~~~~~~~~~~~~~~~ Use this installation mode if you want the version from a particular GitHub branch (e.g main). .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg pip install Cython python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] From source ~~~~~~~~~~~ Use this installation mode if you are contributing to NeMo. .. code-block:: bash apt-get update && apt-get install -y libsndfile1 ffmpeg git clone https://github.com/NVIDIA/NeMo cd NeMo ./reinstall.sh If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` with ``pip install -e .`` when your PWD is the root of the NeMo repository. RNNT ~~~~ Note that RNNT requires numba to be installed from conda. .. code-block:: bash conda remove numba pip uninstall numba conda install -c conda-forge numba NeMo Megatron ~~~~~~~~~~~~~ NeMo Megatron training requires NVIDIA Apex to be installed. Install it manually if not using the NVIDIA PyTorch container. To install Apex, run .. code-block:: bash git clone https://github.com/NVIDIA/apex.git cd apex git checkout 52e18c894223800cb611682dce27d88050edf1de pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: .. code-block:: bash conda install -c nvidia cuda-nvprof=11.8 packaging is also needed: .. code-block:: bash pip install packaging With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. Transformer Engine ~~~~~~~~~~~~~~~~~~ NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. .. code-block:: bash pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. Transformer Engine requires PyTorch to be built with CUDA 11.8. Flash Attention ~~~~~~~~~~~~~~~~~~~~ Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. .. code-block:: bash pip install flash-attn pip install triton==2.0.0.dev20221202 NLP inference UI ~~~~~~~~~~~~~~~~~~~~ To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. .. code-block:: bash pip install gradio==3.34.0 NeMo Text Processing ~~~~~~~~~~~~~~~~~~~~ NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. Docker containers: ~~~~~~~~~~~~~~~~~~ We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. To use built container, please run .. code-block:: bash docker pull nvcr.io/nvidia/nemo:23.06 To build a nemo container with Dockerfile from a branch, please run .. code-block:: bash DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. .. code-block:: bash docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 Examples -------- Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. Contributing ------------ We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. Publications ------------ We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! License ------- NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. [end of README.rst] [start of examples/asr/experimental/k2/align_speech_parallel.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ # Based on examples/asr/transcribe_speech_parallel.py # ASR alignment with multi-GPU/multi-node support for large datasets # It supports both tarred and non-tarred datasets # Arguments # model: path to a nemo/PTL checkpoint file or name of a pretrained model # predict_ds: config of the dataset/dataloader # aligner_args: aligner config # output_path: path to store the predictions # model_stride: model downsampling factor, 8 for Citrinet models and 4 for Conformer models # # Results of each GPU/worker is written into a file named 'predictions_{rank}.json, and aggregated results of all workers are written into 'predictions_all.json' Example for non-tarred datasets: python align_speech_parallel.py \ model=stt_en_conformer_ctc_large \ predict_ds.manifest_filepath=/dataset/manifest_file.json \ predict_ds.batch_size=16 \ output_path=/tmp/ Example for tarred datasets: python align_speech_parallel.py \ predict_ds.is_tarred=true \ predict_ds.manifest_filepath=/tarred_dataset/tarred_audio_manifest.json \ predict_ds.tarred_audio_filepaths=/tarred_dataset/audio__OP_0..127_CL_.tar \ ... By default the trainer uses all the GPUs available and default precision is FP32. By setting the trainer config you may control these configs. For example to do the predictions with AMP on just two GPUs: python align_speech_parallel.py \ trainer.precision=16 \ trainer.gpus=2 \ ... You may control the dataloader's config by setting the predict_ds: python align_speech_parallel.py \ predict_ds.num_workers=8 \ predict_ds.min_duration=2.0 \ predict_ds.sample_rate=16000 \ model=stt_en_conformer_ctc_small \ ... You may control the aligner's config by setting the aligner_args: aligner_args.alignment_type=argmax \ aligner_args.word_output=False \ aligner_args.cpu_decoding=True \ aligner_args.decode_batch_size=8 \ aligner_args.ctc_cfg.prob_suppress_index=-1 \ aligner_args.ctc_cfg.prob_suppress_value=0.5 \ aligner_args.rnnt_cfg.predictor_window_size=10 \ aligner_args.decoder_module_cfg.intersect_pruned=true \ aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ ... """ import os from dataclasses import dataclass, is_dataclass from typing import Optional import pytorch_lightning as ptl import torch from omegaconf import MISSING, OmegaConf from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter from nemo.collections.asr.models import ASRModel from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel from nemo.core.config import TrainerConfig, hydra_runner from nemo.utils import logging from nemo.utils.get_rank import is_global_rank_zero @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() output_path: str = MISSING model_stride: int = 8 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") # there arguments will be ignored return_predictions: bool = False use_cer: bool = False def match_train_config(predict_ds, train_ds): # It copies the important configurations from the train dataset of the model # into the predict_ds to be used for prediction. It is needed to match the training configurations. if train_ds is None: return predict_ds.sample_rate = train_ds.get("sample_rate", 16000) cfg_name_list = [ "int_values", "use_start_end_token", "blank_index", "unk_index", "normalize", "parser", "eos_id", "bos_id", "pad_id", ] if is_dataclass(predict_ds): predict_ds = OmegaConf.structured(predict_ds) for cfg_name in cfg_name_list: if hasattr(train_ds, cfg_name): setattr(predict_ds, cfg_name, getattr(train_ds, cfg_name)) return predict_ds @hydra_runner(config_name="AlignmentConfig", schema=ParallelAlignmentConfig) def main(cfg: ParallelAlignmentConfig): if cfg.model.endswith(".nemo"): logging.info("Attempting to initialize from .nemo file") model = ASRModel.restore_from(restore_path=cfg.model, map_location="cpu") elif cfg.model.endswith(".ckpt"): logging.info("Attempting to initialize from .ckpt file") model = ASRModel.load_from_checkpoint(checkpoint_path=cfg.model, map_location="cpu") else: logging.info( "Attempting to initialize from a pretrained model as the model name does not have the extension of .nemo or .ckpt" ) model = ASRModel.from_pretrained(model_name=cfg.model, map_location="cpu") trainer = ptl.Trainer(**cfg.trainer) cfg.predict_ds.return_sample_id = True cfg.return_predictions = False cfg.use_cer = False cfg.predict_ds = match_train_config(predict_ds=cfg.predict_ds, train_ds=model._cfg.train_ds) data_loader = model._setup_dataloader_from_config(cfg.predict_ds) os.makedirs(cfg.output_path, exist_ok=True) # trainer.global_rank is not valid before predict() is called. Need this hack to find the correct global_rank. global_rank = trainer.node_rank * trainer.num_devices + int(os.environ.get("LOCAL_RANK", 0)) output_file = os.path.join(cfg.output_path, f"predictions_{global_rank}.json") output_ctm_dir = os.path.join(cfg.output_path, "ctm") predictor_writer = ASRCTMPredictionWriter( dataset=data_loader.dataset, output_file=output_file, output_ctm_dir=output_ctm_dir, time_per_frame=cfg.model_stride * model._cfg.preprocessor['window_stride'], ) trainer.callbacks.extend([predictor_writer]) aligner_wrapper = AlignerWrapperModel(model=model, cfg=cfg.aligner_args) trainer.predict(model=aligner_wrapper, dataloaders=data_loader, return_predictions=cfg.return_predictions) samples_num = predictor_writer.close_output_file() logging.info( f"Prediction on rank {global_rank} is done for {samples_num} samples and results are stored in {output_file}." ) if torch.distributed.is_initialized(): torch.distributed.barrier() samples_num = 0 if is_global_rank_zero(): output_file = os.path.join(cfg.output_path, f"predictions_all.json") logging.info(f"Prediction files are being aggregated in {output_file}.") with open(output_file, 'tw', encoding="utf-8") as outf: for rank in range(trainer.world_size): input_file = os.path.join(cfg.output_path, f"predictions_{rank}.json") with open(input_file, 'r', encoding="utf-8") as inpf: lines = inpf.readlines() samples_num += len(lines) outf.writelines(lines) logging.info( f"Prediction is done for {samples_num} samples in total on all workers and results are aggregated in {output_file}." ) if __name__ == '__main__': main() [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import re from abc import abstractmethod from dataclasses import dataclass, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance import numpy as np import torch from omegaconf import OmegaConf from torchmetrics import Metric from nemo.collections.asr.metrics.wer import move_dimension_to_the_front from nemo.collections.asr.parts.submodules import rnnt_beam_decoding as beam_decode from nemo.collections.asr.parts.submodules import rnnt_greedy_decoding as greedy_decode from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses from nemo.utils import logging __all__ = ['RNNTDecoding', 'RNNTWER'] class AbstractRNNTDecoding(ConfidenceMixin): """ Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy, greedy_batch (for greedy decoding). - beam, tsd, alsd (for beam search decoding). compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded tokens as well as the decoded string. Default is False in order to avoid double decoding unless required. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function with the `return_hypotheses` flag set to True. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrete intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. rnnt_timestamp_type: A str value, which represents the types of timestamps that should be calculated. Can take the following values - "char" for character/subword time stamps, "word" for word level time stamps and "all" (default), for both character level and word level time stamps. word_seperator: Str token representing the seperator between words. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of ints. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. The config may further contain the following sub-dictionaries: "greedy": max_symbols: int, describing the maximum number of target tokens to decode per timestep during greedy decoding. Setting to larger values allows longer sentences to be decoded, at the cost of increased execution time. preserve_frame_confidence: Same as above, overrides above value. confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. Set to True by default. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, at increased cost to execution time. alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. If an integer is provided, it can decode sequences of that particular maximum length. If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), where seq_len is the length of the acoustic model output (T). NOTE: If a float is provided, it can be greater than 1! By default, a float of 2.0 is used so that a target sequence can be at most twice as long as the acoustic model output length T. maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 in order to reduce expensive beam search cost later. int >= 0. maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, and affects the speed of inference since large values will perform large beam search in the next step. maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for expansion apart from the "most likely" candidate. Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally tuned on a validation set. softmax_temperature: Scales the logits of the joint prior to computing log_softmax. decoder: The Decoder/Prediction network module. joint: The Joint network module. blank_id: The id of the RNNT blank token. """ def __init__(self, decoding_cfg, decoder, joint, blank_id: int): super(AbstractRNNTDecoding, self).__init__() # Convert dataclass to config object if is_dataclass(decoding_cfg): decoding_cfg = OmegaConf.structured(decoding_cfg) self.cfg = decoding_cfg self.blank_id = blank_id self.num_extra_outputs = joint.num_extra_outputs self.big_blank_durations = self.cfg.get("big_blank_durations", None) self.durations = self.cfg.get("durations", None) self.compute_hypothesis_token_set = self.cfg.get("compute_hypothesis_token_set", False) self.compute_langs = decoding_cfg.get('compute_langs', False) self.preserve_alignments = self.cfg.get('preserve_alignments', None) self.joint_fused_batch_size = self.cfg.get('fused_batch_size', None) self.compute_timestamps = self.cfg.get('compute_timestamps', None) self.word_seperator = self.cfg.get('word_seperator', ' ') if self.durations is not None: # this means it's a TDT model. if blank_id == 0: raise ValueError("blank_id must equal len(non_blank_vocabs) for TDT models") if self.big_blank_durations is not None: raise ValueError("duration and big_blank_durations can't both be not None") if self.cfg.strategy not in ['greedy', 'greedy_batch']: raise ValueError("currently only greedy and greedy_batch inference is supported for TDT models") if self.big_blank_durations is not None: # this means it's a multi-blank model. if blank_id == 0: raise ValueError("blank_id must equal len(vocabs) for multi-blank RNN-T models") if self.cfg.strategy not in ['greedy', 'greedy_batch']: raise ValueError( "currently only greedy and greedy_batch inference is supported for multi-blank models" ) possible_strategies = ['greedy', 'greedy_batch', 'beam', 'tsd', 'alsd', 'maes'] if self.cfg.strategy not in possible_strategies: raise ValueError(f"Decoding strategy must be one of {possible_strategies}") # Update preserve alignments if self.preserve_alignments is None: if self.cfg.strategy in ['greedy', 'greedy_batch']: self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) # Update compute timestamps if self.compute_timestamps is None: if self.cfg.strategy in ['greedy', 'greedy_batch']: self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) # Test if alignments are being preserved for RNNT if self.compute_timestamps is True and self.preserve_alignments is False: raise ValueError("If `compute_timesteps` flag is set, then `preserve_alignments` flag must also be set.") # initialize confidence-related fields self._init_confidence(self.cfg.get('confidence_cfg', None)) # Confidence estimation is not implemented for these strategies if ( not self.preserve_frame_confidence and self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes'] and self.cfg.beam.get('preserve_frame_confidence', False) ): raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") if self.cfg.strategy == 'greedy': if self.big_blank_durations is None: if self.durations is None: self.decoding = greedy_decode.GreedyRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_method_cfg=self.confidence_method_cfg, ) else: self.decoding = greedy_decode.GreedyTDTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, durations=self.durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_method_cfg=self.confidence_method_cfg, ) else: self.decoding = greedy_decode.GreedyMultiblankRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, big_blank_durations=self.big_blank_durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_method_cfg=self.confidence_method_cfg, ) elif self.cfg.strategy == 'greedy_batch': if self.big_blank_durations is None: if self.durations is None: self.decoding = greedy_decode.GreedyBatchedRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_method_cfg=self.confidence_method_cfg, ) else: self.decoding = greedy_decode.GreedyBatchedTDTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, durations=self.durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_method_cfg=self.confidence_method_cfg, ) else: self.decoding = greedy_decode.GreedyBatchedMultiblankRNNTInfer( decoder_model=decoder, joint_model=joint, blank_index=self.blank_id, big_blank_durations=self.big_blank_durations, max_symbols_per_step=( self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) ), preserve_alignments=self.preserve_alignments, preserve_frame_confidence=self.preserve_frame_confidence, confidence_method_cfg=self.confidence_method_cfg, ) elif self.cfg.strategy == 'beam': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='default', score_norm=self.cfg.beam.get('score_norm', True), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ) elif self.cfg.strategy == 'tsd': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='tsd', score_norm=self.cfg.beam.get('score_norm', True), tsd_max_sym_exp_per_step=self.cfg.beam.get('tsd_max_sym_exp', 10), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ) elif self.cfg.strategy == 'alsd': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='alsd', score_norm=self.cfg.beam.get('score_norm', True), alsd_max_target_len=self.cfg.beam.get('alsd_max_target_len', 2), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ) elif self.cfg.strategy == 'maes': self.decoding = beam_decode.BeamRNNTInfer( decoder_model=decoder, joint_model=joint, beam_size=self.cfg.beam.beam_size, return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), search_type='maes', score_norm=self.cfg.beam.get('score_norm', True), maes_num_steps=self.cfg.beam.get('maes_num_steps', 2), maes_prefix_alpha=self.cfg.beam.get('maes_prefix_alpha', 1), maes_expansion_gamma=self.cfg.beam.get('maes_expansion_gamma', 2.3), maes_expansion_beta=self.cfg.beam.get('maes_expansion_beta', 2.0), softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), preserve_alignments=self.preserve_alignments, ngram_lm_model=self.cfg.beam.get('ngram_lm_model', None), ngram_lm_alpha=self.cfg.beam.get('ngram_lm_alpha', 0.0), hat_subtract_ilm=self.cfg.beam.get('hat_subtract_ilm', False), hat_ilm_weight=self.cfg.beam.get('hat_ilm_weight', 0.0), ) else: raise ValueError( f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" f"but was provided {self.cfg.strategy}" ) # Update the joint fused batch size or disable it entirely if needed. self.update_joint_fused_batch_size() def rnnt_decoder_predictions_tensor( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, return_hypotheses: bool = False, partial_hypotheses: Optional[List[Hypothesis]] = None, ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: """ Decode an encoder output by autoregressive decoding of the Decoder+Joint networks. Args: encoder_output: torch.Tensor of shape [B, D, T]. encoded_lengths: torch.Tensor containing lengths of the padded encoder outputs. Shape [B]. return_hypotheses: bool. If set to True it will return list of Hypothesis or NBestHypotheses Returns: If `return_best_hypothesis` is set: A tuple (hypotheses, None): hypotheses - list of Hypothesis (best hypothesis per sample). Look at rnnt_utils.Hypothesis for more information. If `return_best_hypothesis` is not set: A tuple(hypotheses, all_hypotheses) hypotheses - list of Hypothesis (best hypothesis per sample). Look at rnnt_utils.Hypothesis for more information. all_hypotheses - list of NBestHypotheses. Each NBestHypotheses further contains a sorted list of all the hypotheses of the model per sample. Look at rnnt_utils.NBestHypotheses for more information. """ # Compute hypotheses with torch.inference_mode(): hypotheses_list = self.decoding( encoder_output=encoder_output, encoded_lengths=encoded_lengths, partial_hypotheses=partial_hypotheses ) # type: [List[Hypothesis]] # extract the hypotheses hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] prediction_list = hypotheses_list if isinstance(prediction_list[0], NBestHypotheses): hypotheses = [] all_hypotheses = [] for nbest_hyp in prediction_list: # type: NBestHypotheses n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample decoded_hyps = self.decode_hypothesis(n_hyps) # type: List[str] # If computing timestamps if self.compute_timestamps is True: timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') for hyp_idx in range(len(decoded_hyps)): decoded_hyps[hyp_idx] = self.compute_rnnt_timestamps(decoded_hyps[hyp_idx], timestamp_type) hypotheses.append(decoded_hyps[0]) # best hypothesis all_hypotheses.append(decoded_hyps) if return_hypotheses: return hypotheses, all_hypotheses best_hyp_text = [h.text for h in hypotheses] all_hyp_text = [h.text for hh in all_hypotheses for h in hh] return best_hyp_text, all_hyp_text else: hypotheses = self.decode_hypothesis(prediction_list) # type: List[str] # If computing timestamps if self.compute_timestamps is True: timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') for hyp_idx in range(len(hypotheses)): hypotheses[hyp_idx] = self.compute_rnnt_timestamps(hypotheses[hyp_idx], timestamp_type) if return_hypotheses: # greedy decoding, can get high-level confidence scores if self.preserve_frame_confidence and ( self.preserve_word_confidence or self.preserve_token_confidence ): hypotheses = self.compute_confidence(hypotheses) return hypotheses, None best_hyp_text = [h.text for h in hypotheses] return best_hyp_text, None def decode_hypothesis(self, hypotheses_list: List[Hypothesis]) -> List[Union[Hypothesis, NBestHypotheses]]: """ Decode a list of hypotheses into a list of strings. Args: hypotheses_list: List of Hypothesis. Returns: A list of strings. """ for ind in range(len(hypotheses_list)): # Extract the integer encoded hypothesis prediction = hypotheses_list[ind].y_sequence if type(prediction) != list: prediction = prediction.tolist() # RNN-T sample level is already preprocessed by implicit RNNT decoding # Simply remove any blank and possibly big blank tokens if self.big_blank_durations is not None: # multi-blank RNNT num_extra_outputs = len(self.big_blank_durations) prediction = [p for p in prediction if p < self.blank_id - num_extra_outputs] elif self.durations is not None: # TDT model. prediction = [p for p in prediction if p < self.blank_id] else: # standard RNN-T prediction = [p for p in prediction if p != self.blank_id] # De-tokenize the integer tokens; if not computing timestamps if self.compute_timestamps is True: # keep the original predictions, wrap with the number of repetitions per token and alignments # this is done so that `rnnt_decoder_predictions_tensor()` can process this hypothesis # in order to compute exact time stamps. alignments = copy.deepcopy(hypotheses_list[ind].alignments) token_repetitions = [1] * len(alignments) # preserve number of repetitions per token hypothesis = (prediction, alignments, token_repetitions) else: hypothesis = self.decode_tokens_to_str(prediction) # TODO: remove # collapse leading spaces before . , ? for PC models hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) if self.compute_hypothesis_token_set: hypotheses_list[ind].tokens = self.decode_ids_to_tokens(prediction) # De-tokenize the integer tokens hypotheses_list[ind].text = hypothesis return hypotheses_list def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: """ Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. Assumes that `frame_confidence` is present in the hypotheses. Args: hypotheses_list: List of Hypothesis. Returns: A list of hypotheses with high-level confidence scores. """ if self.exclude_blank_from_confidence: for hyp in hypotheses_list: hyp.token_confidence = hyp.non_blank_frame_confidence else: for hyp in hypotheses_list: offset = 0 token_confidence = [] if len(hyp.timestep) > 0: for ts, te in zip(hyp.timestep, hyp.timestep[1:] + [len(hyp.frame_confidence)]): if ts != te: # <blank> tokens are considered to belong to the last non-blank token, if any. token_confidence.append( self._aggregate_confidence( [hyp.frame_confidence[ts][offset]] + [fc[0] for fc in hyp.frame_confidence[ts + 1 : te]] ) ) offset = 0 else: token_confidence.append(hyp.frame_confidence[ts][offset]) offset += 1 hyp.token_confidence = token_confidence if self.preserve_word_confidence: for hyp in hypotheses_list: hyp.word_confidence = self._aggregate_token_confidence(hyp) return hypotheses_list @abstractmethod def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token id list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ raise NotImplementedError() @abstractmethod def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ raise NotImplementedError() @abstractmethod def decode_tokens_to_lang(self, tokens: List[int]) -> str: """ Implemented by subclass in order to compute the most likely language ID (LID) string given the tokens. Args: tokens: List of int representing the token ids. Returns: A decoded LID string. """ raise NotImplementedError() @abstractmethod def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into language ID (LID) list. Args: tokens: List of int representing the token ids. Returns: A list of decoded LIDS. """ raise NotImplementedError() def update_joint_fused_batch_size(self): if self.joint_fused_batch_size is None: # do nothing and let the Joint itself handle setting up of the fused batch return if not hasattr(self.decoding.joint, 'set_fused_batch_size'): logging.warning( "The joint module does not have `set_fused_batch_size(int)` as a setter function.\n" "Ignoring update of joint fused batch size." ) return if not hasattr(self.decoding.joint, 'set_fuse_loss_wer'): logging.warning( "The joint module does not have `set_fuse_loss_wer(bool, RNNTLoss, RNNTWER)` " "as a setter function.\n" "Ignoring update of joint fused batch size." ) return if self.joint_fused_batch_size > 0: self.decoding.joint.set_fused_batch_size(self.joint_fused_batch_size) else: logging.info("Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.") self.decoding.joint.set_fuse_loss_wer(False) def compute_rnnt_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): assert timestamp_type in ['char', 'word', 'all'] # Unpack the temporary storage decoded_prediction, alignments, token_repetitions = hypothesis.text # Retrieve offsets char_offsets = word_offsets = None char_offsets = self._compute_offsets(hypothesis, token_repetitions, self.blank_id) # finally, set the flattened decoded predictions to text field for later text decoding hypothesis.text = decoded_prediction # Assert number of offsets and hypothesis tokens are 1:1 match. num_flattened_tokens = 0 for t in range(len(char_offsets)): # Subtract one here for the extra RNNT BLANK token emitted to designate "End of timestep" num_flattened_tokens += len(char_offsets[t]['char']) - 1 if num_flattened_tokens != len(hypothesis.text): raise ValueError( f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" " have to be of the same length, but are: " f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" f" {len(hypothesis.text)}" ) encoded_char_offsets = copy.deepcopy(char_offsets) # Correctly process the token ids to chars/subwords. for i, offsets in enumerate(char_offsets): decoded_chars = [] for char in offsets['char'][:-1]: # ignore the RNNT Blank token at end of every timestep with -1 subset decoded_chars.append(self.decode_tokens_to_str([int(char)])) char_offsets[i]["char"] = decoded_chars # detect char vs subword models lens = [] for v in char_offsets: tokens = v["char"] # each token may be either 1 unicode token or multiple unicode token # for character based models, only 1 token is used # for subword, more than one token can be used. # Computing max, then summing up total lens is a test to check for char vs subword # For char models, len(lens) == sum(lens) # but this is violated for subword models. max_len = max(len(c) for c in tokens) lens.append(max_len) # array of one or more chars implies subword based model with multiple char emitted per TxU step (via subword) if sum(lens) > len(lens): text_type = 'subword' else: # full array of ones implies character based model with 1 char emitted per TxU step text_type = 'char' # retrieve word offsets from character offsets word_offsets = None if timestamp_type in ['word', 'all']: if text_type == 'char': word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) else: # utilize the copy of char offsets with the correct integer ids for tokens # so as to avoid tokenize -> detokenize -> compare -> merge steps. word_offsets = self._get_word_offsets_subwords_sentencepiece( encoded_char_offsets, hypothesis, decode_ids_to_tokens=self.decode_ids_to_tokens, decode_tokens_to_str=self.decode_tokens_to_str, ) # attach results if len(hypothesis.timestep) > 0: timestep_info = hypothesis.timestep else: timestep_info = [] # Setup defaults hypothesis.timestep = {"timestep": timestep_info} # Add char / subword time stamps if char_offsets is not None and timestamp_type in ['char', 'all']: hypothesis.timestep['char'] = char_offsets # Add word time stamps if word_offsets is not None and timestamp_type in ['word', 'all']: hypothesis.timestep['word'] = word_offsets # Convert the flattened token indices to text hypothesis.text = self.decode_tokens_to_str(hypothesis.text) return hypothesis @staticmethod def _compute_offsets( hypothesis: Hypothesis, token_repetitions: List[int], rnnt_token: int ) -> List[Dict[str, Union[str, int]]]: """ Utility method that calculates the indidual time indices where a token starts and ends. Args: hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token emitted at every time step after rnnt collapse. token_repetitions: A list of ints representing the number of repetitions of each emitted token. rnnt_token: The integer of the rnnt blank token used during rnnt collapse. Returns: """ start_index = 0 # If the exact timestep information is available, utilize the 1st non-rnnt blank token timestep # as the start index. if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: start_index = max(0, hypothesis.timestep[0] - 1) # Construct the start and end indices brackets end_indices = np.asarray(token_repetitions).cumsum() start_indices = np.concatenate(([start_index], end_indices[:-1])) # Process the TxU dangling alignment tensor, containing pairs of (logits, label) alignment_labels = [al_logits_labels for al_logits_labels in hypothesis.text[1]] for t in range(len(alignment_labels)): for u in range(len(alignment_labels[t])): alignment_labels[t][u] = alignment_labels[t][u][1] # pick label from (logit, label) tuple # Merge the results per token into a list of dictionaries offsets = [ {"char": a, "start_offset": s, "end_offset": e} for a, s, e in zip(alignment_labels, start_indices, end_indices) ] # Filter out RNNT token (blank at [t][0] position). This is because blank can only occur at end of a # time step for RNNT, so if 0th token is blank, then that timestep is skipped. offsets = list(filter(lambda offsets: offsets["char"][0] != rnnt_token, offsets)) return offsets @staticmethod def _get_word_offsets_chars( offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of character time stamps. References: This code is a port of the Hugging Face code for word time stamp construction. Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". word_delimiter_char: Character token that represents the word delimiter. By default, " ". Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] last_state = "SPACE" word = "" start_offset = 0 end_offset = 0 for i, offset in enumerate(offsets): chars = offset["char"] for char in chars: state = "SPACE" if char == word_delimiter_char else "WORD" if state == last_state: # If we are in the same state as before, we simply repeat what we've done before end_offset = offset["end_offset"] word += char else: # Switching state if state == "SPACE": # Finishing a word word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) else: # Starting a new word start_offset = offset["start_offset"] end_offset = offset["end_offset"] word = char last_state = state if last_state == "WORD": word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) return word_offsets @staticmethod def _get_word_offsets_subwords_sentencepiece( offsets: Dict[str, Union[str, float]], hypothesis: Hypothesis, decode_ids_to_tokens: Callable[[List[int]], str], decode_tokens_to_str: Callable[[List[int]], str], ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of sub-word time stamps. **Note**: Only supports Sentencepiece based tokenizers ! Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id after rnnt collapse. decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] built_token = [] previous_token_index = 0 # For every offset token for i, offset in enumerate(offsets): # For every subword token in offset token list (ignoring the RNNT Blank token at the end) for char in offset['char'][:-1]: char = int(char) # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). token = decode_ids_to_tokens([char])[0] token_text = decode_tokens_to_str([char]) # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped # after forcing partial text conversion of the token. if token != token_text: # If there are any partially or fully built sub-word token ids, construct to text. # Note: This is "old" subword, that occurs *after* current sub-word has started. if built_token: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[previous_token_index]["start_offset"], "end_offset": offsets[i]["start_offset"], } ) # Prepare list of new sub-word ids built_token.clear() built_token.append(char) previous_token_index = i else: # If the token does not contain any sub-word start mark, then the sub-word has not completed yet # Append to current sub-word list. built_token.append(char) # Inject the start offset of the first token to word offsets # This is because we always skip the delay the injection of the first sub-word due to the loop # condition and check whether built token is ready or not. # Therefore without this forced injection, the start_offset appears as off by 1. # This should only be done when these arrays contain more than one element. if offsets and word_offsets: word_offsets[0]["start_offset"] = offsets[0]["start_offset"] # If there are any remaining tokens left, inject them all into the final word offset. # The start offset of this token is the start time of the next token to process. # The end offset of this token is the end time of the last token from offsets. # Note that built_token is a flat list; but offsets contains a nested list which # may have different dimensionality. # As such, we can't rely on the length of the list of built_token to index offsets. if built_token: # start from the previous token index as this hasn't been committed to word_offsets yet # if we still have content in built_token start_offset = offsets[previous_token_index]["start_offset"] word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": start_offset, "end_offset": offsets[-1]["end_offset"], } ) built_token.clear() return word_offsets class RNNTDecoding(AbstractRNNTDecoding): """ Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy, greedy_batch (for greedy decoding). - beam, tsd, alsd (for beam search decoding). compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded tokens as well as the decoded string. Default is False in order to avoid double decoding unless required. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function with the `return_hypotheses` flag set to True. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. The config may further contain the following sub-dictionaries: "greedy": max_symbols: int, describing the maximum number of target tokens to decode per timestep during greedy decoding. Setting to larger values allows longer sentences to be decoded, at the cost of increased execution time. preserve_frame_confidence: Same as above, overrides above value. confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. Set to True by default. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, at increased cost to execution time. alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. If an integer is provided, it can decode sequences of that particular maximum length. If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), where seq_len is the length of the acoustic model output (T). NOTE: If a float is provided, it can be greater than 1! By default, a float of 2.0 is used so that a target sequence can be at most twice as long as the acoustic model output length T. maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 in order to reduce expensive beam search cost later. int >= 0. maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, and affects the speed of inference since large values will perform large beam search in the next step. maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for expansion apart from the "most likely" candidate. Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally tuned on a validation set. softmax_temperature: Scales the logits of the joint prior to computing log_softmax. decoder: The Decoder/Prediction network module. joint: The Joint network module. vocabulary: The vocabulary (excluding the RNNT blank token) which will be used for decoding. """ def __init__( self, decoding_cfg, decoder, joint, vocabulary, ): # we need to ensure blank is the last token in the vocab for the case of RNNT and Multi-blank RNNT. blank_id = len(vocabulary) + joint.num_extra_outputs if hasattr(decoding_cfg, 'model_type') and decoding_cfg.model_type == 'tdt': blank_id = len(vocabulary) self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) super(RNNTDecoding, self).__init__( decoding_cfg=decoding_cfg, decoder=decoder, joint=joint, blank_id=blank_id, ) if isinstance(self.decoding, beam_decode.BeamRNNTInfer): self.decoding.set_decoding_type('char') def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: """ Implemented by subclass in order to aggregate token confidence to a word-level confidence. Args: hypothesis: Hypothesis Returns: A list of word-level confidence scores. """ return self._aggregate_token_confidence_chars(hypothesis.words, hypothesis.token_confidence) def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) return hypothesis def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ token_list = [self.labels_map[c] for c in tokens if c < self.blank_id - self.num_extra_outputs] return token_list def decode_tokens_to_lang(self, tokens: List[int]) -> str: """ Compute the most likely language ID (LID) string given the tokens. Args: tokens: List of int representing the token ids. Returns: A decoded LID string. """ lang = self.tokenizer.ids_to_lang(tokens) return lang def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: """ Decode a token id list into language ID (LID) list. Args: tokens: List of int representing the token ids. Returns: A list of decoded LIDS. """ lang_list = self.tokenizer.ids_to_text_and_langs(tokens) return lang_list class RNNTWER(Metric): """ This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. When doing distributed training/evaluation the result of res=WER(predictions, targets, target_lengths) calls will be all-reduced between all workers using SUM operations. Here contains two numbers res=[wer_numerator, wer_denominator]. WER=wer_numerator/wer_denominator. If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. Example: def validation_step(self, batch, batch_idx): ... wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} return self.val_outputs def on_validation_epoch_end(self): ... wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} self.val_outputs.clear() # free memory return {'val_loss': val_loss_mean, 'log': tensorboard_logs} Args: decoding: RNNTDecoding object that will perform autoregressive decoding of the RNNT model. batch_dim_index: Index of the batch dimension. use_cer: Whether to use Character Error Rate isntead of Word Error Rate. log_prediction: Whether to log a single decoded sample per call. Returns: res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenshtein's distances for all prediction - reference pairs, total number of words in all references. """ full_state_update = True def __init__( self, decoding: RNNTDecoding, batch_dim_index=0, use_cer=False, log_prediction=True, dist_sync_on_step=False ): super(RNNTWER, self).__init__(dist_sync_on_step=dist_sync_on_step) self.decoding = decoding self.batch_dim_index = batch_dim_index self.use_cer = use_cer self.log_prediction = log_prediction self.blank_id = self.decoding.blank_id self.labels_map = self.decoding.labels_map self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) def update( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, targets: torch.Tensor, target_lengths: torch.Tensor, ) -> torch.Tensor: words = 0 scores = 0 references = [] with torch.no_grad(): # prediction_cpu_tensor = tensors[0].long().cpu() targets_cpu_tensor = targets.long().cpu() targets_cpu_tensor = move_dimension_to_the_front(targets_cpu_tensor, self.batch_dim_index) tgt_lenths_cpu_tensor = target_lengths.long().cpu() # iterate over batch for ind in range(targets_cpu_tensor.shape[0]): tgt_len = tgt_lenths_cpu_tensor[ind].item() target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() reference = self.decoding.decode_tokens_to_str(target) references.append(reference) hypotheses, _ = self.decoding.rnnt_decoder_predictions_tensor(encoder_output, encoded_lengths) if self.log_prediction: logging.info(f"\n") logging.info(f"reference :{references[0]}") logging.info(f"predicted :{hypotheses[0]}") for h, r in zip(hypotheses, references): if self.use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() words += len(r_list) # Compute Levenshtein's distance scores += editdistance.eval(h_list, r_list) self.scores += torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) self.words += torch.tensor(words, device=self.words.device, dtype=self.words.dtype) # return torch.tensor([scores, words]).to(predictions.device) def compute(self): wer = self.scores.float() / self.words return wer, self.scores.detach(), self.words.detach() @dataclass class RNNTDecodingConfig: model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" strategy: str = "greedy_batch" compute_hypothesis_token_set: bool = False # preserve decoding alignments preserve_alignments: Optional[bool] = None # confidence config confidence_cfg: ConfidenceConfig = ConfidenceConfig() # RNNT Joint fused batch size fused_batch_size: Optional[int] = None # compute RNNT time stamps compute_timestamps: Optional[bool] = None # compute language IDs compute_langs: bool = False # token representing word seperator word_seperator: str = " " # type of timestamps to calculate rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() # beam decoding config beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) # can be used to change temperature for decoding temperature: float = 1.0 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/metrics/wer.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import re from abc import abstractmethod from dataclasses import dataclass, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance import jiwer import numpy as np import torch from omegaconf import DictConfig, OmegaConf from torchmetrics import Metric from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses from nemo.utils import logging, logging_mode __all__ = ['word_error_rate', 'word_error_rate_detail', 'WER', 'move_dimension_to_the_front'] def word_error_rate(hypotheses: List[str], references: List[str], use_cer=False) -> float: """ Computes Average Word Error rate between two texts represented as corresponding lists of string. Hypotheses and references must have same length. Args: hypotheses (list): list of hypotheses references(list) : list of references use_cer (bool): set True to enable cer Returns: wer (float): average word error rate """ scores = 0 words = 0 if len(hypotheses) != len(references): raise ValueError( "In word error rate calculation, hypotheses and reference" " lists must have the same number of elements. But I got:" "{0} and {1} correspondingly".format(len(hypotheses), len(references)) ) for h, r in zip(hypotheses, references): if use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() words += len(r_list) # May deprecate using editdistance in future release for here and rest of codebase # once we confirm jiwer is reliable. scores += editdistance.eval(h_list, r_list) if words != 0: wer = 1.0 * scores / words else: wer = float('inf') return wer def word_error_rate_detail( hypotheses: List[str], references: List[str], use_cer=False ) -> Tuple[float, int, float, float, float]: """ Computes Average Word Error Rate with details (insertion rate, deletion rate, substitution rate) between two texts represented as corresponding lists of string. Hypotheses and references must have same length. Args: hypotheses (list): list of hypotheses references(list) : list of references use_cer (bool): set True to enable cer Returns: wer (float): average word error rate words (int): Total number of words/charactors of given reference texts ins_rate (float): average insertion error rate del_rate (float): average deletion error rate sub_rate (float): average substitution error rate """ scores = 0 words = 0 ops_count = {'substitutions': 0, 'insertions': 0, 'deletions': 0} if len(hypotheses) != len(references): raise ValueError( "In word error rate calculation, hypotheses and reference" " lists must have the same number of elements. But I got:" "{0} and {1} correspondingly".format(len(hypotheses), len(references)) ) for h, r in zip(hypotheses, references): if use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() # To get rid of the issue that jiwer does not allow empty string if len(r_list) == 0: if len(h_list) != 0: errors = len(h_list) ops_count['insertions'] += errors else: errors = 0 else: if use_cer: measures = jiwer.cer(r, h, return_dict=True) else: measures = jiwer.compute_measures(r, h) errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] ops_count['insertions'] += measures['insertions'] ops_count['deletions'] += measures['deletions'] ops_count['substitutions'] += measures['substitutions'] scores += errors words += len(r_list) if words != 0: wer = 1.0 * scores / words ins_rate = 1.0 * ops_count['insertions'] / words del_rate = 1.0 * ops_count['deletions'] / words sub_rate = 1.0 * ops_count['substitutions'] / words else: wer, ins_rate, del_rate, sub_rate = float('inf'), float('inf'), float('inf'), float('inf') return wer, words, ins_rate, del_rate, sub_rate def word_error_rate_per_utt(hypotheses: List[str], references: List[str], use_cer=False) -> Tuple[List[float], float]: """ Computes Word Error Rate per utterance and the average WER between two texts represented as corresponding lists of string. Hypotheses and references must have same length. Args: hypotheses (list): list of hypotheses references(list) : list of references use_cer (bool): set True to enable cer Returns: wer_per_utt (List[float]): word error rate per utterance avg_wer (float): average word error rate """ scores = 0 words = 0 wer_per_utt = [] if len(hypotheses) != len(references): raise ValueError( "In word error rate calculation, hypotheses and reference" " lists must have the same number of elements. But I got:" "{0} and {1} correspondingly".format(len(hypotheses), len(references)) ) for h, r in zip(hypotheses, references): if use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() # To get rid of the issue that jiwer does not allow empty string if len(r_list) == 0: if len(h_list) != 0: errors = len(h_list) wer_per_utt.append(float('inf')) else: if use_cer: measures = jiwer.cer(r, h, return_dict=True) er = measures['cer'] else: measures = jiwer.compute_measures(r, h) er = measures['wer'] errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] wer_per_utt.append(er) scores += errors words += len(r_list) if words != 0: avg_wer = 1.0 * scores / words else: avg_wer = float('inf') return wer_per_utt, avg_wer def move_dimension_to_the_front(tensor, dim_index): all_dims = list(range(tensor.ndim)) return tensor.permute(*([dim_index] + all_dims[:dim_index] + all_dims[dim_index + 1 :])) class AbstractCTCDecoding(ConfidenceMixin): """ Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy (for greedy decoding). - beam (for DeepSpeed KenLM based decoding). compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. Can take the following values - "char" for character/subword time stamps, "word" for word level time stamps and "all" (default), for both character level and word level time stamps. word_seperator: Str token representing the seperator between words. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. The config may further contain the following sub-dictionaries: "greedy": preserve_alignments: Same as above, overrides above value. compute_timestamps: Same as above, overrides above value. preserve_frame_confidence: Same as above, overrides above value. confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. beam_alpha: float, the strength of the Language model on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. beam_beta: float, the strength of the sequence length penalty on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). If the path is invalid (file is not found at path), will raise a deferred error at the moment of calculation of beam search, so that users may update / change the decoding strategy to point to the correct file. blank_id: The id of the RNNT blank token. """ def __init__(self, decoding_cfg, blank_id: int): super().__init__() # Convert dataclas to config if is_dataclass(decoding_cfg): decoding_cfg = OmegaConf.structured(decoding_cfg) if not isinstance(decoding_cfg, DictConfig): decoding_cfg = OmegaConf.create(decoding_cfg) OmegaConf.set_struct(decoding_cfg, False) # update minimal config minimal_cfg = ['greedy'] for item in minimal_cfg: if item not in decoding_cfg: decoding_cfg[item] = OmegaConf.create({}) self.cfg = decoding_cfg self.blank_id = blank_id self.preserve_alignments = self.cfg.get('preserve_alignments', None) self.compute_timestamps = self.cfg.get('compute_timestamps', None) self.batch_dim_index = self.cfg.get('batch_dim_index', 0) self.word_seperator = self.cfg.get('word_seperator', ' ') possible_strategies = ['greedy', 'beam', 'pyctcdecode', 'flashlight'] if self.cfg.strategy not in possible_strategies: raise ValueError(f"Decoding strategy must be one of {possible_strategies}. Given {self.cfg.strategy}") # Update preserve alignments if self.preserve_alignments is None: if self.cfg.strategy in ['greedy']: self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) else: self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) # Update compute timestamps if self.compute_timestamps is None: if self.cfg.strategy in ['greedy']: self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) elif self.cfg.strategy in ['beam']: self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) # initialize confidence-related fields self._init_confidence(self.cfg.get('confidence_cfg', None)) # Confidence estimation is not implemented for strategies other than `greedy` if ( not self.preserve_frame_confidence and self.cfg.strategy != 'greedy' and self.cfg.beam.get('preserve_frame_confidence', False) ): raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") # we need timestamps to extract non-blank per-frame confidence if self.compute_timestamps is not None: self.compute_timestamps |= self.preserve_frame_confidence if self.cfg.strategy == 'greedy': self.decoding = ctc_greedy_decoding.GreedyCTCInfer( blank_id=self.blank_id, preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, preserve_frame_confidence=self.preserve_frame_confidence, confidence_method_cfg=self.confidence_method_cfg, ) elif self.cfg.strategy == 'beam': self.decoding = ctc_beam_decoding.BeamCTCInfer( blank_id=blank_id, beam_size=self.cfg.beam.get('beam_size', 1), search_type='default', return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), beam_beta=self.cfg.beam.get('beam_beta', 0.0), kenlm_path=self.cfg.beam.get('kenlm_path', None), ) self.decoding.override_fold_consecutive_value = False elif self.cfg.strategy == 'pyctcdecode': self.decoding = ctc_beam_decoding.BeamCTCInfer( blank_id=blank_id, beam_size=self.cfg.beam.get('beam_size', 1), search_type='pyctcdecode', return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), beam_beta=self.cfg.beam.get('beam_beta', 0.0), kenlm_path=self.cfg.beam.get('kenlm_path', None), pyctcdecode_cfg=self.cfg.beam.get('pyctcdecode_cfg', None), ) self.decoding.override_fold_consecutive_value = False elif self.cfg.strategy == 'flashlight': self.decoding = ctc_beam_decoding.BeamCTCInfer( blank_id=blank_id, beam_size=self.cfg.beam.get('beam_size', 1), search_type='flashlight', return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), preserve_alignments=self.preserve_alignments, compute_timestamps=self.compute_timestamps, beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), beam_beta=self.cfg.beam.get('beam_beta', 0.0), kenlm_path=self.cfg.beam.get('kenlm_path', None), flashlight_cfg=self.cfg.beam.get('flashlight_cfg', None), ) self.decoding.override_fold_consecutive_value = False else: raise ValueError( f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" f"but was provided {self.cfg.strategy}" ) def ctc_decoder_predictions_tensor( self, decoder_outputs: torch.Tensor, decoder_lengths: torch.Tensor = None, fold_consecutive: bool = True, return_hypotheses: bool = False, ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: """ Decodes a sequence of labels to words Args: decoder_outputs: An integer torch.Tensor of shape [Batch, Time, {Vocabulary}] (if ``batch_index_dim == 0``) or [Time, Batch] (if ``batch_index_dim == 1``) of integer indices that correspond to the index of some character in the label set. decoder_lengths: Optional tensor of length `Batch` which contains the integer lengths of the sequence in the padded `predictions` tensor. fold_consecutive: Bool, determine whether to perform "ctc collapse", folding consecutive tokens into a single token. return_hypotheses: Bool flag whether to return just the decoding predictions of the model or a Hypothesis object that holds information such as the decoded `text`, the `alignment` of emited by the CTC Model, and the `length` of the sequence (if available). May also contain the log-probabilities of the decoder (if this method is called via transcribe()) Returns: Either a list of str which represent the CTC decoded strings per sample, or a list of Hypothesis objects containing additional information. """ if isinstance(decoder_outputs, torch.Tensor): decoder_outputs = move_dimension_to_the_front(decoder_outputs, self.batch_dim_index) if ( hasattr(self.decoding, 'override_fold_consecutive_value') and self.decoding.override_fold_consecutive_value is not None ): logging.info( f"Beam search requires that consecutive ctc tokens are not folded. \n" f"Overriding provided value of `fold_consecutive` = {fold_consecutive} to " f"{self.decoding.override_fold_consecutive_value}", mode=logging_mode.ONCE, ) fold_consecutive = self.decoding.override_fold_consecutive_value with torch.inference_mode(): # Resolve the forward step of the decoding strategy hypotheses_list = self.decoding( decoder_output=decoder_outputs, decoder_lengths=decoder_lengths ) # type: List[List[Hypothesis]] # extract the hypotheses hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] if isinstance(hypotheses_list[0], NBestHypotheses): hypotheses = [] all_hypotheses = [] for nbest_hyp in hypotheses_list: # type: NBestHypotheses n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample decoded_hyps = self.decode_hypothesis( n_hyps, fold_consecutive ) # type: List[Union[Hypothesis, NBestHypotheses]] # If computing timestamps if self.compute_timestamps is True: timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') for hyp_idx in range(len(decoded_hyps)): decoded_hyps[hyp_idx] = self.compute_ctc_timestamps(decoded_hyps[hyp_idx], timestamp_type) hypotheses.append(decoded_hyps[0]) # best hypothesis all_hypotheses.append(decoded_hyps) if return_hypotheses: return hypotheses, all_hypotheses best_hyp_text = [h.text for h in hypotheses] all_hyp_text = [h.text for hh in all_hypotheses for h in hh] return best_hyp_text, all_hyp_text else: hypotheses = self.decode_hypothesis( hypotheses_list, fold_consecutive ) # type: List[Union[Hypothesis, NBestHypotheses]] # If computing timestamps if self.compute_timestamps is True: # greedy decoding, can get high-level confidence scores if return_hypotheses and (self.preserve_word_confidence or self.preserve_token_confidence): hypotheses = self.compute_confidence(hypotheses) else: # remove unused token_repetitions from Hypothesis.text for hyp in hypotheses: hyp.text = hyp.text[:2] timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') for hyp_idx in range(len(hypotheses)): hypotheses[hyp_idx] = self.compute_ctc_timestamps(hypotheses[hyp_idx], timestamp_type) if return_hypotheses: return hypotheses, None best_hyp_text = [h.text for h in hypotheses] return best_hyp_text, None def decode_hypothesis( self, hypotheses_list: List[Hypothesis], fold_consecutive: bool ) -> List[Union[Hypothesis, NBestHypotheses]]: """ Decode a list of hypotheses into a list of strings. Args: hypotheses_list: List of Hypothesis. fold_consecutive: Whether to collapse the ctc blank tokens or not. Returns: A list of strings. """ for ind in range(len(hypotheses_list)): # Extract the integer encoded hypothesis hyp = hypotheses_list[ind] prediction = hyp.y_sequence predictions_len = hyp.length if hyp.length > 0 else None if fold_consecutive: if type(prediction) != list: prediction = prediction.numpy().tolist() if predictions_len is not None: prediction = prediction[:predictions_len] # CTC decoding procedure decoded_prediction = [] token_lengths = [] # preserve token lengths token_repetitions = [] # preserve number of repetitions per token previous = self.blank_id last_length = 0 last_repetition = 1 for pidx, p in enumerate(prediction): if (p != previous or previous == self.blank_id) and p != self.blank_id: decoded_prediction.append(p) token_lengths.append(pidx - last_length) last_length = pidx token_repetitions.append(last_repetition) last_repetition = 1 if p == previous and previous != self.blank_id: last_repetition += 1 previous = p if len(token_repetitions) > 0: token_repetitions = token_repetitions[1:] + [last_repetition] else: if predictions_len is not None: prediction = prediction[:predictions_len] decoded_prediction = prediction[prediction != self.blank_id].tolist() token_lengths = [1] * len(decoded_prediction) # preserve number of repetitions per token token_repetitions = [1] * len(decoded_prediction) # preserve number of repetitions per token # De-tokenize the integer tokens; if not computing timestamps if self.compute_timestamps is True: # keep the original predictions, wrap with the number of repetitions per token # this is done so that `ctc_decoder_predictions_tensor()` can process this hypothesis # in order to compute exact time stamps. hypothesis = (decoded_prediction, token_lengths, token_repetitions) else: hypothesis = self.decode_tokens_to_str(decoded_prediction) # TODO: remove # collapse leading spaces before . , ? for PC models hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) # Preserve this wrapped hypothesis or decoded text tokens. hypotheses_list[ind].text = hypothesis return hypotheses_list def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: """ Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. Assumes that `frame_confidence` is present in the hypotheses. Args: hypotheses_list: List of Hypothesis. Returns: A list of hypotheses with high-level confidence scores. """ for hyp in hypotheses_list: if not isinstance(hyp.text, tuple) or len(hyp.text) != 3: # the method must have been called in the wrong place raise ValueError( """Wrong format of the `text` attribute of a hypothesis.\n Expected: (decoded_prediction, token_repetitions)\n The method invocation is expected between .decode_hypothesis() and .compute_ctc_timestamps()""" ) token_repetitions = hyp.text[2] hyp.text = hyp.text[:2] token_confidence = [] if self.exclude_blank_from_confidence: non_blank_frame_confidence = hyp.non_blank_frame_confidence i = 0 for tr in token_repetitions: # token repetition can be zero j = i + tr token_confidence.append(self._aggregate_confidence(non_blank_frame_confidence[i:j])) i = j else: # <blank> tokens are considered to belong to the last non-blank token, if any. token_lengths = hyp.text[1] if len(token_lengths) > 0: ts = token_lengths[0] for tl in token_lengths[1:] + [len(hyp.frame_confidence)]: token_confidence.append(self._aggregate_confidence(hyp.frame_confidence[ts : ts + tl])) ts += tl hyp.token_confidence = token_confidence if self.preserve_word_confidence: for hyp in hypotheses_list: hyp.word_confidence = self._aggregate_token_confidence(hyp) return hypotheses_list @abstractmethod def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token id list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ raise NotImplementedError() @abstractmethod def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ raise NotImplementedError() def compute_ctc_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): """ Method to compute time stamps at char/subword, and word level given some hypothesis. Requires the input hypothesis to contain a `text` field that is the tuple. The tuple contains - the ctc collapsed integer ids, and the number of repetitions of each token. Args: hypothesis: A Hypothesis object, with a wrapped `text` field. The `text` field must contain a tuple with two values - The ctc collapsed integer ids A list of integers that represents the number of repetitions per token. timestamp_type: A str value that represents the type of time stamp calculated. Can be one of "char", "word" or "all" Returns: A Hypothesis object with a modified `timestep` value, which is now a dictionary containing the time stamp information. """ assert timestamp_type in ['char', 'word', 'all'] # Unpack the temporary storage, and set the decoded predictions decoded_prediction, token_lengths = hypothesis.text hypothesis.text = decoded_prediction # Retrieve offsets char_offsets = word_offsets = None char_offsets = self._compute_offsets(hypothesis, token_lengths, self.blank_id) # Assert number of offsets and hypothesis tokens are 1:1 match. if len(char_offsets) != len(hypothesis.text): raise ValueError( f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" " have to be of the same length, but are: " f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" f" {len(hypothesis.text)}" ) # Correctly process the token ids to chars/subwords. for i, char in enumerate(hypothesis.text): char_offsets[i]["char"] = self.decode_tokens_to_str([char]) # detect char vs subword models lens = [len(list(v["char"])) > 1 for v in char_offsets] if any(lens): text_type = 'subword' else: text_type = 'char' # retrieve word offsets from character offsets word_offsets = None if timestamp_type in ['word', 'all']: if text_type == 'char': word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) else: word_offsets = self._get_word_offsets_subwords_sentencepiece( char_offsets, hypothesis, decode_ids_to_tokens=self.decode_ids_to_tokens, decode_tokens_to_str=self.decode_tokens_to_str, ) # attach results if len(hypothesis.timestep) > 0: timestep_info = hypothesis.timestep else: timestep_info = [] # Setup defaults hypothesis.timestep = {"timestep": timestep_info} # Add char / subword time stamps if char_offsets is not None and timestamp_type in ['char', 'all']: hypothesis.timestep['char'] = char_offsets # Add word time stamps if word_offsets is not None and timestamp_type in ['word', 'all']: hypothesis.timestep['word'] = word_offsets # Convert the token indices to text hypothesis.text = self.decode_tokens_to_str(hypothesis.text) return hypothesis @staticmethod def _compute_offsets( hypothesis: Hypothesis, token_lengths: List[int], ctc_token: int ) -> List[Dict[str, Union[str, int]]]: """ Utility method that calculates the indidual time indices where a token starts and ends. Args: hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token emitted at every time step after ctc collapse. token_lengths: A list of ints representing the lengths of each emitted token. ctc_token: The integer of the ctc blank token used during ctc collapse. Returns: """ start_index = 0 # If the exact timestep information is available, utilize the 1st non-ctc blank token timestep # as the start index. if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: start_index = max(0, hypothesis.timestep[0] - 1) # Construct the start and end indices brackets end_indices = np.asarray(token_lengths).cumsum() start_indices = np.concatenate(([start_index], end_indices[:-1])) # Merge the results per token into a list of dictionaries offsets = [ {"char": t, "start_offset": s, "end_offset": e} for t, s, e in zip(hypothesis.text, start_indices, end_indices) ] # Filter out CTC token offsets = list(filter(lambda offsets: offsets["char"] != ctc_token, offsets)) return offsets @staticmethod def _get_word_offsets_chars( offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of character time stamps. References: This code is a port of the Hugging Face code for word time stamp construction. Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". word_delimiter_char: Character token that represents the word delimiter. By default, " ". Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] last_state = "SPACE" word = "" start_offset = 0 end_offset = 0 for i, offset in enumerate(offsets): char = offset["char"] state = "SPACE" if char == word_delimiter_char else "WORD" if state == last_state: # If we are in the same state as before, we simply repeat what we've done before end_offset = offset["end_offset"] word += char else: # Switching state if state == "SPACE": # Finishing a word word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) else: # Starting a new word start_offset = offset["start_offset"] end_offset = offset["end_offset"] word = char last_state = state if last_state == "WORD": word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) return word_offsets @staticmethod def _get_word_offsets_subwords_sentencepiece( offsets: Dict[str, Union[str, float]], hypothesis: Hypothesis, decode_ids_to_tokens: Callable[[List[int]], str], decode_tokens_to_str: Callable[[List[int]], str], ) -> Dict[str, Union[str, float]]: """ Utility method which constructs word time stamps out of sub-word time stamps. **Note**: Only supports Sentencepiece based tokenizers ! Args: offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id after ctc collapse. decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. Returns: A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and "end_offset". """ word_offsets = [] built_token = [] previous_token_index = 0 # For every collapsed sub-word token for i, char in enumerate(hypothesis.text): # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). token = decode_ids_to_tokens([char])[0] token_text = decode_tokens_to_str([char]) # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped # after forcing partial text conversion of the token. if token != token_text: # If there are any partially or fully built sub-word token ids, construct to text. # Note: This is "old" subword, that occurs *after* current sub-word has started. if len(built_token) > 0: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[previous_token_index]["start_offset"], "end_offset": offsets[i]["start_offset"], } ) # Prepare list of new sub-word ids built_token.clear() built_token.append(char) previous_token_index = i else: # If the token does not contain any sub-word start mark, then the sub-word has not completed yet # Append to current sub-word list. built_token.append(char) # Inject the start offset of the first token to word offsets # This is because we always skip the delay the injection of the first sub-word due to the loop # condition and check whether built token is ready or not. # Therefore without this forced injection, the start_offset appears as off by 1. if len(word_offsets) == 0: # alaptev: sometimes word_offsets can be empty if len(built_token) > 0: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[0]["start_offset"], "end_offset": offsets[-1]["end_offset"], } ) built_token.clear() else: word_offsets[0]["start_offset"] = offsets[0]["start_offset"] # If there are any remaining tokens left, inject them all into the final word offset. # Note: The start offset of this token is the start time of the first token inside build_token. # Note: The end offset of this token is the end time of the last token inside build_token if len(built_token) > 0: word_offsets.append( { "word": decode_tokens_to_str(built_token), "start_offset": offsets[-(len(built_token))]["start_offset"], "end_offset": offsets[-1]["end_offset"], } ) built_token.clear() return word_offsets @property def preserve_alignments(self): return self._preserve_alignments @preserve_alignments.setter def preserve_alignments(self, value): self._preserve_alignments = value if hasattr(self, 'decoding'): self.decoding.preserve_alignments = value @property def compute_timestamps(self): return self._compute_timestamps @compute_timestamps.setter def compute_timestamps(self, value): self._compute_timestamps = value if hasattr(self, 'decoding'): self.decoding.compute_timestamps = value @property def preserve_frame_confidence(self): return self._preserve_frame_confidence @preserve_frame_confidence.setter def preserve_frame_confidence(self, value): self._preserve_frame_confidence = value if hasattr(self, 'decoding'): self.decoding.preserve_frame_confidence = value class CTCDecoding(AbstractCTCDecoding): """ Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs for character based models. Args: decoding_cfg: A dict-like object which contains the following key-value pairs. strategy: str value which represents the type of decoding that can occur. Possible values are : - greedy (for greedy decoding). - beam (for DeepSpeed KenLM based decoding). compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. Can take the following values - "char" for character/subword time stamps, "word" for word level time stamps and "all" (default), for both character level and word level time stamps. word_seperator: Str token representing the seperator between words. preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence scores. In order to obtain hypotheses with confidence scores, please utilize `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. The config may further contain the following sub-dictionaries: "greedy": preserve_alignments: Same as above, overrides above value. compute_timestamps: Same as above, overrides above value. preserve_frame_confidence: Same as above, overrides above value. confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. "beam": beam_size: int, defining the beam size for beam search. Must be >= 1. If beam_size == 1, will perform cached greedy search. This might be slightly different results compared to the greedy search above. return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the hypotheses after beam search has concluded. This flag is set by default. beam_alpha: float, the strength of the Language model on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. beam_beta: float, the strength of the sequence length penalty on the final score of a token. final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). If the path is invalid (file is not found at path), will raise a deferred error at the moment of calculation of beam search, so that users may update / change the decoding strategy to point to the correct file. blank_id: The id of the RNNT blank token. """ def __init__( self, decoding_cfg, vocabulary, ): blank_id = len(vocabulary) self.vocabulary = vocabulary self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) super().__init__(decoding_cfg=decoding_cfg, blank_id=blank_id) # Finalize Beam Search Decoding framework if isinstance(self.decoding, ctc_beam_decoding.AbstractBeamCTCInfer): self.decoding.set_vocabulary(self.vocabulary) self.decoding.set_decoding_type('char') def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: """ Implemented by subclass in order to aggregate token confidence to a word-level confidence. Args: hypothesis: Hypothesis Returns: A list of word-level confidence scores. """ return self._aggregate_token_confidence_chars( self.decode_tokens_to_str(hypothesis.text[0]).split(), hypothesis.token_confidence ) def decode_tokens_to_str(self, tokens: List[int]) -> str: """ Implemented by subclass in order to decoder a token list into a string. Args: tokens: List of int representing the token ids. Returns: A decoded string. """ hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) return hypothesis def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: """ Implemented by subclass in order to decode a token id list into a token list. A token list is the string representation of each token id. Args: tokens: List of int representing the token ids. Returns: A list of decoded tokens. """ token_list = [self.labels_map[c] for c in tokens if c != self.blank_id] return token_list class WER(Metric): """ This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. When doing distributed training/evaluation the result of ``res=WER(predictions, targets, target_lengths)`` calls will be all-reduced between all workers using SUM operations. Here ``res`` contains three numbers ``res=[wer, total_levenstein_distance, total_number_of_words]``. If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. Example: def validation_step(self, batch, batch_idx): ... wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} return self.val_outputs def on_validation_epoch_end(self): ... wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} self.val_outputs.clear() # free memory return {'val_loss': val_loss_mean, 'log': tensorboard_logs} Args: decoding: An instance of CTCDecoding. use_cer: Whether to use Character Error Rate instead of Word Error Rate. log_prediction: Whether to log a single decoded sample per call. fold_consecutive: Whether repeated consecutive characters should be folded into one when decoding. Returns: res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenstein's distances for all prediction - reference pairs, total number of words in all references. """ full_state_update: bool = True def __init__( self, decoding: CTCDecoding, use_cer=False, log_prediction=True, fold_consecutive=True, dist_sync_on_step=False, ): super().__init__(dist_sync_on_step=dist_sync_on_step) self.decoding = decoding self.use_cer = use_cer self.log_prediction = log_prediction self.fold_consecutive = fold_consecutive self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) def update( self, predictions: torch.Tensor, targets: torch.Tensor, target_lengths: torch.Tensor, predictions_lengths: torch.Tensor = None, ): """ Updates metric state. Args: predictions: an integer torch.Tensor of shape ``[Batch, Time, {Vocabulary}]`` (if ``batch_dim_index == 0``) or ``[Time, Batch]`` (if ``batch_dim_index == 1``) targets: an integer torch.Tensor of shape ``[Batch, Time]`` (if ``batch_dim_index == 0``) or ``[Time, Batch]`` (if ``batch_dim_index == 1``) target_lengths: an integer torch.Tensor of shape ``[Batch]`` predictions_lengths: an integer torch.Tensor of shape ``[Batch]`` """ words = 0 scores = 0 references = [] with torch.no_grad(): # prediction_cpu_tensor = tensors[0].long().cpu() targets_cpu_tensor = targets.long().cpu() tgt_lenths_cpu_tensor = target_lengths.long().cpu() # iterate over batch for ind in range(targets_cpu_tensor.shape[0]): tgt_len = tgt_lenths_cpu_tensor[ind].item() target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() reference = self.decoding.decode_tokens_to_str(target) references.append(reference) hypotheses, _ = self.decoding.ctc_decoder_predictions_tensor( predictions, predictions_lengths, fold_consecutive=self.fold_consecutive ) if self.log_prediction: logging.info(f"\n") logging.info(f"reference:{references[0]}") logging.info(f"predicted:{hypotheses[0]}") for h, r in zip(hypotheses, references): if self.use_cer: h_list = list(h) r_list = list(r) else: h_list = h.split() r_list = r.split() words += len(r_list) # Compute Levenstein's distance scores += editdistance.eval(h_list, r_list) self.scores = torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) self.words = torch.tensor(words, device=self.words.device, dtype=self.words.dtype) # return torch.tensor([scores, words]).to(predictions.device) def compute(self): scores = self.scores.detach().float() words = self.words.detach().float() return scores / words, scores, words @dataclass class CTCDecodingConfig: strategy: str = "greedy" # preserve decoding alignments preserve_alignments: Optional[bool] = None # compute ctc time stamps compute_timestamps: Optional[bool] = None # token representing word seperator word_seperator: str = " " # type of timestamps to calculate ctc_timestamp_type: str = "all" # can be char, word or all for both # batch dimension batch_dim_index: int = 0 # greedy decoding config greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() # beam decoding config beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) # confidence config confidence_cfg: ConfidenceConfig = ConfidenceConfig() # can be used to change temperature for decoding temperature: float = 1.0 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/asr/models/configs/aligner_config.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @dataclass class AlignerCTCConfig: prob_suppress_index: int = -1 prob_suppress_value: float = 1.0 @dataclass class AlignerRNNTConfig: predictor_window_size: int = 0 predictor_step_size: int = 1 @dataclass class AlignerWrapperModelConfig: alignment_type: str = "forced" word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Dict, List, Optional from omegaconf import MISSING import nemo.core.classes.dataset from nemo.collections.asr.metrics.wer import CTCDecodingConfig from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMelSpectrogramPreprocessorConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig from nemo.core.config import modelPT as model_cfg @dataclass class ASRDatasetConfig(nemo.core.classes.dataset.DatasetConfig): manifest_filepath: Optional[Any] = None sample_rate: int = MISSING labels: List[str] = MISSING trim_silence: bool = False # Tarred dataset support is_tarred: bool = False tarred_audio_filepaths: Optional[Any] = None tarred_shard_strategy: str = "scatter" shard_manifests: bool = False shuffle_n: int = 0 # Optional int_values: Optional[int] = None augmentor: Optional[Dict[str, Any]] = None max_duration: Optional[float] = None min_duration: Optional[float] = None max_utts: int = 0 blank_index: int = -1 unk_index: int = -1 normalize: bool = False trim: bool = True parser: Optional[str] = 'en' eos_id: Optional[int] = None bos_id: Optional[int] = None pad_id: int = 0 use_start_end_token: bool = False return_sample_id: Optional[bool] = False # bucketing params bucketing_strategy: str = "synced_randomized" bucketing_batch_size: Optional[Any] = None bucketing_weights: Optional[List[int]] = None @dataclass class EncDecCTCConfig(model_cfg.ModelConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = False labels: List[str] = MISSING # Dataset configs train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model component configs preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() encoder: ConvASREncoderConfig = ConvASREncoderConfig() decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() decoding: CTCDecodingConfig = CTCDecodingConfig() @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): model: EncDecCTCConfig = EncDecCTCConfig() @dataclass class CacheAwareStreamingConfig: chunk_size: int = 0 # the size of each chunk at each step, it can be a list of two integers to specify different chunk sizes for the first step and others shift_size: int = 0 # the size of the shift in each step, it can be a list of two integers to specify different shift sizes for the first step and others cache_drop_size: int = 0 # the number of steps to drop from the cache last_channel_cache_size: int = 0 # the size of the needed cache for last channel layers valid_out_len: int = 0 # the number of the steps in the final output which are valid (have the same value as in the offline mode) pre_encode_cache_size: int = 0 # the size of the needed cache for the pre-encoding part of the model to avoid caching inside the pre-encoding layers drop_extra_pre_encoded: int = 0 # the number of steps to get dropped after the pre-encoding layer last_channel_num: int = 0 # number of the last channel layers (like MHA layers) which need caching in the model last_time_num: int = 0 # number of the last time layers (like convolutions) which need caching in the model [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Dict, List, Optional from omegaconf import MISSING import nemo.core.classes.dataset from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMFCCPreprocessorConfig, CropOrPadSpectrogramAugmentationConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ConvASRDecoderClassificationConfig, ConvASREncoderConfig from nemo.core.config import modelPT as model_cfg @dataclass class EncDecClassificationDatasetConfig(nemo.core.classes.dataset.DatasetConfig): manifest_filepath: Optional[str] = None sample_rate: int = MISSING labels: List[str] = MISSING trim_silence: bool = False # Tarred dataset support is_tarred: bool = False tarred_audio_filepaths: Optional[str] = None tarred_shard_strategy: str = "scatter" shuffle_n: int = 0 # Optional int_values: Optional[int] = None augmentor: Optional[Dict[str, Any]] = None max_duration: Optional[float] = None min_duration: Optional[float] = None cal_labels_occurrence: Optional[bool] = False # VAD Optional vad_stream: Optional[bool] = None window_length_in_sec: float = 0.31 shift_length_in_sec: float = 0.01 normalize_audio: bool = False is_regression_task: bool = False # bucketing params bucketing_strategy: str = "synced_randomized" bucketing_batch_size: Optional[Any] = None bucketing_weights: Optional[List[int]] = None @dataclass class EncDecClassificationConfig(model_cfg.ModelConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = True kernel_size_factor: float = 1.0 labels: List[str] = MISSING timesteps: int = MISSING # Dataset configs train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=True, trim_silence=False ) validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model component configs preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( audio_length=timesteps ) encoder: ConvASREncoderConfig = ConvASREncoderConfig() decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): model: EncDecClassificationConfig = EncDecClassificationConfig() [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import asdict, dataclass from typing import Any, Dict, Optional, Tuple, Union @dataclass class DiarizerComponentConfig: """Dataclass to imitate HydraConfig dict when accessing parameters.""" def get(self, name: str, default: Optional[Any] = None): return getattr(self, name, default) def __iter__(self): for key in asdict(self): yield key def dict(self) -> Dict: return asdict(self) @dataclass class ASRDiarizerCTCDecoderParams: pretrained_language_model: Optional[str] = None # KenLM model file: .arpa model file or .bin binary file. beam_width: int = 32 alpha: float = 0.5 beta: float = 2.5 @dataclass class ASRRealigningLMParams: # Provide a KenLM language model in .arpa format. arpa_language_model: Optional[str] = None # Min number of words for the left context. min_number_of_words: int = 3 # Max number of words for the right context. max_number_of_words: int = 10 # The threshold for the difference between two log probability values from two hypotheses. logprob_diff_threshold: float = 1.2 @dataclass class ASRDiarizerParams(DiarizerComponentConfig): # if True, speech segmentation for diarization is based on word-timestamps from ASR inference. asr_based_vad: bool = False # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD. asr_based_vad_threshold: float = 1.0 # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null. asr_batch_size: Optional[int] = None # Native decoder delay. null is recommended to use the default values for each ASR model. decoder_delay_in_sec: Optional[float] = None # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. word_ts_anchor_offset: Optional[float] = None # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. word_ts_anchor_pos: str = "start" # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. fix_word_ts_with_VAD: bool = False # If True, use colored text to distinguish speakers in the output transcript. colored_text: bool = False # If True, the start and end time of each speaker turn is printed in the output transcript. print_time: bool = True # If True, the output transcript breaks the line to fix the line width (default is 90 chars) break_lines: bool = False @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" parameters: ASRDiarizerParams = ASRDiarizerParams() ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() @dataclass class VADParams(DiarizerComponentConfig): window_length_in_sec: float = 0.15 # Window length in sec for VAD context input shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech offset: float = 0.1 # Offset threshold for detecting the end of a speech pad_onset: float = 0.1 # Adding durations before each speech segment pad_offset: float = 0 # Adding durations after each speech segment min_duration_on: float = 0 # Threshold for small non_speech deletion min_duration_off: float = 0.2 # Threshold for short speech segment deletion filter_speech_first: bool = True @dataclass class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None parameters: VADParams = VADParams() @dataclass class SpeakerEmbeddingsParams(DiarizerComponentConfig): # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. save_embeddings: bool = True @dataclass class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() @dataclass class ClusteringParams(DiarizerComponentConfig): # If True, use num of speakers value provided in manifest file. oracle_num_speakers: bool = False # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. max_num_speakers: int = 8 # If the number of segments is lower than this number, enhanced speaker counting is activated. enhanced_count_thres: int = 80 # Determines the range of p-value search: 0 < p <= max_rp_threshold. max_rp_threshold: float = 0.25 # The higher the number, the more values will be examined with more time. sparse_search_volume: int = 30 # If True, take a majority vote on multiple p-values to estimate the number of speakers. maj_vote_spk_count: bool = False @dataclass class ClusteringConfig(DiarizerComponentConfig): parameters: ClusteringParams = ClusteringParams() @dataclass class MSDDParams(DiarizerComponentConfig): # If True, use speaker embedding model in checkpoint, else provided speaker embedding model in config will be used. use_speaker_model_from_ckpt: bool = True # Batch size for MSDD inference. infer_batch_size: int = 25 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. sigmoid_threshold: Tuple[float] = (0.7,) # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. seq_eval_mode: bool = False # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. split_infer: bool = True # The length of split short sequence when split_infer is True. diar_window_length: int = 50 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. overlap_infer_spk_limit: int = 5 @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" parameters: MSDDParams = MSDDParams() @dataclass class DiarizerConfig(DiarizerComponentConfig): manifest_filepath: Optional[str] = None out_dir: Optional[str] = None oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring vad: VADConfig = VADConfig() speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() clustering: ClusteringConfig = ClusteringConfig() msdd_model: MSDDConfig = MSDDConfig() asr: ASRDiarizerConfig = ASRDiarizerConfig() @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): diarizer: DiarizerConfig = DiarizerConfig() device: str = "cpu" verbose: bool = False batch_size: int = 64 num_workers: int = 1 sample_rate: int = 16000 name: str = "" @classmethod def init_config(cls, diar_model_path: str, vad_model_path: str, map_location: str, verbose: bool): return NeuralDiarizerInferenceConfig( DiarizerConfig( vad=VADConfig(model_path=vad_model_path), msdd_model=MSDDConfig(model_path=diar_model_path), ), device=map_location, verbose=verbose, ) [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig from nemo.core.config.modelPT import NemoConfig @dataclass class GraphModuleConfig: criterion_type: str = "ml" loss_type: str = "ctc" split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True backend_cfg: BackendConfig = BackendConfig() @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): graph_module_cfg: GraphModuleConfig = GraphModuleConfig() @dataclass class EncDecK2SeqModelConfig(NemoConfig): model: EncDecK2SeqConfig = EncDecK2SeqConfig() [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING from nemo.collections.asr.models.configs import classification_models_config as clf_cfg from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMFCCPreprocessorConfig, CropOrPadSpectrogramAugmentationConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ( ConvASRDecoderClassificationConfig, ConvASREncoderConfig, JasperEncoderConfig, ) from nemo.core.config import modelPT as model_cfg # fmt: off def matchboxnet_3x1x64(): config = [ JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config def matchboxnet_3x1x64_vad(): config = [ JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config # fmt: on @dataclass class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = True kernel_size_factor: float = 1.0 timesteps: int = 128 labels: List[str] = MISSING # Dataset configs train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=True, trim_silence=False ) validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( manifest_filepath=None, shuffle=False ) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model general component configs preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 ) crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( audio_length=128 ) encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() @dataclass class MatchboxNetVADModelConfig(MatchboxNetModelConfig): timesteps: int = 64 labels: List[str] = field(default_factory=lambda: ['background', 'speech']) crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = None class EncDecClassificationModelConfigBuilder(model_cfg.ModelConfigBuilder): VALID_CONFIGS = ['matchboxnet_3x1x64', 'matchboxnet_3x1x64_vad'] def __init__(self, name: str = 'matchboxnet_3x1x64', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): if name not in EncDecClassificationModelConfigBuilder.VALID_CONFIGS: raise ValueError("`name` must be one of : \n" f"{EncDecClassificationModelConfigBuilder.VALID_CONFIGS}") self.name = name if 'matchboxnet_3x1x64_vad' in name: if encoder_cfg_func is None: encoder_cfg_func = matchboxnet_3x1x64_vad model_cfg = MatchboxNetVADModelConfig( repeat=1, separable=True, encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderClassificationConfig(), ) elif 'matchboxnet_3x1x64' in name: if encoder_cfg_func is None: encoder_cfg_func = matchboxnet_3x1x64 model_cfg = MatchboxNetModelConfig( repeat=1, separable=False, spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderClassificationConfig(), ) else: raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") super(EncDecClassificationModelConfigBuilder, self).__init__(model_cfg) self.model_cfg: clf_cfg.EncDecClassificationConfig = model_cfg # enable type hinting def set_labels(self, labels: List[str]): self.model_cfg.labels = labels def set_separable(self, separable: bool): self.model_cfg.separable = separable def set_repeat(self, repeat: int): self.model_cfg.repeat = repeat def set_sample_rate(self, sample_rate: int): self.model_cfg.sample_rate = sample_rate def set_dropout(self, dropout: float = 0.0): self.model_cfg.dropout = dropout def set_timesteps(self, timesteps: int): self.model_cfg.timesteps = timesteps def set_is_regression_task(self, is_regression_task: bool): self.model_cfg.is_regression_task = is_regression_task # Note: Autocomplete for users wont work without these overrides # But practically it is not needed since python will infer at runtime # def set_train_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): # super().set_train_ds(cfg) # # def set_validation_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): # super().set_validation_ds(cfg) # # def set_test_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): # super().set_test_ds(cfg) def _finalize_cfg(self): # propagate labels self.model_cfg.train_ds.labels = self.model_cfg.labels self.model_cfg.validation_ds.labels = self.model_cfg.labels self.model_cfg.test_ds.labels = self.model_cfg.labels self.model_cfg.decoder.vocabulary = self.model_cfg.labels # propagate num classes self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) # propagate sample rate self.model_cfg.sample_rate = self.model_cfg.sample_rate self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate # propagate filters self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters # propagate timeteps if self.model_cfg.crop_or_pad_augment is not None: self.model_cfg.crop_or_pad_augment.audio_length = self.model_cfg.timesteps # propagate separable for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig layer.separable = self.model_cfg.separable # propagate repeat for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig layer.repeat = self.model_cfg.repeat # propagate dropout for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig layer.dropout = self.model_cfg.dropout def build(self) -> clf_cfg.EncDecClassificationConfig: return super().build() [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Callable, List, Optional from omegaconf import MISSING from nemo.collections.asr.models.configs import asr_models_config as ctc_cfg from nemo.collections.asr.modules.audio_preprocessing import ( AudioToMelSpectrogramPreprocessorConfig, SpectrogramAugmentationConfig, ) from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig, JasperEncoderConfig from nemo.core.config import modelPT as model_cfg # fmt: off def qn_15x5(): config = [ JasperEncoderConfig(filters=256, repeat=1, kernel=[33], stride=[2], dilation=[1], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, residual=True, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=1, kernel=[87], stride=[1], dilation=[2], dropout=0.0, residual=False, groups=1, separable=True, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config def jasper_10x5_dr(): config = [ JasperEncoderConfig(filters=256, repeat=1, kernel=[11], stride=[2], dilation=[1], dropout=0.2, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, residual=True, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=896, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.4, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.4, residual=False, groups=1, separable=False, heads=-1, residual_mode='add', residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) ] return config # fmt: on @dataclass class JasperModelConfig(ctc_cfg.EncDecCTCConfig): # Model global arguments sample_rate: int = 16000 repeat: int = 1 dropout: float = 0.0 separable: bool = False labels: List[str] = MISSING # Dataset configs train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( manifest_filepath=None, shuffle=True, trim_silence=True ) validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) # Model general component configs preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() @dataclass class QuartzNetModelConfig(JasperModelConfig): separable: bool = True class EncDecCTCModelConfigBuilder(model_cfg.ModelConfigBuilder): VALID_CONFIGS = ['quartznet_15x5', 'quartznet_15x5_zh', 'jasper_10x5dr'] def __init__(self, name: str = 'quartznet_15x5', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): if name not in EncDecCTCModelConfigBuilder.VALID_CONFIGS: raise ValueError("`name` must be one of : \n" f"{EncDecCTCModelConfigBuilder.VALID_CONFIGS}") self.name = name if 'quartznet_15x5' in name: if encoder_cfg_func is None: encoder_cfg_func = qn_15x5 model_cfg = QuartzNetModelConfig( repeat=5, separable=True, spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderConfig(), ) elif 'jasper_10x5' in name: if encoder_cfg_func is None: encoder_cfg_func = jasper_10x5_dr model_cfg = JasperModelConfig( repeat=5, separable=False, spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), decoder=ConvASRDecoderConfig(), ) else: raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") super(EncDecCTCModelConfigBuilder, self).__init__(model_cfg) self.model_cfg: ctc_cfg.EncDecCTCConfig = model_cfg # enable type hinting if 'zh' in name: self.set_dataset_normalize(normalize=False) def set_labels(self, labels: List[str]): self.model_cfg.labels = labels def set_separable(self, separable: bool): self.model_cfg.separable = separable def set_repeat(self, repeat: int): self.model_cfg.repeat = repeat def set_sample_rate(self, sample_rate: int): self.model_cfg.sample_rate = sample_rate def set_dropout(self, dropout: float = 0.0): self.model_cfg.dropout = dropout def set_dataset_normalize(self, normalize: bool): self.model_cfg.train_ds.normalize = normalize self.model_cfg.validation_ds.normalize = normalize self.model_cfg.test_ds.normalize = normalize # Note: Autocomplete for users wont work without these overrides # But practically it is not needed since python will infer at runtime # def set_train_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): # super().set_train_ds(cfg) # # def set_validation_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): # super().set_validation_ds(cfg) # # def set_test_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): # super().set_test_ds(cfg) def _finalize_cfg(self): # propagate labels self.model_cfg.train_ds.labels = self.model_cfg.labels self.model_cfg.validation_ds.labels = self.model_cfg.labels self.model_cfg.test_ds.labels = self.model_cfg.labels self.model_cfg.decoder.vocabulary = self.model_cfg.labels # propagate num classes self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) # propagate sample rate self.model_cfg.sample_rate = self.model_cfg.sample_rate self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate # propagate filters self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters # propagate separable for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig layer.separable = self.model_cfg.separable # propagate repeat for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig layer.repeat = self.model_cfg.repeat # propagate dropout for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig layer.dropout = self.model_cfg.dropout def build(self) -> ctc_cfg.EncDecCTCConfig: return super().build() [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import random from abc import ABC, abstractmethod from dataclasses import dataclass from typing import Any, Dict, Optional, Tuple import torch from packaging import version from nemo.collections.asr.parts.numba.spec_augment import SpecAugmentNumba, spec_augment_launch_heuristics from nemo.collections.asr.parts.preprocessing.features import ( FilterbankFeatures, FilterbankFeaturesTA, make_seq_mask_like, ) from nemo.collections.asr.parts.submodules.spectr_augment import SpecAugment, SpecCutout from nemo.core.classes import Exportable, NeuralModule, typecheck from nemo.core.neural_types import ( AudioSignal, LengthsType, MelSpectrogramType, MFCCSpectrogramType, NeuralType, SpectrogramType, ) from nemo.core.utils import numba_utils from nemo.core.utils.numba_utils import __NUMBA_MINIMUM_VERSION__ from nemo.utils import logging try: import torchaudio import torchaudio.functional import torchaudio.transforms TORCHAUDIO_VERSION = version.parse(torchaudio.__version__) TORCHAUDIO_VERSION_MIN = version.parse('0.5') HAVE_TORCHAUDIO = True except ModuleNotFoundError: HAVE_TORCHAUDIO = False __all__ = [ 'AudioToMelSpectrogramPreprocessor', 'AudioToSpectrogram', 'SpectrogramToAudio', 'AudioToMFCCPreprocessor', 'SpectrogramAugmentation', 'MaskedPatchAugmentation', 'CropOrPadSpectrogramAugmentation', ] class AudioPreprocessor(NeuralModule, ABC): """ An interface for Neural Modules that performs audio pre-processing, transforming the wav files to features. """ def __init__(self, win_length, hop_length): super().__init__() self.win_length = win_length self.hop_length = hop_length self.torch_windows = { 'hann': torch.hann_window, 'hamming': torch.hamming_window, 'blackman': torch.blackman_window, 'bartlett': torch.bartlett_window, 'ones': torch.ones, None: torch.ones, } @typecheck() @torch.no_grad() def forward(self, input_signal, length): processed_signal, processed_length = self.get_features(input_signal, length) return processed_signal, processed_length @abstractmethod def get_features(self, input_signal, length): # Called by forward(). Subclasses should implement this. pass class AudioToMelSpectrogramPreprocessor(AudioPreprocessor, Exportable): """Featurizer module that converts wavs to mel spectrograms. Args: sample_rate (int): Sample rate of the input audio data. Defaults to 16000 window_size (float): Size of window for fft in seconds Defaults to 0.02 window_stride (float): Stride of window for fft in seconds Defaults to 0.01 n_window_size (int): Size of window for fft in samples Defaults to None. Use one of window_size or n_window_size. n_window_stride (int): Stride of window for fft in samples Defaults to None. Use one of window_stride or n_window_stride. window (str): Windowing function for fft. can be one of ['hann', 'hamming', 'blackman', 'bartlett'] Defaults to "hann" normalize (str): Can be one of ['per_feature', 'all_features']; all other options disable feature normalization. 'all_features' normalizes the entire spectrogram to be mean 0 with std 1. 'pre_features' normalizes per channel / freq instead. Defaults to "per_feature" n_fft (int): Length of FT window. If None, it uses the smallest power of 2 that is larger than n_window_size. Defaults to None preemph (float): Amount of pre emphasis to add to audio. Can be disabled by passing None. Defaults to 0.97 features (int): Number of mel spectrogram freq bins to output. Defaults to 64 lowfreq (int): Lower bound on mel basis in Hz. Defaults to 0 highfreq (int): Lower bound on mel basis in Hz. Defaults to None log (bool): Log features. Defaults to True log_zero_guard_type(str): Need to avoid taking the log of zero. There are two options: "add" or "clamp". Defaults to "add". log_zero_guard_value(float, or str): Add or clamp requires the number to add with or clamp to. log_zero_guard_value can either be a float or "tiny" or "eps". torch.finfo is used if "tiny" or "eps" is passed. Defaults to 2**-24. dither (float): Amount of white-noise dithering. Defaults to 1e-5 pad_to (int): Ensures that the output size of the time dimension is a multiple of pad_to. Defaults to 16 frame_splicing (int): Defaults to 1 exact_pad (bool): If True, sets stft center to False and adds padding, such that num_frames = audio_length // hop_length. Defaults to False. pad_value (float): The value that shorter mels are padded with. Defaults to 0 mag_power (float): The power that the linear spectrogram is raised to prior to multiplication with mel basis. Defaults to 2 for a power spec rng : Random number generator nb_augmentation_prob (float) : Probability with which narrowband augmentation would be applied to samples in the batch. Defaults to 0.0 nb_max_freq (int) : Frequency above which all frequencies will be masked for narrowband augmentation. Defaults to 4000 use_torchaudio: Whether to use the `torchaudio` implementation. mel_norm: Normalization used for mel filterbank weights. Defaults to 'slaney' (area normalization) stft_exact_pad: Deprecated argument, kept for compatibility with older checkpoints. stft_conv: Deprecated argument, kept for compatibility with older checkpoints. """ def save_to(self, save_path: str): pass @classmethod def restore_from(cls, restore_path: str): pass @property def input_types(self): """Returns definitions of module input ports. """ return { "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), "length": NeuralType( tuple('B'), LengthsType() ), # Please note that length should be in samples not seconds. } @property def output_types(self): """Returns definitions of module output ports. processed_signal: 0: AxisType(BatchTag) 1: AxisType(MelSpectrogramSignalTag) 2: AxisType(ProcessedTimeTag) processed_length: 0: AxisType(BatchTag) """ return { "processed_signal": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "processed_length": NeuralType(tuple('B'), LengthsType()), } def __init__( self, sample_rate=16000, window_size=0.02, window_stride=0.01, n_window_size=None, n_window_stride=None, window="hann", normalize="per_feature", n_fft=None, preemph=0.97, features=64, lowfreq=0, highfreq=None, log=True, log_zero_guard_type="add", log_zero_guard_value=2 ** -24, dither=1e-5, pad_to=16, frame_splicing=1, exact_pad=False, pad_value=0, mag_power=2.0, rng=None, nb_augmentation_prob=0.0, nb_max_freq=4000, use_torchaudio: bool = False, mel_norm="slaney", stft_exact_pad=False, # Deprecated arguments; kept for config compatibility stft_conv=False, # Deprecated arguments; kept for config compatibility ): super().__init__(n_window_size, n_window_stride) self._sample_rate = sample_rate if window_size and n_window_size: raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") if window_stride and n_window_stride: raise ValueError( f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." ) if window_size: n_window_size = int(window_size * self._sample_rate) if window_stride: n_window_stride = int(window_stride * self._sample_rate) # Given the long and similar argument list, point to the class and instantiate it by reference if not use_torchaudio: featurizer_class = FilterbankFeatures else: featurizer_class = FilterbankFeaturesTA self.featurizer = featurizer_class( sample_rate=self._sample_rate, n_window_size=n_window_size, n_window_stride=n_window_stride, window=window, normalize=normalize, n_fft=n_fft, preemph=preemph, nfilt=features, lowfreq=lowfreq, highfreq=highfreq, log=log, log_zero_guard_type=log_zero_guard_type, log_zero_guard_value=log_zero_guard_value, dither=dither, pad_to=pad_to, frame_splicing=frame_splicing, exact_pad=exact_pad, pad_value=pad_value, mag_power=mag_power, rng=rng, nb_augmentation_prob=nb_augmentation_prob, nb_max_freq=nb_max_freq, mel_norm=mel_norm, stft_exact_pad=stft_exact_pad, # Deprecated arguments; kept for config compatibility stft_conv=stft_conv, # Deprecated arguments; kept for config compatibility ) def input_example(self, max_batch: int = 8, max_dim: int = 32000, min_length: int = 200): batch_size = torch.randint(low=1, high=max_batch, size=[1]).item() max_length = torch.randint(low=min_length, high=max_dim, size=[1]).item() signals = torch.rand(size=[batch_size, max_length]) * 2 - 1 lengths = torch.randint(low=min_length, high=max_dim, size=[batch_size]) lengths[0] = max_length return signals, lengths def get_features(self, input_signal, length): return self.featurizer(input_signal, length) @property def filter_banks(self): return self.featurizer.filter_banks class AudioToMFCCPreprocessor(AudioPreprocessor): """Preprocessor that converts wavs to MFCCs. Uses torchaudio.transforms.MFCC. Args: sample_rate: The sample rate of the audio. Defaults to 16000. window_size: Size of window for fft in seconds. Used to calculate the win_length arg for mel spectrogram. Defaults to 0.02 window_stride: Stride of window for fft in seconds. Used to caculate the hop_length arg for mel spect. Defaults to 0.01 n_window_size: Size of window for fft in samples Defaults to None. Use one of window_size or n_window_size. n_window_stride: Stride of window for fft in samples Defaults to None. Use one of window_stride or n_window_stride. window: Windowing function for fft. can be one of ['hann', 'hamming', 'blackman', 'bartlett', 'none', 'null']. Defaults to 'hann' n_fft: Length of FT window. If None, it uses the smallest power of 2 that is larger than n_window_size. Defaults to None lowfreq (int): Lower bound on mel basis in Hz. Defaults to 0 highfreq (int): Lower bound on mel basis in Hz. Defaults to None n_mels: Number of mel filterbanks. Defaults to 64 n_mfcc: Number of coefficients to retain Defaults to 64 dct_type: Type of discrete cosine transform to use norm: Type of norm to use log: Whether to use log-mel spectrograms instead of db-scaled. Defaults to True. """ @property def input_types(self): """Returns definitions of module input ports. """ return { "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return { "processed_signal": NeuralType(('B', 'D', 'T'), MFCCSpectrogramType()), "processed_length": NeuralType(tuple('B'), LengthsType()), } def save_to(self, save_path: str): pass @classmethod def restore_from(cls, restore_path: str): pass def __init__( self, sample_rate=16000, window_size=0.02, window_stride=0.01, n_window_size=None, n_window_stride=None, window='hann', n_fft=None, lowfreq=0.0, highfreq=None, n_mels=64, n_mfcc=64, dct_type=2, norm='ortho', log=True, ): self._sample_rate = sample_rate if not HAVE_TORCHAUDIO: logging.error('Could not import torchaudio. Some features might not work.') raise ModuleNotFoundError( "torchaudio is not installed but is necessary for " "AudioToMFCCPreprocessor. We recommend you try " "building it from source for the PyTorch version you have." ) if window_size and n_window_size: raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") if window_stride and n_window_stride: raise ValueError( f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." ) # Get win_length (n_window_size) and hop_length (n_window_stride) if window_size: n_window_size = int(window_size * self._sample_rate) if window_stride: n_window_stride = int(window_stride * self._sample_rate) super().__init__(n_window_size, n_window_stride) mel_kwargs = {} mel_kwargs['f_min'] = lowfreq mel_kwargs['f_max'] = highfreq mel_kwargs['n_mels'] = n_mels mel_kwargs['n_fft'] = n_fft or 2 ** math.ceil(math.log2(n_window_size)) mel_kwargs['win_length'] = n_window_size mel_kwargs['hop_length'] = n_window_stride # Set window_fn. None defaults to torch.ones. window_fn = self.torch_windows.get(window, None) if window_fn is None: raise ValueError( f"Window argument for AudioProcessor is invalid: {window}." f"For no window function, use 'ones' or None." ) mel_kwargs['window_fn'] = window_fn # Use torchaudio's implementation of MFCCs as featurizer self.featurizer = torchaudio.transforms.MFCC( sample_rate=self._sample_rate, n_mfcc=n_mfcc, dct_type=dct_type, norm=norm, log_mels=log, melkwargs=mel_kwargs, ) def get_features(self, input_signal, length): features = self.featurizer(input_signal) seq_len = torch.ceil(length.to(torch.float32) / self.hop_length).to(dtype=torch.long) return features, seq_len class SpectrogramAugmentation(NeuralModule): """ Performs time and freq cuts in one of two ways. SpecAugment zeroes out vertical and horizontal sections as described in SpecAugment (https://arxiv.org/abs/1904.08779). Arguments for use with SpecAugment are `freq_masks`, `time_masks`, `freq_width`, and `time_width`. SpecCutout zeroes out rectangulars as described in Cutout (https://arxiv.org/abs/1708.04552). Arguments for use with Cutout are `rect_masks`, `rect_freq`, and `rect_time`. Args: freq_masks (int): how many frequency segments should be cut. Defaults to 0. time_masks (int): how many time segments should be cut Defaults to 0. freq_width (int): maximum number of frequencies to be cut in one segment. Defaults to 10. time_width (int): maximum number of time steps to be cut in one segment Defaults to 10. rect_masks (int): how many rectangular masks should be cut Defaults to 0. rect_freq (int): maximum size of cut rectangles along the frequency dimension Defaults to 5. rect_time (int): maximum size of cut rectangles along the time dimension Defaults to 25. """ @property def input_types(self): """Returns definitions of module input types """ return { "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output types """ return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} def __init__( self, freq_masks=0, time_masks=0, freq_width=10, time_width=10, rect_masks=0, rect_time=5, rect_freq=20, rng=None, mask_value=0.0, use_numba_spec_augment: bool = True, ): super().__init__() if rect_masks > 0: self.spec_cutout = SpecCutout(rect_masks=rect_masks, rect_time=rect_time, rect_freq=rect_freq, rng=rng,) # self.spec_cutout.to(self._device) else: self.spec_cutout = lambda input_spec: input_spec if freq_masks + time_masks > 0: self.spec_augment = SpecAugment( freq_masks=freq_masks, time_masks=time_masks, freq_width=freq_width, time_width=time_width, rng=rng, mask_value=mask_value, ) else: self.spec_augment = lambda input_spec, length: input_spec # Check if numba is supported, and use a Numba kernel if it is if use_numba_spec_augment and numba_utils.numba_cuda_is_supported(__NUMBA_MINIMUM_VERSION__): logging.info('Numba CUDA SpecAugment kernel is being used') self.spec_augment_numba = SpecAugmentNumba( freq_masks=freq_masks, time_masks=time_masks, freq_width=freq_width, time_width=time_width, rng=rng, mask_value=mask_value, ) else: self.spec_augment_numba = None @typecheck() def forward(self, input_spec, length): augmented_spec = self.spec_cutout(input_spec=input_spec) # To run the Numba kernel, correct numba version is required as well as # tensor must be on GPU and length must be provided if self.spec_augment_numba is not None and spec_augment_launch_heuristics(augmented_spec, length): augmented_spec = self.spec_augment_numba(input_spec=augmented_spec, length=length) else: augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) return augmented_spec class MaskedPatchAugmentation(NeuralModule): """ Zeroes out fixed size time patches of the spectrogram. All samples in batch are guaranteed to have the same amount of masked time steps. Optionally also performs frequency masking in the same way as SpecAugment. Args: patch_size (int): up to how many time steps does one patch consist of. Defaults to 48. mask_patches (float): how many patches should be masked in each sample. if >= 1., interpreted as number of patches (after converting to int) if <1., interpreted as fraction of total tokens to be masked (number of patches is rounded up) Defaults to 10. freq_masks (int): how many frequency segments should be cut. Defaults to 0. freq_width (int): maximum number of frequencies to be cut in a segment. Defaults to 0. """ @property def input_types(self): """Returns definitions of module input types """ return { "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output types """ return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} def __init__( self, patch_size: int = 48, mask_patches: float = 10.0, freq_masks: int = 0, freq_width: int = 0, ): super().__init__() self.patch_size = patch_size if mask_patches >= 1: self.mask_patches = int(mask_patches) elif mask_patches >= 0: self._mask_fraction = mask_patches self.mask_patches = None else: raise ValueError('mask_patches cannot be negative') if freq_masks > 0: self.spec_augment = SpecAugment(freq_masks=freq_masks, time_masks=0, freq_width=freq_width, time_width=0,) else: self.spec_augment = None @typecheck() def forward(self, input_spec, length): augmented_spec = input_spec min_len = torch.min(length) if self.mask_patches is None: # masking specified as fraction len_fraction = int(min_len * self._mask_fraction) mask_patches = len_fraction // self.patch_size + int(len_fraction % self.patch_size != 0) else: mask_patches = self.mask_patches if min_len < self.patch_size * mask_patches: mask_patches = min_len // self.patch_size for idx in range(input_spec.shape[0]): cur_len = length[idx] patches = range(cur_len // self.patch_size) masked_patches = random.sample(patches, mask_patches) for mp in masked_patches: augmented_spec[idx, :, mp * self.patch_size : (mp + 1) * self.patch_size] = 0.0 if self.spec_augment is not None: augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) return augmented_spec class CropOrPadSpectrogramAugmentation(NeuralModule): """ Pad or Crop the incoming Spectrogram to a certain shape. Args: audio_length (int): the final number of timesteps that is required. The signal will be either padded or cropped temporally to this size. """ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length @typecheck() @torch.no_grad() def forward(self, input_signal, length): image = input_signal num_images = image.shape[0] audio_length = self.audio_length image_len = image.shape[-1] # Crop long signal if image_len > audio_length: # randomly slice cutout_images = [] offset = torch.randint(low=0, high=image_len - audio_length + 1, size=[num_images]) for idx, offset in enumerate(offset): cutout_images.append(image[idx : idx + 1, :, offset : offset + audio_length]) image = torch.cat(cutout_images, dim=0) del cutout_images else: # symmetrically pad short signal with zeros pad_left = (audio_length - image_len) // 2 pad_right = (audio_length - image_len) // 2 if (audio_length - image_len) % 2 == 1: pad_right += 1 image = torch.nn.functional.pad(image, [pad_left, pad_right], mode="constant", value=0) # Replace dynamic length sequences with static number of timesteps length = (length * 0) + audio_length return image, length @property def input_types(self): """Returns definitions of module output ports. """ return { "input_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), "length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return { "processed_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), "processed_length": NeuralType(tuple('B'), LengthsType()), } def save_to(self, save_path: str): pass @classmethod def restore_from(cls, restore_path: str): pass class AudioToSpectrogram(NeuralModule): """Transform a batch of input multi-channel signals into a batch of STFT-based spectrograms. Args: fft_length: length of FFT hop_length: length of hops/shifts of the sliding window power: exponent for magnitude spectrogram. Default `None` will return a complex-valued spectrogram """ def __init__(self, fft_length: int, hop_length: int, power: Optional[float] = None): if not HAVE_TORCHAUDIO: logging.error('Could not import torchaudio. Some features might not work.') raise ModuleNotFoundError( "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" ) super().__init__() # For now, assume FFT length is divisible by two if fft_length % 2 != 0: raise ValueError(f'fft_length = {fft_length} must be divisible by 2') self.stft = torchaudio.transforms.Spectrogram( n_fft=fft_length, hop_length=hop_length, power=power, pad_mode='constant' ) # number of subbands self.F = fft_length // 2 + 1 @property def num_subbands(self) -> int: return self.F @property def input_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "input": NeuralType(('B', 'C', 'T'), AudioSignal()), "input_length": NeuralType(('B',), LengthsType(), optional=True), } @property def output_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "output": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), "output_length": NeuralType(('B',), LengthsType()), } @typecheck() def forward( self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None ) -> Tuple[torch.Tensor, torch.Tensor]: """Convert a batch of C-channel input signals into a batch of complex-valued spectrograms. Args: input: Time-domain input signal with C channels, shape (B, C, T) input_length: Length of valid entries along the time dimension, shape (B,) Returns: Output spectrogram with F subbands and N time frames, shape (B, C, F, N) and output length with shape (B,). """ B, T = input.size(0), input.size(-1) input = input.view(B, -1, T) # STFT output (B, C, F, N) with torch.cuda.amp.autocast(enabled=False): output = self.stft(input.float()) if input_length is not None: # Mask padded frames output_length = self.get_output_length(input_length=input_length) length_mask: torch.Tensor = make_seq_mask_like( lengths=output_length, like=output, time_dim=-1, valid_ones=False ) output = output.masked_fill(length_mask, 0.0) else: # Assume all frames are valid for all examples in the batch output_length = output.size(-1) * torch.ones(B, device=output.device).long() return output, output_length def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: """Get length of valid frames for the output. Args: input_length: number of valid samples, shape (B,) Returns: Number of valid frames, shape (B,) """ output_length = input_length.div(self.stft.hop_length, rounding_mode='floor').add(1).long() return output_length class SpectrogramToAudio(NeuralModule): """Transform a batch of input multi-channel spectrograms into a batch of time-domain multi-channel signals. Args: fft_length: length of FFT hop_length: length of hops/shifts of the sliding window power: exponent for magnitude spectrogram. Default `None` will return a complex-valued spectrogram """ def __init__(self, fft_length: int, hop_length: int): if not HAVE_TORCHAUDIO: logging.error('Could not import torchaudio. Some features might not work.') raise ModuleNotFoundError( "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" ) super().__init__() # For now, assume FFT length is divisible by two if fft_length % 2 != 0: raise ValueError(f'fft_length = {fft_length} must be divisible by 2') self.istft = torchaudio.transforms.InverseSpectrogram( n_fft=fft_length, hop_length=hop_length, pad_mode='constant' ) self.F = fft_length // 2 + 1 @property def num_subbands(self) -> int: return self.F @property def input_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "input": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), "input_length": NeuralType(('B',), LengthsType(), optional=True), } @property def output_types(self) -> Dict[str, NeuralType]: """Returns definitions of module output ports. """ return { "output": NeuralType(('B', 'C', 'T'), AudioSignal()), "output_length": NeuralType(('B',), LengthsType()), } @typecheck() def forward(self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None) -> torch.Tensor: """Convert input complex-valued spectrogram to a time-domain signal. Multi-channel IO is supported. Args: input: Input spectrogram for C channels, shape (B, C, F, N) input_length: Length of valid entries along the time dimension, shape (B,) Returns: Time-domain signal with T time-domain samples and C channels, (B, C, T) and output length with shape (B,). """ B, F, N = input.size(0), input.size(-2), input.size(-1) assert F == self.F, f'Number of subbands F={F} not matching self.F={self.F}' input = input.view(B, -1, F, N) # iSTFT output (B, C, T) with torch.cuda.amp.autocast(enabled=False): output = self.istft(input.cfloat()) if input_length is not None: # Mask padded samples output_length = self.get_output_length(input_length=input_length) length_mask: torch.Tensor = make_seq_mask_like( lengths=output_length, like=output, time_dim=-1, valid_ones=False ) output = output.masked_fill(length_mask, 0.0) else: # Assume all frames are valid for all examples in the batch output_length = output.size(-1) * torch.ones(B, device=output.device).long() return output, output_length def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: """Get length of valid samples for the output. Args: input_length: number of valid frames, shape (B,) Returns: Number of valid samples, shape (B,) """ output_length = input_length.sub(1).mul(self.istft.hop_length).long() return output_length @dataclass class AudioToMelSpectrogramPreprocessorConfig: _target_: str = "nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor" sample_rate: int = 16000 window_size: float = 0.02 window_stride: float = 0.01 n_window_size: Optional[int] = None n_window_stride: Optional[int] = None window: str = "hann" normalize: str = "per_feature" n_fft: Optional[int] = None preemph: float = 0.97 features: int = 64 lowfreq: int = 0 highfreq: Optional[int] = None log: bool = True log_zero_guard_type: str = "add" log_zero_guard_value: float = 2 ** -24 dither: float = 1e-5 pad_to: int = 16 frame_splicing: int = 1 exact_pad: bool = False pad_value: int = 0 mag_power: float = 2.0 rng: Optional[str] = None nb_augmentation_prob: float = 0.0 nb_max_freq: int = 4000 use_torchaudio: bool = False mel_norm: str = "slaney" stft_exact_pad: bool = False # Deprecated argument, kept for compatibility with older checkpoints. stft_conv: bool = False # Deprecated argument, kept for compatibility with older checkpoints. @dataclass class AudioToMFCCPreprocessorConfig: _target_: str = 'nemo.collections.asr.modules.AudioToMFCCPreprocessor' sample_rate: int = 16000 window_size: float = 0.02 window_stride: float = 0.01 n_window_size: Optional[int] = None n_window_stride: Optional[int] = None window: str = 'hann' n_fft: Optional[int] = None lowfreq: Optional[float] = 0.0 highfreq: Optional[float] = None n_mels: int = 64 n_mfcc: int = 64 dct_type: int = 2 norm: str = 'ortho' log: bool = True @dataclass class SpectrogramAugmentationConfig: _target_: str = "nemo.collections.asr.modules.SpectrogramAugmentation" freq_masks: int = 0 time_masks: int = 0 freq_width: int = 0 time_width: Optional[Any] = 0 rect_masks: int = 0 rect_time: int = 0 rect_freq: int = 0 mask_value: float = 0 rng: Optional[Any] = None # random.Random() type use_numba_spec_augment: bool = True @dataclass class CropOrPadSpectrogramAugmentationConfig: audio_length: int _target_: str = "nemo.collections.asr.modules.CropOrPadSpectrogramAugmentation" @dataclass class MaskedPatchAugmentationConfig: patch_size: int = 48 mask_patches: float = 10.0 freq_masks: int = 0 freq_width: int = 0 _target_: str = "nemo.collections.asr.modules.MaskedPatchAugmentation" [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/parts/k2/classes.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from abc import ABC from dataclasses import dataclass from typing import Any, Optional, Tuple import torch from omegaconf import DictConfig from nemo.utils import logging @dataclass class GraphIntersectDenseConfig: """Graph dense intersection config. """ search_beam: float = 20.0 output_beam: float = 10.0 min_active_states: int = 30 max_active_states: int = 10000 @dataclass class GraphModuleConfig: """Config for graph modules. Typically used with graph losses and decoders. """ topo_type: str = "default" topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 class ASRK2Mixin(ABC): """k2 Mixin class that simplifies the construction of various models with k2-based losses. It does the following: - Sets up the graph loss and decoder (methods _init_k2 and update_k2_modules). - Registers external graphs, if needed. - Augments forward(...) with optional graph decoding to get accurate predictions. """ def _init_k2(self): """ k2-related initialization implementation. This method is expected to run after the __init__ which sets self._cfg self._cfg is expected to have the attribute graph_module_cfg """ if not hasattr(self, "_cfg"): raise ValueError("self._cfg must be set before calling _init_k2().") if not hasattr(self._cfg, "graph_module_cfg") or self._cfg.graph_module_cfg is None: raise ValueError("self._cfg.graph_module_cfg must be set and cannot be None.") self.graph_module_cfg = self._cfg.graph_module_cfg # register token_lm for MAPLoss criterion_type = self.graph_module_cfg.get("criterion_type", "ml") self.use_graph_lm = criterion_type == "map" if self.use_graph_lm: token_lm_path = self.graph_module_cfg.backend_cfg.get("token_lm", None) if token_lm_path is None: raise ValueError( f"graph_module_cfg.backend_cfg.token_lm is empty. It must be set for criterion_type == `{criterion_type}`" ) token_lm_path = self.register_artifact('graph_module_cfg.backend_cfg.token_lm', token_lm_path) self.graph_module_cfg.backend_cfg["token_lm"] = token_lm_path self.update_k2_modules(self.graph_module_cfg) def update_k2_modules(self, input_cfg: DictConfig): """ Helper function to initialize or update k2 loss and transcribe_decoder. Args: input_cfg: DictConfig to take new parameters from. Schema is expected as in nemo.collections.asr.models.configs.k2_sequence_models_config.GraphModuleConfig """ del self.loss if hasattr(self, "transcribe_decoder"): del self.transcribe_decoder if hasattr(self, "joint"): # RNNT num_classes = self.joint.num_classes_with_blank - 1 else: # CTC, MMI, ... num_classes = self.decoder.num_classes_with_blank - 1 remove_consecutive = input_cfg.backend_cfg.get("topo_with_self_loops", True) and input_cfg.backend_cfg.get( "topo_type", "default" ) not in ["forced_blank", "identity",] self._wer.remove_consecutive = remove_consecutive from nemo.collections.asr.losses.lattice_losses import LatticeLoss self.loss = LatticeLoss( num_classes=num_classes, reduction=self._cfg.get("ctc_reduction", "mean_batch"), backend="k2", criterion_type=input_cfg.get("criterion_type", "ml"), loss_type=input_cfg.get("loss_type", "ctc"), split_batch_size=input_cfg.get("split_batch_size", 0), graph_module_cfg=input_cfg.backend_cfg, ) criterion_type = self.loss.criterion_type self.use_graph_lm = criterion_type == "map" transcribe_training = input_cfg.get("transcribe_training", False) if transcribe_training and criterion_type == "ml": logging.warning( f"""You do not need to use transcribe_training=`{transcribe_training}` with criterion_type=`{criterion_type}`. transcribe_training will be set to False.""" ) transcribe_training = False self.transcribe_training = transcribe_training if self.use_graph_lm: from nemo.collections.asr.modules.graph_decoder import ViterbiDecoderWithGraph self.transcribe_decoder = ViterbiDecoderWithGraph( num_classes=num_classes, backend="k2", dec_type="token_lm", return_type="1best", return_ilabels=True, output_aligned=True, split_batch_size=input_cfg.get("split_batch_size", 0), graph_module_cfg=input_cfg.backend_cfg, ) def _forward_k2_post_processing( self, log_probs: torch.Tensor, encoded_length: torch.Tensor, greedy_predictions: torch.Tensor ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """ k2-related post-processing parf of .forward() Args: log_probs: The log probabilities tensor of shape [B, T, D]. encoded_length: The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. greedy_predictions: The greedy token predictions of the model of shape [B, T] Returns: A tuple of 3 elements - 1) The log probabilities tensor of shape [B, T, D]. 2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 3) The greedy token predictions of the model of shape [B, T] (via argmax) """ # greedy_predictions from .forward() are incorrect for criterion_type=`map` # getting correct greedy_predictions, if needed if self.use_graph_lm and (not self.training or self.transcribe_training): greedy_predictions, encoded_length, _ = self.transcribe_decoder.forward( log_probs=log_probs, log_probs_length=encoded_length ) return log_probs, encoded_length, greedy_predictions [end of nemo/collections/asr/parts/k2/classes.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math from dataclasses import dataclass from typing import Any, Optional import torch from torch import nn as nn from nemo.collections.asr.parts.submodules import multi_head_attention as mha from nemo.collections.common.parts import adapter_modules from nemo.core.classes.mixins import adapter_mixin_strategies class MHAResidualAddAdapterStrategy(adapter_mixin_strategies.ResidualAddAdapterStrategy): """ An implementation of residual addition of an adapter module with its input for the MHA Adapters. """ def forward(self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin'): """ A basic strategy, comprising of a residual connection over the input, after forward pass by the underlying adapter. Additional work is done to pack and unpack the dictionary of inputs and outputs. Note: The `value` tensor is added to the output of the attention adapter as the residual connection. Args: input: A dictionary of multiple input arguments for the adapter module. `query`, `key`, `value`: Original output tensor of the module, or the output of the previous adapter (if more than one adapters are enabled). `mask`: Attention mask. `pos_emb`: Optional positional embedding for relative encoding. adapter: The adapter module that is currently required to perform the forward pass. module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, therefore the strategy can access all other adapters in this module via `module.adapter_layer`. Returns: The result tensor, after one of the active adapters has finished its forward passes. """ out = self.compute_output(input, adapter, module=module) # If not in training mode, or probability of stochastic depth is 0, skip step. p = self.stochastic_depth if not module.training or p == 0.0: pass else: out = self.apply_stochastic_depth(out, input['value'], adapter, module=module) # Return the residual connection output = input + adapter(input) result = input['value'] + out # If l2_lambda is activated, register the loss value self.compute_auxiliary_losses(result, input['value'], adapter, module=module) return result def compute_output( self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin' ) -> torch.Tensor: """ Compute the output of a single adapter to some input. Args: input: Original output tensor of the module, or the output of the previous adapter (if more than one adapters are enabled). adapter: The adapter module that is currently required to perform the forward pass. module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, therefore the strategy can access all other adapters in this module via `module.adapter_layer`. Returns: The result tensor, after one of the active adapters has finished its forward passes. """ if isinstance(input, (list, tuple)): out = adapter(*input) elif isinstance(input, dict): out = adapter(**input) else: out = adapter(input) return out @dataclass class MHAResidualAddAdapterStrategyConfig(adapter_mixin_strategies.ResidualAddAdapterStrategyConfig): _target_: str = "{0}.{1}".format( MHAResidualAddAdapterStrategy.__module__, MHAResidualAddAdapterStrategy.__name__ ) # mandatory field class MultiHeadAttentionAdapter(mha.MultiHeadAttention, adapter_modules.AdapterModuleUtil): """Multi-Head Attention layer of Transformer. Args: n_head (int): number of heads n_feat (int): size of the features dropout_rate (float): dropout rate proj_dim (int, optional): Optional integer value for projection before computing attention. If None, then there is no projection (equivalent to proj_dim = n_feat). If > 0, then will project the n_feat to proj_dim before calculating attention. If <0, then will equal n_head, so that each head has a projected dimension of 1. adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, n_head: int, n_feat: int, dropout_rate: float, proj_dim: Optional[int] = None, adapter_strategy: MHAResidualAddAdapterStrategy = None, ): super().__init__(n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, max_cache_len=0) self.pre_norm = nn.LayerNorm(n_feat) # Set the projection dim to number of heads automatically if proj_dim is not None and proj_dim < 1: proj_dim = n_head self.proj_dim = proj_dim # Recompute weights for projection dim if self.proj_dim is not None: if self.proj_dim % n_head != 0: raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") self.d_k = self.proj_dim // n_head self.s_d_k = math.sqrt(self.d_k) self.linear_q = nn.Linear(n_feat, self.proj_dim) self.linear_k = nn.Linear(n_feat, self.proj_dim) self.linear_v = nn.Linear(n_feat, self.proj_dim) self.linear_out = nn.Linear(self.proj_dim, n_feat) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) # reset parameters for Q to be identity operation self.reset_parameters() def forward(self, query, key, value, mask, pos_emb=None, cache=None): """Compute 'Scaled Dot Product Attention'. Args: query (torch.Tensor): (batch, time1, size) key (torch.Tensor): (batch, time2, size) value(torch.Tensor): (batch, time2, size) mask (torch.Tensor): (batch, time1, time2) cache (torch.Tensor) : (batch, time_cache, size) returns: output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention cache (torch.Tensor) : (batch, time_cache_next, size) """ # Need to perform duplicate computations as at this point the tensors have been # separated by the adapter forward query = self.pre_norm(query) key = self.pre_norm(key) value = self.pre_norm(value) return super().forward(query, key, value, mask, pos_emb, cache=cache) def reset_parameters(self): with torch.no_grad(): nn.init.zeros_(self.linear_out.weight) nn.init.zeros_(self.linear_out.bias) def get_default_strategy_config(self) -> 'dataclass': return MHAResidualAddAdapterStrategyConfig() @dataclass class MultiHeadAttentionAdapterConfig: n_head: int n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): """Multi-Head Attention layer of Transformer-XL with support of relative positional encoding. Paper: https://arxiv.org/abs/1901.02860 Args: n_head (int): number of heads n_feat (int): size of the features dropout_rate (float): dropout rate proj_dim (int, optional): Optional integer value for projection before computing attention. If None, then there is no projection (equivalent to proj_dim = n_feat). If > 0, then will project the n_feat to proj_dim before calculating attention. If <0, then will equal n_head, so that each head has a projected dimension of 1. adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, n_head: int, n_feat: int, dropout_rate: float, proj_dim: Optional[int] = None, adapter_strategy: MHAResidualAddAdapterStrategyConfig = None, ): super().__init__( n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, pos_bias_u=None, pos_bias_v=None, max_cache_len=0 ) self.pre_norm = nn.LayerNorm(n_feat) # Set the projection dim to number of heads automatically if proj_dim is not None and proj_dim < 1: proj_dim = n_head self.proj_dim = proj_dim # Recompute weights for projection dim if self.proj_dim is not None: if self.proj_dim % n_head != 0: raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") self.d_k = self.proj_dim // n_head self.s_d_k = math.sqrt(self.d_k) self.linear_q = nn.Linear(n_feat, self.proj_dim) self.linear_k = nn.Linear(n_feat, self.proj_dim) self.linear_v = nn.Linear(n_feat, self.proj_dim) self.linear_out = nn.Linear(self.proj_dim, n_feat) self.linear_pos = nn.Linear(n_feat, self.proj_dim, bias=False) self.pos_bias_u = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) self.pos_bias_v = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) # reset parameters for Q to be identity operation self.reset_parameters() def forward(self, query, key, value, mask, pos_emb, cache=None): """Compute 'Scaled Dot Product Attention' with rel. positional encoding. Args: query (torch.Tensor): (batch, time1, size) key (torch.Tensor): (batch, time2, size) value(torch.Tensor): (batch, time2, size) mask (torch.Tensor): (batch, time1, time2) pos_emb (torch.Tensor) : (batch, time1, size) cache (torch.Tensor) : (batch, time_cache, size) Returns: output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention cache_next (torch.Tensor) : (batch, time_cache_next, size) """ # Need to perform duplicate computations as at this point the tensors have been # separated by the adapter forward query = self.pre_norm(query) key = self.pre_norm(key) value = self.pre_norm(value) return super().forward(query, key, value, mask, pos_emb, cache=cache) def reset_parameters(self): with torch.no_grad(): nn.init.zeros_(self.linear_out.weight) nn.init.zeros_(self.linear_out.bias) # NOTE: This exact procedure apparently highly important. # Above operation is safe to do as self.linear_out.weight *= 0.0 (similar for bias) # However: # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 # For some reason at init sometimes it will cause the value of the tensor to become NaN # All operations to compute matrix_ac and matrix_bd will then fail. nn.init.zeros_(self.pos_bias_u) nn.init.zeros_(self.pos_bias_v) def get_default_strategy_config(self) -> 'dataclass': return MHAResidualAddAdapterStrategyConfig() @dataclass class RelPositionMultiHeadAttentionAdapterConfig: n_head: int n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) class PositionalEncodingAdapter(mha.PositionalEncoding, adapter_modules.AdapterModuleUtil): """ Absolute positional embedding adapter. .. note:: Absolute positional embedding value is added to the input tensor *without residual connection* ! Therefore, the input is changed, if you only require the positional embedding, drop the returned `x` ! Args: d_model (int): The input dimension of x. max_len (int): The max sequence length. xscale (float): The input scaling factor. Defaults to 1.0. adapter_strategy (AbstractAdapterStrategy): By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. NOTE: Since this is a positional encoding, it will not add a residual ! """ def __init__( self, d_model: int, max_len: int = 5000, xscale=1.0, adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, ): super().__init__( d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, ) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) def get_default_strategy_config(self) -> 'dataclass': return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() @dataclass class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): """ Relative positional encoding for TransformerXL's layers See : Appendix B in https://arxiv.org/abs/1901.02860 .. note:: Relative positional embedding value is **not** added to the input tensor ! Therefore, the input should be updated changed, if you only require the positional embedding, drop the returned `x` ! Args: d_model (int): embedding dim max_len (int): maximum input length xscale (bool): whether to scale the input by sqrt(d_model) adapter_strategy: By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, d_model: int, max_len: int = 5000, xscale=1.0, adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, ): super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) def get_default_strategy_config(self) -> 'dataclass': return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() @dataclass class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import os from dataclasses import dataclass from typing import List, Optional, Tuple, Union import torch from nemo.collections.asr.parts.utils import rnnt_utils from nemo.collections.common.tokenizers.tokenizer_spec import TokenizerSpec from nemo.core.classes import Typing, typecheck from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType from nemo.utils import logging DEFAULT_TOKEN_OFFSET = 100 def pack_hypotheses( hypotheses: List[rnnt_utils.NBestHypotheses], logitlen: torch.Tensor, ) -> List[rnnt_utils.NBestHypotheses]: if logitlen is not None: if hasattr(logitlen, 'cpu'): logitlen_cpu = logitlen.to('cpu') else: logitlen_cpu = logitlen for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.NBestHypotheses for candidate_idx, cand in enumerate(hyp.n_best_hypotheses): cand.y_sequence = torch.tensor(cand.y_sequence, dtype=torch.long) if logitlen is not None: cand.length = logitlen_cpu[idx] if cand.dec_state is not None: cand.dec_state = _states_to_device(cand.dec_state) return hypotheses def _states_to_device(dec_state, device='cpu'): if torch.is_tensor(dec_state): dec_state = dec_state.to(device) elif isinstance(dec_state, (list, tuple)): dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) return dec_state class AbstractBeamCTCInfer(Typing): """A beam CTC decoder. Provides a common abstraction for sample level beam decoding. Args: blank_id: int, index of the blank token. Can be 0 or len(vocabulary). beam_size: int, size of the beam used in the underlying beam search engine. """ @property def input_types(self): """Returns definitions of module input ports. """ return { "decoder_output": NeuralType(('B', 'T', 'D'), LogprobsType()), "decoder_lengths": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return {"predictions": [NeuralType(elements_type=HypothesisType())]} def __init__(self, blank_id: int, beam_size: int): self.blank_id = blank_id if beam_size < 1: raise ValueError("Beam search size cannot be less than 1!") self.beam_size = beam_size # Variables set by corresponding setter methods self.vocab = None self.decoding_type = None self.tokenizer = None # Utility maps for vocabulary self.vocab_index_map = None self.index_vocab_map = None # Internal variable, used to prevent double reduction of consecutive tokens (ctc collapse) self.override_fold_consecutive_value = None def set_vocabulary(self, vocab: List[str]): """ Set the vocabulary of the decoding framework. Args: vocab: List of str. Each token corresponds to its location in the vocabulary emitted by the model. Note that this vocabulary must NOT contain the "BLANK" token. """ self.vocab = vocab self.vocab_index_map = {v: i for i, v in enumerate(vocab)} self.index_vocab_map = {i: v for i, v in enumerate(vocab)} def set_decoding_type(self, decoding_type: str): """ Sets the decoding type of the framework. Can support either char or subword models. Args: decoding_type: Str corresponding to decoding type. Only supports "char" and "subword". """ decoding_type = decoding_type.lower() supported_types = ['char', 'subword'] if decoding_type not in supported_types: raise ValueError( f"Unsupported decoding type. Supported types = {supported_types}.\n" f"Given = {decoding_type}" ) self.decoding_type = decoding_type def set_tokenizer(self, tokenizer: TokenizerSpec): """ Set the tokenizer of the decoding framework. Args: tokenizer: NeMo tokenizer object, which inherits from TokenizerSpec. """ self.tokenizer = tokenizer @typecheck() def forward( self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-repressively. Args: decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). decoder_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ raise NotImplementedError() def __call__(self, *args, **kwargs): return self.forward(*args, **kwargs) class BeamCTCInfer(AbstractBeamCTCInfer): """A greedy CTC decoder. Provides a common abstraction for sample level and batch level greedy decoding. Args: blank_index: int index of the blank token. Can be 0 or len(vocabulary). preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. """ def __init__( self, blank_id: int, beam_size: int, search_type: str = "default", return_best_hypothesis: bool = True, preserve_alignments: bool = False, compute_timestamps: bool = False, beam_alpha: float = 1.0, beam_beta: float = 0.0, kenlm_path: str = None, flashlight_cfg: Optional['FlashlightConfig'] = None, pyctcdecode_cfg: Optional['PyCTCDecodeConfig'] = None, ): super().__init__(blank_id=blank_id, beam_size=beam_size) self.search_type = search_type self.return_best_hypothesis = return_best_hypothesis self.preserve_alignments = preserve_alignments self.compute_timestamps = compute_timestamps if self.compute_timestamps: raise ValueError(f"Currently this flag is not supported for beam search algorithms.") self.vocab = None # This must be set by specific method by user before calling forward() ! if search_type == "default" or search_type == "nemo": self.search_algorithm = self.default_beam_search elif search_type == "pyctcdecode": self.search_algorithm = self._pyctcdecode_beam_search elif search_type == "flashlight": self.search_algorithm = self.flashlight_beam_search else: raise NotImplementedError( f"The search type ({search_type}) supplied is not supported!\n" f"Please use one of : (default, nemo, pyctcdecode)" ) # Log the beam search algorithm logging.info(f"Beam search algorithm: {search_type}") self.beam_alpha = beam_alpha self.beam_beta = beam_beta # Default beam search args self.kenlm_path = kenlm_path # PyCTCDecode params if pyctcdecode_cfg is None: pyctcdecode_cfg = PyCTCDecodeConfig() self.pyctcdecode_cfg = pyctcdecode_cfg # type: PyCTCDecodeConfig if flashlight_cfg is None: flashlight_cfg = FlashlightConfig() self.flashlight_cfg = flashlight_cfg # Default beam search scorer functions self.default_beam_scorer = None self.pyctcdecode_beam_scorer = None self.flashlight_beam_scorer = None self.token_offset = 0 @typecheck() def forward( self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-repressively. Args: decoder_output: A tensor of size (batch, timesteps, features). decoder_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ if self.vocab is None: raise RuntimeError("Please set the vocabulary with `set_vocabulary()` before calling this function.") if self.decoding_type is None: raise ValueError("Please set the decoding type with `set_decoding_type()` before calling this function.") with torch.no_grad(), torch.inference_mode(): # Process each sequence independently prediction_tensor = decoder_output if prediction_tensor.ndim != 3: raise ValueError( f"`decoder_output` must be a tensor of shape [B, T, V] (log probs, float). " f"Provided shape = {prediction_tensor.shape}" ) # determine type of input - logprobs or labels out_len = decoder_lengths if decoder_lengths is not None else None hypotheses = self.search_algorithm(prediction_tensor, out_len) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, decoder_lengths) # Pack the result if self.return_best_hypothesis and isinstance(packed_result[0], rnnt_utils.NBestHypotheses): packed_result = [res.n_best_hypotheses[0] for res in packed_result] # type: Hypothesis return (packed_result,) @torch.no_grad() def default_beam_search( self, x: torch.Tensor, out_len: torch.Tensor ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: """ Open Seq2Seq Beam Search Algorithm (DeepSpeed) Args: x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, and V is the vocabulary size. The tensor contains log-probabilities. out_len: Tensor of shape [B], contains lengths of each sequence in the batch. Returns: A list of NBestHypotheses objects, one for each sequence in the batch. """ if self.compute_timestamps: raise ValueError( f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" ) if self.default_beam_scorer is None: # Check for filepath if self.kenlm_path is None or not os.path.exists(self.kenlm_path): raise FileNotFoundError( f"KenLM binary file not found at : {self.kenlm_path}. " f"Please set a valid path in the decoding config." ) # perform token offset for subword models if self.decoding_type == 'subword': vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] else: # char models vocab = self.vocab # Must import at runtime to avoid circular dependency due to module level import. from nemo.collections.asr.modules.beam_search_decoder import BeamSearchDecoderWithLM self.default_beam_scorer = BeamSearchDecoderWithLM( vocab=vocab, lm_path=self.kenlm_path, beam_width=self.beam_size, alpha=self.beam_alpha, beta=self.beam_beta, num_cpus=max(1, os.cpu_count()), input_tensor=False, ) x = x.to('cpu') with typecheck.disable_checks(): data = [x[sample_id, : out_len[sample_id], :].softmax(dim=-1) for sample_id in range(len(x))] beams_batch = self.default_beam_scorer.forward(log_probs=data, log_probs_length=None) # For each sample in the batch nbest_hypotheses = [] for beams_idx, beams in enumerate(beams_batch): # For each beam candidate / hypothesis in each sample hypotheses = [] for candidate_idx, candidate in enumerate(beams): hypothesis = rnnt_utils.Hypothesis( score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None ) # For subword encoding, NeMo will double encode the subword (multiple tokens) into a # singular unicode id. In doing so, we preserve the semantic of the unicode token, and # compress the size of the final KenLM ARPA / Binary file. # In order to do double encoding, we shift the subword by some token offset. # This step is ignored for character based models. if self.decoding_type == 'subword': pred_token_ids = [ord(c) - self.token_offset for c in candidate[1]] else: # Char models pred_token_ids = [self.vocab_index_map[c] for c in candidate[1]] # We preserve the token ids and the score for this hypothesis hypothesis.y_sequence = pred_token_ids hypothesis.score = candidate[0] # If alignment must be preserved, we preserve a view of the output logprobs. # Note this view is shared amongst all beams within the sample, be sure to clone it if you # require specific processing for each sample in the beam. # This is done to preserve memory. if self.preserve_alignments: hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] hypotheses.append(hypothesis) # Wrap the result in NBestHypothesis. hypotheses = rnnt_utils.NBestHypotheses(hypotheses) nbest_hypotheses.append(hypotheses) return nbest_hypotheses @torch.no_grad() def _pyctcdecode_beam_search( self, x: torch.Tensor, out_len: torch.Tensor ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: """ PyCTCDecode Beam Search Algorithm. Should support Char and Subword models. Args: x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, and V is the vocabulary size. The tensor contains log-probabilities. out_len: Tensor of shape [B], contains lengths of each sequence in the batch. Returns: A list of NBestHypotheses objects, one for each sequence in the batch. """ if self.compute_timestamps: raise ValueError( f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" ) try: import pyctcdecode except (ImportError, ModuleNotFoundError): raise ImportError( f"Could not load `pyctcdecode` library. Please install it from pip using :\n" f"pip install --upgrade pyctcdecode" ) if self.pyctcdecode_beam_scorer is None: self.pyctcdecode_beam_scorer = pyctcdecode.build_ctcdecoder( labels=self.vocab, kenlm_model_path=self.kenlm_path, alpha=self.beam_alpha, beta=self.beam_beta ) # type: pyctcdecode.BeamSearchDecoderCTC x = x.to('cpu').numpy() with typecheck.disable_checks(): beams_batch = [] for sample_id in range(len(x)): logprobs = x[sample_id, : out_len[sample_id], :] result = self.pyctcdecode_beam_scorer.decode_beams( logprobs, beam_width=self.beam_size, beam_prune_logp=self.pyctcdecode_cfg.beam_prune_logp, token_min_logp=self.pyctcdecode_cfg.token_min_logp, prune_history=self.pyctcdecode_cfg.prune_history, hotwords=self.pyctcdecode_cfg.hotwords, hotword_weight=self.pyctcdecode_cfg.hotword_weight, lm_start_state=None, ) # Output format: text, last_lm_state, text_frames, logit_score, lm_score beams_batch.append(result) nbest_hypotheses = [] for beams_idx, beams in enumerate(beams_batch): hypotheses = [] for candidate_idx, candidate in enumerate(beams): # Candidate = (text, last_lm_state, text_frames, logit_score, lm_score) hypothesis = rnnt_utils.Hypothesis( score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None ) # TODO: Requires token ids to be returned rather than text. if self.decoding_type == 'subword': if self.tokenizer is None: raise ValueError("Tokenizer must be provided for subword decoding. Use set_tokenizer().") pred_token_ids = self.tokenizer.text_to_ids(candidate[0]) else: if self.vocab is None: raise ValueError("Vocab must be provided for character decoding. Use set_vocab().") chars = list(candidate[0]) pred_token_ids = [self.vocab_index_map[c] for c in chars] hypothesis.y_sequence = pred_token_ids hypothesis.text = candidate[0] # text hypothesis.score = candidate[4] # score # Inject word level timestamps hypothesis.timestep = candidate[2] # text_frames if self.preserve_alignments: hypothesis.alignments = torch.from_numpy(x[beams_idx][: out_len[beams_idx]]) hypotheses.append(hypothesis) hypotheses = rnnt_utils.NBestHypotheses(hypotheses) nbest_hypotheses.append(hypotheses) return nbest_hypotheses @torch.no_grad() def flashlight_beam_search( self, x: torch.Tensor, out_len: torch.Tensor ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: """ Flashlight Beam Search Algorithm. Should support Char and Subword models. Args: x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, and V is the vocabulary size. The tensor contains log-probabilities. out_len: Tensor of shape [B], contains lengths of each sequence in the batch. Returns: A list of NBestHypotheses objects, one for each sequence in the batch. """ if self.compute_timestamps: raise ValueError( f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" ) if self.flashlight_beam_scorer is None: # Check for filepath if self.kenlm_path is None or not os.path.exists(self.kenlm_path): raise FileNotFoundError( f"KenLM binary file not found at : {self.kenlm_path}. " f"Please set a valid path in the decoding config." ) # perform token offset for subword models # if self.decoding_type == 'subword': # vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] # else: # # char models # vocab = self.vocab # Must import at runtime to avoid circular dependency due to module level import. from nemo.collections.asr.modules.flashlight_decoder import FlashLightKenLMBeamSearchDecoder self.flashlight_beam_scorer = FlashLightKenLMBeamSearchDecoder( lm_path=self.kenlm_path, vocabulary=self.vocab, tokenizer=self.tokenizer, lexicon_path=self.flashlight_cfg.lexicon_path, boost_path=self.flashlight_cfg.boost_path, beam_size=self.beam_size, beam_size_token=self.flashlight_cfg.beam_size_token, beam_threshold=self.flashlight_cfg.beam_threshold, lm_weight=self.beam_alpha, word_score=self.beam_beta, unk_weight=self.flashlight_cfg.unk_weight, sil_weight=self.flashlight_cfg.sil_weight, ) x = x.to('cpu') with typecheck.disable_checks(): beams_batch = self.flashlight_beam_scorer.forward(log_probs=x) # For each sample in the batch nbest_hypotheses = [] for beams_idx, beams in enumerate(beams_batch): # For each beam candidate / hypothesis in each sample hypotheses = [] for candidate_idx, candidate in enumerate(beams): hypothesis = rnnt_utils.Hypothesis( score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None ) # We preserve the token ids and the score for this hypothesis hypothesis.y_sequence = candidate['tokens'].tolist() hypothesis.score = candidate['score'] # If alignment must be preserved, we preserve a view of the output logprobs. # Note this view is shared amongst all beams within the sample, be sure to clone it if you # require specific processing for each sample in the beam. # This is done to preserve memory. if self.preserve_alignments: hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] hypotheses.append(hypothesis) # Wrap the result in NBestHypothesis. hypotheses = rnnt_utils.NBestHypotheses(hypotheses) nbest_hypotheses.append(hypotheses) return nbest_hypotheses def set_decoding_type(self, decoding_type: str): super().set_decoding_type(decoding_type) # Please check train_kenlm.py in scripts/asr_language_modeling/ to find out why we need # TOKEN_OFFSET for BPE-based models if self.decoding_type == 'subword': self.token_offset = DEFAULT_TOKEN_OFFSET @dataclass class PyCTCDecodeConfig: # These arguments cannot be imported from pyctcdecode (optional dependency) # Therefore we copy the values explicitly # Taken from pyctcdecode.constant beam_prune_logp: float = -10.0 token_min_logp: float = -5.0 prune_history: bool = False hotwords: Optional[List[str]] = None hotword_weight: float = 10.0 @dataclass class FlashlightConfig: lexicon_path: Optional[str] = None boost_path: Optional[str] = None beam_size_token: int = 16 beam_threshold: float = 20.0 unk_weight: float = -math.inf sil_weight: float = 0.0 @dataclass class BeamCTCInferConfig: beam_size: int search_type: str = 'default' preserve_alignments: bool = False compute_timestamps: bool = False return_best_hypothesis: bool = True beam_alpha: float = 1.0 beam_beta: float = 0.0 kenlm_path: Optional[str] = None flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import List, Optional import torch from omegaconf import DictConfig, OmegaConf from nemo.collections.asr.parts.utils import rnnt_utils from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMethodConfig, ConfidenceMethodMixin from nemo.core.classes import Typing, typecheck from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType from nemo.utils import logging def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: if logitlen is not None: if hasattr(logitlen, 'cpu'): logitlen_cpu = logitlen.to('cpu') else: logitlen_cpu = logitlen for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) if logitlen is not None: hyp.length = logitlen_cpu[idx] if hyp.dec_state is not None: hyp.dec_state = _states_to_device(hyp.dec_state) return hypotheses def _states_to_device(dec_state, device='cpu'): if torch.is_tensor(dec_state): dec_state = dec_state.to(device) elif isinstance(dec_state, (list, tuple)): dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) return dec_state class GreedyCTCInfer(Typing, ConfidenceMethodMixin): """A greedy CTC decoder. Provides a common abstraction for sample level and batch level greedy decoding. Args: blank_index: int index of the blank token. Can be 0 or len(vocabulary). preserve_alignments: Bool flag which preserves the history of logprobs generated during decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. compute_timestamps: A bool flag, which determines whether to compute the character/subword, or word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. The timestamps will be available in the returned Hypothesis.timestep as a dictionary. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ @property def input_types(self): """Returns definitions of module input ports. """ # Input can be of dimention - # ('B', 'T', 'D') [Log probs] or ('B', 'T') [Labels] return { "decoder_output": NeuralType(None, LogprobsType()), "decoder_lengths": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self): """Returns definitions of module output ports. """ return {"predictions": [NeuralType(elements_type=HypothesisType())]} def __init__( self, blank_id: int, preserve_alignments: bool = False, compute_timestamps: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__() self.blank_id = blank_id self.preserve_alignments = preserve_alignments # we need timestamps to extract non-blank per-frame confidence self.compute_timestamps = compute_timestamps | preserve_frame_confidence self.preserve_frame_confidence = preserve_frame_confidence # set confidence calculation method self._init_confidence_method(confidence_method_cfg) @typecheck() def forward( self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-repressively. Args: decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). decoder_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ with torch.inference_mode(): hypotheses = [] # Process each sequence independently prediction_cpu_tensor = decoder_output.cpu() if prediction_cpu_tensor.ndim < 2 or prediction_cpu_tensor.ndim > 3: raise ValueError( f"`decoder_output` must be a tensor of shape [B, T] (labels, int) or " f"[B, T, V] (log probs, float). Provided shape = {prediction_cpu_tensor.shape}" ) # determine type of input - logprobs or labels if prediction_cpu_tensor.ndim == 2: # labels greedy_decode = self._greedy_decode_labels else: greedy_decode = self._greedy_decode_logprobs for ind in range(prediction_cpu_tensor.shape[0]): out_len = decoder_lengths[ind] if decoder_lengths is not None else None hypothesis = greedy_decode(prediction_cpu_tensor[ind], out_len) hypotheses.append(hypothesis) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, decoder_lengths) return (packed_result,) @torch.no_grad() def _greedy_decode_logprobs(self, x: torch.Tensor, out_len: torch.Tensor): # x: [T, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) prediction = x.detach().cpu() if out_len is not None: prediction = prediction[:out_len] prediction_logprobs, prediction_labels = prediction.max(dim=-1) non_blank_ids = prediction_labels != self.blank_id hypothesis.y_sequence = prediction_labels.numpy().tolist() hypothesis.score = (prediction_logprobs[non_blank_ids]).sum() if self.preserve_alignments: # Preserve the logprobs, as well as labels after argmax hypothesis.alignments = (prediction.clone(), prediction_labels.clone()) if self.compute_timestamps: hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() if self.preserve_frame_confidence: hypothesis.frame_confidence = self._get_confidence(prediction) return hypothesis @torch.no_grad() def _greedy_decode_labels(self, x: torch.Tensor, out_len: torch.Tensor): # x: [T] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) prediction_labels = x.detach().cpu() if out_len is not None: prediction_labels = prediction_labels[:out_len] non_blank_ids = prediction_labels != self.blank_id hypothesis.y_sequence = prediction_labels.numpy().tolist() hypothesis.score = -1.0 if self.preserve_alignments: raise ValueError("Requested for alignments, but predictions provided were labels, not log probabilities.") if self.compute_timestamps: hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() if self.preserve_frame_confidence: raise ValueError( "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." ) return hypothesis def __call__(self, *args, **kwargs): return self.forward(*args, **kwargs) @dataclass class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.confidence_method_cfg = OmegaConf.structured( self.confidence_method_cfg if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) else ConfidenceMethodConfig(**self.confidence_method_cfg) ) [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # Copyright 2017 Johns Hopkins University (Shinji Watanabe) # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import List, Optional, Tuple, Union import numpy as np import torch from omegaconf import DictConfig, OmegaConf from nemo.collections.asr.modules import rnnt_abstract from nemo.collections.asr.parts.utils import rnnt_utils from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMethodConfig, ConfidenceMethodMixin from nemo.collections.common.parts.rnn import label_collate from nemo.core.classes import Typing, typecheck from nemo.core.neural_types import AcousticEncodedRepresentation, ElementType, HypothesisType, LengthsType, NeuralType from nemo.utils import logging def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: if hasattr(logitlen, 'cpu'): logitlen_cpu = logitlen.to('cpu') else: logitlen_cpu = logitlen for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) hyp.length = logitlen_cpu[idx] if hyp.dec_state is not None: hyp.dec_state = _states_to_device(hyp.dec_state) return hypotheses def _states_to_device(dec_state, device='cpu'): if torch.is_tensor(dec_state): dec_state = dec_state.to(device) elif isinstance(dec_state, (list, tuple)): dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) return dec_state class _GreedyRNNTInfer(Typing, ConfidenceMethodMixin): """A greedy transducer decoder. Provides a common abstraction for sample level and batch level greedy decoding. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Can be 0 or len(vocabulary). max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ @property def input_types(self): """Returns definitions of module input ports. """ return { "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), "encoded_lengths": NeuralType(tuple('B'), LengthsType()), "partial_hypotheses": [NeuralType(elements_type=HypothesisType(), optional=True)], # must always be last } @property def output_types(self): """Returns definitions of module output ports. """ return {"predictions": [NeuralType(elements_type=HypothesisType())]} def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__() self.decoder = decoder_model self.joint = joint_model self._blank_index = blank_index self._SOS = blank_index # Start of single index self.max_symbols = max_symbols_per_step self.preserve_alignments = preserve_alignments self.preserve_frame_confidence = preserve_frame_confidence # set confidence calculation method self._init_confidence_method(confidence_method_cfg) def __call__(self, *args, **kwargs): return self.forward(*args, **kwargs) @torch.no_grad() def _pred_step( self, label: Union[torch.Tensor, int], hidden: Optional[torch.Tensor], add_sos: bool = False, batch_size: Optional[int] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: """ Common prediction step based on the AbstractRNNTDecoder implementation. Args: label: (int/torch.Tensor): Label or "Start-of-Signal" token. hidden: (Optional torch.Tensor): RNN State vector add_sos (bool): Whether to add a zero vector at the begging as "start of sentence" token. batch_size: Batch size of the output tensor. Returns: g: (B, U, H) if add_sos is false, else (B, U + 1, H) hid: (h, c) where h is the final sequence hidden state and c is the final cell state: h (tensor), shape (L, B, H) c (tensor), shape (L, B, H) """ if isinstance(label, torch.Tensor): # label: [batch, 1] if label.dtype != torch.long: label = label.long() else: # Label is an integer if label == self._SOS: return self.decoder.predict(None, hidden, add_sos=add_sos, batch_size=batch_size) label = label_collate([[label]]) # output: [B, 1, K] return self.decoder.predict(label, hidden, add_sos=add_sos, batch_size=batch_size) def _joint_step(self, enc, pred, log_normalize: Optional[bool] = None): """ Common joint step based on AbstractRNNTJoint implementation. Args: enc: Output of the Encoder model. A torch.Tensor of shape [B, 1, H1] pred: Output of the Decoder model. A torch.Tensor of shape [B, 1, H2] log_normalize: Whether to log normalize or not. None will log normalize only for CPU. Returns: logits of shape (B, T=1, U=1, V + 1) """ with torch.no_grad(): logits = self.joint.joint(enc, pred) if log_normalize is None: if not logits.is_cuda: # Use log softmax only if on CPU logits = logits.log_softmax(dim=len(logits.shape) - 1) else: if log_normalize: logits = logits.log_softmax(dim=len(logits.shape) - 1) return logits class GreedyRNNTInfer(_GreedyRNNTInfer): """A greedy transducer decoder. Sequence level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Can be 0 or len(vocabulary). max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_method_cfg=confidence_method_cfg, ) @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) self.decoder.eval() self.joint.eval() hypotheses = [] # Process each sequence independently with self.decoder.as_frozen(), self.joint.as_frozen(): for batch_idx in range(encoder_output.size(0)): inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] logitlen = encoded_lengths[batch_idx] partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) hypotheses.append(hypothesis) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, encoded_lengths) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) @torch.no_grad() def _greedy_decode( self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None ): # x: [T, 1, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) if partial_hypotheses is not None: hypothesis.last_token = partial_hypotheses.last_token hypothesis.y_sequence = ( partial_hypotheses.y_sequence.cpu().tolist() if isinstance(partial_hypotheses.y_sequence, torch.Tensor) else partial_hypotheses.y_sequence ) if partial_hypotheses.dec_state is not None: hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) if self.preserve_alignments: # Alignments is a 2-dimensional dangling list representing T x U hypothesis.alignments = [[]] if self.preserve_frame_confidence: hypothesis.frame_confidence = [[]] # For timestep t in X_t for time_idx in range(out_len): # Extract encoder embedding at timestep t # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] f = x.narrow(dim=0, start=time_idx, length=1) # Setup exit flags and counter not_blank = True symbols_added = 0 # While blank is not predicted, or we dont run out of max symbols per timestep while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # In the first timestep, we initialize the network with RNNT Blank # In later timesteps, we provide previous predicted label as input. if hypothesis.last_token is None and hypothesis.dec_state is None: last_label = self._SOS else: last_label = label_collate([[hypothesis.last_token]]) # Perform prediction network and joint network steps. g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 0, 0, 0, : ] del g # torch.max(0) op doesnt exist for FP 16. if logp.dtype != torch.float32: logp = logp.float() # get index k, of max prob v, k = logp.max(0) k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. if self.preserve_alignments: # insert logprobs into last timestep hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) if self.preserve_frame_confidence: # insert confidence into last timestep hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) del logp # If blank token is predicted, exit inner loop, move onto next timestep t if k == self._blank_index: not_blank = False else: # Append token to label set, update RNN state. hypothesis.y_sequence.append(k) hypothesis.score += float(v) hypothesis.timestep.append(time_idx) hypothesis.dec_state = hidden_prime hypothesis.last_token = k # Increment token counter. symbols_added += 1 if self.preserve_alignments: # convert Ti-th logits into a torch array hypothesis.alignments.append([]) # blank buffer for next timestep if self.preserve_frame_confidence: hypothesis.frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of Alignments if self.preserve_alignments: if len(hypothesis.alignments[-1]) == 0: del hypothesis.alignments[-1] # Remove trailing empty list of per-frame confidence if self.preserve_frame_confidence: if len(hypothesis.frame_confidence[-1]) == 0: del hypothesis.frame_confidence[-1] # Unpack the hidden states hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) return hypothesis class GreedyBatchedRNNTInfer(_GreedyRNNTInfer): """A batch level greedy transducer decoder. Batch level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Can be 0 or len(vocabulary). max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_method_cfg=confidence_method_cfg, ) # Depending on availability of `blank_as_pad` support # switch between more efficient batch decoding technique if self.decoder.blank_as_pad: self._greedy_decode = self._greedy_decode_blank_as_pad else: self._greedy_decode = self._greedy_decode_masked @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) logitlen = encoded_lengths self.decoder.eval() self.joint.eval() with self.decoder.as_frozen(), self.joint.as_frozen(): inseq = encoder_output # [B, T, D] hypotheses = self._greedy_decode( inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses ) # Pack the hypotheses results packed_result = pack_hypotheses(hypotheses, logitlen) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) def _greedy_decode_blank_as_pad( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") with torch.inference_mode(): # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize list of Hypothesis batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a dangling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] # If confidence scores need to be preserved, register a dangling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() for time_idx in range(max_out_len): f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) all_blanks = torch.all(blank_mask) del k_is_blank # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx, is_blank in enumerate(blank_mask): # we only want to update non-blanks, unless we are at the last step in the loop where # all elements produced blanks, otherwise there will be duplicate predictions # saved in alignments if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx, is_blank in enumerate(blank_mask): if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if all_blanks: not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.clone().view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses def _greedy_decode_masked( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize state batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] else: alignments = None # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) last_label_without_blank = last_label.clone() # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() with torch.inference_mode(): for time_idx in range(max_out_len): f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Set a dummy label for the blank value # This value will be overwritten by "blank" again the last label update below # This is done as vocabulary of prediction network does not contain "blank" token of RNNT last_label_without_blank_mask = last_label == self._blank_index last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label last_label_without_blank[~last_label_without_blank_mask] = last_label[ ~last_label_without_blank_mask ] # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) all_blanks = torch.all(blank_mask) # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx, is_blank in enumerate(blank_mask): # we only want to update non-blanks, unless we are at the last step in the loop where # all elements produced blanks, otherwise there will be duplicate predictions # saved in alignments if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx, is_blank in enumerate(blank_mask): if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses class ExportedModelGreedyBatchedRNNTInfer: def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = None): self.encoder_model_path = encoder_model self.decoder_joint_model_path = decoder_joint_model self.max_symbols_per_step = max_symbols_per_step # Will be populated at runtime self._blank_index = None def __call__(self, audio_signal: torch.Tensor, length: torch.Tensor): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ with torch.no_grad(): # Apply optional preprocessing encoder_output, encoded_lengths = self.run_encoder(audio_signal=audio_signal, length=length) if torch.is_tensor(encoder_output): encoder_output = encoder_output.transpose(1, 2) else: encoder_output = encoder_output.transpose([0, 2, 1]) # (B, T, D) logitlen = encoded_lengths inseq = encoder_output # [B, T, D] hypotheses, timestamps = self._greedy_decode(inseq, logitlen) # Pack the hypotheses results packed_result = [rnnt_utils.Hypothesis(score=-1.0, y_sequence=[]) for _ in range(len(hypotheses))] for i in range(len(packed_result)): packed_result[i].y_sequence = torch.tensor(hypotheses[i], dtype=torch.long) packed_result[i].length = timestamps[i] del hypotheses return packed_result def _greedy_decode(self, x, out_len): # x: [B, T, D] # out_len: [B] # Initialize state batchsize = x.shape[0] hidden = self._get_initial_states(batchsize) target_lengths = torch.ones(batchsize, dtype=torch.int32) # Output string buffer label = [[] for _ in range(batchsize)] timesteps = [[] for _ in range(batchsize)] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long).numpy() if torch.is_tensor(x): last_label = torch.from_numpy(last_label).to(self.device) # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool).numpy() # Get max sequence length max_out_len = out_len.max() for time_idx in range(max_out_len): f = x[:, time_idx : time_idx + 1, :] # [B, 1, D] if torch.is_tensor(f): f = f.transpose(1, 2) else: f = f.transpose([0, 2, 1]) # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask *= False # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols_per_step is None or symbols_added < self.max_symbols_per_step): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0: g = torch.tensor([self._blank_index] * batchsize, dtype=torch.int32).view(-1, 1) else: if torch.is_tensor(last_label): g = last_label.type(torch.int32) else: g = last_label.astype(np.int32) # Batched joint step - Output = [B, V + 1] joint_out, hidden_prime = self.run_decoder_joint(f, g, target_lengths, *hidden) logp, pred_lengths = joint_out logp = logp[:, 0, 0, :] # Get index k, of max prob for batch if torch.is_tensor(logp): v, k = logp.max(1) else: k = np.argmax(logp, axis=1).astype(np.int32) # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask |= k_is_blank del k_is_blank del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past if torch.is_tensor(blank_mask): blank_indices = (blank_mask == 1).nonzero(as_tuple=False) else: blank_indices = blank_mask.astype(np.int32).nonzero() if type(blank_indices) in (list, tuple): blank_indices = blank_indices[0] # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states for state_id in range(len(hidden)): hidden_prime[state_id][:, blank_indices, :] = hidden[state_id][:, blank_indices, :] elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states for state_id in range(len(hidden_prime)): hidden_prime[state_id][:, blank_indices, :] *= 0.0 # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration if torch.is_tensor(k): last_label = k.clone().reshape(-1, 1) else: last_label = k.copy().reshape(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: label[kidx].append(ki) timesteps[kidx].append(time_idx) symbols_added += 1 return label, timesteps def _setup_blank_index(self): raise NotImplementedError() def run_encoder(self, audio_signal, length): raise NotImplementedError() def run_decoder_joint(self, enc_logits, targets, target_length, *states): raise NotImplementedError() def _get_initial_states(self, batchsize): raise NotImplementedError() class ONNXGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = 10): super().__init__( encoder_model=encoder_model, decoder_joint_model=decoder_joint_model, max_symbols_per_step=max_symbols_per_step, ) try: import onnx import onnxruntime except (ModuleNotFoundError, ImportError): raise ImportError(f"`onnx` or `onnxruntime` could not be imported, please install the libraries.\n") if torch.cuda.is_available(): # Try to use onnxruntime-gpu providers = ['TensorrtExecutionProvider', 'CUDAExecutionProvider'] else: # Fall back to CPU and onnxruntime-cpu providers = ['CPUExecutionProvider'] onnx_session_opt = onnxruntime.SessionOptions() onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL onnx_model = onnx.load(self.encoder_model_path) onnx.checker.check_model(onnx_model, full_check=True) self.encoder_model = onnx_model self.encoder = onnxruntime.InferenceSession( onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt ) onnx_model = onnx.load(self.decoder_joint_model_path) onnx.checker.check_model(onnx_model, full_check=True) self.decoder_joint_model = onnx_model self.decoder_joint = onnxruntime.InferenceSession( onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt ) logging.info("Successfully loaded encoder, decoder and joint onnx models !") # Will be populated at runtime self._blank_index = None self.max_symbols_per_step = max_symbols_per_step self._setup_encoder_input_output_keys() self._setup_decoder_joint_input_output_keys() self._setup_blank_index() def _setup_encoder_input_output_keys(self): self.encoder_inputs = list(self.encoder_model.graph.input) self.encoder_outputs = list(self.encoder_model.graph.output) def _setup_decoder_joint_input_output_keys(self): self.decoder_joint_inputs = list(self.decoder_joint_model.graph.input) self.decoder_joint_outputs = list(self.decoder_joint_model.graph.output) def _setup_blank_index(self): # ASSUME: Single input with no time length information dynamic_dim = 257 shapes = self.encoder_inputs[0].type.tensor_type.shape.dim ip_shape = [] for shape in shapes: if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: ip_shape.append(dynamic_dim) # replace dynamic axes with constant else: ip_shape.append(int(shape.dim_value)) enc_logits, encoded_length = self.run_encoder( audio_signal=torch.randn(*ip_shape), length=torch.randint(0, 1, size=(dynamic_dim,)) ) # prepare states states = self._get_initial_states(batchsize=dynamic_dim) # run decoder 1 step joint_out, states = self.run_decoder_joint(enc_logits, None, None, *states) log_probs, lengths = joint_out self._blank_index = log_probs.shape[-1] - 1 # last token of vocab size is blank token logging.info( f"Enc-Dec-Joint step was evaluated, blank token id = {self._blank_index}; vocab size = {log_probs.shape[-1]}" ) def run_encoder(self, audio_signal, length): if hasattr(audio_signal, 'cpu'): audio_signal = audio_signal.cpu().numpy() if hasattr(length, 'cpu'): length = length.cpu().numpy() ip = { self.encoder_inputs[0].name: audio_signal, self.encoder_inputs[1].name: length, } enc_out = self.encoder.run(None, ip) enc_out, encoded_length = enc_out # ASSUME: single output return enc_out, encoded_length def run_decoder_joint(self, enc_logits, targets, target_length, *states): # ASSUME: Decoder is RNN Transducer if targets is None: targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32) target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32) if hasattr(targets, 'cpu'): targets = targets.cpu().numpy() if hasattr(target_length, 'cpu'): target_length = target_length.cpu().numpy() ip = { self.decoder_joint_inputs[0].name: enc_logits, self.decoder_joint_inputs[1].name: targets, self.decoder_joint_inputs[2].name: target_length, } num_states = 0 if states is not None and len(states) > 0: num_states = len(states) for idx, state in enumerate(states): if hasattr(state, 'cpu'): state = state.cpu().numpy() ip[self.decoder_joint_inputs[len(ip)].name] = state dec_out = self.decoder_joint.run(None, ip) # unpack dec output if num_states > 0: new_states = dec_out[-num_states:] dec_out = dec_out[:-num_states] else: new_states = None return dec_out, new_states def _get_initial_states(self, batchsize): # ASSUME: LSTM STATES of shape (layers, batchsize, dim) input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] num_states = len(input_state_nodes) if num_states == 0: return input_states = [] for state_id in range(num_states): node = input_state_nodes[state_id] ip_shape = [] for shape_idx, shape in enumerate(node.type.tensor_type.shape.dim): if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: ip_shape.append(batchsize) # replace dynamic axes with constant else: ip_shape.append(int(shape.dim_value)) input_states.append(torch.zeros(*ip_shape)) return input_states class TorchscriptGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): def __init__( self, encoder_model: str, decoder_joint_model: str, cfg: DictConfig, device: str, max_symbols_per_step: Optional[int] = 10, ): super().__init__( encoder_model=encoder_model, decoder_joint_model=decoder_joint_model, max_symbols_per_step=max_symbols_per_step, ) self.cfg = cfg self.device = device self.encoder = torch.jit.load(self.encoder_model_path, map_location=self.device) self.decoder_joint = torch.jit.load(self.decoder_joint_model_path, map_location=self.device) logging.info("Successfully loaded encoder, decoder and joint torchscript models !") # Will be populated at runtime self._blank_index = None self.max_symbols_per_step = max_symbols_per_step self._setup_encoder_input_keys() self._setup_decoder_joint_input_keys() self._setup_blank_index() def _setup_encoder_input_keys(self): arguments = self.encoder.forward.schema.arguments[1:] self.encoder_inputs = [arg for arg in arguments] def _setup_decoder_joint_input_keys(self): arguments = self.decoder_joint.forward.schema.arguments[1:] self.decoder_joint_inputs = [arg for arg in arguments] def _setup_blank_index(self): self._blank_index = len(self.cfg.joint.vocabulary) logging.info(f"Blank token id = {self._blank_index}; vocab size = {len(self.cfg.joint.vocabulary) + 1}") def run_encoder(self, audio_signal, length): enc_out = self.encoder(audio_signal, length) enc_out, encoded_length = enc_out # ASSUME: single output return enc_out, encoded_length def run_decoder_joint(self, enc_logits, targets, target_length, *states): # ASSUME: Decoder is RNN Transducer if targets is None: targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32, device=enc_logits.device) target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32, device=enc_logits.device) num_states = 0 if states is not None and len(states) > 0: num_states = len(states) dec_out = self.decoder_joint(enc_logits, targets, target_length, *states) # unpack dec output if num_states > 0: new_states = dec_out[-num_states:] dec_out = dec_out[:-num_states] else: new_states = None return dec_out, new_states def _get_initial_states(self, batchsize): # ASSUME: LSTM STATES of shape (layers, batchsize, dim) input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] num_states = len(input_state_nodes) if num_states == 0: return input_states = [] for state_id in range(num_states): # Hardcode shape size for LSTM (1 is for num layers in LSTM, which is flattened for export) ip_shape = [1, batchsize, self.cfg.model_defaults.pred_hidden] input_states.append(torch.zeros(*ip_shape, device=self.device)) return input_states class GreedyMultiblankRNNTInfer(GreedyRNNTInfer): """A greedy transducer decoder for multi-blank RNN-T. Sequence level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. big_blank_durations: a list containing durations for big blanks the model supports. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, big_blank_durations: list, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_method_cfg=confidence_method_cfg, ) self.big_blank_durations = big_blank_durations self._SOS = blank_index - len(big_blank_durations) @torch.no_grad() def _greedy_decode( self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None ): # x: [T, 1, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) if partial_hypotheses is not None: hypothesis.last_token = partial_hypotheses.last_token hypothesis.y_sequence = ( partial_hypotheses.y_sequence.cpu().tolist() if isinstance(partial_hypotheses.y_sequence, torch.Tensor) else partial_hypotheses.y_sequence ) if partial_hypotheses.dec_state is not None: hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) if self.preserve_alignments: # Alignments is a 2-dimensional dangling list representing T x U hypothesis.alignments = [[]] if self.preserve_frame_confidence: hypothesis.frame_confidence = [[]] # if this variable is > 1, it means the last emission was a big-blank and we need to skip frames. big_blank_duration = 1 # For timestep t in X_t for time_idx in range(out_len): if big_blank_duration > 1: # skip frames until big_blank_duration == 1. big_blank_duration -= 1 continue # Extract encoder embedding at timestep t # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] f = x.narrow(dim=0, start=time_idx, length=1) # Setup exit flags and counter not_blank = True symbols_added = 0 # While blank is not predicted, or we dont run out of max symbols per timestep while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # In the first timestep, we initialize the network with RNNT Blank # In later timesteps, we provide previous predicted label as input. if hypothesis.last_token is None and hypothesis.dec_state is None: last_label = self._SOS else: last_label = label_collate([[hypothesis.last_token]]) # Perform prediction network and joint network steps. g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 0, 0, 0, : ] del g # torch.max(0) op doesnt exist for FP 16. if logp.dtype != torch.float32: logp = logp.float() # get index k, of max prob v, k = logp.max(0) k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. # Note, we have non-blanks in the vocab first, followed by big blanks, and standard blank at last. # here we check if it's a big blank and if yes, set the duration variable. if k >= self._blank_index - len(self.big_blank_durations) and k < self._blank_index: big_blank_duration = self.big_blank_durations[self._blank_index - k - 1] if self.preserve_alignments: # insert logprobs into last timestep hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) if self.preserve_frame_confidence: # insert confidence into last timestep hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) del logp # If any type of blank token is predicted, exit inner loop, move onto next timestep t if k >= self._blank_index - len(self.big_blank_durations): not_blank = False else: # Append token to label set, update RNN state. hypothesis.y_sequence.append(k) hypothesis.score += float(v) hypothesis.timestep.append(time_idx) hypothesis.dec_state = hidden_prime hypothesis.last_token = k # Increment token counter. symbols_added += 1 if self.preserve_alignments: # convert Ti-th logits into a torch array hypothesis.alignments.append([]) # blank buffer for next timestep if self.preserve_frame_confidence: hypothesis.frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of Alignments if self.preserve_alignments: if len(hypothesis.alignments[-1]) == 0: del hypothesis.alignments[-1] # Remove trailing empty list of per-frame confidence if self.preserve_frame_confidence: if len(hypothesis.frame_confidence[-1]) == 0: del hypothesis.frame_confidence[-1] # Unpack the hidden states hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) return hypothesis class GreedyBatchedMultiblankRNNTInfer(GreedyBatchedRNNTInfer): """A batch level greedy transducer decoder. Batch level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. big_blank_durations: a list containing durations for big blanks the model supports. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, big_blank_durations: List[int], max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_method_cfg=confidence_method_cfg, ) self.big_blank_durations = big_blank_durations # Depending on availability of `blank_as_pad` support # switch between more efficient batch decoding technique if self.decoder.blank_as_pad: self._greedy_decode = self._greedy_decode_blank_as_pad else: self._greedy_decode = self._greedy_decode_masked self._SOS = blank_index - len(big_blank_durations) def _greedy_decode_blank_as_pad( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") with torch.inference_mode(): # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize list of Hypothesis batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._SOS, dtype=torch.long, device=device) # this mask is true for if the emission is *any type* of blank. blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() # We have a mask for each big blank. A mask is "true" means: the previous emission is exactly the big-blank # with the corresponding duration, or has larger duration. E.g., for big_blank_mask for duration 2, it will # be set true if the previous emission was a big blank with duration 4, or 3 or 2; but false if prevoius # emission was a standard blank (with duration = 1). big_blank_masks = [torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device)] * len( self.big_blank_durations ) # if this variable > 1, it means the previous emission was big-blank and we need to skip frames. big_blank_duration = 1 for time_idx in range(max_out_len): if big_blank_duration > 1: # skip frames until big_blank_duration == 1 big_blank_duration -= 1 continue f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset all blank masks blank_mask.mul_(False) for i in range(len(big_blank_masks)): big_blank_masks[i].mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len for i in range(len(big_blank_masks)): big_blank_masks[i] = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1 + num-big-blanks] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k >= self._blank_index - len(self.big_blank_durations) blank_mask.bitwise_or_(k_is_blank) for i in range(len(big_blank_masks)): # using <= since as we mentioned before, the mask doesn't store exact matches. # instead, it is True when the predicted blank's duration is >= the duration that the # mask corresponds to. k_is_big_blank = k <= self._blank_index - 1 - i # need to do a bitwise_and since it could also be a non-blank. k_is_big_blank.bitwise_and_(k_is_blank) big_blank_masks[i].bitwise_or_(k_is_big_blank) del k_is_blank # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.clone().view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 for i in range(len(big_blank_masks) + 1): # The task here is find the shortest blank duration of all batches. # so we start from the shortest blank duration and go up, # and stop once we found the duration whose corresponding mask isn't all True. if i == len(big_blank_masks) or not big_blank_masks[i].all(): big_blank_duration = self.big_blank_durations[i - 1] if i > 0 else 1 break # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses def _greedy_decode_masked( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") if self.big_blank_durations != [1] * len(self.big_blank_durations): raise NotImplementedError( "Efficient frame-skipping version for multi-blank masked decoding is not supported." ) # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize state batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] else: hyp.alignments = None # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) last_label_without_blank = last_label.clone() # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() with torch.inference_mode(): for time_idx in range(max_out_len): f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # Prepare t timestamp batch variables not_blank = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Set a dummy label for the blank value # This value will be overwritten by "blank" again the last label update below # This is done as vocabulary of prediction network does not contain "blank" token of RNNT last_label_without_blank_mask = last_label >= self._blank_index last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label last_label_without_blank[~last_label_without_blank_mask] = last_label[ ~last_label_without_blank_mask ] # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1 + num-big-blanks] # If preserving per-frame confidence, log_normalize must be true logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ :, 0, 0, : ] if logp.dtype != torch.float32: logp = logp.float() # Get index k, of max prob for batch v, k = logp.max(1) del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) # If preserving alignments, check if sequence length of sample has been reached # before adding alignment if self.preserve_alignments: # Insert logprobs into last timestep per sample logp_vals = logp.to('cpu') logp_ids = logp_vals.max(1)[1] for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].alignments[-1].append( (logp_vals[batch_idx], logp_ids[batch_idx]) ) del logp_vals # If preserving per-frame confidence, check if sequence length of sample has been reached # before adding confidence scores if self.preserve_frame_confidence: # Insert probabilities into last timestep per sample confidence = self._get_confidence(logp) for batch_idx in range(batchsize): if time_idx < out_len[batch_idx]: hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) del logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if blank_mask.all(): not_blank = False else: # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # If preserving alignments, convert the current Uj alignments into a torch.Tensor # Then preserve U at current timestep Ti # Finally, forward the timestep history to Ti+1 for that sample # All of this should only be done iff the current time index <= sample-level AM length. # Otherwise ignore and move to next sample / next timestep. if self.preserve_alignments: # convert Ti-th logits into a torch array for batch_idx in range(batchsize): # this checks if current timestep <= sample-level AM length # If current timestep > sample-level AM length, no alignments will be added # Therefore the list of Uj alignments is empty here. if len(hypotheses[batch_idx].alignments[-1]) > 0: hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep # Do the same if preserving per-frame confidence if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses @dataclass class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.confidence_method_cfg = OmegaConf.structured( self.confidence_method_cfg if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) else ConfidenceMethodConfig(**self.confidence_method_cfg) ) @dataclass class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.confidence_method_cfg = OmegaConf.structured( self.confidence_method_cfg if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) else ConfidenceMethodConfig(**self.confidence_method_cfg) ) class GreedyTDTInfer(_GreedyRNNTInfer): """A greedy TDT decoder. Sequence level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. durations: a list containing durations for TDT. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, durations: list, max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_method_cfg=confidence_method_cfg, ) self.durations = durations @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) self.decoder.eval() self.joint.eval() hypotheses = [] # Process each sequence independently with self.decoder.as_frozen(), self.joint.as_frozen(): for batch_idx in range(encoder_output.size(0)): inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] logitlen = encoded_lengths[batch_idx] partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) hypotheses.append(hypothesis) # Pack results into Hypotheses packed_result = pack_hypotheses(hypotheses, encoded_lengths) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) @torch.no_grad() def _greedy_decode( self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None ): # x: [T, 1, D] # out_len: [seq_len] # Initialize blank state and empty label set in Hypothesis hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) if partial_hypotheses is not None: hypothesis.last_token = partial_hypotheses.last_token hypothesis.y_sequence = ( partial_hypotheses.y_sequence.cpu().tolist() if isinstance(partial_hypotheses.y_sequence, torch.Tensor) else partial_hypotheses.y_sequence ) if partial_hypotheses.dec_state is not None: hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) if self.preserve_alignments: # Alignments is a 2-dimensional dangling list representing T x U hypothesis.alignments = [[]] if self.preserve_frame_confidence: hypothesis.frame_confidence = [[]] time_idx = 0 while time_idx < out_len: # Extract encoder embedding at timestep t # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] f = x.narrow(dim=0, start=time_idx, length=1) # Setup exit flags and counter not_blank = True symbols_added = 0 need_loop = True # While blank is not predicted, or we dont run out of max symbols per timestep while need_loop and (self.max_symbols is None or symbols_added < self.max_symbols): # In the first timestep, we initialize the network with RNNT Blank # In later timesteps, we provide previous predicted label as input. if hypothesis.last_token is None and hypothesis.dec_state is None: last_label = self._SOS else: last_label = label_collate([[hypothesis.last_token]]) # Perform prediction network and joint network steps. g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) # If preserving per-frame confidence, log_normalize must be true logits = self._joint_step(f, g, log_normalize=False) logp = logits[0, 0, 0, : -len(self.durations)] if self.preserve_frame_confidence: logp = torch.log_softmax(logp, -1) duration_logp = torch.log_softmax(logits[0, 0, 0, -len(self.durations) :], dim=-1) del g # torch.max(0) op doesnt exist for FP 16. if logp.dtype != torch.float32: logp = logp.float() # get index k, of max prob v, k = logp.max(0) k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. d_v, d_k = duration_logp.max(0) d_k = d_k.item() skip = self.durations[d_k] if self.preserve_alignments: # insert logprobs into last timestep hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) if self.preserve_frame_confidence: # insert confidence into last timestep hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) del logp # If blank token is predicted, exit inner loop, move onto next timestep t if k == self._blank_index: not_blank = False else: # Append token to label set, update RNN state. hypothesis.y_sequence.append(k) hypothesis.score += float(v) hypothesis.timestep.append(time_idx) hypothesis.dec_state = hidden_prime hypothesis.last_token = k # Increment token counter. symbols_added += 1 time_idx += skip need_loop = skip == 0 # this rarely happens, but we manually increment the `skip` number # if blank is emitted and duration=0 is predicted. This prevents possible # infinite loops. if skip == 0: skip = 1 if self.preserve_alignments: # convert Ti-th logits into a torch array hypothesis.alignments.append([]) # blank buffer for next timestep if self.preserve_frame_confidence: hypothesis.frame_confidence.append([]) # blank buffer for next timestep if symbols_added == self.max_symbols: time_idx += 1 # Remove trailing empty list of Alignments if self.preserve_alignments: if len(hypothesis.alignments[-1]) == 0: del hypothesis.alignments[-1] # Remove trailing empty list of per-frame confidence if self.preserve_frame_confidence: if len(hypothesis.frame_confidence[-1]) == 0: del hypothesis.frame_confidence[-1] # Unpack the hidden states hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) return hypothesis class GreedyBatchedTDTInfer(_GreedyRNNTInfer): """A batch level greedy TDT decoder. Batch level greedy decoding, performed auto-regressively. Args: decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. joint_model: rnnt_utils.AbstractRNNTJoint implementation. blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. durations: a list containing durations. max_symbols_per_step: Optional int. The maximum number of symbols that can be added to a sequence in a single time step; if set to None then there is no limit. preserve_alignments: Bool flag which preserves the history of alignments generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `alignments` in it. Here, `alignments` is a List of List of Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. U is the number of target tokens for the current timestep Ti. preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. The length of the list corresponds to the Acoustic Length (T). Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. U is the number of target tokens for the current timestep Ti. confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ def __init__( self, decoder_model: rnnt_abstract.AbstractRNNTDecoder, joint_model: rnnt_abstract.AbstractRNNTJoint, blank_index: int, durations: List[int], max_symbols_per_step: Optional[int] = None, preserve_alignments: bool = False, preserve_frame_confidence: bool = False, confidence_method_cfg: Optional[DictConfig] = None, ): super().__init__( decoder_model=decoder_model, joint_model=joint_model, blank_index=blank_index, max_symbols_per_step=max_symbols_per_step, preserve_alignments=preserve_alignments, preserve_frame_confidence=preserve_frame_confidence, confidence_method_cfg=confidence_method_cfg, ) self.durations = durations # Depending on availability of `blank_as_pad` support # switch between more efficient batch decoding technique if self.decoder.blank_as_pad: self._greedy_decode = self._greedy_decode_blank_as_pad else: self._greedy_decode = self._greedy_decode_masked @typecheck() def forward( self, encoder_output: torch.Tensor, encoded_lengths: torch.Tensor, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): """Returns a list of hypotheses given an input batch of the encoder hidden embedding. Output token is generated auto-regressively. Args: encoder_output: A tensor of size (batch, features, timesteps). encoded_lengths: list of int representing the length of each sequence output sequence. Returns: packed list containing batch number of sentences (Hypotheses). """ # Preserve decoder and joint training state decoder_training_state = self.decoder.training joint_training_state = self.joint.training with torch.inference_mode(): # Apply optional preprocessing encoder_output = encoder_output.transpose(1, 2) # (B, T, D) logitlen = encoded_lengths self.decoder.eval() self.joint.eval() with self.decoder.as_frozen(), self.joint.as_frozen(): inseq = encoder_output # [B, T, D] hypotheses = self._greedy_decode( inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses ) # Pack the hypotheses results packed_result = pack_hypotheses(hypotheses, logitlen) self.decoder.train(decoder_training_state) self.joint.train(joint_training_state) return (packed_result,) def _greedy_decode_blank_as_pad( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): if partial_hypotheses is not None: raise NotImplementedError("`partial_hypotheses` support is not supported") with torch.inference_mode(): # x: [B, T, D] # out_len: [B] # device: torch.device # Initialize list of Hypothesis batchsize = x.shape[0] hypotheses = [ rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) ] # Initialize Hidden state matrix (shared by entire batch) hidden = None # If alignments need to be preserved, register a danling list to hold the values if self.preserve_alignments: # alignments is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.alignments = [[]] # If confidence scores need to be preserved, register a danling list to hold the values if self.preserve_frame_confidence: # frame_confidence is a 3-dimensional dangling list representing B x T x U for hyp in hypotheses: hyp.frame_confidence = [[]] # Last Label buffer + Last Label without blank buffer # batch level equivalent of the last_label last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) # Mask buffers blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) # Get max sequence length max_out_len = out_len.max() # skip means the number of frames the next decoding step should "jump" to. When skip == 1 # it means the next decoding step will just use the next input frame. skip = 1 for time_idx in range(max_out_len): if skip > 1: # if skip > 1 at the current step, we decrement it and skip the current frame. skip -= 1 continue f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] # need_to_stay is a boolean indicates whether the next decoding step should remain in the same frame. need_to_stay = True symbols_added = 0 # Reset blank mask blank_mask.mul_(False) # Update blank mask with time mask # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len blank_mask = time_idx >= out_len # Start inner loop while need_to_stay and (self.max_symbols is None or symbols_added < self.max_symbols): # Batch prediction and joint network steps # If very first prediction step, submit SOS tag (blank) to pred_step. # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state if time_idx == 0 and symbols_added == 0 and hidden is None: g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) else: # Perform batch step prediction of decoder, getting new states and scores ("g") g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) # Batched joint step - Output = [B, V + 1 + num-big-blanks] # Note: log_normalize must not be True here since the joiner output is contanetation of both token logits and duration logits, # and they need to be normalized independently. joined = self._joint_step(f, g, log_normalize=None) logp = joined[:, 0, 0, : -len(self.durations)] duration_logp = joined[:, 0, 0, -len(self.durations) :] if logp.dtype != torch.float32: logp = logp.float() duration_logp = duration_logp.float() # get the max for both token and duration predictions. v, k = logp.max(1) dv, dk = duration_logp.max(1) # here we set the skip value to be the minimum of all predicted durations, hense the "torch.min(dk)" call there. # Please refer to Section 5.2 of our paper https://arxiv.org/pdf/2304.06795.pdf for explanation of this. skip = self.durations[int(torch.min(dk))] # this is a special case: if all batches emit blanks, we require that skip be at least 1 # so we don't loop forever at the current frame. if blank_mask.all(): if skip == 0: skip = 1 need_to_stay = skip == 0 del g # Update blank mask with current predicted blanks # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) k_is_blank = k == self._blank_index blank_mask.bitwise_or_(k_is_blank) del k_is_blank del logp, duration_logp # If all samples predict / have predicted prior blanks, exit loop early # This is equivalent to if single sample predicted k if not blank_mask.all(): # Collect batch indices where blanks occurred now/past blank_indices = (blank_mask == 1).nonzero(as_tuple=False) # Recover prior state for all samples which predicted blank now/past if hidden is not None: hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) elif len(blank_indices) > 0 and hidden is None: # Reset state if there were some blank and other non-blank predictions in batch # Original state is filled with zeros so we just multiply # LSTM has 2 states hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) # Recover prior predicted label for all samples which predicted blank now/past k[blank_indices] = last_label[blank_indices, 0] # Update new label and hidden state for next iteration last_label = k.clone().view(-1, 1) hidden = hidden_prime # Update predicted labels, accounting for time mask # If blank was predicted even once, now or in the past, # Force the current predicted label to also be blank # This ensures that blanks propogate across all timesteps # once they have occured (normally stopping condition of sample level loop). for kidx, ki in enumerate(k): if blank_mask[kidx] == 0: hypotheses[kidx].y_sequence.append(ki) hypotheses[kidx].timestep.append(time_idx) hypotheses[kidx].score += float(v[kidx]) symbols_added += 1 # Remove trailing empty list of alignments at T_{am-len} x Uj if self.preserve_alignments: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].alignments[-1]) == 0: del hypotheses[batch_idx].alignments[-1] # Remove trailing empty list of confidence scores at T_{am-len} x Uj if self.preserve_frame_confidence: for batch_idx in range(batchsize): if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: del hypotheses[batch_idx].frame_confidence[-1] # Preserve states for batch_idx in range(batchsize): hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) return hypotheses def _greedy_decode_masked( self, x: torch.Tensor, out_len: torch.Tensor, device: torch.device, partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, ): raise NotImplementedError("masked greedy-batched decode is not supported for TDT models.") [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math from abc import ABC, abstractmethod from dataclasses import dataclass from functools import partial from typing import List, Optional import torch from omegaconf import DictConfig, OmegaConf from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis from nemo.utils import logging class ConfidenceMethodConstants: NAMES = ("max_prob", "entropy") ENTROPY_TYPES = ("gibbs", "tsallis", "renyi") ENTROPY_NORMS = ("lin", "exp") @classmethod def print(cls): return ( cls.__name__ + ": " + str({"NAMES": cls.NAMES, "ENTROPY_TYPES": cls.ENTROPY_TYPES, "ENTROPY_NORMS": cls.ENTROPY_NORMS}) ) class ConfidenceConstants: AGGREGATIONS = ("mean", "min", "max", "prod") @classmethod def print(cls): return cls.__name__ + ": " + str({"AGGREGATIONS": cls.AGGREGATIONS}) @dataclass class ConfidenceMethodConfig: """A Config which contains the method name and settings to compute per-frame confidence scores. Args: name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ name: str = "entropy" entropy_type: str = "tsallis" alpha: float = 0.33 entropy_norm: str = "exp" temperature: str = "DEPRECATED" def __post_init__(self): if self.temperature != "DEPRECATED": # self.temperature has type str self.alpha = float(self.temperature) self.temperature = "DEPRECATED" if self.name not in ConfidenceMethodConstants.NAMES: raise ValueError( f"`name` must be one of the following: " f"{'`' + '`, `'.join(ConfidenceMethodConstants.NAMES) + '`'}. Provided: `{self.name}`" ) if self.entropy_type not in ConfidenceMethodConstants.ENTROPY_TYPES: raise ValueError( f"`entropy_type` must be one of the following: " f"{'`' + '`, `'.join(ConfidenceMethodConstants.ENTROPY_TYPES) + '`'}. Provided: `{self.entropy_type}`" ) if self.alpha <= 0.0: raise ValueError(f"`alpha` must be > 0. Provided: {self.alpha}") if self.entropy_norm not in ConfidenceMethodConstants.ENTROPY_NORMS: raise ValueError( f"`entropy_norm` must be one of the following: " f"{'`' + '`, `'.join(ConfidenceMethodConstants.ENTROPY_NORMS) + '`'}. Provided: `{self.entropy_norm}`" ) @dataclass class ConfidenceConfig: """A config which contains the following key-value pairs related to confidence scores. Args: preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated during decoding. When set to true, the Hypothesis will contain the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. The length of the list corresponds to the number of recognized tokens. preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. The length of the list corresponds to the number of recognized words. exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded from the `token_confidence`. aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. Valid options are `mean`, `min`, `max`, `prod`. method_cfg: A dict-like object which contains the method name and settings to compute per-frame confidence scores. name: The method name (str). Supported values: - 'max_prob' for using the maximum token probability as a confidence. - 'entropy' for using a normalized entropy of a log-likelihood vector. entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. Supported values: - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). Note that for this entropy, the alpha should comply the following inequality: (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) where V is the model vocabulary size. - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/Tsallis_entropy - 'renyi' for the Rรฉnyi entropy. Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. When the alpha equals one, scaling is not applied to 'max_prob', and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) entropy_norm: A mapping of the entropy value to the interval [0,1]. Supported values: - 'lin' for using the linear mapping. - 'exp' for using exponential mapping with linear shift. """ preserve_frame_confidence: bool = False preserve_token_confidence: bool = False preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed self.method_cfg = OmegaConf.structured( self.method_cfg if isinstance(self.method_cfg, ConfidenceMethodConfig) else ConfidenceMethodConfig(**self.method_cfg) ) if self.aggregation not in ConfidenceConstants.AGGREGATIONS: raise ValueError( f"`aggregation` has to be one of the following: " f"{'`' + '`, `'.join(ConfidenceConstants.AGGREGATIONS) + '`'}. Provided: `{self.aggregation}`" ) def get_confidence_measure_bank(): """Generate a dictionary with confidence measure functionals. Supported confidence measures: max_prob: normalized maximum probability entropy_gibbs_lin: Gibbs entropy with linear normalization entropy_gibbs_exp: Gibbs entropy with exponential normalization entropy_tsallis_lin: Tsallis entropy with linear normalization entropy_tsallis_exp: Tsallis entropy with exponential normalization entropy_renyi_lin: Rรฉnyi entropy with linear normalization entropy_renyi_exp: Rรฉnyi entropy with exponential normalization Returns: dictionary with lambda functions. """ # helper functions # Gibbs entropy is implemented without alpha neg_entropy_gibbs = lambda x: (x.exp() * x).sum(-1) neg_entropy_alpha = lambda x, t: (x * t).exp().sum(-1) neg_entropy_alpha_gibbs = lambda x, t: ((x * t).exp() * x).sum(-1) # too big for a lambda def entropy_tsallis_exp(x, v, t): exp_neg_max_ent = math.exp((1 - math.pow(v, 1 - t)) / (1 - t)) return (((1 - neg_entropy_alpha(x, t)) / (1 - t)).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) def entropy_gibbs_exp(x, v, t): exp_neg_max_ent = math.pow(v, -t * math.pow(v, 1 - t)) return ((neg_entropy_alpha_gibbs(x, t) * t).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) # use Gibbs entropies for Tsallis and Rรฉnyi with t == 1.0 entropy_gibbs_lin_baseline = lambda x, v: 1 + neg_entropy_gibbs(x) / math.log(v) entropy_gibbs_exp_baseline = lambda x, v: (neg_entropy_gibbs(x).exp() * v - 1) / (v - 1) # fill the measure bank confidence_measure_bank = {} # Maximum probability measure is implemented without alpha confidence_measure_bank["max_prob"] = ( lambda x, v, t: (x.max(dim=-1)[0].exp() * v - 1) / (v - 1) if t == 1.0 else ((x.max(dim=-1)[0] * t).exp() * math.pow(v, t) - 1) / (math.pow(v, t) - 1) ) confidence_measure_bank["entropy_gibbs_lin"] = ( lambda x, v, t: entropy_gibbs_lin_baseline(x, v) if t == 1.0 else 1 + neg_entropy_alpha_gibbs(x, t) / math.log(v) / math.pow(v, 1 - t) ) confidence_measure_bank["entropy_gibbs_exp"] = ( lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_gibbs_exp(x, v, t) ) confidence_measure_bank["entropy_tsallis_lin"] = ( lambda x, v, t: entropy_gibbs_lin_baseline(x, v) if t == 1.0 else 1 + (1 - neg_entropy_alpha(x, t)) / (math.pow(v, 1 - t) - 1) ) confidence_measure_bank["entropy_tsallis_exp"] = ( lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_tsallis_exp(x, v, t) ) confidence_measure_bank["entropy_renyi_lin"] = ( lambda x, v, t: entropy_gibbs_lin_baseline(x, v) if t == 1.0 else 1 + neg_entropy_alpha(x, t).log2() / (t - 1) / math.log(v, 2) ) confidence_measure_bank["entropy_renyi_exp"] = ( lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else (neg_entropy_alpha(x, t).pow(1 / (t - 1)) * v - 1) / (v - 1) ) return confidence_measure_bank def get_confidence_aggregation_bank(): """Generate a dictionary with confidence aggregation functions. Supported confidence aggregation functions: min: minimum max: maximum mean: arithmetic mean prod: product Returns: dictionary with functions. """ confidence_aggregation_bank = {"mean": lambda x: sum(x) / len(x), "min": min, "max": max} # python 3.7 and earlier do not have math.prod if hasattr(math, "prod"): confidence_aggregation_bank["prod"] = math.prod else: import operator from functools import reduce confidence_aggregation_bank["prod"] = lambda x: reduce(operator.mul, x, 1) return confidence_aggregation_bank class ConfidenceMethodMixin(ABC): """Confidence Method Mixin class. It initializes per-frame confidence method. """ def _init_confidence_method(self, confidence_method_cfg: Optional[DictConfig] = None): """Initialize per-frame confidence method from config. """ # OmegaConf.structured ensures that post_init check is always executed confidence_method_cfg = OmegaConf.structured( ConfidenceMethodConfig() if confidence_method_cfg is None else ConfidenceMethodConfig(**confidence_method_cfg) ) # set confidence calculation method # we suppose that self.blank_id == len(vocabulary) self.num_tokens = (self.blank_id if hasattr(self, "blank_id") else self._blank_index) + 1 self.alpha = confidence_method_cfg.alpha # init confidence measure bank self.confidence_measure_bank = get_confidence_measure_bank() measure = None # construct measure_name measure_name = "" if confidence_method_cfg.name == "max_prob": measure_name = "max_prob" elif confidence_method_cfg.name == "entropy": measure_name = '_'.join( [confidence_method_cfg.name, confidence_method_cfg.entropy_type, confidence_method_cfg.entropy_norm] ) else: raise ValueError(f"Unsupported `confidence_method_cfg.name`: `{confidence_method_cfg.name}`") if measure_name not in self.confidence_measure_bank: raise ValueError(f"Unsupported measure setup: `{measure_name}`") measure = partial(self.confidence_measure_bank[measure_name], v=self.num_tokens, t=self.alpha) self._get_confidence = lambda x: measure(torch.nan_to_num(x)).tolist() class ConfidenceMixin(ABC): """Confidence Mixin class. It is responsible for confidence estimation method initialization and high-level confidence score calculation. """ def _init_confidence(self, confidence_cfg: Optional[DictConfig] = None): """Initialize confidence-related fields and confidence aggregation function from config. """ # OmegaConf.structured ensures that post_init check is always executed confidence_cfg = OmegaConf.structured( ConfidenceConfig() if confidence_cfg is None else ConfidenceConfig(**confidence_cfg) ) self.confidence_method_cfg = confidence_cfg.method_cfg # extract the config self.preserve_word_confidence = confidence_cfg.get('preserve_word_confidence', False) # set preserve_frame_confidence and preserve_token_confidence to True # if preserve_word_confidence is True self.preserve_token_confidence = ( confidence_cfg.get('preserve_token_confidence', False) | self.preserve_word_confidence ) # set preserve_frame_confidence to True if preserve_token_confidence is True self.preserve_frame_confidence = ( confidence_cfg.get('preserve_frame_confidence', False) | self.preserve_token_confidence ) self.exclude_blank_from_confidence = confidence_cfg.get('exclude_blank', True) self.word_confidence_aggregation = confidence_cfg.get('aggregation', "min") # define aggregation functions self.confidence_aggregation_bank = get_confidence_aggregation_bank() self._aggregate_confidence = self.confidence_aggregation_bank[self.word_confidence_aggregation] # Update preserve frame confidence if self.preserve_frame_confidence is False: if self.cfg.strategy in ['greedy', 'greedy_batch']: self.preserve_frame_confidence = self.cfg.greedy.get('preserve_frame_confidence', False) # OmegaConf.structured ensures that post_init check is always executed confidence_method_cfg = OmegaConf.structured(self.cfg.greedy).get('confidence_method_cfg', None) self.confidence_method_cfg = ( OmegaConf.structured(ConfidenceMethodConfig()) if confidence_method_cfg is None else OmegaConf.structured(ConfidenceMethodConfig(**confidence_method_cfg)) ) @abstractmethod def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: """Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. Assumes that `frame_confidence` is present in the hypotheses. Args: hypotheses_list: List of Hypothesis. Returns: A list of hypotheses with high-level confidence scores. """ raise NotImplementedError() @abstractmethod def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: """Implemented by subclass in order to aggregate token confidence to a word-level confidence. Args: hypothesis: Hypothesis Returns: A list of word-level confidence scores. """ raise NotImplementedError() def _aggregate_token_confidence_chars(self, words: List[str], token_confidence: List[float]) -> List[float]: """Implementation of token confidence aggregation for character-based models. Args: words: List of words of a hypothesis. token_confidence: List of token-level confidence scores of a hypothesis. Returns: A list of word-level confidence scores. """ word_confidence = [] i = 0 for word in words: word_len = len(word) word_confidence.append(self._aggregate_confidence(token_confidence[i : i + word_len])) # we assume that there is exactly one space token between words and exclude it from word confidence i += word_len + 1 return word_confidence def _aggregate_token_confidence_subwords_sentencepiece( self, words: List[str], token_confidence: List[float], token_ids: List[int] ) -> List[float]: """Implementation of token confidence aggregation for subword-based models. **Note**: Only supports Sentencepiece based tokenizers ! Args: words: List of words of a hypothesis. token_confidence: List of token-level confidence scores of a hypothesis. token_ids: List of token ids of a hypothesis. Returns: A list of word-level confidence scores. """ word_confidence = [] # run only if there are final words if len(words) > 0: j = 0 prev_unk = False prev_underline = False for i, token_id in enumerate(token_ids): token = self.decode_ids_to_tokens([int(token_id)])[0] token_text = self.decode_tokens_to_str([int(token_id)]) # treat `<unk>` as a separate word regardless of the next token # to match the result of `tokenizer.ids_to_text` if (token != token_text or prev_unk) and i > j: # do not add confidence for `โ–` if the current token starts with `โ–` # to match the result of `tokenizer.ids_to_text` if not prev_underline: word_confidence.append(self._aggregate_confidence(token_confidence[j:i])) j = i prev_unk = token == '<unk>' prev_underline = token == 'โ–' if not prev_underline: word_confidence.append(self._aggregate_confidence(token_confidence[j : len(token_ids)])) if len(words) != len(word_confidence): raise RuntimeError( f"""Something went wrong with word-level confidence aggregation.\n Please check these values for debugging:\n len(words): {len(words)},\n len(word_confidence): {len(word_confidence)},\n recognized text: `{' '.join(words)}`""" ) return word_confidence [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of nemo/collections/common/parts/adapter_modules.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass, is_dataclass from typing import Any, Optional from hydra.utils import instantiate from omegaconf import OmegaConf from torch import nn as nn from nemo.collections.common.parts.utils import activation_registry from nemo.core.classes.mixins import access_mixins, adapter_mixin_strategies class AdapterModuleUtil(access_mixins.AccessMixin): """ Base class of Adapter Modules, providing common functionality to all Adapter Modules. """ def setup_adapter_strategy(self, adapter_strategy: Optional[adapter_mixin_strategies.AbstractAdapterStrategy]): """ Setup adapter strategy of this class, enabling dynamic change in the way the adapter output is merged with the input. When called successfully, will assign the variable `adapter_strategy` to the module. Args: adapter_strategy: Can be a None or an implementation of AbstractAdapterStrategy. """ # set default adapter strategy if adapter_strategy is None: adapter_strategy = self.get_default_strategy_config() if is_dataclass(adapter_strategy): adapter_strategy = OmegaConf.structured(adapter_strategy) OmegaConf.set_struct(adapter_strategy, False) # The config must have the `_target_` field pointing to the actual adapter strategy class # which will load that strategy dynamically to this module. if isinstance(adapter_strategy, dict) or OmegaConf.is_config(adapter_strategy): self.adapter_strategy = instantiate(adapter_strategy) elif isinstance(adapter_strategy, adapter_mixin_strategies.AbstractAdapterStrategy): self.adapter_strategy = adapter_strategy else: raise AttributeError(f'`adapter_strategy` provided is invalid : {adapter_strategy}') def get_default_strategy_config(self) -> 'dataclass': """ Returns a default adapter module strategy. """ return adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() def adapter_unfreeze(self,): """ Sets the requires grad for all parameters in the adapter to True. This method should be overridden for any custom unfreeze behavior that is required. For example, if not all params of the adapter should be unfrozen. """ for param in self.parameters(): param.requires_grad_(True) class LinearAdapter(nn.Module, AdapterModuleUtil): """ Simple Linear Feedforward Adapter module with LayerNorm and singe hidden layer with activation function. Note: The adapter explicitly initializes its final layer with all zeros in order to avoid affecting the original model when all adapters are disabled. Args: in_features: Input dimension of the module. Note that for adapters, input_dim == output_dim. dim: Hidden dimension of the feed forward network. activation: Str name for an activation function. norm_position: Str, can be `pre` or `post`. Defaults to `pre`. Determines whether the normalization will occur in the first layer or the last layer. Certain architectures may prefer one over the other. dropout: float value, whether to perform dropout on the output of the last layer of the adapter. adapter_strategy: By default, ResidualAddAdapterStrategyConfig. An adapter composition function object. """ def __init__( self, in_features: int, dim: int, activation: str = 'swish', norm_position: str = 'pre', dropout: float = 0.0, adapter_strategy: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig = None, ): super().__init__() activation = activation_registry[activation]() # If the activation can be executed in place, do so. if hasattr(activation, 'inplace'): activation.inplace = True assert norm_position in ['pre', 'post'] self.norm_position = norm_position if norm_position == 'pre': self.module = nn.Sequential( nn.LayerNorm(in_features), nn.Linear(in_features, dim, bias=False), activation, nn.Linear(dim, in_features, bias=False), ) elif norm_position == 'post': self.module = nn.Sequential( nn.Linear(in_features, dim, bias=False), activation, nn.Linear(dim, in_features, bias=False), nn.LayerNorm(in_features), ) if dropout > 0.0: self.dropout = nn.Dropout(dropout) else: self.dropout = None # Setup adapter strategy self.setup_adapter_strategy(adapter_strategy) # reset parameters self.reset_parameters() def reset_parameters(self): # Final layer initializations must be 0 if self.norm_position == 'pre': self.module[-1].weight.data *= 0 elif self.norm_position == 'post': self.module[-1].weight.data *= 0 self.module[-1].bias.data *= 0 def forward(self, x): x = self.module(x) # Add dropout if available if self.dropout is not None: x = self.dropout(x) return x @dataclass class LinearAdapterConfig: in_features: int dim: int activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import re from typing import List import ipadic import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer class EnJaProcessor: """ Tokenizer, Detokenizer and Normalizer utilities for Japanese & English Args: lang_id: One of ['en', 'ja']. """ def __init__(self, lang_id: str): self.lang_id = lang_id self.moses_tokenizer = MosesTokenizer(lang=lang_id) self.moses_detokenizer = MosesDetokenizer(lang=lang_id) self.normalizer = MosesPunctNormalizer( lang=lang_id, pre_replace_unicode_punct=True, post_remove_control_chars=True ) def detokenize(self, tokens: List[str]) -> str: """ Detokenizes a list of tokens Args: tokens: list of strings as tokens Returns: detokenized Japanese or English string """ return self.moses_detokenizer.detokenize(tokens) def tokenize(self, text) -> str: """ Tokenizes text using Moses. Returns a string of tokens. """ tokens = self.moses_tokenizer.tokenize(text) return ' '.join(tokens) def normalize(self, text) -> str: # Normalization doesn't handle Japanese periods correctly; # 'ใ€‚'becomes '.'. if self.lang_id == 'en': return self.normalizer.normalize(text) else: return text class JaMecabProcessor: """ Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English """ def __init__(self): self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: RE_WS_IN_FW = re.compile( r'([\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])\s+(?=[\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])' ) detokenize = lambda s: spacing(RE_WS_IN_FW.sub(r'\1', s)).strip() return detokenize(' '.join(text)) def tokenize(self, text) -> str: """ Tokenizes text using Moses. Returns a string of tokens. """ return self.mecab_tokenizer.parse(text).strip() def normalize(self, text) -> str: return text [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING from nemo.collections.nlp.data.machine_translation.machine_translation_dataset import TranslationDataConfig from nemo.collections.nlp.models.enc_dec_nlp_model import EncDecNLPModelConfig from nemo.collections.nlp.modules.common.token_classifier import TokenClassifierConfig from nemo.collections.nlp.modules.common.tokenizer_utils import TokenizerConfig from nemo.collections.nlp.modules.common.transformer.transformer import ( NeMoTransformerConfig, NeMoTransformerEncoderConfig, ) from nemo.collections.nlp.modules.common.transformer.transformer_bottleneck import ( NeMoTransformerBottleneckDecoderConfig, NeMoTransformerBottleneckEncoderConfig, ) from nemo.core.config.modelPT import OptimConfig, SchedConfig @dataclass class MTSchedConfig(SchedConfig): name: str = 'InverseSquareRootAnnealing' warmup_ratio: Optional[float] = None last_epoch: int = -1 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). @dataclass class MTOptimConfig(OptimConfig): name: str = 'adam' lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 sched: Optional[MTSchedConfig] = MTSchedConfig() @dataclass class MTEncDecModelConfig(EncDecNLPModelConfig): # machine translation configurations num_val_examples: int = 3 num_test_examples: int = 3 max_generation_delta: int = 10 label_smoothing: Optional[float] = 0.0 beam_size: int = 4 len_pen: float = 0.0 src_language: Any = 'en' # Any = str or List[str] tgt_language: Any = 'en' # Any = str or List[str] find_unused_parameters: Optional[bool] = True shared_tokenizer: Optional[bool] = True multilingual: Optional[bool] = False preproc_out_dir: Optional[str] = None validate_input_ids: Optional[bool] = True shared_embeddings: bool = False # network architecture configuration encoder_tokenizer: Any = MISSING encoder: Any = MISSING decoder_tokenizer: Any = MISSING decoder: Any = MISSING head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) # dataset configurations train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( src_file_name=MISSING, tgt_file_name=MISSING, tokens_in_batch=512, clean=True, shuffle=True, cache_ids=False, use_cache=False, ) validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( src_file_name=MISSING, tgt_file_name=MISSING, tokens_in_batch=512, clean=False, shuffle=False, cache_ids=False, use_cache=False, ) test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( src_file_name=MISSING, tgt_file_name=MISSING, tokens_in_batch=512, clean=False, shuffle=False, cache_ids=False, use_cache=False, ) optim: Optional[OptimConfig] = MTOptimConfig() @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( library='nemo', model_name=None, pretrained=False, hidden_size=512, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, ) decoder: NeMoTransformerConfig = NeMoTransformerConfig( library='nemo', model_name=None, pretrained=False, hidden_size=512, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, ) @dataclass class MTBottleneckModelConfig(AAYNBaseConfig): model_type: str = 'nll' min_logv: float = -6 latent_size: int = -1 # -1 will take value of encoder hidden non_recon_warmup_batches: int = 200000 recon_per_token: bool = True log_timing: bool = True encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( library='nemo', model_name=None, pretrained=False, hidden_size=512, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, arch='seq2seq', hidden_steps=32, hidden_blocks=1, hidden_init_method='params', ) decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( library='nemo', model_name=None, pretrained=False, inner_size=2048, num_layers=6, num_attention_heads=8, ffn_dropout=0.1, attn_score_dropout=0.1, attn_layer_dropout=0.1, arch='seq2seq', ) [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict from nemo.collections.common.parts.adapter_modules import LinearAdapterConfig from nemo.collections.nlp.data.token_classification.punctuation_capitalization_dataset import ( PunctuationCapitalizationEvalDataConfig, PunctuationCapitalizationTrainDataConfig, legacy_data_config_to_new_data_config, ) from nemo.core.config import TrainerConfig from nemo.core.config.modelPT import NemoConfig from nemo.utils.exp_manager import ExpManagerConfig @dataclass class FreezeConfig: is_enabled: bool = False """Freeze audio encoder weight and add Conformer Layers on top of it""" d_model: Optional[int] = 256 """`d_model` parameter of ``ConformerLayer``""" d_ff: Optional[int] = 1024 """``d_ff`` parameter of ``ConformerLayer``""" num_layers: Optional[int] = 8 """``num_layers`` number of ``ConformerLayer`` modules to add on top of audio encoder""" @dataclass class AdapterConfig: config: Optional[LinearAdapterConfig] = None """Linear adapter config see ``collections.common.parts.LinearAdapterConfig``""" enable: bool = False """Use adapters for audio encoder""" @dataclass class FusionConfig: num_layers: Optional[int] = 4 """"Number of layers to use in fusion""" num_attention_heads: Optional[int] = 4 """Number of attention heads to use in fusion""" inner_size: Optional[int] = 2048 """Fusion inner size""" @dataclass class AudioEncoderConfig: pretrained_model: str = MISSING """A configuration for restoring pretrained audio encoder""" freeze: Optional[FreezeConfig] = None adapter: Optional[AdapterConfig] = None fusion: Optional[FusionConfig] = None @dataclass class TokenizerConfig: """A structure and default values of source text tokenizer.""" vocab_file: Optional[str] = None """A path to vocabulary file which is used in ``'word'``, ``'char'``, and HuggingFace tokenizers""" tokenizer_name: str = MISSING """A name of the tokenizer used for tokenization of source sequences. Possible options are ``'sentencepiece'``, ``'word'``, ``'char'``, HuggingFace tokenizers (e.g. ``'bert-base-uncased'``). For more options see function ``nemo.collections.nlp.modules.common.get_tokenizer``. The tokenizer must have properties ``cls_id``, ``pad_id``, ``sep_id``, ``unk_id``.""" special_tokens: Optional[Dict[str, str]] = None """A dictionary with special tokens passed to constructors of ``'char'``, ``'word'``, ``'sentencepiece'``, and various HuggingFace tokenizers.""" tokenizer_model: Optional[str] = None """A path to a tokenizer model required for ``'sentencepiece'`` tokenizer.""" @dataclass class LanguageModelConfig: """ A structure and default values of language model configuration of punctuation and capitalization model. BERT like HuggingFace models are supported. Provide a valid ``pretrained_model_name`` and, optionally, you may reinitialize model via ``config_file`` or ``config``. Alternatively you can initialize the language model using ``lm_checkpoint``. This config is a part of :class:`PunctuationCapitalizationModelConfig` config. """ pretrained_model_name: str = MISSING """A mandatory parameter containing name of HuggingFace pretrained model. For example, ``'bert-base-uncased'``.""" config_file: Optional[str] = None """A path to a file with HuggingFace model config which is used to reinitialize language model.""" config: Optional[Dict] = None """A HuggingFace config which is used to reinitialize language model.""" lm_checkpoint: Optional[str] = None """A path to a ``torch`` checkpoint of a language model.""" @dataclass class HeadConfig: """ A structure and default values of configuration of capitalization or punctuation model head. This config defines a multilayer perceptron which is applied to output of a language model. Number of units in the hidden layer is equal to the dimension of the language model. This config is a part of :class:`PunctuationCapitalizationModelConfig` config. """ num_fc_layers: int = 1 """A number of hidden layers in a multilayer perceptron.""" fc_dropout: float = 0.1 """A dropout used in an MLP.""" activation: str = 'relu' """An activation used in hidden layers.""" use_transformer_init: bool = True """Whether to initialize the weights of the classifier head with the approach that was used for language model initialization.""" @dataclass class ClassLabelsConfig: """ A structure and default values of a mandatory part of config which contains names of files which are saved in .nemo checkpoint. These files can also be used for passing label vocabulary to the model. For using them as label vocabularies you will need to provide path these files in parameter ``model.common_dataset_parameters.label_vocab_dir``. Each line in labels files contains 1 label. The values are sorted, ``<line number>==<label id>``, starting from ``0``. A label with ``0`` id must contain neutral label which must be equal to ``model.common_dataset_parameters.pad_label``. This config is a part of :class:`~CommonDatasetParametersConfig`. """ punct_labels_file: str = MISSING """A name of punctuation labels file.""" capit_labels_file: str = MISSING """A name of capitalization labels file.""" @dataclass class CommonDatasetParametersConfig: """ A structure and default values of common dataset parameters config which includes label and loss mask information. If you omit parameters ``punct_label_ids``, ``capit_label_ids``, ``label_vocab_dir``, then labels will be inferred from a training dataset or loaded from a checkpoint. Parameters ``ignore_extra_tokens`` and ``ignore_start_end`` are responsible for forming loss mask. A loss mask defines on which tokens loss is computed. This parameter is a part of config :class:`~PunctuationCapitalizationModelConfig`. """ pad_label: str = MISSING """A mandatory parameter which should contain label used for punctuation and capitalization label padding. It also serves as a neutral label for both punctuation and capitalization. If any of ``punct_label_ids``, ``capit_label_ids`` parameters is provided, then ``pad_label`` must have ``0`` id in them. In addition, if ``label_vocab_dir`` is provided, then ``pad_label`` must be on the first lines in files ``class_labels.punct_labels_file`` and ``class_labels.capit_labels_file``.""" ignore_extra_tokens: bool = False """Whether to compute loss on not first tokens in words. If this parameter is ``True``, then loss mask is ``False`` for all tokens in a word except the first.""" ignore_start_end: bool = True """If ``False``, then loss is computed on [CLS] and [SEP] tokens.""" punct_label_ids: Optional[Dict[str, int]] = None """A dictionary with punctuation label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this parameter and pass label ids through ``class_labels.punct_labels_file`` or let the model to infer label ids from dataset or load them from checkpoint.""" capit_label_ids: Optional[Dict[str, int]] = None """A dictionary with capitalization label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this parameter and pass label ids through ``class_labels.capit_labels_file`` or let model to infer label ids from dataset or load them from checkpoint.""" label_vocab_dir: Optional[str] = None """A path to directory which contains class labels files. See :class:`ClassLabelsConfig`. If this parameter is provided, then labels will be loaded from files which are located in ``label_vocab_dir`` and have names specified in ``model.class_labels`` configuration section. A label specified in ``pad_label`` has to be on the first lines of ``model.class_labels`` files.""" @dataclass class PunctuationCapitalizationModelConfig: """ A configuration of :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model. See an example of model config in `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ This config is a part of :class:`~PunctuationCapitalizationConfig`. """ class_labels: ClassLabelsConfig = ClassLabelsConfig() """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None """A configuration for creating training dataset and data loader.""" validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating validation datasets and data loaders.""" test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" punct_head: HeadConfig = HeadConfig() """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" capit_head: HeadConfig = HeadConfig() """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" tokenizer: Any = TokenizerConfig() """A configuration for source text tokenizer.""" language_model: LanguageModelConfig = LanguageModelConfig() """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For description see `Optimizers <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/core/core.html#optimizers>`_ section in documentation and `primer <https://github.com/NVIDIA/NeMo/blob/main/tutorials/00_NeMo_Primer.ipynb>_ tutorial.""" @dataclass class PunctuationCapitalizationLexicalAudioModelConfig(PunctuationCapitalizationModelConfig): """ A configuration of :class:`~nemo.collections.nlp.models.token_classification.punctuation_lexical_audio_capitalization_model.PunctuationCapitalizationLexicalAudioModel` model. See an example of model config in `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_lexical_audio_config.yaml>`_ Audio encoder can be frozen during training with ``freeze_audio_encoder`` parameter. Adapter can be added to audio encoder with ``use_adapters`` and ``adapter_config`` parameters. More conformer layers can be added on top of pretrained audio encoder with ``frozen_conf_d_model``, ``frozen_conf_d_ff`` and ``frozen_conf_num_layers`` parameters. """ train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None """A configuration for creating training dataset and data loader.""" validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating validation datasets and data loaders.""" test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" audio_encoder: Optional[AudioEncoderConfig] = None restore_lexical_encoder_from: Optional[str] = None """"Path to .nemo checkpoint to load weights from""" # add more comments use_weighted_loss: Optional[bool] = False """If set to ``True`` CrossEntropyLoss will be weighted""" @dataclass class PunctuationCapitalizationConfig(NemoConfig): """ A config for punctuation model training and testing. See an example of full config in `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ """ pretrained_model: Optional[str] = None """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options by calling method :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. """ name: Optional[str] = 'Punctuation_and_Capitalization' """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" do_training: bool = True """Whether to perform training of the model.""" do_testing: bool = False """Whether ot perform testing of the model.""" model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" trainer: Optional[TrainerConfig] = TrainerConfig() """Contains ``Trainer`` Lightning class constructor parameters.""" exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() def is_legacy_model_config(model_cfg: DictConfig) -> bool: """ Test if model config is old style config. Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support tarred datasets. Args: model_cfg: model configuration Returns: whether ``model_config`` is legacy """ return 'common_dataset_parameters' not in model_cfg def legacy_model_config_to_new_model_config(model_cfg: DictConfig) -> DictConfig: """ Transform old style config into :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig`. Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support tarred datasets. Args: model_cfg: old style config Returns: model config which follows dataclass :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig` """ train_ds = model_cfg.get('train_ds') validation_ds = model_cfg.get('validation_ds') test_ds = model_cfg.get('test_ds') dataset = model_cfg.dataset punct_head_config = model_cfg.get('punct_head', {}) capit_head_config = model_cfg.get('capit_head', {}) omega_conf = OmegaConf.structured( PunctuationCapitalizationModelConfig( class_labels=model_cfg.class_labels, common_dataset_parameters=CommonDatasetParametersConfig( pad_label=dataset.pad_label, ignore_extra_tokens=dataset.get( 'ignore_extra_tokens', CommonDatasetParametersConfig.ignore_extra_tokens ), ignore_start_end=dataset.get('ignore_start_end', CommonDatasetParametersConfig.ignore_start_end), punct_label_ids=model_cfg.punct_label_ids, capit_label_ids=model_cfg.capit_label_ids, ), train_ds=None if train_ds is None else legacy_data_config_to_new_data_config(train_ds, dataset, train=True), validation_ds=None if validation_ds is None else legacy_data_config_to_new_data_config(validation_ds, dataset, train=False), test_ds=None if test_ds is None else legacy_data_config_to_new_data_config(test_ds, dataset, train=False), punct_head=HeadConfig( num_fc_layers=punct_head_config.get('punct_num_fc_layers', HeadConfig.num_fc_layers), fc_dropout=punct_head_config.get('fc_dropout', HeadConfig.fc_dropout), activation=punct_head_config.get('activation', HeadConfig.activation), use_transformer_init=punct_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), ), capit_head=HeadConfig( num_fc_layers=capit_head_config.get('capit_num_fc_layers', HeadConfig.num_fc_layers), fc_dropout=capit_head_config.get('fc_dropout', HeadConfig.fc_dropout), activation=capit_head_config.get('activation', HeadConfig.activation), use_transformer_init=capit_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), ), tokenizer=model_cfg.tokenizer, language_model=model_cfg.language_model, optim=model_cfg.optim, ) ) with open_dict(omega_conf): retain_during_legacy_conversion = model_cfg.get('retain_during_legacy_conversion', {}) for key in retain_during_legacy_conversion.keys(): omega_conf[key] = retain_during_legacy_conversion[key] return omega_conf [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Transformer based language model.""" from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( MegatronRetrievalTransformerEncoderModule, ) from nemo.collections.nlp.modules.common.megatron.utils import ( ApexGuardDefaults, init_method_normal, scaled_init_method_normal, ) try: from apex.transformer.enums import AttnMaskType, ModelType HAVE_APEX = True except (ImportError, ModuleNotFoundError): HAVE_APEX = False # fake missing classes with None attributes AttnMaskType = ApexGuardDefaults() ModelType = ApexGuardDefaults() try: from megatron.core import ModelParallelConfig HAVE_MEGATRON_CORE = True except (ImportError, ModuleNotFoundError): ModelParallelConfig = ApexGuardDefaults HAVE_MEGATRON_CORE = False __all__ = [] AVAILABLE_ENCODERS = ["transformer", "perceiver", "retro"] def get_encoder_model( config: ModelParallelConfig, arch, hidden_size, ffn_hidden_size, num_layers, num_attention_heads, apply_query_key_layer_scaling=False, kv_channels=None, init_method=None, scaled_init_method=None, encoder_attn_mask_type=AttnMaskType.padding, pre_process=True, post_process=True, init_method_std=0.02, megatron_amp_O2=False, hidden_dropout=0.1, attention_dropout=0.1, ffn_dropout=0.0, precision=16, fp32_residual_connection=False, activations_checkpoint_method=None, activations_checkpoint_num_layers=1, activations_checkpoint_granularity=None, layernorm_epsilon=1e-5, bias_activation_fusion=True, bias_dropout_add_fusion=True, masked_softmax_fusion=True, persist_layer_norm=False, openai_gelu=False, activation="gelu", onnx_safe=False, bias=True, normalization="layernorm", headscale=False, transformer_block_type="pre_ln", hidden_steps=32, parent_model_type=ModelType.encoder_or_decoder, layer_type=None, chunk_size=64, num_self_attention_per_cross_attention=1, layer_number_offset=0, # this is use only for attention norm_factor scaling megatron_legacy=False, normalize_attention_scores=True, sequence_parallel=False, num_moe_experts=1, moe_frequency=1, moe_dropout=0.0, turn_off_rop=False, # turn off the RoP positional embedding version=1, # model version position_embedding_type='learned_absolute', use_flash_attention=False, ): """Build language model and return along with the key to save.""" if kv_channels is None: assert ( hidden_size % num_attention_heads == 0 ), 'hidden_size must be divisible by num_attention_heads if kv_channels is None' kv_channels = hidden_size // num_attention_heads if init_method is None: init_method = init_method_normal(init_method_std) if scaled_init_method is None: scaled_init_method = scaled_init_method_normal(init_method_std, num_layers) if arch == "transformer": # Language encoder. encoder = MegatronTransformerEncoderModule( config=config, init_method=init_method, output_layer_init_method=scaled_init_method, hidden_size=hidden_size, num_layers=num_layers, num_attention_heads=num_attention_heads, apply_query_key_layer_scaling=apply_query_key_layer_scaling, kv_channels=kv_channels, ffn_hidden_size=ffn_hidden_size, encoder_attn_mask_type=encoder_attn_mask_type, pre_process=pre_process, post_process=post_process, megatron_amp_O2=megatron_amp_O2, hidden_dropout=hidden_dropout, attention_dropout=attention_dropout, ffn_dropout=ffn_dropout, precision=precision, fp32_residual_connection=fp32_residual_connection, activations_checkpoint_method=activations_checkpoint_method, activations_checkpoint_num_layers=activations_checkpoint_num_layers, activations_checkpoint_granularity=activations_checkpoint_granularity, layernorm_epsilon=layernorm_epsilon, bias_activation_fusion=bias_activation_fusion, bias_dropout_add_fusion=bias_dropout_add_fusion, masked_softmax_fusion=masked_softmax_fusion, persist_layer_norm=persist_layer_norm, openai_gelu=openai_gelu, onnx_safe=onnx_safe, activation=activation, bias=bias, normalization=normalization, transformer_block_type=transformer_block_type, headscale=headscale, parent_model_type=parent_model_type, megatron_legacy=megatron_legacy, normalize_attention_scores=normalize_attention_scores, num_moe_experts=num_moe_experts, moe_frequency=moe_frequency, moe_dropout=moe_dropout, position_embedding_type=position_embedding_type, use_flash_attention=use_flash_attention, ) elif arch == "retro": encoder = MegatronRetrievalTransformerEncoderModule( config=config, init_method=init_method, output_layer_init_method=scaled_init_method, hidden_size=hidden_size, num_layers=num_layers, num_attention_heads=num_attention_heads, apply_query_key_layer_scaling=apply_query_key_layer_scaling, kv_channels=kv_channels, layer_type=layer_type, ffn_hidden_size=ffn_hidden_size, pre_process=pre_process, post_process=post_process, megatron_amp_O2=megatron_amp_O2, hidden_dropout=hidden_dropout, attention_dropout=attention_dropout, precision=precision, fp32_residual_connection=fp32_residual_connection, activations_checkpoint_method=activations_checkpoint_method, activations_checkpoint_num_layers=activations_checkpoint_num_layers, activations_checkpoint_granularity=activations_checkpoint_granularity, layernorm_epsilon=layernorm_epsilon, bias_activation_fusion=bias_activation_fusion, bias_dropout_add_fusion=bias_dropout_add_fusion, masked_softmax_fusion=masked_softmax_fusion, persist_layer_norm=persist_layer_norm, openai_gelu=openai_gelu, onnx_safe=onnx_safe, activation=activation, bias=bias, normalization=normalization, transformer_block_type=transformer_block_type, parent_model_type=parent_model_type, chunk_size=chunk_size, layer_number_offset=layer_number_offset, megatron_legacy=megatron_legacy, normalize_attention_scores=normalize_attention_scores, turn_off_rop=turn_off_rop, version=version, ) elif arch == "perceiver": encoder = MegatronPerceiverEncoderModule( config=config, init_method=init_method, output_layer_init_method=scaled_init_method, hidden_size=hidden_size, num_layers=num_layers, num_attention_heads=num_attention_heads, apply_query_key_layer_scaling=apply_query_key_layer_scaling, kv_channels=kv_channels, ffn_hidden_size=ffn_hidden_size, encoder_attn_mask_type=encoder_attn_mask_type, pre_process=pre_process, post_process=post_process, megatron_amp_O2=megatron_amp_O2, hidden_dropout=hidden_dropout, attention_dropout=attention_dropout, ffn_dropout=ffn_dropout, precision=precision, fp32_residual_connection=fp32_residual_connection, activations_checkpoint_method=activations_checkpoint_method, activations_checkpoint_num_layers=activations_checkpoint_num_layers, activations_checkpoint_granularity=activations_checkpoint_granularity, layernorm_epsilon=layernorm_epsilon, bias_activation_fusion=bias_activation_fusion, bias_dropout_add_fusion=bias_dropout_add_fusion, masked_softmax_fusion=masked_softmax_fusion, persist_layer_norm=persist_layer_norm, openai_gelu=openai_gelu, onnx_safe=onnx_safe, activation=activation, bias=bias, normalization=normalization, transformer_block_type=transformer_block_type, headscale=headscale, parent_model_type=parent_model_type, hidden_steps=hidden_steps, num_self_attention_per_cross_attention=num_self_attention_per_cross_attention, megatron_legacy=megatron_legacy, normalize_attention_scores=normalize_attention_scores, ) else: raise ValueError(f"Unknown encoder arch = {arch}. Available encoder arch = {AVAILABLE_ENCODERS}") return encoder [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/tts/models/fastpitch.py] # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import contextlib from dataclasses import dataclass from pathlib import Path from typing import List, Optional import torch from hydra.utils import instantiate from omegaconf import DictConfig, OmegaConf, open_dict from pytorch_lightning import Trainer from pytorch_lightning.loggers import TensorBoardLogger from nemo.collections.common.parts.preprocessing import parsers from nemo.collections.tts.losses.aligner_loss import BinLoss, ForwardSumLoss from nemo.collections.tts.losses.fastpitchloss import DurationLoss, EnergyLoss, MelLoss, PitchLoss from nemo.collections.tts.models.base import SpectrogramGenerator from nemo.collections.tts.modules.fastpitch import FastPitchModule from nemo.collections.tts.parts.mixins import FastPitchAdapterModelMixin from nemo.collections.tts.parts.utils.callbacks import LoggingCallback from nemo.collections.tts.parts.utils.helpers import ( batch_from_ragged, g2p_backward_compatible_support, plot_alignment_to_numpy, plot_spectrogram_to_numpy, process_batch, sample_tts_input, ) from nemo.core.classes import Exportable from nemo.core.classes.common import PretrainedModelInfo, typecheck from nemo.core.neural_types.elements import ( Index, LengthsType, MelSpectrogramType, ProbsType, RegressionValuesType, TokenDurationType, TokenIndex, TokenLogDurationType, ) from nemo.core.neural_types.neural_type import NeuralType from nemo.utils import logging, model_utils @dataclass class G2PConfig: _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" phoneme_probability: float = 0.5 @dataclass class TextTokenizer: _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" punct: bool = True stresses: bool = True chars: bool = True apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True g2p: G2PConfig = G2PConfig() @dataclass class TextTokenizerConfig: text_tokenizer: TextTokenizer = TextTokenizer() class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" def __init__(self, cfg: DictConfig, trainer: Trainer = None): # Convert to Hydra 1.0 compatible DictConfig cfg = model_utils.convert_model_config_to_dict_config(cfg) cfg = model_utils.maybe_update_config_version(cfg) # Setup normalizer self.normalizer = None self.text_normalizer_call = None self.text_normalizer_call_kwargs = {} self._setup_normalizer(cfg) self.learn_alignment = cfg.get("learn_alignment", False) # Setup vocabulary (=tokenizer) and input_fft_kwargs (supported only with self.learn_alignment=True) input_fft_kwargs = {} if self.learn_alignment: self.vocab = None self.ds_class = cfg.train_ds.dataset._target_ self.ds_class_name = self.ds_class.split(".")[-1] if not self.ds_class in [ "nemo.collections.tts.data.dataset.TTSDataset", "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset", "nemo.collections.tts.torch.data.TTSDataset", ]: raise ValueError(f"Unknown dataset class: {self.ds_class}.") self._setup_tokenizer(cfg) assert self.vocab is not None input_fft_kwargs["n_embed"] = len(self.vocab.tokens) input_fft_kwargs["padding_idx"] = self.vocab.pad self._parser = None self._tb_logger = None super().__init__(cfg=cfg, trainer=trainer) self.bin_loss_warmup_epochs = cfg.get("bin_loss_warmup_epochs", 100) self.log_images = cfg.get("log_images", False) self.log_train_images = False default_prosody_loss_scale = 0.1 if self.learn_alignment else 1.0 dur_loss_scale = cfg.get("dur_loss_scale", default_prosody_loss_scale) pitch_loss_scale = cfg.get("pitch_loss_scale", default_prosody_loss_scale) energy_loss_scale = cfg.get("energy_loss_scale", default_prosody_loss_scale) self.mel_loss_fn = MelLoss() self.pitch_loss_fn = PitchLoss(loss_scale=pitch_loss_scale) self.duration_loss_fn = DurationLoss(loss_scale=dur_loss_scale) self.energy_loss_fn = EnergyLoss(loss_scale=energy_loss_scale) self.aligner = None if self.learn_alignment: aligner_loss_scale = cfg.get("aligner_loss_scale", 1.0) self.aligner = instantiate(self._cfg.alignment_module) self.forward_sum_loss_fn = ForwardSumLoss(loss_scale=aligner_loss_scale) self.bin_loss_fn = BinLoss(loss_scale=aligner_loss_scale) self.preprocessor = instantiate(self._cfg.preprocessor) input_fft = instantiate(self._cfg.input_fft, **input_fft_kwargs) output_fft = instantiate(self._cfg.output_fft) duration_predictor = instantiate(self._cfg.duration_predictor) pitch_predictor = instantiate(self._cfg.pitch_predictor) speaker_encoder = instantiate(self._cfg.get("speaker_encoder", None)) energy_embedding_kernel_size = cfg.get("energy_embedding_kernel_size", 0) energy_predictor = instantiate(self._cfg.get("energy_predictor", None)) # [TODO] may remove if we change the pre-trained config # cfg: condition_types = [ "add" ] n_speakers = cfg.get("n_speakers", 0) speaker_emb_condition_prosody = cfg.get("speaker_emb_condition_prosody", False) speaker_emb_condition_decoder = cfg.get("speaker_emb_condition_decoder", False) speaker_emb_condition_aligner = cfg.get("speaker_emb_condition_aligner", False) min_token_duration = cfg.get("min_token_duration", 0) use_log_energy = cfg.get("use_log_energy", True) if n_speakers > 1 and "add" not in input_fft.cond_input.condition_types: input_fft.cond_input.condition_types.append("add") if speaker_emb_condition_prosody: duration_predictor.cond_input.condition_types.append("add") pitch_predictor.cond_input.condition_types.append("add") if speaker_emb_condition_decoder: output_fft.cond_input.condition_types.append("add") if speaker_emb_condition_aligner and self.aligner is not None: self.aligner.cond_input.condition_types.append("add") self.fastpitch = FastPitchModule( input_fft, output_fft, duration_predictor, pitch_predictor, energy_predictor, self.aligner, speaker_encoder, n_speakers, cfg.symbols_embedding_dim, cfg.pitch_embedding_kernel_size, energy_embedding_kernel_size, cfg.n_mel_channels, min_token_duration, cfg.max_token_duration, use_log_energy, ) self._input_types = self._output_types = None self.export_config = { "emb_range": (0, self.fastpitch.encoder.word_emb.num_embeddings), "enable_volume": False, "enable_ragged_batches": False, } if self.fastpitch.speaker_emb is not None: self.export_config["num_speakers"] = cfg.n_speakers self.log_config = cfg.get("log_config", None) # Adapter modules setup (from FastPitchAdapterModelMixin) self.setup_adapters() def _get_default_text_tokenizer_conf(self): text_tokenizer: TextTokenizerConfig = TextTokenizerConfig() return OmegaConf.create(OmegaConf.to_yaml(text_tokenizer)) def _setup_normalizer(self, cfg): if "text_normalizer" in cfg: normalizer_kwargs = {} if "whitelist" in cfg.text_normalizer: normalizer_kwargs["whitelist"] = self.register_artifact( 'text_normalizer.whitelist', cfg.text_normalizer.whitelist ) try: import nemo_text_processing self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) except Exception as e: logging.error(e) raise ImportError( "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" ) self.text_normalizer_call = self.normalizer.normalize if "text_normalizer_call_kwargs" in cfg: self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs def _setup_tokenizer(self, cfg): text_tokenizer_kwargs = {} if "g2p" in cfg.text_tokenizer: # for backward compatibility if ( self._is_model_being_restored() and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") ): cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( cfg.text_tokenizer.g2p["_target_"] ) g2p_kwargs = {} if "phoneme_dict" in cfg.text_tokenizer.g2p: g2p_kwargs["phoneme_dict"] = self.register_artifact( 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, ) if "heteronyms" in cfg.text_tokenizer.g2p: g2p_kwargs["heteronyms"] = self.register_artifact( 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, ) # for backward compatability text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) # TODO @xueyang: rename the instance of tokenizer because vocab is misleading. self.vocab = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) @property def tb_logger(self): if self._tb_logger is None: if self.logger is None and self.logger.experiment is None: return None tb_logger = self.logger.experiment for logger in self.trainer.loggers: if isinstance(logger, TensorBoardLogger): tb_logger = logger.experiment break self._tb_logger = tb_logger return self._tb_logger @property def parser(self): if self._parser is not None: return self._parser if self.learn_alignment: self._parser = self.vocab.encode else: self._parser = parsers.make_parser( labels=self._cfg.labels, name='en', unk_id=-1, blank_id=-1, do_normalize=True, abbreviation_version="fastpitch", make_table=False, ) return self._parser def parse(self, str_input: str, normalize=True) -> torch.tensor: if self.training: logging.warning("parse() is meant to be called in eval mode.") if normalize and self.text_normalizer_call is not None: str_input = self.text_normalizer_call(str_input, **self.text_normalizer_call_kwargs) if self.learn_alignment: eval_phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): eval_phon_mode = self.vocab.set_phone_prob(prob=1.0) # Disable mixed g2p representation if necessary with eval_phon_mode: tokens = self.parser(str_input) else: tokens = self.parser(str_input) x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device) return x @typecheck( input_types={ "text": NeuralType(('B', 'T_text'), TokenIndex()), "durs": NeuralType(('B', 'T_text'), TokenDurationType()), "pitch": NeuralType(('B', 'T_audio'), RegressionValuesType()), "energy": NeuralType(('B', 'T_audio'), RegressionValuesType(), optional=True), "speaker": NeuralType(('B'), Index(), optional=True), "pace": NeuralType(optional=True), "spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), "attn_prior": NeuralType(('B', 'T_spec', 'T_text'), ProbsType(), optional=True), "mel_lens": NeuralType(('B'), LengthsType(), optional=True), "input_lens": NeuralType(('B'), LengthsType(), optional=True), # reference_* data is used for multi-speaker FastPitch training "reference_spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), "reference_spec_lens": NeuralType(('B'), LengthsType(), optional=True), } ) def forward( self, *, text, durs=None, pitch=None, energy=None, speaker=None, pace=1.0, spec=None, attn_prior=None, mel_lens=None, input_lens=None, reference_spec=None, reference_spec_lens=None, ): return self.fastpitch( text=text, durs=durs, pitch=pitch, energy=energy, speaker=speaker, pace=pace, spec=spec, attn_prior=attn_prior, mel_lens=mel_lens, input_lens=input_lens, reference_spec=reference_spec, reference_spec_lens=reference_spec_lens, ) @typecheck(output_types={"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType())}) def generate_spectrogram( self, tokens: 'torch.tensor', speaker: Optional[int] = None, pace: float = 1.0, reference_spec: Optional['torch.tensor'] = None, reference_spec_lens: Optional['torch.tensor'] = None, ) -> torch.tensor: if self.training: logging.warning("generate_spectrogram() is meant to be called in eval mode.") if isinstance(speaker, int): speaker = torch.tensor([speaker]).to(self.device) spect, *_ = self( text=tokens, durs=None, pitch=None, speaker=speaker, pace=pace, reference_spec=reference_spec, reference_spec_lens=reference_spec_lens, ) return spect def training_step(self, batch, batch_idx): attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( None, None, None, None, None, None, ) if self.learn_alignment: if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": batch_dict = batch else: batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) audio = batch_dict.get("audio") audio_lens = batch_dict.get("audio_lens") text = batch_dict.get("text") text_lens = batch_dict.get("text_lens") attn_prior = batch_dict.get("align_prior_matrix", None) pitch = batch_dict.get("pitch", None) energy = batch_dict.get("energy", None) speaker = batch_dict.get("speaker_id", None) reference_audio = batch_dict.get("reference_audio", None) reference_audio_len = batch_dict.get("reference_audio_lens", None) else: audio, audio_lens, text, text_lens, durs, pitch, speaker = batch mels, spec_len = self.preprocessor(input_signal=audio, length=audio_lens) reference_spec, reference_spec_len = None, None if reference_audio is not None: reference_spec, reference_spec_len = self.preprocessor( input_signal=reference_audio, length=reference_audio_len ) ( mels_pred, _, _, log_durs_pred, pitch_pred, attn_soft, attn_logprob, attn_hard, attn_hard_dur, pitch, energy_pred, energy_tgt, ) = self( text=text, durs=durs, pitch=pitch, energy=energy, speaker=speaker, pace=1.0, spec=mels if self.learn_alignment else None, reference_spec=reference_spec, reference_spec_lens=reference_spec_len, attn_prior=attn_prior, mel_lens=spec_len, input_lens=text_lens, ) if durs is None: durs = attn_hard_dur mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) loss = mel_loss + dur_loss if self.learn_alignment: ctc_loss = self.forward_sum_loss_fn(attn_logprob=attn_logprob, in_lens=text_lens, out_lens=spec_len) bin_loss_weight = min(self.current_epoch / self.bin_loss_warmup_epochs, 1.0) * 1.0 bin_loss = self.bin_loss_fn(hard_attention=attn_hard, soft_attention=attn_soft) * bin_loss_weight loss += ctc_loss + bin_loss pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) loss += pitch_loss + energy_loss self.log("t_loss", loss) self.log("t_mel_loss", mel_loss) self.log("t_dur_loss", dur_loss) self.log("t_pitch_loss", pitch_loss) if energy_tgt is not None: self.log("t_energy_loss", energy_loss) if self.learn_alignment: self.log("t_ctc_loss", ctc_loss) self.log("t_bin_loss", bin_loss) # Log images to tensorboard if self.log_images and self.log_train_images and isinstance(self.logger, TensorBoardLogger): self.log_train_images = False self.tb_logger.add_image( "train_mel_target", plot_spectrogram_to_numpy(mels[0].data.cpu().float().numpy()), self.global_step, dataformats="HWC", ) spec_predict = mels_pred[0].data.cpu().float().numpy() self.tb_logger.add_image( "train_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", ) if self.learn_alignment: attn = attn_hard[0].data.cpu().float().numpy().squeeze() self.tb_logger.add_image( "train_attn", plot_alignment_to_numpy(attn.T), self.global_step, dataformats="HWC", ) soft_attn = attn_soft[0].data.cpu().float().numpy().squeeze() self.tb_logger.add_image( "train_soft_attn", plot_alignment_to_numpy(soft_attn.T), self.global_step, dataformats="HWC", ) return loss def validation_step(self, batch, batch_idx): attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( None, None, None, None, None, None, ) if self.learn_alignment: if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": batch_dict = batch else: batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) audio = batch_dict.get("audio") audio_lens = batch_dict.get("audio_lens") text = batch_dict.get("text") text_lens = batch_dict.get("text_lens") attn_prior = batch_dict.get("align_prior_matrix", None) pitch = batch_dict.get("pitch", None) energy = batch_dict.get("energy", None) speaker = batch_dict.get("speaker_id", None) reference_audio = batch_dict.get("reference_audio", None) reference_audio_len = batch_dict.get("reference_audio_lens", None) else: audio, audio_lens, text, text_lens, durs, pitch, speaker = batch mels, mel_lens = self.preprocessor(input_signal=audio, length=audio_lens) reference_spec, reference_spec_len = None, None if reference_audio is not None: reference_spec, reference_spec_len = self.preprocessor( input_signal=reference_audio, length=reference_audio_len ) # Calculate val loss on ground truth durations to better align L2 loss in time (mels_pred, _, _, log_durs_pred, pitch_pred, _, _, _, attn_hard_dur, pitch, energy_pred, energy_tgt,) = self( text=text, durs=durs, pitch=pitch, energy=energy, speaker=speaker, pace=1.0, spec=mels if self.learn_alignment else None, reference_spec=reference_spec, reference_spec_lens=reference_spec_len, attn_prior=attn_prior, mel_lens=mel_lens, input_lens=text_lens, ) if durs is None: durs = attn_hard_dur mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) loss = mel_loss + dur_loss + pitch_loss + energy_loss val_outputs = { "val_loss": loss, "mel_loss": mel_loss, "dur_loss": dur_loss, "pitch_loss": pitch_loss, "energy_loss": energy_loss if energy_tgt is not None else None, "mel_target": mels if batch_idx == 0 else None, "mel_pred": mels_pred if batch_idx == 0 else None, } self.validation_step_outputs.append(val_outputs) return val_outputs def on_validation_epoch_end(self): collect = lambda key: torch.stack([x[key] for x in self.validation_step_outputs]).mean() val_loss = collect("val_loss") mel_loss = collect("mel_loss") dur_loss = collect("dur_loss") pitch_loss = collect("pitch_loss") self.log("val_loss", val_loss, sync_dist=True) self.log("val_mel_loss", mel_loss, sync_dist=True) self.log("val_dur_loss", dur_loss, sync_dist=True) self.log("val_pitch_loss", pitch_loss, sync_dist=True) if self.validation_step_outputs[0]["energy_loss"] is not None: energy_loss = collect("energy_loss") self.log("val_energy_loss", energy_loss, sync_dist=True) _, _, _, _, _, spec_target, spec_predict = self.validation_step_outputs[0].values() if self.log_images and isinstance(self.logger, TensorBoardLogger): self.tb_logger.add_image( "val_mel_target", plot_spectrogram_to_numpy(spec_target[0].data.cpu().float().numpy()), self.global_step, dataformats="HWC", ) spec_predict = spec_predict[0].data.cpu().float().numpy() self.tb_logger.add_image( "val_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", ) self.log_train_images = True self.validation_step_outputs.clear() # free memory) def _setup_train_dataloader(self, cfg): phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): phon_mode = self.vocab.set_phone_prob(self.vocab.phoneme_probability) with phon_mode: dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) sampler = dataset.get_sampler(cfg.dataloader_params.batch_size) return torch.utils.data.DataLoader( dataset, collate_fn=dataset.collate_fn, sampler=sampler, **cfg.dataloader_params ) def _setup_test_dataloader(self, cfg): phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): phon_mode = self.vocab.set_phone_prob(0.0) with phon_mode: dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): raise ValueError(f"No dataset for {name}") if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): raise ValueError(f"No dataloader_params for {name}") if shuffle_should_be: if 'shuffle' not in cfg.dataloader_params: logging.warning( f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " "config. Manually setting to True" ) with open_dict(cfg.dataloader_params): cfg.dataloader_params.shuffle = True elif not cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") elif cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") if self.ds_class == "nemo.collections.tts.data.dataset.TTSDataset": phon_mode = contextlib.nullcontext() if hasattr(self.vocab, "set_phone_prob"): phon_mode = self.vocab.set_phone_prob(prob=None if name == "val" else self.vocab.phoneme_probability) with phon_mode: dataset = instantiate( cfg.dataset, text_normalizer=self.normalizer, text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, text_tokenizer=self.vocab, ) else: dataset = instantiate(cfg.dataset) return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) def setup_training_data(self, cfg): if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": self._train_dl = self._setup_train_dataloader(cfg) else: self._train_dl = self.__setup_dataloader_from_config(cfg) def setup_validation_data(self, cfg): if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": self._validation_dl = self._setup_test_dataloader(cfg) else: self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="val") def setup_test_data(self, cfg): """Omitted.""" pass def configure_callbacks(self): if not self.log_config: return [] sample_ds_class = self.log_config.dataset._target_ if sample_ds_class != "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": raise ValueError(f"Logging callback only supported for TextToSpeechDataset, got {sample_ds_class}") data_loader = self._setup_test_dataloader(self.log_config) generators = instantiate(self.log_config.generators) log_dir = Path(self.log_config.log_dir) if self.log_config.log_dir else None log_callback = LoggingCallback( generators=generators, data_loader=data_loader, log_epochs=self.log_config.log_epochs, epoch_frequency=self.log_config.epoch_frequency, output_dir=log_dir, loggers=self.trainer.loggers, log_tensorboard=self.log_config.log_tensorboard, log_wandb=self.log_config.log_wandb, ) return [log_callback] @classmethod def list_available_models(cls) -> 'List[PretrainedModelInfo]': """ This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. Returns: List of available pre-trained models. """ list_of_models = [] # en-US, single speaker, 22050Hz, LJSpeech (ARPABET). model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/1.8.1/files/tts_en_fastpitch_align.nemo", description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is ARPABET-based.", class_=cls, ) list_of_models.append(model) # en-US, single speaker, 22050Hz, LJSpeech (IPA). model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch_ipa", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/IPA_1.13.0/files/tts_en_fastpitch_align_ipa.nemo", description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is IPA-based.", class_=cls, ) list_of_models.append(model) # en-US, multi-speaker, 44100Hz, HiFiTTS. model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch_multispeaker", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_multispeaker_fastpitchhifigan/versions/1.10.0/files/tts_en_fastpitch_multispeaker.nemo", description="This model is trained on HiFITTS sampled at 44100Hz with and can be used to generate male and female English voices with an American accent.", class_=cls, ) list_of_models.append(model) # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 21.02 model = PretrainedModelInfo( pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2102", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2102.nemo", description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 21.02 Dataset sampled at 22050Hz and can be used to generate male German voices.", class_=cls, ) list_of_models.append(model) # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 22.10 model = PretrainedModelInfo( pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2210", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2210.nemo", description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 22.10 Dataset sampled at 22050Hz and can be used to generate male German voices.", class_=cls, ) list_of_models.append(model) # de-DE, multi-speaker, 5 speakers, 44100 Hz, HUI-Audio-Corpus-German Clean. model = PretrainedModelInfo( pretrained_model_name="tts_de_fastpitch_multispeaker_5", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitch_multispeaker_5/versions/1.11.0/files/tts_de_fastpitch_multispeaker_5.nemo", description="This model is trained on 5 speakers in HUI-Audio-Corpus-German clean subset sampled at 44100Hz with and can be used to generate male and female German voices.", class_=cls, ) list_of_models.append(model) # es, 174 speakers, 44100Hz, OpenSLR (IPA) model = PretrainedModelInfo( pretrained_model_name="tts_es_fastpitch_multispeaker", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_es_multispeaker_fastpitchhifigan/versions/1.15.0/files/tts_es_fastpitch_multispeaker.nemo", description="This model is trained on 174 speakers in 6 crowdsourced Latin American Spanish OpenSLR datasets sampled at 44100Hz and can be used to generate male and female Spanish voices with Latin American accents.", class_=cls, ) list_of_models.append(model) # zh, single female speaker, 22050Hz, SFSpeech Bilingual Chinese/English dataset, improved model using richer # dict and jieba word segmenter for polyphone disambiguation. model = PretrainedModelInfo( pretrained_model_name="tts_zh_fastpitch_sfspeech", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_zh_fastpitch_hifigan_sfspeech/versions/1.15.0/files/tts_zh_fastpitch_sfspeech.nemo", description="This model is trained on a single female speaker in SFSpeech Bilingual Chinese/English dataset" " sampled at 22050Hz and can be used to generate female Mandarin Chinese voices. It is improved" " using richer dict and jieba word segmenter for polyphone disambiguation.", class_=cls, ) list_of_models.append(model) # en, multi speaker, LibriTTS, 16000 Hz # stft 25ms 10ms matching ASR params # for use during Enhlish ASR training/adaptation model = PretrainedModelInfo( pretrained_model_name="tts_en_fastpitch_for_asr_finetuning", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch_spectrogram_enhancer_for_asr_finetuning/versions/1.20.0/files/tts_en_fastpitch_for_asr_finetuning.nemo", description="This model is trained on LibriSpeech, train-960 subset." " STFT parameters follow those commonly used in ASR: 25 ms window, 10 ms hop." " This model is supposed to be used with its companion SpetrogramEnhancer for " " ASR fine-tuning. Usage for regular TTS tasks is not advised.", class_=cls, ) list_of_models.append(model) return list_of_models # Methods for model exportability def _prepare_for_export(self, **kwargs): super()._prepare_for_export(**kwargs) tensor_shape = ('T') if self.export_config["enable_ragged_batches"] else ('B', 'T') # Define input_types and output_types as required by export() self._input_types = { "text": NeuralType(tensor_shape, TokenIndex()), "pitch": NeuralType(tensor_shape, RegressionValuesType()), "pace": NeuralType(tensor_shape), "volume": NeuralType(tensor_shape, optional=True), "batch_lengths": NeuralType(('B'), optional=True), "speaker": NeuralType(('B'), Index(), optional=True), } self._output_types = { "spect": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "num_frames": NeuralType(('B'), TokenDurationType()), "durs_predicted": NeuralType(('B', 'T'), TokenDurationType()), "log_durs_predicted": NeuralType(('B', 'T'), TokenLogDurationType()), "pitch_predicted": NeuralType(('B', 'T'), RegressionValuesType()), } if self.export_config["enable_volume"]: self._output_types["volume_aligned"] = NeuralType(('B', 'T'), RegressionValuesType()) def _export_teardown(self): self._input_types = self._output_types = None @property def disabled_deployment_input_names(self): """Implement this method to return a set of input names disabled for export""" disabled_inputs = set() if self.fastpitch.speaker_emb is None: disabled_inputs.add("speaker") if not self.export_config["enable_ragged_batches"]: disabled_inputs.add("batch_lengths") if not self.export_config["enable_volume"]: disabled_inputs.add("volume") return disabled_inputs @property def input_types(self): return self._input_types @property def output_types(self): return self._output_types def input_example(self, max_batch=1, max_dim=44): """ Generates input examples for tracing etc. Returns: A tuple of input examples. """ par = next(self.fastpitch.parameters()) inputs = sample_tts_input(self.export_config, par.device, max_batch=max_batch, max_dim=max_dim) if 'enable_ragged_batches' not in self.export_config: inputs.pop('batch_lengths', None) return (inputs,) def forward_for_export(self, text, pitch, pace, volume=None, batch_lengths=None, speaker=None): if self.export_config["enable_ragged_batches"]: text, pitch, pace, volume_tensor, lens = batch_from_ragged( text, pitch, pace, batch_lengths, padding_idx=self.fastpitch.encoder.padding_idx, volume=volume ) if volume is not None: volume = volume_tensor return self.fastpitch.infer(text=text, pitch=pitch, pace=pace, volume=volume, speaker=speaker) def interpolate_speaker( self, original_speaker_1, original_speaker_2, weight_speaker_1, weight_speaker_2, new_speaker_id ): """ This method performs speaker interpolation between two original speakers the model is trained on. Inputs: original_speaker_1: Integer speaker ID of first existing speaker in the model original_speaker_2: Integer speaker ID of second existing speaker in the model weight_speaker_1: Floating point weight associated in to first speaker during weight combination weight_speaker_2: Floating point weight associated in to second speaker during weight combination new_speaker_id: Integer speaker ID of new interpolated speaker in the model """ if self.fastpitch.speaker_emb is None: raise Exception( "Current FastPitch model is not a multi-speaker FastPitch model. Speaker interpolation can only \ be performed with a multi-speaker model" ) n_speakers = self.fastpitch.speaker_emb.weight.data.size()[0] if original_speaker_1 >= n_speakers or original_speaker_2 >= n_speakers or new_speaker_id >= n_speakers: raise Exception( f"Parameters original_speaker_1, original_speaker_2, new_speaker_id should be less than the total \ total number of speakers FastPitch was trained on (n_speakers = {n_speakers})." ) speaker_emb_1 = ( self.fastpitch.speaker_emb(torch.tensor(original_speaker_1, dtype=torch.int32).cuda()).clone().detach() ) speaker_emb_2 = ( self.fastpitch.speaker_emb(torch.tensor(original_speaker_2, dtype=torch.int32).cuda()).clone().detach() ) new_speaker_emb = weight_speaker_1 * speaker_emb_1 + weight_speaker_2 * speaker_emb_2 self.fastpitch.speaker_emb.weight.data[new_speaker_id] = new_speaker_emb [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/tts/models/tacotron2.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import contextlib from dataclasses import dataclass from typing import Any, Dict, List, Optional import torch from hydra.utils import instantiate from omegaconf import MISSING, DictConfig, OmegaConf, open_dict from omegaconf.errors import ConfigAttributeError from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger from torch import nn from nemo.collections.common.parts.preprocessing import parsers from nemo.collections.tts.losses.tacotron2loss import Tacotron2Loss from nemo.collections.tts.models.base import SpectrogramGenerator from nemo.collections.tts.parts.utils.helpers import ( g2p_backward_compatible_support, get_mask_from_lengths, tacotron2_log_to_tb_func, tacotron2_log_to_wandb_func, ) from nemo.core.classes.common import PretrainedModelInfo, typecheck from nemo.core.neural_types.elements import ( AudioSignal, EmbeddedTextType, LengthsType, LogitsType, MelSpectrogramType, SequenceToSequenceAlignmentType, ) from nemo.core.neural_types.neural_type import NeuralType from nemo.utils import logging, model_utils @dataclass class Preprocessor: _target_: str = MISSING pad_value: float = MISSING @dataclass class Tacotron2Config: preprocessor: Preprocessor = Preprocessor() encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING labels: List = MISSING train_ds: Optional[Dict[Any, Any]] = None validation_ds: Optional[Dict[Any, Any]] = None class Tacotron2Model(SpectrogramGenerator): """Tacotron 2 Model that is used to generate mel spectrograms from text""" def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None): # Convert to Hydra 1.0 compatible DictConfig cfg = model_utils.convert_model_config_to_dict_config(cfg) cfg = model_utils.maybe_update_config_version(cfg) # setup normalizer self.normalizer = None self.text_normalizer_call = None self.text_normalizer_call_kwargs = {} self._setup_normalizer(cfg) # setup tokenizer self.tokenizer = None if hasattr(cfg, 'text_tokenizer'): self._setup_tokenizer(cfg) self.num_tokens = len(self.tokenizer.tokens) self.tokenizer_pad = self.tokenizer.pad self.tokenizer_unk = self.tokenizer.oov # assert self.tokenizer is not None else: self.num_tokens = len(cfg.labels) + 3 super().__init__(cfg=cfg, trainer=trainer) schema = OmegaConf.structured(Tacotron2Config) # ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes if isinstance(cfg, dict): cfg = OmegaConf.create(cfg) elif not isinstance(cfg, DictConfig): raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") # Ensure passed cfg is compliant with schema try: OmegaConf.merge(cfg, schema) self.pad_value = cfg.preprocessor.pad_value except ConfigAttributeError: self.pad_value = cfg.preprocessor.params.pad_value logging.warning( "Your config is using an old NeMo yaml configuration. Please ensure that the yaml matches the " "current version in the main branch for future compatibility." ) self._parser = None self.audio_to_melspec_precessor = instantiate(cfg.preprocessor) self.text_embedding = nn.Embedding(self.num_tokens, 512) self.encoder = instantiate(self._cfg.encoder) self.decoder = instantiate(self._cfg.decoder) self.postnet = instantiate(self._cfg.postnet) self.loss = Tacotron2Loss() self.calculate_loss = True @property def parser(self): if self._parser is not None: return self._parser ds_class_name = self._cfg.train_ds.dataset._target_.split(".")[-1] if ds_class_name == "TTSDataset": self._parser = None elif hasattr(self._cfg, "labels"): self._parser = parsers.make_parser( labels=self._cfg.labels, name='en', unk_id=-1, blank_id=-1, do_normalize=True, abbreviation_version="fastpitch", make_table=False, ) else: raise ValueError("Wanted to setup parser, but model does not have necessary paramaters") return self._parser def parse(self, text: str, normalize=True) -> torch.Tensor: if self.training: logging.warning("parse() is meant to be called in eval mode.") if normalize and self.text_normalizer_call is not None: text = self.text_normalizer_call(text, **self.text_normalizer_call_kwargs) eval_phon_mode = contextlib.nullcontext() if hasattr(self.tokenizer, "set_phone_prob"): eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0) with eval_phon_mode: if self.tokenizer is not None: tokens = self.tokenizer.encode(text) else: tokens = self.parser(text) # Old parser doesn't add bos and eos ids, so maunally add it tokens = [len(self._cfg.labels)] + tokens + [len(self._cfg.labels) + 1] tokens_tensor = torch.tensor(tokens).unsqueeze_(0).to(self.device) return tokens_tensor @property def input_types(self): if self.training: return { "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), "token_len": NeuralType(('B'), LengthsType()), "audio": NeuralType(('B', 'T'), AudioSignal()), "audio_len": NeuralType(('B'), LengthsType()), } else: return { "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), "token_len": NeuralType(('B'), LengthsType()), "audio": NeuralType(('B', 'T'), AudioSignal(), optional=True), "audio_len": NeuralType(('B'), LengthsType(), optional=True), } @property def output_types(self): if not self.calculate_loss and not self.training: return { "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "gate_pred": NeuralType(('B', 'T'), LogitsType()), "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), "pred_length": NeuralType(('B'), LengthsType()), } return { "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "gate_pred": NeuralType(('B', 'T'), LogitsType()), "spec_target": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), "spec_target_len": NeuralType(('B'), LengthsType()), "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), } @typecheck() def forward(self, *, tokens, token_len, audio=None, audio_len=None): if audio is not None and audio_len is not None: spec_target, spec_target_len = self.audio_to_melspec_precessor(audio, audio_len) else: if self.training or self.calculate_loss: raise ValueError( f"'audio' and 'audio_len' can not be None when either 'self.training' or 'self.calculate_loss' is True." ) token_embedding = self.text_embedding(tokens).transpose(1, 2) encoder_embedding = self.encoder(token_embedding=token_embedding, token_len=token_len) if self.training: spec_pred_dec, gate_pred, alignments = self.decoder( memory=encoder_embedding, decoder_inputs=spec_target, memory_lengths=token_len ) else: spec_pred_dec, gate_pred, alignments, pred_length = self.decoder( memory=encoder_embedding, memory_lengths=token_len ) spec_pred_postnet = self.postnet(mel_spec=spec_pred_dec) if not self.calculate_loss and not self.training: return spec_pred_dec, spec_pred_postnet, gate_pred, alignments, pred_length return spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments @typecheck( input_types={"tokens": NeuralType(('B', 'T'), EmbeddedTextType())}, output_types={"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType())}, ) def generate_spectrogram(self, *, tokens): self.eval() self.calculate_loss = False token_len = torch.tensor([len(i) for i in tokens]).to(self.device) tensors = self(tokens=tokens, token_len=token_len) spectrogram_pred = tensors[1] if spectrogram_pred.shape[0] > 1: # Silence all frames past the predicted end mask = ~get_mask_from_lengths(tensors[-1]) mask = mask.expand(spectrogram_pred.shape[1], mask.size(0), mask.size(1)) mask = mask.permute(1, 0, 2) spectrogram_pred.data.masked_fill_(mask, self.pad_value) return spectrogram_pred def training_step(self, batch, batch_idx): audio, audio_len, tokens, token_len = batch spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, _ = self.forward( audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len ) loss, _ = self.loss( spec_pred_dec=spec_pred_dec, spec_pred_postnet=spec_pred_postnet, gate_pred=gate_pred, spec_target=spec_target, spec_target_len=spec_target_len, pad_value=self.pad_value, ) output = { 'loss': loss, 'progress_bar': {'training_loss': loss}, 'log': {'loss': loss}, } return output def validation_step(self, batch, batch_idx): audio, audio_len, tokens, token_len = batch spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments = self.forward( audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len ) loss, gate_target = self.loss( spec_pred_dec=spec_pred_dec, spec_pred_postnet=spec_pred_postnet, gate_pred=gate_pred, spec_target=spec_target, spec_target_len=spec_target_len, pad_value=self.pad_value, ) loss = { "val_loss": loss, "mel_target": spec_target, "mel_postnet": spec_pred_postnet, "gate": gate_pred, "gate_target": gate_target, "alignments": alignments, } self.validation_step_outputs.append(loss) return loss def on_validation_epoch_end(self): if self.logger is not None and self.logger.experiment is not None: logger = self.logger.experiment for logger in self.trainer.loggers: if isinstance(logger, TensorBoardLogger): logger = logger.experiment break if isinstance(logger, TensorBoardLogger): tacotron2_log_to_tb_func( logger, self.validation_step_outputs[0].values(), self.global_step, tag="val", log_images=True, add_audio=False, ) elif isinstance(logger, WandbLogger): tacotron2_log_to_wandb_func( logger, self.validation_step_outputs[0].values(), self.global_step, tag="val", log_images=True, add_audio=False, ) avg_loss = torch.stack( [x['val_loss'] for x in self.validation_step_outputs] ).mean() # This reduces across batches, not workers! self.log('val_loss', avg_loss) self.validation_step_outputs.clear() # free memory def _setup_normalizer(self, cfg): if "text_normalizer" in cfg: normalizer_kwargs = {} if "whitelist" in cfg.text_normalizer: normalizer_kwargs["whitelist"] = self.register_artifact( 'text_normalizer.whitelist', cfg.text_normalizer.whitelist ) try: import nemo_text_processing self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) except Exception as e: logging.error(e) raise ImportError( "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" ) self.text_normalizer_call = self.normalizer.normalize if "text_normalizer_call_kwargs" in cfg: self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs def _setup_tokenizer(self, cfg): text_tokenizer_kwargs = {} if "g2p" in cfg.text_tokenizer and cfg.text_tokenizer.g2p is not None: # for backward compatibility if ( self._is_model_being_restored() and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") ): cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( cfg.text_tokenizer.g2p["_target_"] ) g2p_kwargs = {} if "phoneme_dict" in cfg.text_tokenizer.g2p: g2p_kwargs["phoneme_dict"] = self.register_artifact( 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, ) if "heteronyms" in cfg.text_tokenizer.g2p: g2p_kwargs["heteronyms"] = self.register_artifact( 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, ) text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) self.tokenizer = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): raise ValueError(f"No dataset for {name}") if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): raise ValueError(f"No dataloder_params for {name}") if shuffle_should_be: if 'shuffle' not in cfg.dataloader_params: logging.warning( f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " "config. Manually setting to True" ) with open_dict(cfg.dataloader_params): cfg.dataloader_params.shuffle = True elif not cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") elif not shuffle_should_be and cfg.dataloader_params.shuffle: logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") dataset = instantiate( cfg.dataset, text_normalizer=self.normalizer, text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, text_tokenizer=self.tokenizer, ) return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) def setup_training_data(self, cfg): self._train_dl = self.__setup_dataloader_from_config(cfg) def setup_validation_data(self, cfg): self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="validation") @classmethod def list_available_models(cls) -> 'List[PretrainedModelInfo]': """ This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. Returns: List of available pre-trained models. """ list_of_models = [] model = PretrainedModelInfo( pretrained_model_name="tts_en_tacotron2", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_tacotron2/versions/1.10.0/files/tts_en_tacotron2.nemo", description="This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.", class_=cls, aliases=["Tacotron2-22050Hz"], ) list_of_models.append(model) return list_of_models [end of nemo/collections/tts/models/tacotron2.py] [start of nemo/core/config/modelPT.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf import MISSING from nemo.core import config from nemo.core.classes.dataset import DatasetConfig from nemo.utils import exp_manager @dataclass class SchedConfig: name: str = MISSING min_lr: float = 0.0 last_epoch: int = -1 @dataclass class OptimConfig: name: str = MISSING sched: Optional[SchedConfig] = None @dataclass class ModelConfig: """ Model component inside ModelPT """ # ... train_ds: Optional[DatasetConfig] = None validation_ds: Optional[DatasetConfig] = None test_ds: Optional[DatasetConfig] = None optim: Optional[OptimConfig] = None @dataclass class HydraConfig: run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) @dataclass class NemoConfig: name: str = MISSING model: ModelConfig = MISSING trainer: config.TrainerConfig = config.TrainerConfig( strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' ) exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() hydra: HydraConfig = HydraConfig() class ModelConfigBuilder: def __init__(self, model_cfg: ModelConfig): """ Base class for any Model Config Builder. A Model Config Builder is a utility class that accepts a ModelConfig dataclass, and via a set of utility methods (that are implemented by the subclassed ModelConfigBuilder), builds a finalized ModelConfig that can be supplied to a NemoModel dataclass as the `model` component. Subclasses *must* implement the private method `_finalize_cfg`. Inside this method, they must update `self.model_cfg` with all interdependent config options that need to be set (either updated by user explicitly or with their default value). The updated model config must then be preserved in `self.model_cfg`. Example: # Create the config builder config_builder = <subclass>ModelConfigBuilder() # Update the components of the config that are modifiable config_builder.set_X(X) config_builder.set_Y(Y) # Create a "finalized" config dataclass that will contain all the updates # that were specified by the builder model_config = config_builder.build() # Use model config as is (or further update values), then create a new Model model = nemo.<domain>.models.<ModelName>Model(cfg=model_config, trainer=Trainer()) Supported build methods: - set_train_ds: All model configs can accept a subclass of `DatasetConfig` as their training config. Subclasses can override this method to enable auto-complete by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. - set_validation_ds: All model configs can accept a subclass of `DatasetConfig` as their validation config. Subclasses can override this method to enable auto-complete by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. - set_test_ds: All model configs can accept a subclass of `DatasetConfig` as their test config. Subclasses can override this method to enable auto-complete by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. - set_optim: A build method that supports changes to the Optimizer (and optionally, the Scheduler) used for training the model. The function accepts two inputs - `cfg`: A subclass of `OptimizerParams` - any OptimizerParams subclass can be used, in order to select an appropriate Optimizer. Examples: AdamParams. `sched_cfg`: A subclass of `SchedulerParams` - any SchedulerParams subclass can be used, in order to select an appropriate Scheduler. Examples: CosineAnnealingParams. Note that this argument is optional. - build(): The method which should return a "finalized" ModelConfig dataclass. Subclasses *should* always override this method, and update the signature of this method with the return type of the Dataclass, so that it enables autocomplete for the user. Example: def build(self) -> EncDecCTCConfig: return super().build() Any additional build methods must be added by subclasses of ModelConfigBuilder. Args: model_cfg: """ self.model_cfg = model_cfg self.train_ds_cfg = None self.validation_ds_cfg = None self.test_ds_cfg = None self.optim_cfg = None def set_train_ds(self, cfg: Optional[DatasetConfig] = None): self.model_cfg.train_ds = cfg def set_validation_ds(self, cfg: Optional[DatasetConfig] = None): self.model_cfg.validation_ds = cfg def set_test_ds(self, cfg: Optional[DatasetConfig] = None): self.model_cfg.test_ds = cfg def set_optim(self, cfg: config.OptimizerParams, sched_cfg: Optional[config.SchedulerParams] = None): @dataclass class WrappedOptimConfig(OptimConfig, cfg.__class__): pass # Setup optim optim_name = cfg.__class__.__name__.replace("Params", "").lower() wrapped_cfg = WrappedOptimConfig(name=optim_name, sched=None, **vars(cfg)) if sched_cfg is not None: @dataclass class WrappedSchedConfig(SchedConfig, sched_cfg.__class__): pass # Setup scheduler sched_name = sched_cfg.__class__.__name__.replace("Params", "") wrapped_sched_cfg = WrappedSchedConfig(name=sched_name, **vars(sched_cfg)) wrapped_cfg.sched = wrapped_sched_cfg self.model_cfg.optim = wrapped_cfg def _finalize_cfg(self): raise NotImplementedError() def build(self) -> ModelConfig: # validate config self._finalize_cfg() return self.model_cfg [end of nemo/core/config/modelPT.py] [start of nemo/utils/exp_manager.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import glob import os import subprocess import sys import time import warnings from dataclasses import dataclass from datetime import timedelta from pathlib import Path from shutil import copy, move from typing import Any, Dict, List, Optional, Tuple, Union import pytorch_lightning import torch from hydra.core.hydra_config import HydraConfig from hydra.utils import get_original_cwd from omegaconf import DictConfig, OmegaConf, open_dict from pytorch_lightning.callbacks import Callback, ModelCheckpoint from pytorch_lightning.callbacks.early_stopping import EarlyStopping from pytorch_lightning.callbacks.timer import Interval, Timer from pytorch_lightning.loggers import MLFlowLogger, TensorBoardLogger, WandbLogger from pytorch_lightning.loops import _TrainingEpochLoop from pytorch_lightning.strategies.ddp import DDPStrategy from nemo.collections.common.callbacks import EMA from nemo.constants import NEMO_ENV_VARNAME_TESTING, NEMO_ENV_VARNAME_VERSION from nemo.utils import logging, timers from nemo.utils.app_state import AppState from nemo.utils.callbacks import NeMoModelCheckpoint, PreemptionCallback from nemo.utils.env_var_parsing import get_envbool from nemo.utils.exceptions import NeMoBaseException from nemo.utils.get_rank import is_global_rank_zero from nemo.utils.lightning_logger_patch import add_filehandlers_to_pl_logger from nemo.utils.loggers import ClearMLLogger, ClearMLParams, DLLogger, DLLoggerParams, MLFlowParams from nemo.utils.model_utils import uninject_model_parallel_rank class NotFoundError(NeMoBaseException): """ Raised when a file or folder is not found""" class LoggerMisconfigurationError(NeMoBaseException): """ Raised when a mismatch between trainer.logger and exp_manager occurs""" def __init__(self, message): message = ( message + " You can disable lighning's trainer from creating a logger by passing logger=False to its constructor." ) super().__init__(message) class CheckpointMisconfigurationError(NeMoBaseException): """ Raised when a mismatch between trainer.callbacks and exp_manager occurs""" @dataclass class EarlyStoppingParams: monitor: str = "val_loss" # The metric that early stopping should consider. mode: str = "min" # inform early stopping whether to look for increase or decrease in monitor. min_delta: float = 0.001 # smallest change to consider as improvement. patience: int = 10 # how many (continuous) validation cycles to wait with no improvement and stopping training. verbose: bool = True strict: bool = True check_finite: bool = True stopping_threshold: Optional[float] = None divergence_threshold: Optional[float] = None check_on_train_epoch_end: Optional[bool] = None log_rank_zero_only: bool = False @dataclass class CallbackParams: filepath: Optional[str] = None # Deprecated dirpath: Optional[str] = None # If None, exp_manager will attempt to handle the filepath filename: Optional[str] = None # If None, exp_manager will attempt to handle the filepath monitor: Optional[str] = "val_loss" verbose: Optional[bool] = True save_last: Optional[bool] = True save_top_k: Optional[int] = 3 save_weights_only: Optional[bool] = False mode: Optional[str] = "min" auto_insert_metric_name: bool = True every_n_epochs: Optional[int] = 1 every_n_train_steps: Optional[int] = None train_time_interval: Optional[str] = None prefix: Optional[str] = None # If None, exp_manager will attempt to handle the filepath postfix: str = ".nemo" save_best_model: bool = False always_save_nemo: bool = False save_nemo_on_train_end: Optional[bool] = True # Whether to automatically save .nemo file durin on_train_end hook model_parallel_size: Optional[int] = None # tensor parallel size * pipeline parallel size save_on_train_epoch_end: Optional[bool] = False # Save after training, not after validation @dataclass class StepTimingParams: reduction: Optional[str] = "mean" # if True torch.cuda.synchronize() is called on start/stop sync_cuda: Optional[bool] = False # if positive, defines the size of a sliding window for computing mean buffer_size: Optional[int] = 1 @dataclass class EMAParams: enable: Optional[bool] = False decay: Optional[float] = 0.999 cpu_offload: Optional[bool] = False validate_original_weights: Optional[bool] = False every_n_steps: int = 1 @dataclass class ExpManagerConfig: """Experiment Manager config for validation of passed arguments. """ # Log dir creation parameters explicit_log_dir: Optional[str] = None exp_dir: Optional[str] = None name: Optional[str] = None version: Optional[str] = None use_datetime_version: Optional[bool] = True resume_if_exists: Optional[bool] = False resume_past_end: Optional[bool] = False resume_ignore_no_checkpoint: Optional[bool] = False resume_from_checkpoint: Optional[str] = None # Logging parameters create_tensorboard_logger: Optional[bool] = True summary_writer_kwargs: Optional[Dict[Any, Any]] = None create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() create_dllogger_logger: Optional[bool] = False dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() create_clearml_logger: Optional[bool] = False clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() create_early_stopping_callback: Optional[bool] = False early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True ema: Optional[EMAParams] = EMAParams() # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization seconds_to_sleep: float = 5 class TimingCallback(Callback): """ Logs execution time of train/val/test steps """ def __init__(self, timer_kwargs={}): self.timer = timers.NamedTimer(**timer_kwargs) def _on_batch_start(self, name): # reset only if we do not return mean of a sliding window if self.timer.buffer_size <= 0: self.timer.reset(name) self.timer.start(name) def _on_batch_end(self, name, pl_module): self.timer.stop(name) # Set the `batch_size=1` as WAR for `dataloader_iter`, which is not used for any metric pl_module.log( name + ' in s', self.timer[name], on_step=True, on_epoch=False, batch_size=1, prog_bar=(name == "train_step_timing"), ) def on_train_batch_start(self, trainer, pl_module, batch, batch_idx): self._on_batch_start("train_step_timing") def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): self._on_batch_end("train_step_timing", pl_module) def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): self._on_batch_start("validation_step_timing") def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): self._on_batch_end("validation_step_timing", pl_module) def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): self._on_batch_start("test_step_timing") def on_test_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): self._on_batch_end("test_step_timing", pl_module) def on_before_backward(self, trainer, pl_module, loss): self._on_batch_start("train_backward_timing") def on_after_backward(self, trainer, pl_module): self._on_batch_end("train_backward_timing", pl_module) def exp_manager(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None) -> Optional[Path]: """ exp_manager is a helper function used to manage folders for experiments. It follows the pytorch lightning paradigm of exp_dir/model_or_experiment_name/version. If the lightning trainer has a logger, exp_manager will get exp_dir, name, and version from the logger. Otherwise it will use the exp_dir and name arguments to create the logging directory. exp_manager also allows for explicit folder creation via explicit_log_dir. The version can be a datetime string or an integer. Datestime version can be disabled if use_datetime_version is set to False. It optionally creates TensorBoardLogger, WandBLogger, DLLogger, MLFlowLogger, ClearMLLogger, ModelCheckpoint objects from pytorch lightning. It copies sys.argv, and git information if available to the logging directory. It creates a log file for each process to log their output into. exp_manager additionally has a resume feature (resume_if_exists) which can be used to continuing training from the constructed log_dir. When you need to continue the training repeatedly (like on a cluster which you need multiple consecutive jobs), you need to avoid creating the version folders. Therefore from v1.0.0, when resume_if_exists is set to True, creating the version folders is ignored. Args: trainer (pytorch_lightning.Trainer): The lightning trainer. cfg (DictConfig, dict): Can have the following keys: - explicit_log_dir (str, Path): Can be used to override exp_dir/name/version folder creation. Defaults to None, which will use exp_dir, name, and version to construct the logging directory. - exp_dir (str, Path): The base directory to create the logging directory. Defaults to None, which logs to ./nemo_experiments. - name (str): The name of the experiment. Defaults to None which turns into "default" via name = name or "default". - version (str): The version of the experiment. Defaults to None which uses either a datetime string or lightning's TensorboardLogger system of using version_{int}. - use_datetime_version (bool): Whether to use a datetime string for version. Defaults to True. - resume_if_exists (bool): Whether this experiment is resuming from a previous run. If True, it sets trainer._checkpoint_connector._ckpt_path so that the trainer should auto-resume. exp_manager will move files under log_dir to log_dir/run_{int}. Defaults to False. From v1.0.0, when resume_if_exists is True, we would not create version folders to make it easier to find the log folder for next runs. - resume_past_end (bool): exp_manager errors out if resume_if_exists is True and a checkpoint matching ``*end.ckpt`` indicating a previous training run fully completed. This behaviour can be disabled, in which case the ``*end.ckpt`` will be loaded by setting resume_past_end to True. Defaults to False. - resume_ignore_no_checkpoint (bool): exp_manager errors out if resume_if_exists is True and no checkpoint could be found. This behaviour can be disabled, in which case exp_manager will print a message and continue without restoring, by setting resume_ignore_no_checkpoint to True. Defaults to False. - resume_from_checkpoint (str): Can be used to specify a path to a specific checkpoint file to load from. This will override any checkpoint found when resume_if_exists is True. Defaults to None. - create_tensorboard_logger (bool): Whether to create a tensorboard logger and attach it to the pytorch lightning trainer. Defaults to True. - summary_writer_kwargs (dict): A dictionary of kwargs that can be passed to lightning's TensorboardLogger class. Note that log_dir is passed by exp_manager and cannot exist in this dict. Defaults to None. - create_wandb_logger (bool): Whether to create a Weights and Baises logger and attach it to the pytorch lightning trainer. Defaults to False. - wandb_logger_kwargs (dict): A dictionary of kwargs that can be passed to lightning's WandBLogger class. Note that name and project are required parameters if create_wandb_logger is True. Defaults to None. - create_mlflow_logger (bool): Whether to create an MLFlow logger and attach it to the pytorch lightning training. Defaults to False - mlflow_logger_kwargs (dict): optional parameters for the MLFlow logger - create_dllogger_logger (bool): Whether to create an DLLogger logger and attach it to the pytorch lightning training. Defaults to False - dllogger_logger_kwargs (dict): optional parameters for the DLLogger logger - create_clearml_logger (bool): Whether to create an ClearML logger and attach it to the pytorch lightning training. Defaults to False - clearml_logger_kwargs (dict): optional parameters for the ClearML logger - create_checkpoint_callback (bool): Whether to create a ModelCheckpoint callback and attach it to the pytorch lightning trainer. The ModelCheckpoint saves the top 3 models with the best "val_loss", the most recent checkpoint under ``*last.ckpt``, and the final checkpoint after training completes under ``*end.ckpt``. Defaults to True. - create_early_stopping_callback (bool): Flag to decide if early stopping should be used to stop training. Default is False. See EarlyStoppingParams dataclass above. - create_preemption_callback (bool): Flag to decide whether to enable preemption callback to save checkpoints and exit training immediately upon preemption. Default is True. - files_to_copy (list): A list of files to copy to the experiment logging directory. Defaults to None which copies no files. - log_local_rank_0_only (bool): Whether to only create log files for local rank 0. Defaults to False. Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. - log_global_rank_0_only (bool): Whether to only create log files for global rank 0. Defaults to False. Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. - max_time (str): The maximum wall clock time *per run*. This is intended to be used on clusters where you want a checkpoint to be saved after this specified time and be able to resume from that checkpoint. Defaults to None. - seconds_to_sleep (float): seconds to sleep non rank 0 processes for. Used to give enough time for rank 0 to initialize returns: log_dir (Path): The final logging directory where logging files are saved. Usually the concatenation of exp_dir, name, and version. """ # Add rank information to logger # Note: trainer.global_rank and trainer.is_global_zero are not set until trainer.fit, so have to hack around it local_rank = int(os.environ.get("LOCAL_RANK", 0)) global_rank = trainer.node_rank * trainer.num_devices + local_rank logging.rank = global_rank if cfg is None: logging.error("exp_manager did not receive a cfg argument. It will be disabled.") return if trainer.fast_dev_run: logging.info("Trainer was called with fast_dev_run. exp_manager will return without any functionality.") return # Ensure passed cfg is compliant with ExpManagerConfig schema = OmegaConf.structured(ExpManagerConfig) if isinstance(cfg, dict): cfg = OmegaConf.create(cfg) elif not isinstance(cfg, DictConfig): raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True)) cfg = OmegaConf.merge(schema, cfg) error_checks(trainer, cfg) # Ensures that trainer options are compliant with NeMo and exp_manager arguments log_dir, exp_dir, name, version = get_log_dir( trainer=trainer, exp_dir=cfg.exp_dir, name=cfg.name, version=cfg.version, explicit_log_dir=cfg.explicit_log_dir, use_datetime_version=cfg.use_datetime_version, resume_if_exists=cfg.resume_if_exists, ) check_resume( trainer, log_dir, cfg.resume_if_exists, cfg.resume_past_end, cfg.resume_ignore_no_checkpoint, cfg.checkpoint_callback_params.dirpath, cfg.resume_from_checkpoint, ) checkpoint_name = name # If name returned from get_log_dir is "", use cfg.name for checkpointing if checkpoint_name is None or checkpoint_name == '': checkpoint_name = cfg.name or "default" # Set mlflow name if it's not set, before the main name is erased if cfg.create_mlflow_logger and (not cfg.mlflow_logger_kwargs.get("experiment_name", None)): cfg.mlflow_logger_kwargs.experiment_name = cfg.name logging.warning( 'mlflow logger specified but no experiment name set. Using the same as Tensorboard: %s', cfg.mlflow_logger_kwargs.experiment_name, ) cfg.name = name # Used for configure_loggers so that the log_dir is properly set even if name is "" cfg.version = version # update app_state with log_dir, exp_dir, etc app_state = AppState() app_state.log_dir = log_dir app_state.exp_dir = exp_dir app_state.name = name app_state.version = version app_state.checkpoint_name = checkpoint_name app_state.create_checkpoint_callback = cfg.create_checkpoint_callback app_state.checkpoint_callback_params = cfg.checkpoint_callback_params # Create the logging directory if it does not exist os.makedirs(log_dir, exist_ok=True) # Cannot limit creation to global zero as all ranks write to own log file logging.info(f'Experiments will be logged at {log_dir}') trainer._default_root_dir = log_dir if cfg.log_local_rank_0_only is True and cfg.log_global_rank_0_only is True: raise ValueError( f"Cannot set both log_local_rank_0_only and log_global_rank_0_only to True. Please set either one or neither." ) # This is set if the env var NEMO_TESTING is set to True. nemo_testing = get_envbool(NEMO_ENV_VARNAME_TESTING, False) # Handle logging to file log_file = log_dir / f'nemo_log_globalrank-{global_rank}_localrank-{local_rank}.txt' if cfg.log_local_rank_0_only is True and not nemo_testing: if local_rank == 0: logging.add_file_handler(log_file) elif cfg.log_global_rank_0_only is True and not nemo_testing: if global_rank == 0: logging.add_file_handler(log_file) else: # Logs on all ranks. logging.add_file_handler(log_file) # For some reason, LearningRateLogger requires trainer to have a logger. Safer to create logger on all ranks # not just global rank 0. if ( cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger or cfg.create_dllogger_logger or cfg.create_clearml_logger ): configure_loggers( trainer, exp_dir, log_dir, cfg.name, cfg.version, cfg.checkpoint_callback_params, cfg.create_tensorboard_logger, cfg.summary_writer_kwargs, cfg.create_wandb_logger, cfg.wandb_logger_kwargs, cfg.create_mlflow_logger, cfg.mlflow_logger_kwargs, cfg.create_dllogger_logger, cfg.dllogger_logger_kwargs, cfg.create_clearml_logger, cfg.clearml_logger_kwargs, ) # add loggers timing callbacks if cfg.log_step_timing: timing_callback = TimingCallback(timer_kwargs=cfg.step_timing_kwargs or {}) trainer.callbacks.insert(0, timing_callback) if cfg.ema.enable: ema_callback = EMA( decay=cfg.ema.decay, validate_original_weights=cfg.ema.validate_original_weights, cpu_offload=cfg.ema.cpu_offload, every_n_steps=cfg.ema.every_n_steps, ) trainer.callbacks.append(ema_callback) if cfg.create_early_stopping_callback: early_stop_callback = EarlyStopping(**cfg.early_stopping_callback_params) trainer.callbacks.append(early_stop_callback) if cfg.create_checkpoint_callback: configure_checkpointing( trainer, log_dir, checkpoint_name, cfg.resume_if_exists, cfg.checkpoint_callback_params, cfg.create_preemption_callback, ) if cfg.disable_validation_on_resume: # extend training loop to skip initial validation when resuming from checkpoint configure_no_restart_validation_training_loop(trainer) # Setup a stateless timer for use on clusters. if cfg.max_time_per_run is not None: found_ptl_timer = False for idx, callback in enumerate(trainer.callbacks): if isinstance(callback, Timer): # NOTE: PTL does not expose a `trainer.max_time`. By the time we are in this function, PTL has already setup a timer if the user specifies `trainer.max_time` so best we can do is replace that. # Working: If only `trainer.max_time` is set - it behaves as a normal PTL timer. If only `exp_manager.max_time_per_run` is set - it behaves as a StateLessTimer. If both are set, it also behaves as a StateLessTimer. logging.warning( f'Found a PTL Timer callback, replacing with a StatelessTimer callback. This will happen if you set trainer.max_time as well as exp_manager.max_time_per_run.' ) trainer.callbacks[idx] = StatelessTimer(cfg.max_time_per_run) found_ptl_timer = True break if not found_ptl_timer: trainer.max_time = cfg.max_time_per_run trainer.callbacks.append(StatelessTimer(cfg.max_time_per_run)) if is_global_rank_zero(): # Move files_to_copy to folder and add git information if present if cfg.files_to_copy: for _file in cfg.files_to_copy: copy(Path(_file), log_dir) # Create files for cmd args and git info with open(log_dir / 'cmd-args.log', 'w', encoding='utf-8') as _file: _file.write(" ".join(sys.argv)) # Try to get git hash git_repo, git_hash = get_git_hash() if git_repo: with open(log_dir / 'git-info.log', 'w', encoding='utf-8') as _file: _file.write(f'commit hash: {git_hash}') _file.write(get_git_diff()) # Add err_file logging to global_rank zero logging.add_err_file_handler(log_dir / 'nemo_error_log.txt') # Add lightning file logging to global_rank zero add_filehandlers_to_pl_logger(log_dir / 'lightning_logs.txt', log_dir / 'nemo_error_log.txt') elif trainer.num_nodes * trainer.num_devices > 1: # sleep other ranks so rank 0 can finish # doing the initialization such as moving files time.sleep(cfg.seconds_to_sleep) return log_dir def error_checks(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None): """ Checks that the passed trainer is compliant with NeMo and exp_manager's passed configuration. Checks that: - Throws error when hydra has changed the working directory. This causes issues with lightning's DDP - Throws error when trainer has loggers defined but create_tensorboard_logger or create_wandB_logger or create_mlflow_logger or create_dllogger_logger is True - Prints error messages when 1) run on multi-node and not Slurm, and 2) run on multi-gpu without DDP """ if HydraConfig.initialized() and get_original_cwd() != os.getcwd(): raise ValueError( "Hydra changed the working directory. This interferes with ExpManger's functionality. Please pass " "hydra.run.dir=. to your python script." ) if trainer.logger is not None and ( cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger ): raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger, and either " f"create_tensorboard_logger: {cfg.create_tensorboard_logger} or create_wandb_logger: " f"{cfg.create_wandb_logger} or create_mlflow_logger: {cfg.create_mlflow_logger}" f"or create_dllogger_logger: {cfg.create_mlflow_logger} was set to True. " "These can only be used if trainer does not already have a logger." ) if trainer.num_nodes > 1 and not check_slurm(trainer): logging.error( "You are running multi-node training without SLURM handling the processes." " Please note that this is not tested in NeMo and could result in errors." ) if trainer.num_devices > 1 and not isinstance(trainer.strategy, DDPStrategy): logging.error( "You are running multi-gpu without ddp.Please note that this is not tested in NeMo and could result in " "errors." ) def check_resume( trainer: 'pytorch_lightning.Trainer', log_dir: str, resume_if_exists: bool = False, resume_past_end: bool = False, resume_ignore_no_checkpoint: bool = False, dirpath: str = None, resume_from_checkpoint: str = None, ): """Checks that resume=True was used correctly with the arguments pass to exp_manager. Sets trainer._checkpoint_connector._ckpt_path as necessary. Returns: log_dir (Path): The log_dir exp_dir (str): The base exp_dir without name nor version name (str): The name of the experiment version (str): The version of the experiment Raises: NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. ValueError: If resume is True, and there were more than 1 checkpoint could found. """ if not log_dir: raise ValueError(f"Resuming requires the log_dir {log_dir} to be passed to exp_manager") checkpoint = None if resume_from_checkpoint: checkpoint = resume_from_checkpoint if resume_if_exists: # Use <log_dir>/checkpoints/ unless `dirpath` is set checkpoint_dir = Path(dirpath) if dirpath else Path(Path(log_dir) / "checkpoints") # when using distributed checkpointing, checkpoint_dir is a directory of directories # we check for this here dist_checkpoints = [d for d in list(checkpoint_dir.glob("*")) if d.is_dir()] end_dist_checkpoints = [d for d in dist_checkpoints if d.match("*end")] last_dist_checkpoints = [d for d in dist_checkpoints if d.match("*last")] end_checkpoints = end_dist_checkpoints if end_dist_checkpoints else list(checkpoint_dir.rglob("*end.ckpt")) last_checkpoints = last_dist_checkpoints if last_dist_checkpoints else list(checkpoint_dir.rglob("*last.ckpt")) if not checkpoint_dir.exists() or (not len(end_checkpoints) > 0 and not len(last_checkpoints) > 0): if resume_ignore_no_checkpoint: warn = f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. " if checkpoint is None: warn += "Training from scratch." elif checkpoint == resume_from_checkpoint: warn += f"Training from {resume_from_checkpoint}." logging.warning(warn) else: raise NotFoundError( f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. Cannot resume." ) elif len(end_checkpoints) > 0: if resume_past_end: if len(end_checkpoints) > 1: if 'mp_rank' in str(end_checkpoints[0]): checkpoint = end_checkpoints[0] else: raise ValueError(f"Multiple checkpoints {end_checkpoints} that matches *end.ckpt.") else: raise ValueError( f"Found {end_checkpoints[0]} indicating that the last training run has already completed." ) elif len(last_checkpoints) > 1: if 'mp_rank' in str(last_checkpoints[0]) or 'tp_rank' in str(last_checkpoints[0]): checkpoint = last_checkpoints[0] checkpoint = uninject_model_parallel_rank(checkpoint) else: raise ValueError(f"Multiple checkpoints {last_checkpoints} that matches *last.ckpt.") else: checkpoint = last_checkpoints[0] # PTL 2.0 supports ckpt_path instead of resume_from_checkpoint as the trainer flag if checkpoint is not None: trainer.ckpt_path = str(checkpoint) logging.info(f'Resuming training from checkpoint: {trainer.ckpt_path}') if is_global_rank_zero(): # Check to see if any files exist that need to be moved files_to_move = [] if Path(log_dir).exists(): for child in Path(log_dir).iterdir(): if child.is_file(): files_to_move.append(child) if len(files_to_move) > 0: # Move old files to a new folder other_run_dirs = Path(log_dir).glob("run_*") run_count = 0 for fold in other_run_dirs: if fold.is_dir(): run_count += 1 new_run_dir = Path(Path(log_dir) / f"run_{run_count}") new_run_dir.mkdir() for _file in files_to_move: move(str(_file), str(new_run_dir)) def check_explicit_log_dir( trainer: 'pytorch_lightning.Trainer', explicit_log_dir: Union[Path, str], exp_dir: str, name: str, version: str ) -> Tuple[Path, str, str, str]: """ Checks that the passed arguments are compatible with explicit_log_dir. Returns: log_dir (Path): the log_dir exp_dir (str): the base exp_dir without name nor version name (str): The name of the experiment version (str): The version of the experiment Raise: LoggerMisconfigurationError """ if trainer.logger is not None: raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger and explicit_log_dir: " f"{explicit_log_dir} was pass to exp_manager. Please remove the logger from the lightning trainer." ) # Checking only (explicit_log_dir) vs (exp_dir and version). # The `name` will be used as the actual name of checkpoint/archive. if exp_dir or version: logging.error( f"exp_manager received explicit_log_dir: {explicit_log_dir} and at least one of exp_dir: {exp_dir}, " f"or version: {version}. Please note that exp_dir, name, and version will be ignored." ) if is_global_rank_zero() and Path(explicit_log_dir).exists(): logging.warning(f"Exp_manager is logging to {explicit_log_dir}, but it already exists.") return Path(explicit_log_dir), str(explicit_log_dir), "", "" def get_log_dir( trainer: 'pytorch_lightning.Trainer', exp_dir: str = None, name: str = None, version: str = None, explicit_log_dir: str = None, use_datetime_version: bool = True, resume_if_exists: bool = False, ) -> Tuple[Path, str, str, str]: """ Obtains the log_dir used for exp_manager. Returns: log_dir (Path): the log_dir exp_dir (str): the base exp_dir without name nor version name (str): The name of the experiment version (str): The version of the experiment explicit_log_dir (str): The explicit path to the log folder. Defaults to False. use_datetime_version (bool): Uses date and time as the version of the log folder. Defaults to True. resume_if_exists (bool): if resume_if_exists of the exp_manager's config is enabled or not. When enabled, the version folders would not get created. Raise: LoggerMisconfigurationError: If trainer is incompatible with arguments NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. ValueError: If resume is True, and there were more than 1 checkpoint could found. """ if explicit_log_dir: # If explicit log_dir was passed, short circuit return check_explicit_log_dir(trainer, explicit_log_dir, exp_dir, name, version) # Default exp_dir to ./nemo_experiments if None was passed _exp_dir = exp_dir if exp_dir is None: _exp_dir = str(Path.cwd() / 'nemo_experiments') # If the user has already defined a logger for the trainer, use the logger defaults for logging directory if trainer.logger is not None: if trainer.logger.save_dir: if exp_dir: raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger, the logger's " f"save_dir was not None, and exp_dir ({exp_dir}) was not None. If trainer.logger.save_dir " "exists, exp_manager will use trainer.logger.save_dir as the logging directory and exp_dir " "must be None." ) _exp_dir = trainer.logger.save_dir if name: raise LoggerMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a logger, and name: " f"{name} was also passed to exp_manager. If the trainer contains a " "logger, exp_manager will use trainer.logger.name, and name passed to exp_manager must be None." ) name = trainer.logger.name version = f"version_{trainer.logger.version}" # Use user-defined exp_dir, project_name, exp_name, and versioning options else: name = name or "default" version = version or os.environ.get(NEMO_ENV_VARNAME_VERSION, None) if not version: if resume_if_exists: logging.warning( "No version folders would be created under the log folder as 'resume_if_exists' is enabled." ) version = None elif is_global_rank_zero(): if use_datetime_version: version = time.strftime('%Y-%m-%d_%H-%M-%S') else: tensorboard_logger = TensorBoardLogger(save_dir=Path(_exp_dir), name=name, version=version) version = f"version_{tensorboard_logger.version}" os.environ[NEMO_ENV_VARNAME_VERSION] = "" if version is None else version log_dir = Path(_exp_dir) / Path(str(name)) / Path("" if version is None else str(version)) return log_dir, str(_exp_dir), name, version def get_git_hash(): """ Helper function that tries to get the commit hash if running inside a git folder returns: Bool: Whether the git subprocess ran without error str: git subprocess output or error message """ try: return ( True, subprocess.check_output(['git', 'rev-parse', 'HEAD'], stderr=subprocess.STDOUT).decode(), ) except subprocess.CalledProcessError as err: return False, "{}\n".format(err.output.decode("utf-8")) def get_git_diff(): """ Helper function that tries to get the git diff if running inside a git folder returns: Bool: Whether the git subprocess ran without error str: git subprocess output or error message """ try: return subprocess.check_output(['git', 'diff'], stderr=subprocess.STDOUT).decode() except subprocess.CalledProcessError as err: return "{}\n".format(err.output.decode("utf-8")) def configure_loggers( trainer: 'pytorch_lightning.Trainer', exp_dir: [Path, str], log_dir: [Path, str], name: str, version: str, checkpoint_callback_params: dict, create_tensorboard_logger: bool, summary_writer_kwargs: dict, create_wandb_logger: bool, wandb_kwargs: dict, create_mlflow_logger: bool, mlflow_kwargs: dict, create_dllogger_logger: bool, dllogger_kwargs: dict, create_clearml_logger: bool, clearml_kwargs: dict, ): """ Creates TensorboardLogger and/or WandBLogger / MLFlowLogger / DLlogger / ClearMLLogger and attach them to trainer. Raises ValueError if summary_writer_kwargs or wandb_kwargs are misconfigured. """ # Potentially create tensorboard logger and/or WandBLogger / MLFlowLogger / DLLogger logger_list = [] if create_tensorboard_logger: if summary_writer_kwargs is None: summary_writer_kwargs = {} elif "log_dir" in summary_writer_kwargs: raise ValueError( "You cannot pass `log_dir` as part of `summary_writer_kwargs`. `log_dir` is handled by lightning's " "TensorBoardLogger logger." ) tensorboard_logger = TensorBoardLogger(save_dir=exp_dir, name=name, version=version, **summary_writer_kwargs) logger_list.append(tensorboard_logger) logging.info("TensorboardLogger has been set up") if create_wandb_logger: if wandb_kwargs is None: wandb_kwargs = {} if "name" not in wandb_kwargs and "project" not in wandb_kwargs: raise ValueError("name and project are required for wandb_logger") # Update the wandb save_dir if wandb_kwargs.get('save_dir', None) is None: wandb_kwargs['save_dir'] = exp_dir os.makedirs(wandb_kwargs['save_dir'], exist_ok=True) wandb_logger = WandbLogger(version=version, **wandb_kwargs) logger_list.append(wandb_logger) logging.info("WandBLogger has been set up") if create_mlflow_logger: mlflow_logger = MLFlowLogger(run_name=version, **mlflow_kwargs) logger_list.append(mlflow_logger) logging.info("MLFlowLogger has been set up") if create_dllogger_logger: dllogger_logger = DLLogger(**dllogger_kwargs) logger_list.append(dllogger_logger) logging.info("DLLogger has been set up") if create_clearml_logger: clearml_logger = ClearMLLogger( clearml_cfg=clearml_kwargs, log_dir=log_dir, prefix=name, save_best_model=checkpoint_callback_params.save_best_model, ) logger_list.append(clearml_logger) logging.info("ClearMLLogger has been set up") trainer._logger_connector.configure_logger(logger_list) def configure_checkpointing( trainer: 'pytorch_lightning.Trainer', log_dir: Path, name: str, resume: bool, params: 'DictConfig', create_preemption_callback: bool, ): """ Adds ModelCheckpoint to trainer. Raises CheckpointMisconfigurationError if trainer already has a ModelCheckpoint callback """ for callback in trainer.callbacks: if isinstance(callback, ModelCheckpoint): raise CheckpointMisconfigurationError( "The pytorch lightning trainer that was passed to exp_manager contained a ModelCheckpoint " "and create_checkpoint_callback was set to True. Please either set create_checkpoint_callback " "to False, or remove ModelCheckpoint from the lightning trainer" ) # Create the callback and attach it to trainer if "filepath" in params: if params.filepath is not None: logging.warning("filepath is deprecated. Please switch to dirpath and filename instead") if params.dirpath is None: params.dirpath = Path(params.filepath).parent if params.filename is None: params.filename = Path(params.filepath).name with open_dict(params): del params["filepath"] if params.dirpath is None: params.dirpath = Path(log_dir / 'checkpoints') if params.filename is None: params.filename = f'{name}--{{{params.monitor}:.4f}}-{{epoch}}' if params.prefix is None: params.prefix = name NeMoModelCheckpoint.CHECKPOINT_NAME_LAST = params.filename + '-last' logging.debug(params.dirpath) logging.debug(params.filename) logging.debug(params.prefix) if "val" in params.monitor: if ( trainer.max_epochs is not None and trainer.max_epochs != -1 and trainer.max_epochs < trainer.check_val_every_n_epoch ): logging.error( "The checkpoint callback was told to monitor a validation value but trainer.max_epochs(" f"{trainer.max_epochs}) was less than trainer.check_val_every_n_epoch({trainer.check_val_every_n_epoch}" f"). It is very likely this run will fail with ModelCheckpoint(monitor='{params.monitor}') not found " "in the returned metrics. Please ensure that validation is run within trainer.max_epochs." ) elif trainer.max_steps is not None and trainer.max_steps != -1: logging.warning( "The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to " f"{trainer.max_steps}. Please ensure that max_steps will run for at least " f"{trainer.check_val_every_n_epoch} epochs to ensure that checkpointing will not error out." ) checkpoint_callback = NeMoModelCheckpoint(n_resume=resume, **params) checkpoint_callback.last_model_path = trainer.ckpt_path or "" if 'mp_rank' in checkpoint_callback.last_model_path or 'tp_rank' in checkpoint_callback.last_model_path: checkpoint_callback.last_model_path = uninject_model_parallel_rank(checkpoint_callback.last_model_path) trainer.callbacks.append(checkpoint_callback) if create_preemption_callback: # Check if cuda is avialable as preemption is supported only on GPUs if torch.cuda.is_available(): ## By default PreemptionCallback handles SIGTERM. To handle other signals pass the signal in the call as below: ## PreemptionCallback(checkpoint_callback, signal.SIGCHLD) preemption_callback = PreemptionCallback(checkpoint_callback) trainer.callbacks.append(preemption_callback) else: logging.info("Preemption is supported only on GPUs, disabling preemption") def check_slurm(trainer): try: return trainer.accelerator_connector.is_slurm_managing_tasks except AttributeError: return False class StatelessTimer(Timer): """Extension of PTL timers to be per run.""" def __init__(self, duration: timedelta = None, interval: str = Interval.step, verbose: bool = True,) -> None: super().__init__(duration, interval, verbose) # Override PTL Timer's state dict to not store elapsed time information so that we can restore and continue training. def state_dict(self) -> Dict[str, Any]: return {} def load_state_dict(self, state_dict: Dict[str, Any]) -> None: return def configure_no_restart_validation_training_loop(trainer: pytorch_lightning.Trainer) -> None: if type(trainer.fit_loop.epoch_loop) != _TrainingEpochLoop: warnings.warn("Detected custom epoch loop. Skipping no validation on restart support.", UserWarning) return ## Pass trainer object to avoid trainer getting overwritten as None loop = SkipResumeTrainingValidationLoop(trainer, trainer.min_steps, trainer.max_steps) trainer.fit_loop.epoch_loop = loop class SkipResumeTrainingValidationLoop(_TrainingEpochLoop): """ Extend the PTL Epoch loop to skip validating when resuming. This happens when resuming a checkpoint that has already run validation, but loading restores the training state before validation has run. """ def _should_check_val_fx(self) -> bool: if self.restarting and self.global_step % self.trainer.val_check_batch == 0: return False return super()._should_check_val_fx() def clean_exp_ckpt(exp_log_dir: Union[str, Path], remove_ckpt: bool = True, remove_nemo: bool = False): """ Helper method that removes Pytorch Lightning .ckpt files or NeMo .nemo files from the checkpoint directory Args: exp_log_dir: str path to the root directory of the current experiment. remove_ckpt: bool, whether to remove all *.ckpt files in the checkpoints directory. remove_nemo: bool, whether to remove all *.nemo files in the checkpoints directory. """ exp_log_dir = str(exp_log_dir) if remove_ckpt: logging.info("Deleting *.ckpt files ...") ckpt_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.ckpt")) for filepath in ckpt_files: os.remove(filepath) logging.info(f"Deleted file : {filepath}") if remove_nemo: logging.info("Deleting *.nemo files ...") nemo_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.nemo")) for filepath in nemo_files: os.remove(filepath) logging.info(f"Deleted file : {filepath}") [end of nemo/utils/exp_manager.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # """ # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in # fusion with beam search decoders on top of a trained ASR model with CTC decoder. To evaluate a model with # Transducer (RNN-T) decoder use another script 'scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py'. # NeMo's beam search decoders are capable of using the KenLM's N-gram models # to find the best candidates. This script supports both character level and BPE level # encodings and models which is detected automatically from the type of the model. # You may train the LM model with 'scripts/asr_language_modeling/ngram_lm/train_kenlm.py'. # Config Help To discover all arguments of the script, please run : python eval_beamsearch_ngram.py --help python eval_beamsearch_ngram.py --cfg job # USAGE python eval_beamsearch_ngram.py nemo_model_file=<path to the .nemo file of the model> \ input_manifest=<path to the evaluation JSON manifest file> \ kenlm_model_file=<path to the binary KenLM model> \ beam_width=[<list of the beam widths, separated with commas>] \ beam_alpha=[<list of the beam alphas, separated with commas>] \ beam_beta=[<list of the beam betas, separated with commas>] \ preds_output_folder=<optional folder to store the predictions> \ probs_cache_file=null \ decoding_mode=beamsearch_ngram ... # Grid Search for Hyper parameters For grid search, you can provide a list of arguments as follows - beam_width=[4,8,16,....] \ beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ beam_beta=[-1.0,-0.5,0.0,...,1.0] \ # You may find more info on how to use this script at: # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html """ import contextlib import json import os import pickle from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import List, Optional import editdistance import numpy as np import torch from omegaconf import MISSING, OmegaConf from sklearn.model_selection import ParameterGrid from tqdm.auto import tqdm import nemo.collections.asr as nemo_asr from nemo.collections.asr.models import EncDecHybridRNNTCTCModel from nemo.collections.asr.parts.submodules import ctc_beam_decoding from nemo.collections.asr.parts.utils.transcribe_utils import PunctuationCapitalization, TextProcessingConfig from nemo.core.config import hydra_runner from nemo.utils import logging # fmt: off @dataclass class EvalBeamSearchNGramConfig: """ Evaluate an ASR model with beam search decoding and n-gram KenLM language model. """ # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) nemo_model_file: str = MISSING # File paths input_manifest: str = MISSING # The manifest file of the evaluation set kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model # Parameters for inference acoustic_batch_size: int = 16 # The batch size to calculate log probabilities beam_batch_size: int = 128 # The batch size to be used for beam search decoding device: str = "cuda" # The device to load the model onto to calculate log probabilities use_amp: bool = False # Whether to use AMP if available to calculate log probabilities # Beam Search hyperparameters # The decoding scheme to be used for evaluation. # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] decoding_mode: str = "beamsearch_ngram" beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, ) # fmt: on def beam_search_eval( model: nemo_asr.models.ASRModel, cfg: EvalBeamSearchNGramConfig, all_probs: List[torch.Tensor], target_transcripts: List[str], preds_output_file: str = None, lm_path: str = None, beam_alpha: float = 1.0, beam_beta: float = 0.0, beam_width: int = 128, beam_batch_size: int = 128, progress_bar: bool = True, punctuation_capitalization: PunctuationCapitalization = None, ): level = logging.getEffectiveLevel() logging.setLevel(logging.CRITICAL) # Reset config model.change_decoding_strategy(None) # Override the beam search config with current search candidate configuration cfg.decoding.beam_size = beam_width cfg.decoding.beam_alpha = beam_alpha cfg.decoding.beam_beta = beam_beta cfg.decoding.return_best_hypothesis = False cfg.decoding.kenlm_path = cfg.kenlm_model_file # Update model's decoding strategy config model.cfg.decoding.strategy = cfg.decoding_strategy model.cfg.decoding.beam = cfg.decoding # Update model's decoding strategy if isinstance(model, EncDecHybridRNNTCTCModel): model.change_decoding_strategy(model.cfg.decoding, decoder_type='ctc') decoding = model.ctc_decoding else: model.change_decoding_strategy(model.cfg.decoding) decoding = model.decoding logging.setLevel(level) wer_dist_first = cer_dist_first = 0 wer_dist_best = cer_dist_best = 0 words_count = 0 chars_count = 0 sample_idx = 0 if preds_output_file: out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') if progress_bar: it = tqdm( range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=f"Beam search decoding with width={beam_width}, alpha={beam_alpha}, beta={beam_beta}", ncols=120, ) else: it = range(int(np.ceil(len(all_probs) / beam_batch_size))) for batch_idx in it: # disabling type checking probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] probs_lens = torch.tensor([prob.shape[0] for prob in probs_batch]) with torch.no_grad(): packed_batch = torch.zeros(len(probs_batch), max(probs_lens), probs_batch[0].shape[-1], device='cpu') for prob_index in range(len(probs_batch)): packed_batch[prob_index, : probs_lens[prob_index], :] = torch.tensor( probs_batch[prob_index], device=packed_batch.device, dtype=packed_batch.dtype ) _, beams_batch = decoding.ctc_decoder_predictions_tensor( packed_batch, decoder_lengths=probs_lens, return_hypotheses=True, ) for beams_idx, beams in enumerate(beams_batch): target = target_transcripts[sample_idx + beams_idx] target_split_w = target.split() target_split_c = list(target) words_count += len(target_split_w) chars_count += len(target_split_c) wer_dist_min = cer_dist_min = 10000 for candidate_idx, candidate in enumerate(beams): # type: (int, ctc_beam_decoding.rnnt_utils.Hypothesis) pred_text = candidate.text if cfg.text_processing.do_lowercase: pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] if cfg.text_processing.rm_punctuation: pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] if cfg.text_processing.separate_punctuation: pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] pred_split_w = pred_text.split() wer_dist = editdistance.eval(target_split_w, pred_split_w) pred_split_c = list(pred_text) cer_dist = editdistance.eval(target_split_c, pred_split_c) wer_dist_min = min(wer_dist_min, wer_dist) cer_dist_min = min(cer_dist_min, cer_dist) if candidate_idx == 0: # first candidate wer_dist_first += wer_dist cer_dist_first += cer_dist score = candidate.score if preds_output_file: out_file.write('{}\t{}\n'.format(pred_text, score)) wer_dist_best += wer_dist_min cer_dist_best += cer_dist_min sample_idx += len(probs_batch) if preds_output_file: out_file.close() logging.info(f"Stored the predictions of beam search decoding at '{preds_output_file}'.") if lm_path: logging.info( 'WER/CER with beam search decoding and N-gram model = {:.2%}/{:.2%}'.format( wer_dist_first / words_count, cer_dist_first / chars_count ) ) else: logging.info( 'WER/CER with beam search decoding = {:.2%}/{:.2%}'.format( wer_dist_first / words_count, cer_dist_first / chars_count ) ) logging.info( 'Oracle WER/CER in candidates with perfect LM= {:.2%}/{:.2%}'.format( wer_dist_best / words_count, cer_dist_best / chars_count ) ) logging.info(f"=================================================================================") return wer_dist_first / words_count, cer_dist_first / chars_count @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) def main(cfg: EvalBeamSearchNGramConfig): logging.warning("This file will be renamed to eval_beamsearch_ngram_ctc.py in the future NeMo (1.21) release.") if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig valid_decoding_modes = ["greedy", "beamsearch", "beamsearch_ngram"] if cfg.decoding_mode not in valid_decoding_modes: raise ValueError( f"Given decoding_mode={cfg.decoding_mode} is invalid. Available options are :\n" f"{valid_decoding_modes}" ) if cfg.nemo_model_file.endswith('.nemo'): asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) else: logging.warning( "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." ) asr_model = nemo_asr.models.ASRModel.from_pretrained( cfg.nemo_model_file, map_location=torch.device(cfg.device) ) target_transcripts = [] manifest_dir = Path(cfg.input_manifest).parent with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: audio_file_paths = [] for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): data = json.loads(line) audio_file = Path(data['audio_filepath']) if not audio_file.is_file() and not audio_file.is_absolute(): audio_file = manifest_dir / audio_file target_transcripts.append(data['text']) audio_file_paths.append(str(audio_file.absolute())) punctuation_capitalization = PunctuationCapitalization(cfg.text_processing.punctuation_marks) if cfg.text_processing.do_lowercase: target_transcripts = punctuation_capitalization.do_lowercase(target_transcripts) if cfg.text_processing.rm_punctuation: target_transcripts = punctuation_capitalization.rm_punctuation(target_transcripts) if cfg.text_processing.separate_punctuation: target_transcripts = punctuation_capitalization.separate_punctuation(target_transcripts) if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") with open(cfg.probs_cache_file, 'rb') as probs_file: all_probs = pickle.load(probs_file) if len(all_probs) != len(audio_file_paths): raise ValueError( f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " f"match the manifest file. You may need to delete the probabilities cached file." ) else: @contextlib.contextmanager def default_autocast(): yield if cfg.use_amp: if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): logging.info("AMP is enabled!\n") autocast = torch.cuda.amp.autocast else: autocast = default_autocast else: autocast = default_autocast with autocast(): with torch.no_grad(): if isinstance(asr_model, EncDecHybridRNNTCTCModel): asr_model.cur_decoder = 'ctc' all_logits = asr_model.transcribe(audio_file_paths, batch_size=cfg.acoustic_batch_size, logprobs=True) all_probs = all_logits if cfg.probs_cache_file: logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") with open(cfg.probs_cache_file, 'wb') as f_dump: pickle.dump(all_probs, f_dump) wer_dist_greedy = 0 cer_dist_greedy = 0 words_count = 0 chars_count = 0 for batch_idx, probs in enumerate(all_probs): preds = np.argmax(probs, axis=1) preds_tensor = torch.tensor(preds, device='cpu').unsqueeze(0) if isinstance(asr_model, EncDecHybridRNNTCTCModel): pred_text = asr_model.ctc_decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] else: pred_text = asr_model._wer.decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] if cfg.text_processing.do_lowercase: pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] if cfg.text_processing.rm_punctuation: pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] if cfg.text_processing.separate_punctuation: pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] pred_split_w = pred_text.split() target_split_w = target_transcripts[batch_idx].split() pred_split_c = list(pred_text) target_split_c = list(target_transcripts[batch_idx]) wer_dist = editdistance.eval(target_split_w, pred_split_w) cer_dist = editdistance.eval(target_split_c, pred_split_c) wer_dist_greedy += wer_dist cer_dist_greedy += cer_dist words_count += len(target_split_w) chars_count += len(target_split_c) logging.info('Greedy WER/CER = {:.2%}/{:.2%}'.format(wer_dist_greedy / words_count, cer_dist_greedy / chars_count)) asr_model = asr_model.to('cpu') if cfg.decoding_mode == "beamsearch_ngram": if not os.path.exists(cfg.kenlm_model_file): raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") lm_path = cfg.kenlm_model_file else: lm_path = None # 'greedy' decoding_mode would skip the beam search decoding if cfg.decoding_mode in ["beamsearch_ngram", "beamsearch"]: if cfg.beam_width is None or cfg.beam_alpha is None or cfg.beam_beta is None: raise ValueError("beam_width, beam_alpha and beam_beta are needed to perform beam search decoding.") params = {'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'beam_beta': cfg.beam_beta} hp_grid = ParameterGrid(params) hp_grid = list(hp_grid) best_wer_beam_size, best_cer_beam_size = None, None best_wer_alpha, best_cer_alpha = None, None best_wer_beta, best_cer_beta = None, None best_wer, best_cer = 1e6, 1e6 logging.info(f"==============================Starting the beam search decoding===============================") logging.info(f"Grid search size: {len(hp_grid)}") logging.info(f"It may take some time...") logging.info(f"==============================================================================================") if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): os.mkdir(cfg.preds_output_folder) for hp in hp_grid: if cfg.preds_output_folder: preds_output_file = os.path.join( cfg.preds_output_folder, f"preds_out_width{hp['beam_width']}_alpha{hp['beam_alpha']}_beta{hp['beam_beta']}.tsv", ) else: preds_output_file = None candidate_wer, candidate_cer = beam_search_eval( asr_model, cfg, all_probs=all_probs, target_transcripts=target_transcripts, preds_output_file=preds_output_file, lm_path=lm_path, beam_width=hp["beam_width"], beam_alpha=hp["beam_alpha"], beam_beta=hp["beam_beta"], beam_batch_size=cfg.beam_batch_size, progress_bar=True, punctuation_capitalization=punctuation_capitalization, ) if candidate_cer < best_cer: best_cer_beam_size = hp["beam_width"] best_cer_alpha = hp["beam_alpha"] best_cer_beta = hp["beam_beta"] best_cer = candidate_cer if candidate_wer < best_wer: best_wer_beam_size = hp["beam_width"] best_wer_alpha = hp["beam_alpha"] best_wer_beta = hp["beam_beta"] best_wer = candidate_wer logging.info( f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' f'Beam alpha = {best_wer_alpha}, Beam beta = {best_wer_beta}' ) logging.info( f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' f'Beam alpha = {best_cer_alpha}, Beam beta = {best_cer_beta}' ) logging.info(f"=================================================================================") if __name__ == '__main__': main() [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # """ # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in # fusion with beam search decoders on top of a trained ASR Transducer model. NeMo's beam search decoders are capable of using the # KenLM's N-gram models to find the best candidates. This script supports both character level and BPE level # encodings and models which is detected automatically from the type of the model. # You may train the LM model with 'scripts/ngram_lm/train_kenlm.py'. # Config Help To discover all arguments of the script, please run : python eval_beamsearch_ngram.py --help python eval_beamsearch_ngram.py --cfg job # USAGE python eval_beamsearch_ngram_transducer.py nemo_model_file=<path to the .nemo file of the model> \ input_manifest=<path to the evaluation JSON manifest file \ kenlm_model_file=<path to the binary KenLM model> \ beam_width=[<list of the beam widths, separated with commas>] \ beam_alpha=[<list of the beam alphas, separated with commas>] \ preds_output_folder=<optional folder to store the predictions> \ probs_cache_file=null \ decoding_strategy=<greedy_batch or maes decoding> maes_prefix_alpha=[<list of the maes prefix alphas, separated with commas>] \ maes_expansion_gamma=[<list of the maes expansion gammas, separated with commas>] \ hat_subtract_ilm=<in case of HAT model: subtract internal LM or not> \ hat_ilm_weight=[<in case of HAT model: list of the HAT internal LM weights, separated with commas>] \ ... # Grid Search for Hyper parameters For grid search, you can provide a list of arguments as follows - beam_width=[4,8,16,....] \ beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ # You may find more info on how to use this script at: # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html """ import contextlib import json import os import pickle import tempfile from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import List, Optional import editdistance import numpy as np import torch from omegaconf import MISSING, OmegaConf from sklearn.model_selection import ParameterGrid from tqdm.auto import tqdm import nemo.collections.asr as nemo_asr from nemo.collections.asr.parts.submodules import rnnt_beam_decoding from nemo.core.config import hydra_runner from nemo.utils import logging # fmt: off @dataclass class EvalBeamSearchNGramConfig: """ Evaluate an ASR model with beam search decoding and n-gram KenLM language model. """ # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) nemo_model_file: str = MISSING # File paths input_manifest: str = MISSING # The manifest file of the evaluation set kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model # Parameters for inference acoustic_batch_size: int = 128 # The batch size to calculate log probabilities beam_batch_size: int = 128 # The batch size to be used for beam search decoding device: str = "cuda" # The device to load the model onto to calculate log probabilities use_amp: bool = False # Whether to use AMP if available to calculate log probabilities num_workers: int = 1 # Number of workers for DataLoader # The decoding scheme to be used for evaluation decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] # Beam Search hyperparameters beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding # HAT related parameters (only for internal lm subtraction) hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) # fmt: on def decoding_step( model: nemo_asr.models.ASRModel, cfg: EvalBeamSearchNGramConfig, all_probs: List[torch.Tensor], target_transcripts: List[str], preds_output_file: str = None, beam_batch_size: int = 128, progress_bar: bool = True, ): level = logging.getEffectiveLevel() logging.setLevel(logging.CRITICAL) # Reset config model.change_decoding_strategy(None) cfg.decoding.hat_ilm_weight = cfg.decoding.hat_ilm_weight * cfg.hat_subtract_ilm # Override the beam search config with current search candidate configuration cfg.decoding.return_best_hypothesis = False cfg.decoding.ngram_lm_model = cfg.kenlm_model_file cfg.decoding.hat_subtract_ilm = cfg.hat_subtract_ilm # Update model's decoding strategy config model.cfg.decoding.strategy = cfg.decoding_strategy model.cfg.decoding.beam = cfg.decoding # Update model's decoding strategy model.change_decoding_strategy(model.cfg.decoding) logging.setLevel(level) wer_dist_first = cer_dist_first = 0 wer_dist_best = cer_dist_best = 0 words_count = 0 chars_count = 0 sample_idx = 0 if preds_output_file: out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') if progress_bar: if cfg.decoding_strategy == "greedy_batch": description = "Greedy_batch decoding.." else: description = f"{cfg.decoding_strategy} decoding with bw={cfg.decoding.beam_size}, ba={cfg.decoding.ngram_lm_alpha}, ma={cfg.decoding.maes_prefix_alpha}, mg={cfg.decoding.maes_expansion_gamma}, hat_ilmw={cfg.decoding.hat_ilm_weight}" it = tqdm(range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=description, ncols=120) else: it = range(int(np.ceil(len(all_probs) / beam_batch_size))) for batch_idx in it: # disabling type checking probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] probs_lens = torch.tensor([prob.shape[-1] for prob in probs_batch]) with torch.no_grad(): packed_batch = torch.zeros(len(probs_batch), probs_batch[0].shape[0], max(probs_lens), device='cpu') for prob_index in range(len(probs_batch)): packed_batch[prob_index, :, : probs_lens[prob_index]] = torch.tensor( probs_batch[prob_index].unsqueeze(0), device=packed_batch.device, dtype=packed_batch.dtype ) best_hyp_batch, beams_batch = model.decoding.rnnt_decoder_predictions_tensor( packed_batch, probs_lens, return_hypotheses=True, ) if cfg.decoding_strategy == "greedy_batch": beams_batch = [[x] for x in best_hyp_batch] for beams_idx, beams in enumerate(beams_batch): target = target_transcripts[sample_idx + beams_idx] target_split_w = target.split() target_split_c = list(target) words_count += len(target_split_w) chars_count += len(target_split_c) wer_dist_min = cer_dist_min = 10000 for candidate_idx, candidate in enumerate(beams): # type: (int, rnnt_beam_decoding.rnnt_utils.Hypothesis) pred_text = candidate.text pred_split_w = pred_text.split() wer_dist = editdistance.eval(target_split_w, pred_split_w) pred_split_c = list(pred_text) cer_dist = editdistance.eval(target_split_c, pred_split_c) wer_dist_min = min(wer_dist_min, wer_dist) cer_dist_min = min(cer_dist_min, cer_dist) if candidate_idx == 0: # first candidate wer_dist_first += wer_dist cer_dist_first += cer_dist score = candidate.score if preds_output_file: out_file.write('{}\t{}\n'.format(pred_text, score)) wer_dist_best += wer_dist_min cer_dist_best += cer_dist_min sample_idx += len(probs_batch) if cfg.decoding_strategy == "greedy_batch": return wer_dist_first / words_count, cer_dist_first / chars_count if preds_output_file: out_file.close() logging.info(f"Stored the predictions of {cfg.decoding_strategy} decoding at '{preds_output_file}'.") if cfg.decoding.ngram_lm_model: logging.info( f"WER/CER with {cfg.decoding_strategy} decoding and N-gram model = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" ) else: logging.info( f"WER/CER with {cfg.decoding_strategy} decoding = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" ) logging.info( f"Oracle WER/CER in candidates with perfect LM= {wer_dist_best / words_count:.2%}/{cer_dist_best / chars_count:.2%}" ) logging.info(f"=================================================================================") return wer_dist_first / words_count, cer_dist_first / chars_count @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) def main(cfg: EvalBeamSearchNGramConfig): if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig valid_decoding_strategis = ["greedy_batch", "beam", "tsd", "alsd", "maes"] if cfg.decoding_strategy not in valid_decoding_strategis: raise ValueError( f"Given decoding_strategy={cfg.decoding_strategy} is invalid. Available options are :\n" f"{valid_decoding_strategis}" ) if cfg.nemo_model_file.endswith('.nemo'): asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) else: logging.warning( "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." ) asr_model = nemo_asr.models.ASRModel.from_pretrained( cfg.nemo_model_file, map_location=torch.device(cfg.device) ) if cfg.kenlm_model_file: if not os.path.exists(cfg.kenlm_model_file): raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") if cfg.decoding_strategy != "maes": raise ValueError(f"Decoding with kenlm model is supported only for maes decoding algorithm.") lm_path = cfg.kenlm_model_file else: lm_path = None cfg.beam_alpha = [0.0] if cfg.hat_subtract_ilm: assert lm_path, "kenlm must be set for hat internal lm subtraction" if cfg.decoding_strategy != "maes": cfg.maes_prefix_alpha, cfg.maes_expansion_gamma, cfg.hat_ilm_weight = [0], [0], [0] target_transcripts = [] manifest_dir = Path(cfg.input_manifest).parent with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: audio_file_paths = [] for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): data = json.loads(line) audio_file = Path(data['audio_filepath']) if not audio_file.is_file() and not audio_file.is_absolute(): audio_file = manifest_dir / audio_file target_transcripts.append(data['text']) audio_file_paths.append(str(audio_file.absolute())) if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") with open(cfg.probs_cache_file, 'rb') as probs_file: all_probs = pickle.load(probs_file) if len(all_probs) != len(audio_file_paths): raise ValueError( f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " f"match the manifest file. You may need to delete the probabilities cached file." ) else: @contextlib.contextmanager def default_autocast(): yield if cfg.use_amp: if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): logging.info("AMP is enabled!\n") autocast = torch.cuda.amp.autocast else: autocast = default_autocast else: autocast = default_autocast # manual calculation of encoder_embeddings with autocast(): with torch.no_grad(): asr_model.eval() asr_model.encoder.freeze() device = next(asr_model.parameters()).device all_probs = [] with tempfile.TemporaryDirectory() as tmpdir: with open(os.path.join(tmpdir, 'manifest.json'), 'w', encoding='utf-8') as fp: for audio_file in audio_file_paths: entry = {'audio_filepath': audio_file, 'duration': 100000, 'text': ''} fp.write(json.dumps(entry) + '\n') config = { 'paths2audio_files': audio_file_paths, 'batch_size': cfg.acoustic_batch_size, 'temp_dir': tmpdir, 'num_workers': cfg.num_workers, 'channel_selector': None, 'augmentor': None, } temporary_datalayer = asr_model._setup_transcribe_dataloader(config) for test_batch in tqdm(temporary_datalayer, desc="Transcribing", disable=True): encoded, encoded_len = asr_model.forward( input_signal=test_batch[0].to(device), input_signal_length=test_batch[1].to(device) ) # dump encoder embeddings per file for idx in range(encoded.shape[0]): encoded_no_pad = encoded[idx, :, : encoded_len[idx]] all_probs.append(encoded_no_pad) if cfg.probs_cache_file: logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") with open(cfg.probs_cache_file, 'wb') as f_dump: pickle.dump(all_probs, f_dump) if cfg.decoding_strategy == "greedy_batch": asr_model = asr_model.to('cpu') candidate_wer, candidate_cer = decoding_step( asr_model, cfg, all_probs=all_probs, target_transcripts=target_transcripts, beam_batch_size=cfg.beam_batch_size, progress_bar=True, ) logging.info(f"Greedy batch WER/CER = {candidate_wer:.2%}/{candidate_cer:.2%}") asr_model = asr_model.to('cpu') # 'greedy_batch' decoding_strategy would skip the beam search decoding if cfg.decoding_strategy in ["beam", "tsd", "alsd", "maes"]: if cfg.beam_width is None or cfg.beam_alpha is None: raise ValueError("beam_width and beam_alpha are needed to perform beam search decoding.") params = { 'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'maes_prefix_alpha': cfg.maes_prefix_alpha, 'maes_expansion_gamma': cfg.maes_expansion_gamma, 'hat_ilm_weight': cfg.hat_ilm_weight, } hp_grid = ParameterGrid(params) hp_grid = list(hp_grid) best_wer_beam_size, best_cer_beam_size = None, None best_wer_alpha, best_cer_alpha = None, None best_wer, best_cer = 1e6, 1e6 logging.info( f"==============================Starting the {cfg.decoding_strategy} decoding===============================" ) logging.info(f"Grid search size: {len(hp_grid)}") logging.info(f"It may take some time...") logging.info(f"==============================================================================================") if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): os.mkdir(cfg.preds_output_folder) for hp in hp_grid: if cfg.preds_output_folder: results_file = f"preds_out_{cfg.decoding_strategy}_bw{hp['beam_width']}" if cfg.decoding_strategy == "maes": results_file = f"{results_file}_ma{hp['maes_prefix_alpha']}_mg{hp['maes_expansion_gamma']}" if cfg.kenlm_model_file: results_file = f"{results_file}_ba{hp['beam_alpha']}" if cfg.hat_subtract_ilm: results_file = f"{results_file}_hat_ilmw{hp['hat_ilm_weight']}" preds_output_file = os.path.join(cfg.preds_output_folder, f"{results_file}.tsv") else: preds_output_file = None cfg.decoding.beam_size = hp["beam_width"] cfg.decoding.ngram_lm_alpha = hp["beam_alpha"] cfg.decoding.maes_prefix_alpha = hp["maes_prefix_alpha"] cfg.decoding.maes_expansion_gamma = hp["maes_expansion_gamma"] cfg.decoding.hat_ilm_weight = hp["hat_ilm_weight"] candidate_wer, candidate_cer = decoding_step( asr_model, cfg, all_probs=all_probs, target_transcripts=target_transcripts, preds_output_file=preds_output_file, beam_batch_size=cfg.beam_batch_size, progress_bar=True, ) if candidate_cer < best_cer: best_cer_beam_size = hp["beam_width"] best_cer_alpha = hp["beam_alpha"] best_cer_ma = hp["maes_prefix_alpha"] best_cer_mg = hp["maes_expansion_gamma"] best_cer_hat_ilm_weight = hp["hat_ilm_weight"] best_cer = candidate_cer if candidate_wer < best_wer: best_wer_beam_size = hp["beam_width"] best_wer_alpha = hp["beam_alpha"] best_wer_ma = hp["maes_prefix_alpha"] best_wer_ga = hp["maes_expansion_gamma"] best_wer_hat_ilm_weight = hp["hat_ilm_weight"] best_wer = candidate_wer wer_hat_parameter = "" if cfg.hat_subtract_ilm: wer_hat_parameter = f"HAT ilm weight = {best_wer_hat_ilm_weight}, " logging.info( f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' f'Beam alpha = {best_wer_alpha}, {wer_hat_parameter}' f'maes_prefix_alpha = {best_wer_ma}, maes_expansion_gamma = {best_wer_ga} ' ) cer_hat_parameter = "" if cfg.hat_subtract_ilm: cer_hat_parameter = f"HAT ilm weight = {best_cer_hat_ilm_weight}" logging.info( f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' f'Beam alpha = {best_cer_alpha}, {cer_hat_parameter} ' f'maes_prefix_alpha = {best_cer_ma}, maes_expansion_gamma = {best_cer_mg}' ) logging.info(f"=================================================================================") if __name__ == '__main__': main() [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of scripts/confidence_ensembles/build_ensemble.py] # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ This script provides a functionality to create confidence-based ensembles from a collection of pretrained models. For more details see the paper https://arxiv.org/abs/2306.15824 or tutorial in tutorials/asr/Confidence_Ensembles.ipynb You would typically use this script by providing a yaml config file or overriding default options from command line. Usage examples: 1. Building an ensemble of two monolingual models with default settings (no confidence tuning). python build_ensemble.py --config-path=. --config-name=ensemble_config.yaml ensemble.0.model=stt_it_conformer_ctc_large ensemble.0.training_manifest=<path to the Italian data of 100+ utterances (no transcription required)> ensemble.1.model=stt_es_conformer_ctc_large ensemble.1.training_manifest=<path to the Spanish data of 100+ utterances (no transcription required)> output_path=<path to the desired location of the .nemo checkpoint> You can have more than 2 models and can control transcription settings (e.g., batch size) with ``transcription.<any argument of examples/asr/transcribe_speech.py>`` parameters. 2. If you want to get improved results, you can enable tuning of the confidence and logistic regression (LR) parameters. E.g. python build_ensemble.py <all arguments like in the previous example> ensemble.0.dev_manifest=<path to the dev data that's required for tuning> ... # IMPORTANT: see the note below if you use > 2 models! ensemble.N.dev_manifest=<path to the dev data that's required for tuning> tune_confidence=True # to allow confidence tuning. LR is tuned by default As with any tuning, it is recommended to have reasonably large validation set for each model, otherwise you might overfit to the validation data. Note that if you add additional models (> 2) you will need to modify ensemble_config.yaml or create a new one with added models in there. While it's theoretically possible to fully override such parameters from commandline, hydra is very unfriendly for such use-cases, so it's strongly recommended to be creating new configs. 3. If you want to precisely control tuning grid search, you can do that with python build_ensemble.py <all arguments as in the previous examples> tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set You can check the dataclasses in this file for the full list of supported arguments and their default values. """ import atexit # using default logging to be able to silence unnecessary messages from nemo import logging import os import random import sys import tempfile from copy import deepcopy from dataclasses import dataclass from pathlib import Path from typing import Dict, List, Optional, Tuple import joblib import numpy as np import pytorch_lightning as pl from omegaconf import MISSING, DictConfig, OmegaConf from sklearn.linear_model import LogisticRegression from sklearn.metrics import confusion_matrix from sklearn.pipeline import Pipeline, make_pipeline from sklearn.preprocessing import StandardScaler from tqdm import tqdm from nemo.collections.asr.models.confidence_ensemble import ( ConfidenceEnsembleModel, ConfidenceSpec, compute_confidence, get_filtered_logprobs, ) from nemo.collections.asr.parts.utils.asr_confidence_utils import ( ConfidenceConfig, ConfidenceMethodConfig, get_confidence_aggregation_bank, get_confidence_measure_bank, ) from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis from nemo.core.config import hydra_runner LOG = logging.getLogger(__file__) # adding Python path. If not found, asking user to get the file try: sys.path.append(str(Path(__file__).parents[2] / "examples" / "asr")) import transcribe_speech except ImportError: # if users run script normally from nemo repo, this shouldn't be triggered as # we modify the path above. But if they downloaded the build_ensemble.py as # an isolated script, we'd ask them to also download corresponding version # of the transcribe_speech.py print( "Current script depends on 'examples/asr/transcribe_speech.py', but can't find it. " "If it's not present, download it from the NeMo github manually and put inside this folder." ) @dataclass class EnsembleConfig: # .nemo path or pretrained name model: str = MISSING # path to the training data manifest (non-tarred) training_manifest: str = MISSING # specify to limit the number of training samples # 100 is most likely enough, but setting higher default just in case max_training_samples: int = 1000 # specify to provide dev data manifest for HP tuning dev_manifest: Optional[str] = None @dataclass class TuneConfidenceConfig: # important parameter, so should always be tuned exclude_blank: Tuple[bool] = (True, False) # prod is pretty much always worse, so not including by default aggregation: Tuple[str] = ("mean", "min", "max") # not including max prob, as there is always an entropy-based metric # that's better but otherwise including everything confidence_type: Tuple[str] = ( "entropy_renyi_exp", "entropy_renyi_lin", "entropy_tsallis_exp", "entropy_tsallis_lin", "entropy_gibbs_lin", "entropy_gibbs_exp", ) # TODO: currently it's not possible to efficiently tune temperature, as we always # apply log-softmax in the decoder, so to try different values it will be required # to rerun the decoding, which is very slow. To support this for one-off experiments # it's possible to modify the code of CTC decoder / Transducer joint to # remove log-softmax and then apply it directly in this script with the temperature # # Alternatively, one can run this script multiple times with different values of # temperature and pick the best performing ensemble. Note that this will increase # tuning time by the number of temperature values tried. On the other hand, # the above approach is a lot more efficient and will only slightly increase # the total tuning runtime. # very important to tune for max prob, but for entropy metrics 1.0 is almost always best # temperature: Tuple[float] = (1.0,) # not that important, but can sometimes make a small difference alpha: Tuple[float] = (0.25, 0.33, 0.5, 1.0) def get_grid_size(self) -> int: """Returns the total number of points in the search space.""" if "max_prob" in self.confidence_type: return ( len(self.exclude_blank) * len(self.aggregation) * ((len(self.confidence_type) - 1) * len(self.alpha) + 1) ) return len(self.exclude_blank) * len(self.aggregation) * len(self.confidence_type) * len(self.alpha) @dataclass class TuneLogisticRegressionConfig: # will have log-uniform grid over this range with that many points # note that a value of 10000.0 (not regularization) is always added C_num_points: int = 10 C_min: float = 0.0001 C_max: float = 10.0 # not too important multi_class: Tuple[str] = ("ovr", "multinomial") # should try to include weights directly if the data is too imbalanced class_weight: Tuple = (None, "balanced") # increase if getting many warnings that algorithm didn't converge max_iter: int = 1000 @dataclass class BuildEnsembleConfig: # where to save the resulting ensemble model output_path: str = MISSING # each model specification ensemble: List[EnsembleConfig] = MISSING random_seed: int = 0 # for reproducibility # default confidence, can override confidence: ConfidenceConfig = ConfidenceConfig( # we keep frame confidences and apply aggregation manually to get full-utterance confidence preserve_frame_confidence=True, exclude_blank=True, aggregation="mean", method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() # set to True to tune the confidence. # requires dev manifests to be specified for each model tune_confidence: bool = False # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() def __post_init__(self): """Checking that if any dev data is provided, all are provided. Will also auto-set tune_logistic_regression to False if no dev data is available. If tune_confidence is set to True (user choice) and no dev data is provided, will raise an error. """ num_dev_data = 0 for ensemble_cfg in self.ensemble: num_dev_data += ensemble_cfg.dev_manifest is not None if num_dev_data == 0: if self.tune_confidence: raise ValueError("tune_confidence is set to True, but no dev data is provided") LOG.info("Setting tune_logistic_regression = False since no dev data is provided") self.tune_logistic_regression = False return if num_dev_data < len(self.ensemble): raise ValueError( "Some ensemble configs specify dev data, but some don't. Either all have to specify it or none!" ) def calculate_score(features: np.ndarray, labels: np.ndarray, pipe: Pipeline) -> Tuple[float, np.ndarray]: """Score is always calculated as mean of the per-class scores. This is done to account for possible class imbalances. Args: features: numpy array of features of shape [N x D], where N is the number of objects (typically a total number of utterances in all datasets) and D is the total number of confidence scores used to train the model (typically = number of models). labels: numpy array of shape [N] contatining ground-truth model indices. pipe: classification pipeline (currently, standardization + logistic regression). Returns: tuple: score value in [0, 1] and full classification confusion matrix. """ predictions = pipe.predict(features) conf_m = confusion_matrix(labels, predictions) score = np.diag(conf_m).sum() / conf_m.sum() return score, conf_m def train_model_selection( training_features: np.ndarray, training_labels: np.ndarray, dev_features: Optional[np.ndarray] = None, dev_labels: Optional[np.ndarray] = None, tune_lr: bool = False, tune_lr_cfg: Optional[TuneLogisticRegressionConfig] = None, verbose: bool = False, ) -> Tuple[Pipeline, float]: """Trains model selection block with an (optional) tuning of the parameters. Returns a pipeline consisting of feature standardization and logistic regression. If tune_lr is set to True, dev features/labels will be used to tune the hyperparameters of the logistic regression with the grid search that's defined via ``tune_lr_cfg``. If no tuning is requested, uses the following parameters:: best_pipe = make_pipeline( StandardScaler(), LogisticRegression( multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced", ), ) Args: training_features: numpy array of features of shape [N x D], where N is the number of objects (typically a total number of utterances in all training datasets) and D is the total number of confidence scores used to train the model (typically = number of models). training_labels: numpy array of shape [N] contatining ground-truth model indices. dev_features: same as training, but for the validation subset. dev_labels: same as training, but for the validation subset. tune_lr: controls whether tuning of LR hyperparameters is performed. If set to True, it's required to also provide dev features/labels. tune_lr_cfg: specifies what values of LR hyperparameters to try. verbose: if True, will output final training/dev scores. Returns: tuple: trained model selection pipeline, best score (or -1 if no tuning was done). """ if not tune_lr: # default parameters: C=10000.0 disables regularization best_pipe = make_pipeline( StandardScaler(), LogisticRegression(multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced"), ) max_score = -1 else: C_pms = np.append( np.exp(np.linspace(np.log(tune_lr_cfg.C_min), np.log(tune_lr_cfg.C_max), tune_lr_cfg.C_num_points)), 10000.0, ) max_score = 0 best_pipe = None for class_weight in tune_lr_cfg.class_weight: for multi_class in tune_lr_cfg.multi_class: for C in C_pms: pipe = make_pipeline( StandardScaler(), LogisticRegression( multi_class=multi_class, C=C, max_iter=tune_lr_cfg.max_iter, class_weight=class_weight ), ) pipe.fit(training_features, training_labels) score, confusion = calculate_score(dev_features, dev_labels, pipe) if score > max_score: max_score = score best_pipe = pipe best_pipe.fit(training_features, training_labels) if verbose: accuracy, confusion = calculate_score(training_features, training_labels, best_pipe) LOG.info("Training fit accuracy: %.4f", accuracy * 100.0) LOG.info("Training confusion matrix:\n%s", str(confusion)) if dev_features is not None and verbose: accuracy, confusion = calculate_score(dev_features, dev_labels, best_pipe) LOG.info("Dev fit accuracy: %.4f", accuracy * 100.0) LOG.info("Dev confusion matrix:\n%s", str(confusion)) return best_pipe, max_score def subsample_manifest(manifest_file: str, max_samples: int) -> str: """Will save a subsampled version of the manifest to the same folder. Have to save to the same folder to support relative paths. Args: manifest_file: path to the manifest file that needs subsampling. max_samples: how many samples to retain. Will randomly select that many lines from the manifest. Returns: str: the path to the subsampled manifest file. """ with open(manifest_file, "rt", encoding="utf-8") as fin: lines = fin.readlines() if max_samples < len(lines): lines = random.sample(lines, max_samples) output_file = manifest_file + "-subsampled" with open(output_file, "wt", encoding="utf-8") as fout: fout.write("".join(lines)) return output_file def cleanup_subsampled_manifests(subsampled_manifests: List[str]): """Removes all generated subsamples manifests.""" for manifest in subsampled_manifests: os.remove(manifest) def compute_all_confidences( hypothesis: Hypothesis, tune_confidence_cfg: TuneConfidenceConfig ) -> Dict[ConfidenceSpec, float]: """Computes a set of confidence scores from a given hypothesis. Works with the output of both CTC and Transducer decoding. Args: hypothesis: generated hypothesis as returned from the transcribe method of the ASR model. tune_confidence_cfg: config specifying what confidence scores to compute. Returns: dict: dictionary with confidenct spec -> confidence score mapping. """ conf_values = {} for exclude_blank in tune_confidence_cfg.exclude_blank: filtered_logprobs = get_filtered_logprobs(hypothesis, exclude_blank) vocab_size = filtered_logprobs.shape[1] for aggregation in tune_confidence_cfg.aggregation: aggr_func = get_confidence_aggregation_bank()[aggregation] for conf_type in tune_confidence_cfg.confidence_type: conf_func = get_confidence_measure_bank()[conf_type] if conf_type == "max_prob": # skipping alpha in this case conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=1.0)).cpu().item() conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, 1.0)] = conf_value else: for alpha in tune_confidence_cfg.alpha: conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=alpha)).cpu().item() conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, alpha)] = conf_value return conf_values def find_best_confidence( train_confidences: List[List[Dict[ConfidenceSpec, float]]], train_labels: List[int], dev_confidences: List[List[Dict[ConfidenceSpec, float]]], dev_labels: List[int], tune_lr: bool, tune_lr_config: TuneConfidenceConfig, ) -> Tuple[ConfidenceConfig, Pipeline]: """Finds the best confidence configuration for model selection. Will loop over all values in the confidence dictionary and fit the LR model (optionally tuning its HPs). The best performing confidence (on the dev set) will be used for the final LR model. Args: train_confidences: this is an object of type ``List[List[Dict[ConfidenceSpec, float]]]``. The shape of this object is [M, N, S], where M: number of models N: number of utterances in all training sets S: number of confidence scores to try This argument will be used to construct np.array objects for each of the confidence scores with the shape [M, N] train_labels: ground-truth labels of the correct model for each data points. This is a list of size [N] dev_confidences: same as training, but for the validation subset. dev_labels: same as training, but for the validation subset. tune_lr: controls whether tuning of LR hyperparameters is performed. tune_lr_cfg: specifies what values of LR hyperparameters to try. Returns: tuple: best confidence config, best model selection pipeline """ max_score = 0 best_pipe = None best_conf_spec = None LOG.info("Evaluation all confidences. Total grid size: %d", len(train_confidences[0][0].keys())) for conf_spec in tqdm(train_confidences[0][0].keys()): cur_train_confidences = [] for model_confs in train_confidences: cur_train_confidences.append([]) for model_conf in model_confs: cur_train_confidences[-1].append(model_conf[conf_spec]) cur_dev_confidences = [] for model_confs in dev_confidences: cur_dev_confidences.append([]) for model_conf in model_confs: cur_dev_confidences[-1].append(model_conf[conf_spec]) # transposing with zip(*list) training_features = np.array(list(zip(*cur_train_confidences))) training_labels = np.array(train_labels) dev_features = np.array(list(zip(*cur_dev_confidences))) dev_labels = np.array(dev_labels) pipe, score = train_model_selection( training_features, training_labels, dev_features, dev_labels, tune_lr, tune_lr_config, ) if max_score < score: max_score = score best_pipe = pipe best_conf_spec = conf_spec LOG.info("Found better parameters: %s. New score: %.4f", str(conf_spec), max_score) return best_conf_spec.to_confidence_config(), best_pipe @hydra_runner(config_name="BuildEnsembleConfig", schema=BuildEnsembleConfig) def main(cfg: BuildEnsembleConfig): # silencing all messages from nemo/ptl to avoid dumping tons of configs to the stdout logging.getLogger('pytorch_lightning').setLevel(logging.CRITICAL) logging.getLogger('nemo_logger').setLevel(logging.CRITICAL) LOG.info(f'Build ensemble config:\n{OmegaConf.to_yaml(cfg)}') # to ensure post init is called cfg = BuildEnsembleConfig(**cfg) pl.seed_everything(cfg.random_seed) cfg.transcription.random_seed = None # seed is already applied cfg.transcription.return_transcriptions = True cfg.transcription.preserve_alignment = True cfg.transcription.ctc_decoding.temperature = cfg.temperature cfg.transcription.rnnt_decoding.temperature = cfg.temperature # this ensures that generated output is after log-softmax for consistency with CTC train_confidences = [] dev_confidences = [] train_labels = [] dev_labels = [] # registering clean-up function that will hold on to this list and # should clean up even if there is partial error in some of the transcribe # calls subsampled_manifests = [] atexit.register(cleanup_subsampled_manifests, subsampled_manifests) # note that we loop over the same config. # This is intentional, as we need to run all models on all datasets # this loop will do the following things: # 1. Goes through each model X each training dataset # 2. Computes predictions by directly calling transcribe_speech.main # 3. Converts transcription to the confidence score(s) as specified in the config # 4. If dev sets are provided, computes the same for them # 5. Creates a list of ground-truth model indices by mapping each model # to its own training dataset as specified in the config. # 6. After the loop, we either run tuning over all confidence scores or # directly use a single score to fit logistic regression and save the # final ensemble model. for model_idx, model_cfg in enumerate(cfg.ensemble): train_model_confidences = [] dev_model_confidences = [] for data_idx, data_cfg in enumerate(cfg.ensemble): if model_idx == 0: # generating subsampled manifests only one time subsampled_manifests.append( subsample_manifest(data_cfg.training_manifest, data_cfg.max_training_samples) ) subsampled_manifest = subsampled_manifests[data_idx] if model_cfg.model.endswith(".nemo"): cfg.transcription.model_path = model_cfg.model else: # assuming pretrained model cfg.transcription.pretrained_name = model_cfg.model cfg.transcription.dataset_manifest = subsampled_manifest # training with tempfile.NamedTemporaryFile() as output_file: cfg.transcription.output_filename = output_file.name LOG.info("Transcribing training dataset %d with model %d", data_idx, model_idx) transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) LOG.info("Generating confidence scores") # TODO: parallelize this loop? for transcription in tqdm(transcriptions): if cfg.tune_confidence: train_model_confidences.append( compute_all_confidences(transcription, cfg.tune_confidence_config) ) else: train_model_confidences.append(compute_confidence(transcription, cfg.confidence)) if model_idx == 0: # labels are the same for all models train_labels.append(data_idx) # optional dev if data_cfg.dev_manifest is not None: cfg.transcription.dataset_manifest = data_cfg.dev_manifest with tempfile.NamedTemporaryFile() as output_file: cfg.transcription.output_filename = output_file.name LOG.info("Transcribing dev dataset %d with model %d", data_idx, model_idx) transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) LOG.info("Generating confidence scores") for transcription in tqdm(transcriptions): if cfg.tune_confidence: dev_model_confidences.append( compute_all_confidences(transcription, cfg.tune_confidence_config) ) else: dev_model_confidences.append(compute_confidence(transcription, cfg.confidence)) if model_idx == 0: # labels are the same for all models dev_labels.append(data_idx) train_confidences.append(train_model_confidences) if dev_model_confidences: dev_confidences.append(dev_model_confidences) if cfg.tune_confidence: best_confidence, model_selection_block = find_best_confidence( train_confidences, train_labels, dev_confidences, dev_labels, cfg.tune_logistic_regression, cfg.tune_logistic_regression_config, ) else: best_confidence = cfg.confidence # transposing with zip(*list) training_features = np.array(list(zip(*train_confidences))) training_labels = np.array(train_labels) if dev_confidences: dev_features = np.array(list(zip(*dev_confidences))) dev_labels = np.array(dev_labels) else: dev_features = None dev_labels = None model_selection_block, _ = train_model_selection( training_features, training_labels, dev_features, dev_labels, cfg.tune_logistic_regression, cfg.tune_logistic_regression_config, verbose=True, ) with tempfile.TemporaryDirectory() as tmpdir: model_selection_block_path = os.path.join(tmpdir, 'model_selection_block.pkl') joblib.dump(model_selection_block, model_selection_block_path) # creating ensemble checkpoint ensemble_model = ConfidenceEnsembleModel( cfg=DictConfig( { 'model_selection_block': model_selection_block_path, 'confidence': best_confidence, 'temperature': cfg.temperature, 'load_models': [model_cfg.model for model_cfg in cfg.ensemble], } ), trainer=None, ) ensemble_model.save_to(cfg.output_path) if __name__ == '__main__': main() [end of scripts/confidence_ensembles/build_ensemble.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os from dataclasses import dataclass, is_dataclass from pathlib import Path from typing import Optional import pytorch_lightning as pl import torch from omegaconf import MISSING, OmegaConf from sklearn.model_selection import ParameterGrid from nemo.collections.asr.metrics.rnnt_wer import RNNTDecodingConfig from nemo.collections.asr.metrics.wer import CTCDecodingConfig from nemo.collections.asr.models import ASRModel, EncDecRNNTModel from nemo.collections.asr.parts.utils.asr_confidence_benchmarking_utils import ( apply_confidence_parameters, run_confidence_benchmark, ) from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig from nemo.core.config import hydra_runner from nemo.utils import logging, model_utils """ Get confidence metrics and curve plots for a given model, dataset, and confidence parameters. # Arguments model_path: Path to .nemo ASR checkpoint pretrained_name: Name of pretrained ASR model (from NGC registry) dataset_manifest: Path to dataset JSON manifest file (in NeMo format) output_dir: Output directory to store a report and curve plot directories batch_size: batch size during inference num_workers: number of workers during inference cuda: Optional int to enable or disable execution of model on certain CUDA device amp: Bool to decide if Automatic Mixed Precision should be used during inference audio_type: Str filetype of the audio. Supported = wav, flac, mp3 target_level: Word- or token-level confidence. Supported = word, token, auto (for computing both word and token) confidence_cfg: Config with confidence parameters grid_params: Dictionary with lists of parameters to iteratively benchmark on # Usage ASR model can be specified by either "model_path" or "pretrained_name". Data for transcription are defined with "dataset_manifest". Results are returned as a benchmark report and curve plots. python benchmark_asr_confidence.py \ model_path=null \ pretrained_name=null \ dataset_manifest="" \ output_dir="" \ batch_size=64 \ num_workers=8 \ cuda=0 \ amp=True \ target_level="word" \ confidence_cfg.exclude_blank=False \ 'grid_params="{\"aggregation\": [\"min\", \"prod\"], \"alpha\": [0.33, 0.5]}"' """ def get_experiment_params(cfg): """Get experiment parameters from a confidence config and generate the experiment name. Returns: List of experiment parameters. String with the experiment name. """ blank = "no_blank" if cfg.exclude_blank else "blank" aggregation = cfg.aggregation method_name = cfg.method_cfg.name alpha = cfg.method_cfg.alpha if method_name == "entropy": entropy_type = cfg.method_cfg.entropy_type entropy_norm = cfg.method_cfg.entropy_norm experiment_param_list = [ aggregation, str(cfg.exclude_blank), method_name, entropy_type, entropy_norm, str(alpha), ] experiment_str = "-".join([aggregation, blank, method_name, entropy_type, entropy_norm, str(alpha)]) else: experiment_param_list = [aggregation, str(cfg.exclude_blank), method_name, "-", "-", str(alpha)] experiment_str = "-".join([aggregation, blank, method_name, str(alpha)]) return experiment_param_list, experiment_str @dataclass class ConfidenceBenchmarkingConfig: # Required configs model_path: Optional[str] = None # Path to a .nemo file pretrained_name: Optional[str] = None # Name of a pretrained model dataset_manifest: str = MISSING output_dir: str = MISSING # General configs batch_size: int = 32 num_workers: int = 4 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA # device anyway, and do inference on CPU only if CUDA device is not found. # If `cuda` is a negative number, inference will be on CPU only. cuda: Optional[int] = None amp: bool = False audio_type: str = "wav" # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) def main(cfg: ConfidenceBenchmarkingConfig): torch.set_grad_enabled(False) logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) if cfg.model_path is None and cfg.pretrained_name is None: raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None!") # setup GPU if cfg.cuda is None: if torch.cuda.is_available(): device = [0] # use 0th CUDA device accelerator = 'gpu' else: device = 1 accelerator = 'cpu' else: device = [cfg.cuda] accelerator = 'gpu' map_location = torch.device('cuda:{}'.format(device[0]) if accelerator == 'gpu' else 'cpu') # setup model if cfg.model_path is not None: # restore model from .nemo file path model_cfg = ASRModel.restore_from(restore_path=cfg.model_path, return_config=True) classpath = model_cfg.target # original class path imported_class = model_utils.import_class_by_path(classpath) # type: ASRModel logging.info(f"Restoring model : {imported_class.__name__}") asr_model = imported_class.restore_from( restore_path=cfg.model_path, map_location=map_location ) # type: ASRModel else: # restore model by name asr_model = ASRModel.from_pretrained( model_name=cfg.pretrained_name, map_location=map_location ) # type: ASRModel trainer = pl.Trainer(devices=device, accelerator=accelerator) asr_model.set_trainer(trainer) asr_model = asr_model.eval() # Check if ctc or rnnt model is_rnnt = isinstance(asr_model, EncDecRNNTModel) # Check that the model has the `change_decoding_strategy` method if not hasattr(asr_model, 'change_decoding_strategy'): raise RuntimeError("The asr_model you are using must have the `change_decoding_strategy` method.") # get filenames and reference texts from manifest filepaths = [] reference_texts = [] if os.stat(cfg.dataset_manifest).st_size == 0: logging.error(f"The input dataset_manifest {cfg.dataset_manifest} is empty. Exiting!") return None manifest_dir = Path(cfg.dataset_manifest).parent with open(cfg.dataset_manifest, 'r') as f: for line in f: item = json.loads(line) audio_file = Path(item['audio_filepath']) if not audio_file.is_file() and not audio_file.is_absolute(): audio_file = manifest_dir / audio_file filepaths.append(str(audio_file.absolute())) reference_texts.append(item['text']) # setup AMP (optional) autocast = None if cfg.amp and torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): logging.info("AMP enabled!\n") autocast = torch.cuda.amp.autocast # do grid-based benchmarking if grid_params is provided, otherwise a regular one work_dir = Path(cfg.output_dir) os.makedirs(work_dir, exist_ok=True) report_legend = ( ",".join( [ "model_type", "aggregation", "blank", "method_name", "entropy_type", "entropy_norm", "alpha", "target_level", "auc_roc", "auc_pr", "auc_nt", "nce", "ece", "auc_yc", "std_yc", "max_yc", ] ) + "\n" ) model_typename = "RNNT" if is_rnnt else "CTC" report_file = work_dir / Path("report.csv") if cfg.grid_params: asr_model.change_decoding_strategy( RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) if is_rnnt else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) ) params = json.loads(cfg.grid_params) hp_grid = ParameterGrid(params) hp_grid = list(hp_grid) logging.info(f"==============================Running a benchmarking with grid search=========================") logging.info(f"Grid search size: {len(hp_grid)}") logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directories near the file") logging.info(f"==============================================================================================") with open(report_file, "tw", encoding="utf-8") as f: f.write(report_legend) f.flush() for i, hp in enumerate(hp_grid): logging.info(f"Run # {i + 1}, grid: `{hp}`") asr_model.change_decoding_strategy(apply_confidence_parameters(asr_model.cfg.decoding, hp)) param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) plot_dir = work_dir / Path(experiment_name) results = run_confidence_benchmark( asr_model, cfg.target_level, filepaths, reference_texts, cfg.batch_size, cfg.num_workers, plot_dir, autocast, ) for level, result in results.items(): f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") f.flush() else: asr_model.change_decoding_strategy( RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) if is_rnnt else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) ) param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) plot_dir = work_dir / Path(experiment_name) logging.info(f"==============================Running a single benchmarking===================================") logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directory `{plot_dir}`") with open(report_file, "tw", encoding="utf-8") as f: f.write(report_legend) f.flush() results = run_confidence_benchmark( asr_model, cfg.batch_size, cfg.num_workers, cfg.target_level, filepaths, reference_texts, plot_dir, autocast, ) for level, result in results.items(): f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") logging.info(f"===========================================Done===============================================") if __name__ == '__main__': main() [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ # This script converts an existing audio dataset with a manifest to # a tarred and sharded audio dataset that can be read by the # TarredAudioToTextDataLayer. # Please make sure your audio_filepath DOES NOT CONTAIN '-sub'! # Because we will use it to handle files which have duplicate filenames but with different offsets # (see function create_shard for details) # Bucketing can help to improve the training speed. You may use --buckets_num to specify the number of buckets. # It creates multiple tarred datasets, one per bucket, based on the audio durations. # The range of [min_duration, max_duration) is split into equal sized buckets. # Recommend to use --sort_in_shards to speedup the training by reducing the paddings in the batches # More info on how to use bucketing feature: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/datasets.html # If valid NVIDIA DALI version is installed, will also generate the corresponding DALI index files that need to be # supplied to the config in order to utilize webdataset for efficient large dataset handling. # NOTE: DALI + Webdataset is NOT compatible with Bucketing support ! # Usage: 1) Creating a new tarfile dataset python convert_to_tarred_audio_dataset.py \ --manifest_path=<path to the manifest file> \ --target_dir=<path to output directory> \ --num_shards=<number of tarfiles that will contain the audio> \ --max_duration=<float representing maximum duration of audio samples> \ --min_duration=<float representing minimum duration of audio samples> \ --shuffle --shuffle_seed=1 \ --sort_in_shards \ --workers=-1 2) Concatenating more tarfiles to a pre-existing tarred dataset python convert_to_tarred_audio_dataset.py \ --manifest_path=<path to the tarred manifest file> \ --metadata_path=<path to the metadata.yaml (or metadata_version_{X}.yaml) file> \ --target_dir=<path to output directory where the original tarfiles are contained> \ --max_duration=<float representing maximum duration of audio samples> \ --min_duration=<float representing minimum duration of audio samples> \ --shuffle --shuffle_seed=1 \ --sort_in_shards \ --workers=-1 \ --concat_manifest_paths \ <space separated paths to 1 or more manifest files to concatenate into the original tarred dataset> 3) Writing an empty metadata file python convert_to_tarred_audio_dataset.py \ --target_dir=<path to output directory> \ # any other optional argument --num_shards=8 \ --max_duration=16.7 \ --min_duration=0.01 \ --shuffle \ --workers=-1 \ --sort_in_shards \ --shuffle_seed=1 \ --write_metadata """ import argparse import copy import json import os import random import tarfile from collections import defaultdict from dataclasses import dataclass, field from datetime import datetime from typing import Any, List, Optional from joblib import Parallel, delayed from omegaconf import DictConfig, OmegaConf, open_dict try: import create_dali_tarred_dataset_index as dali_index DALI_INDEX_SCRIPT_AVAILABLE = True except (ImportError, ModuleNotFoundError, FileNotFoundError): DALI_INDEX_SCRIPT_AVAILABLE = False parser = argparse.ArgumentParser( description="Convert an existing ASR dataset to tarballs compatible with TarredAudioToTextDataLayer." ) parser.add_argument( "--manifest_path", default=None, type=str, required=False, help="Path to the existing dataset's manifest." ) parser.add_argument( '--concat_manifest_paths', nargs='+', default=None, type=str, required=False, help="Path to the additional dataset's manifests that will be concatenated with base dataset.", ) # Optional arguments parser.add_argument( "--target_dir", default='./tarred', type=str, help="Target directory for resulting tarballs and manifest. Defaults to `./tarred`. Creates the path if necessary.", ) parser.add_argument( "--metadata_path", required=False, default=None, type=str, help="Path to metadata file for the dataset.", ) parser.add_argument( "--num_shards", default=-1, type=int, help="Number of shards (tarballs) to create. Used for partitioning data among workers.", ) parser.add_argument( '--max_duration', default=None, required=True, type=float, help='Maximum duration of audio clip in the dataset. By default, it is None and is required to be set.', ) parser.add_argument( '--min_duration', default=None, type=float, help='Minimum duration of audio clip in the dataset. By default, it is None and will not filter files.', ) parser.add_argument( "--shuffle", action='store_true', help="Whether or not to randomly shuffle the samples in the manifest before tarring/sharding.", ) parser.add_argument( "--keep_files_together", action='store_true', help="Whether or not to keep entries from the same file (but different offsets) together when sorting before tarring/sharding.", ) parser.add_argument( "--sort_in_shards", action='store_true', help="Whether or not to sort samples inside the shards based on their duration.", ) parser.add_argument( "--buckets_num", type=int, default=1, help="Number of buckets to create based on duration.", ) parser.add_argument("--shuffle_seed", type=int, default=None, help="Random seed for use if shuffling is enabled.") parser.add_argument( '--write_metadata', action='store_true', help=( "Flag to write a blank metadata with the current call config. " "Note that the metadata will not contain the number of shards, " "and it must be filled out by the user." ), ) parser.add_argument( "--no_shard_manifests", action='store_true', help="Do not write sharded manifests along with the aggregated manifest.", ) parser.add_argument('--workers', type=int, default=1, help='Number of worker processes') args = parser.parse_args() @dataclass class ASRTarredDatasetConfig: num_shards: int = -1 shuffle: bool = False max_duration: Optional[float] = None min_duration: Optional[float] = None shuffle_seed: Optional[int] = None sort_in_shards: bool = True shard_manifests: bool = True keep_files_together: bool = False @dataclass class ASRTarredDatasetMetadata: created_datetime: Optional[str] = None version: int = 0 num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): self.created_datetime = self.get_current_datetime() def get_current_datetime(self): return datetime.now().strftime("%m-%d-%Y %H-%M-%S") @classmethod def from_config(cls, config: DictConfig): obj = cls() obj.__dict__.update(**config) return obj @classmethod def from_file(cls, filepath: str): config = OmegaConf.load(filepath) return ASRTarredDatasetMetadata.from_config(config=config) class ASRTarredDatasetBuilder: """ Helper class that constructs a tarred dataset from scratch, or concatenates tarred datasets together and constructs manifests for them. """ def __init__(self): self.config = None def configure(self, config: ASRTarredDatasetConfig): """ Sets the config generated from command line overrides. Args: config: ASRTarredDatasetConfig dataclass object. """ self.config = config # type: ASRTarredDatasetConfig if self.config.num_shards < 0: raise ValueError("`num_shards` must be > 0. Please fill in the metadata information correctly.") def create_new_dataset(self, manifest_path: str, target_dir: str = "./tarred/", num_workers: int = 0): """ Creates a new tarred dataset from a given manifest file. Args: manifest_path: Path to the original ASR manifest. target_dir: Output directory. num_workers: Integer denoting number of parallel worker processes which will write tarfiles. Defaults to 1 - which denotes sequential worker process. Output: Writes tarfiles, along with the tarred dataset compatible manifest file. Also preserves a record of the metadata used to construct this tarred dataset. """ if self.config is None: raise ValueError("Config has not been set. Please call `configure(config: ASRTarredDatasetConfig)`") if manifest_path is None: raise FileNotFoundError("Manifest filepath cannot be None !") config = self.config # type: ASRTarredDatasetConfig if not os.path.exists(target_dir): os.makedirs(target_dir) # Read the existing manifest entries, total_duration, filtered_entries, filtered_duration = self._read_manifest(manifest_path, config) if len(filtered_entries) > 0: print(f"Filtered {len(filtered_entries)} files which amounts to {filtered_duration} seconds of audio.") print( f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." ) if len(entries) == 0: print("No tarred dataset was created as there were 0 valid samples after filtering!") return if config.shuffle: random.seed(config.shuffle_seed) print("Shuffling...") if config.keep_files_together: filename_entries = defaultdict(list) for ent in entries: filename_entries[ent["audio_filepath"]].append(ent) filenames = list(filename_entries.keys()) random.shuffle(filenames) shuffled_entries = [] for filename in filenames: shuffled_entries += filename_entries[filename] entries = shuffled_entries else: random.shuffle(entries) # Create shards and updated manifest entries print(f"Number of samples added : {len(entries)}") print(f"Remainder: {len(entries) % config.num_shards}") start_indices = [] end_indices = [] # Build indices for i in range(config.num_shards): start_idx = (len(entries) // config.num_shards) * i end_idx = start_idx + (len(entries) // config.num_shards) print(f"Shard {i} has entries {start_idx} ~ {end_idx}") files = set() for ent_id in range(start_idx, end_idx): files.add(entries[ent_id]["audio_filepath"]) print(f"Shard {i} contains {len(files)} files") if i == config.num_shards - 1: # We discard in order to have the same number of entries per shard. print(f"Have {len(entries) - end_idx} entries left over that will be discarded.") start_indices.append(start_idx) end_indices.append(end_idx) manifest_folder, _ = os.path.split(manifest_path) with Parallel(n_jobs=num_workers, verbose=config.num_shards) as parallel: # Call parallel tarfile construction new_entries_list = parallel( delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, i, manifest_folder) for i, (start_idx, end_idx) in enumerate(zip(start_indices, end_indices)) ) if config.shard_manifests: sharded_manifests_dir = target_dir + '/sharded_manifests' if not os.path.exists(sharded_manifests_dir): os.makedirs(sharded_manifests_dir) for manifest in new_entries_list: shard_id = manifest[0]['shard_id'] new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: for entry in manifest: json.dump(entry, m2) m2.write('\n') # Flatten the list of list of entries to a list of entries new_entries = [sample for manifest in new_entries_list for sample in manifest] del new_entries_list print("Total number of entries in manifest :", len(new_entries)) # Write manifest new_manifest_path = os.path.join(target_dir, 'tarred_audio_manifest.json') with open(new_manifest_path, 'w', encoding='utf-8') as m2: for entry in new_entries: json.dump(entry, m2) m2.write('\n') # Write metadata (default metadata for new datasets) new_metadata_path = os.path.join(target_dir, 'metadata.yaml') metadata = ASRTarredDatasetMetadata() # Update metadata metadata.dataset_config = config metadata.num_samples_per_shard = len(new_entries) // config.num_shards # Write metadata metadata_yaml = OmegaConf.structured(metadata) OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) def create_concatenated_dataset( self, base_manifest_path: str, manifest_paths: List[str], metadata: ASRTarredDatasetMetadata, target_dir: str = "./tarred_concatenated/", num_workers: int = 1, ): """ Creates new tarfiles in order to create a concatenated dataset, whose manifest contains the data for both the original dataset as well as the new data submitted in manifest paths. Args: base_manifest_path: Path to the manifest file which contains the information for the original tarred dataset (with flattened paths). manifest_paths: List of one or more paths to manifest files that will be concatenated with above base tarred dataset. metadata: ASRTarredDatasetMetadata dataclass instance with overrides from command line. target_dir: Output directory Output: Writes tarfiles which with indices mapping to a "concatenated" tarred dataset, along with the tarred dataset compatible manifest file which includes information about all the datasets that comprise the concatenated dataset. Also preserves a record of the metadata used to construct this tarred dataset. """ if not os.path.exists(target_dir): os.makedirs(target_dir) if base_manifest_path is None: raise FileNotFoundError("Base manifest filepath cannot be None !") if manifest_paths is None or len(manifest_paths) == 0: raise FileNotFoundError("List of additional manifest filepaths cannot be None !") config = ASRTarredDatasetConfig(**(metadata.dataset_config)) # Read the existing manifest (no filtering here) base_entries, _, _, _ = self._read_manifest(base_manifest_path, config) print(f"Read base manifest containing {len(base_entries)} samples.") # Precompute number of samples per shard if metadata.num_samples_per_shard is None: num_samples_per_shard = len(base_entries) // config.num_shards else: num_samples_per_shard = metadata.num_samples_per_shard print("Number of samples per shard :", num_samples_per_shard) # Compute min and max duration and update config (if no metadata passed) print(f"Selected max duration : {config.max_duration}") print(f"Selected min duration : {config.min_duration}") entries = [] for new_manifest_idx in range(len(manifest_paths)): new_entries, total_duration, filtered_new_entries, filtered_duration = self._read_manifest( manifest_paths[new_manifest_idx], config ) if len(filtered_new_entries) > 0: print( f"Filtered {len(filtered_new_entries)} files which amounts to {filtered_duration:0.2f}" f" seconds of audio from manifest {manifest_paths[new_manifest_idx]}." ) print( f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." ) entries.extend(new_entries) if len(entries) == 0: print("No tarred dataset was created as there were 0 valid samples after filtering!") return if config.shuffle: random.seed(config.shuffle_seed) print("Shuffling...") random.shuffle(entries) # Drop last section of samples that cannot be added onto a chunk drop_count = len(entries) % num_samples_per_shard total_new_entries = len(entries) entries = entries[:-drop_count] print( f"Dropping {drop_count} samples from total new samples {total_new_entries} since they cannot " f"be added into a uniformly sized chunk." ) # Create shards and updated manifest entries num_added_shards = len(entries) // num_samples_per_shard print(f"Number of samples in base dataset : {len(base_entries)}") print(f"Number of samples in additional datasets : {len(entries)}") print(f"Number of added shards : {num_added_shards}") print(f"Remainder: {len(entries) % num_samples_per_shard}") start_indices = [] end_indices = [] shard_indices = [] for i in range(num_added_shards): start_idx = (len(entries) // num_added_shards) * i end_idx = start_idx + (len(entries) // num_added_shards) shard_idx = i + config.num_shards print(f"Shard {shard_idx} has entries {start_idx + len(base_entries)} ~ {end_idx + len(base_entries)}") start_indices.append(start_idx) end_indices.append(end_idx) shard_indices.append(shard_idx) manifest_folder, _ = os.path.split(base_manifest_path) with Parallel(n_jobs=num_workers, verbose=num_added_shards) as parallel: # Call parallel tarfile construction new_entries_list = parallel( delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, shard_idx, manifest_folder) for i, (start_idx, end_idx, shard_idx) in enumerate(zip(start_indices, end_indices, shard_indices)) ) if config.shard_manifests: sharded_manifests_dir = target_dir + '/sharded_manifests' if not os.path.exists(sharded_manifests_dir): os.makedirs(sharded_manifests_dir) for manifest in new_entries_list: shard_id = manifest[0]['shard_id'] new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: for entry in manifest: json.dump(entry, m2) m2.write('\n') # Flatten the list of list of entries to a list of entries new_entries = [sample for manifest in new_entries_list for sample in manifest] del new_entries_list # Write manifest if metadata is None: new_version = 1 # start with `1`, where `0` indicates the base manifest + dataset else: new_version = metadata.version + 1 print("Total number of entries in manifest :", len(base_entries) + len(new_entries)) new_manifest_path = os.path.join(target_dir, f'tarred_audio_manifest_version_{new_version}.json') with open(new_manifest_path, 'w', encoding='utf-8') as m2: # First write all the entries of base manifest for entry in base_entries: json.dump(entry, m2) m2.write('\n') # Finally write the new entries for entry in new_entries: json.dump(entry, m2) m2.write('\n') # Preserve historical metadata base_metadata = metadata # Write metadata (updated metadata for concatenated datasets) new_metadata_path = os.path.join(target_dir, f'metadata_version_{new_version}.yaml') metadata = ASRTarredDatasetMetadata() # Update config config.num_shards = config.num_shards + num_added_shards # Update metadata metadata.version = new_version metadata.dataset_config = config metadata.num_samples_per_shard = num_samples_per_shard metadata.is_concatenated_manifest = True metadata.created_datetime = metadata.get_current_datetime() # Attach history current_metadata = OmegaConf.structured(base_metadata.history) metadata.history = current_metadata # Write metadata metadata_yaml = OmegaConf.structured(metadata) OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) def _read_manifest(self, manifest_path: str, config: ASRTarredDatasetConfig): """Read and filters data from the manifest""" # Read the existing manifest entries = [] total_duration = 0.0 filtered_entries = [] filtered_duration = 0.0 with open(manifest_path, 'r', encoding='utf-8') as m: for line in m: entry = json.loads(line) if (config.max_duration is None or entry['duration'] < config.max_duration) and ( config.min_duration is None or entry['duration'] >= config.min_duration ): entries.append(entry) total_duration += entry["duration"] else: filtered_entries.append(entry) filtered_duration += entry['duration'] return entries, total_duration, filtered_entries, filtered_duration def _create_shard(self, entries, target_dir, shard_id, manifest_folder): """Creates a tarball containing the audio files from `entries`. """ if self.config.sort_in_shards: entries.sort(key=lambda x: x["duration"], reverse=False) new_entries = [] tar = tarfile.open(os.path.join(target_dir, f'audio_{shard_id}.tar'), mode='w', dereference=True) count = dict() for entry in entries: # We squash the filename since we do not preserve directory structure of audio files in the tarball. if os.path.exists(entry["audio_filepath"]): audio_filepath = entry["audio_filepath"] else: audio_filepath = os.path.join(manifest_folder, entry["audio_filepath"]) if not os.path.exists(audio_filepath): raise FileNotFoundError(f"Could not find {entry['audio_filepath']}!") base, ext = os.path.splitext(audio_filepath) base = base.replace('/', '_') # Need the following replacement as long as WebDataset splits on first period base = base.replace('.', '_') squashed_filename = f'{base}{ext}' if squashed_filename not in count: tar.add(audio_filepath, arcname=squashed_filename) to_write = squashed_filename count[squashed_filename] = 1 else: to_write = base + "-sub" + str(count[squashed_filename]) + ext count[squashed_filename] += 1 new_entry = { 'audio_filepath': to_write, 'duration': entry['duration'], 'shard_id': shard_id, # Keep shard ID for recordkeeping } if 'label' in entry: new_entry['label'] = entry['label'] if 'text' in entry: new_entry['text'] = entry['text'] if 'offset' in entry: new_entry['offset'] = entry['offset'] if 'lang' in entry: new_entry['lang'] = entry['lang'] new_entries.append(new_entry) tar.close() return new_entries @classmethod def setup_history(cls, base_metadata: ASRTarredDatasetMetadata, history: List[Any]): if 'history' in base_metadata.keys(): for history_val in base_metadata.history: cls.setup_history(history_val, history) if base_metadata is not None: metadata_copy = copy.deepcopy(base_metadata) with open_dict(metadata_copy): metadata_copy.pop('history', None) history.append(metadata_copy) def main(): if args.buckets_num > 1: bucket_length = (args.max_duration - args.min_duration) / float(args.buckets_num) for i in range(args.buckets_num): min_duration = args.min_duration + i * bucket_length max_duration = min_duration + bucket_length if i == args.buckets_num - 1: # add a small number to cover the samples with exactly duration of max_duration in the last bucket. max_duration += 1e-5 target_dir = os.path.join(args.target_dir, f"bucket{i+1}") print(f"Creating bucket {i+1} with min_duration={min_duration} and max_duration={max_duration} ...") print(f"Results are being saved at: {target_dir}.") create_tar_datasets(min_duration=min_duration, max_duration=max_duration, target_dir=target_dir) print(f"Bucket {i+1} is created.") else: create_tar_datasets(min_duration=args.min_duration, max_duration=args.max_duration, target_dir=args.target_dir) def create_tar_datasets(min_duration: float, max_duration: float, target_dir: str): builder = ASRTarredDatasetBuilder() shard_manifests = False if args.no_shard_manifests else True if args.write_metadata: metadata = ASRTarredDatasetMetadata() dataset_cfg = ASRTarredDatasetConfig( num_shards=args.num_shards, shuffle=args.shuffle, max_duration=max_duration, min_duration=min_duration, shuffle_seed=args.shuffle_seed, sort_in_shards=args.sort_in_shards, shard_manifests=shard_manifests, keep_files_together=args.keep_files_together, ) metadata.dataset_config = dataset_cfg output_path = os.path.join(target_dir, 'default_metadata.yaml') OmegaConf.save(metadata, output_path, resolve=True) print(f"Default metadata written to {output_path}") exit(0) if args.concat_manifest_paths is None or len(args.concat_manifest_paths) == 0: print("Creating new tarred dataset ...") # Create a tarred dataset from scratch config = ASRTarredDatasetConfig( num_shards=args.num_shards, shuffle=args.shuffle, max_duration=max_duration, min_duration=min_duration, shuffle_seed=args.shuffle_seed, sort_in_shards=args.sort_in_shards, shard_manifests=shard_manifests, keep_files_together=args.keep_files_together, ) builder.configure(config) builder.create_new_dataset(manifest_path=args.manifest_path, target_dir=target_dir, num_workers=args.workers) else: if args.buckets_num > 1: raise ValueError("Concatenation feature does not support buckets_num > 1.") print("Concatenating multiple tarred datasets ...") # Implicitly update config from base details if args.metadata_path is not None: metadata = ASRTarredDatasetMetadata.from_file(args.metadata_path) else: raise ValueError("`metadata` yaml file path must be provided!") # Preserve history history = [] builder.setup_history(OmegaConf.structured(metadata), history) metadata.history = history # Add command line overrides (everything other than num_shards) metadata.dataset_config.max_duration = max_duration metadata.dataset_config.min_duration = min_duration metadata.dataset_config.shuffle = args.shuffle metadata.dataset_config.shuffle_seed = args.shuffle_seed metadata.dataset_config.sort_in_shards = args.sort_in_shards metadata.dataset_config.shard_manifests = shard_manifests builder.configure(metadata.dataset_config) # Concatenate a tarred dataset onto a previous one builder.create_concatenated_dataset( base_manifest_path=args.manifest_path, manifest_paths=args.concat_manifest_paths, metadata=metadata, target_dir=target_dir, num_workers=args.workers, ) if DALI_INDEX_SCRIPT_AVAILABLE and dali_index.INDEX_CREATOR_AVAILABLE: print("Constructing DALI Tarfile Index - ", target_dir) index_config = dali_index.DALITarredIndexConfig(tar_dir=target_dir, workers=args.workers) dali_index.main(index_config) if __name__ == "__main__": main() [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of tools/nemo_forced_aligner/align.py] # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import math import os from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import List, Optional import torch from omegaconf import OmegaConf from utils.data_prep import ( add_t_start_end_to_utt_obj, get_batch_starts_ends, get_batch_variables, get_manifest_lines_batch, is_entry_in_all_lines, is_entry_in_any_lines, ) from utils.make_ass_files import make_ass_files from utils.make_ctm_files import make_ctm_files from utils.make_output_manifest import write_manifest_out_line from utils.viterbi_decoding import viterbi_decoding from nemo.collections.asr.models.ctc_models import EncDecCTCModel from nemo.collections.asr.models.hybrid_rnnt_ctc_models import EncDecHybridRNNTCTCModel from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchASR from nemo.collections.asr.parts.utils.transcribe_utils import setup_model from nemo.core.config import hydra_runner from nemo.utils import logging """ Align the utterances in manifest_filepath. Results are saved in ctm files in output_dir. Arguments: pretrained_name: string specifying the name of a CTC NeMo ASR model which will be automatically downloaded from NGC and used for generating the log-probs which we will use to do alignment. Note: NFA can only use CTC models (not Transducer models) at the moment. model_path: string specifying the local filepath to a CTC NeMo ASR model which will be used to generate the log-probs which we will use to do alignment. Note: NFA can only use CTC models (not Transducer models) at the moment. Note: if a model_path is provided, it will override the pretrained_name. manifest_filepath: filepath to the manifest of the data you want to align, containing 'audio_filepath' and 'text' fields. output_dir: the folder where output CTM files and new JSON manifest will be saved. align_using_pred_text: if True, will transcribe the audio using the specified model and then use that transcription as the reference text for the forced alignment. transcribe_device: None, or a string specifying the device that will be used for generating log-probs (i.e. "transcribing"). The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available (otherwise will set it to 'cpu'). viterbi_device: None, or string specifying the device that will be used for doing Viterbi decoding. The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available (otherwise will set it to 'cpu'). batch_size: int specifying batch size that will be used for generating log-probs and doing Viterbi decoding. use_local_attention: boolean flag specifying whether to try to use local attention for the ASR Model (will only work if the ASR Model is a Conformer model). If local attention is used, we will set the local attention context size to [64,64]. additional_segment_grouping_separator: an optional string used to separate the text into smaller segments. If this is not specified, then the whole text will be treated as a single segment. remove_blank_tokens_from_ctm: a boolean denoting whether to remove <blank> tokens from token-level output CTMs. audio_filepath_parts_in_utt_id: int specifying how many of the 'parts' of the audio_filepath we will use (starting from the final part of the audio_filepath) to determine the utt_id that will be used in the CTM files. Note also that any spaces that are present in the audio_filepath will be replaced with dashes, so as not to change the number of space-separated elements in the CTM files. e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 1 => utt_id will be "e1" e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 2 => utt_id will be "d_e1" e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 3 => utt_id will be "c_d_e1" use_buffered_infer: False, if set True, using streaming to do get the logits for alignment This flag is useful when aligning large audio file. However, currently the chunk streaming inference does not support batch inference, which means even you set batch_size > 1, it will only infer one by one instead of doing the whole batch inference together. chunk_len_in_secs: float chunk length in seconds total_buffer_in_secs: float Length of buffer (chunk + left and right padding) in seconds chunk_batch_size: int batch size for buffered chunk inference, which will cut one audio into segments and do inference on chunk_batch_size segments at a time simulate_cache_aware_streaming: False, if set True, using cache aware streaming to do get the logits for alignment save_output_file_formats: List of strings specifying what type of output files to save (default: ["ctm", "ass"]) ctm_file_config: CTMFileConfig to specify the configuration of the output CTM files ass_file_config: ASSFileConfig to specify the configuration of the output ASS files """ @dataclass class CTMFileConfig: remove_blank_tokens: bool = False # minimum duration (in seconds) for timestamps in the CTM.If any line in the CTM has a # duration lower than this, it will be enlarged from the middle outwards until it # meets the minimum_timestamp_duration, or reaches the beginning or end of the audio file. # Note that this may cause timestamps to overlap. minimum_timestamp_duration: float = 0 @dataclass class ASSFileConfig: fontsize: int = 20 vertical_alignment: str = "center" # if resegment_text_to_fill_space is True, the ASS files will use new segments # such that each segment will not take up more than (approximately) max_lines_per_segment # when the ASS file is applied to a video resegment_text_to_fill_space: bool = False max_lines_per_segment: int = 2 text_already_spoken_rgb: List[int] = field(default_factory=lambda: [49, 46, 61]) # dark gray text_being_spoken_rgb: List[int] = field(default_factory=lambda: [57, 171, 9]) # dark green text_not_yet_spoken_rgb: List[int] = field(default_factory=lambda: [194, 193, 199]) # light gray @dataclass class AlignmentConfig: # Required configs pretrained_name: Optional[str] = None model_path: Optional[str] = None manifest_filepath: Optional[str] = None output_dir: Optional[str] = None # General configs align_using_pred_text: bool = False transcribe_device: Optional[str] = None viterbi_device: Optional[str] = None batch_size: int = 1 use_local_attention: bool = True additional_segment_grouping_separator: Optional[str] = None audio_filepath_parts_in_utt_id: int = 1 # Buffered chunked streaming configs use_buffered_chunked_streaming: bool = False chunk_len_in_secs: float = 1.6 total_buffer_in_secs: float = 4.0 chunk_batch_size: int = 32 # Cache aware streaming configs simulate_cache_aware_streaming: Optional[bool] = False # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) ctm_file_config: CTMFileConfig = CTMFileConfig() ass_file_config: ASSFileConfig = ASSFileConfig() @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) def main(cfg: AlignmentConfig): logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') if is_dataclass(cfg): cfg = OmegaConf.structured(cfg) # Validate config if cfg.model_path is None and cfg.pretrained_name is None: raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None") if cfg.model_path is not None and cfg.pretrained_name is not None: raise ValueError("One of cfg.model_path and cfg.pretrained_name must be None") if cfg.manifest_filepath is None: raise ValueError("cfg.manifest_filepath must be specified") if cfg.output_dir is None: raise ValueError("cfg.output_dir must be specified") if cfg.batch_size < 1: raise ValueError("cfg.batch_size cannot be zero or a negative number") if cfg.additional_segment_grouping_separator == "" or cfg.additional_segment_grouping_separator == " ": raise ValueError("cfg.additional_grouping_separator cannot be empty string or space character") if cfg.ctm_file_config.minimum_timestamp_duration < 0: raise ValueError("cfg.minimum_timestamp_duration cannot be a negative number") if cfg.ass_file_config.vertical_alignment not in ["top", "center", "bottom"]: raise ValueError("cfg.ass_file_config.vertical_alignment must be one of 'top', 'center' or 'bottom'") for rgb_list in [ cfg.ass_file_config.text_already_spoken_rgb, cfg.ass_file_config.text_already_spoken_rgb, cfg.ass_file_config.text_already_spoken_rgb, ]: if len(rgb_list) != 3: raise ValueError( "cfg.ass_file_config.text_already_spoken_rgb," " cfg.ass_file_config.text_being_spoken_rgb," " and cfg.ass_file_config.text_already_spoken_rgb all need to contain" " exactly 3 elements." ) # Validate manifest contents if not is_entry_in_all_lines(cfg.manifest_filepath, "audio_filepath"): raise RuntimeError( "At least one line in cfg.manifest_filepath does not contain an 'audio_filepath' entry. " "All lines must contain an 'audio_filepath' entry." ) if cfg.align_using_pred_text: if is_entry_in_any_lines(cfg.manifest_filepath, "pred_text"): raise RuntimeError( "Cannot specify cfg.align_using_pred_text=True when the manifest at cfg.manifest_filepath " "contains 'pred_text' entries. This is because the audio will be transcribed and may produce " "a different 'pred_text'. This may cause confusion." ) else: if not is_entry_in_all_lines(cfg.manifest_filepath, "text"): raise RuntimeError( "At least one line in cfg.manifest_filepath does not contain a 'text' entry. " "NFA requires all lines to contain a 'text' entry when cfg.align_using_pred_text=False." ) # init devices if cfg.transcribe_device is None: transcribe_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") else: transcribe_device = torch.device(cfg.transcribe_device) logging.info(f"Device to be used for transcription step (`transcribe_device`) is {transcribe_device}") if cfg.viterbi_device is None: viterbi_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") else: viterbi_device = torch.device(cfg.viterbi_device) logging.info(f"Device to be used for viterbi step (`viterbi_device`) is {viterbi_device}") if transcribe_device.type == 'cuda' or viterbi_device.type == 'cuda': logging.warning( 'One or both of transcribe_device and viterbi_device are GPUs. If you run into OOM errors ' 'it may help to change both devices to be the CPU.' ) # load model model, _ = setup_model(cfg, transcribe_device) model.eval() if isinstance(model, EncDecHybridRNNTCTCModel): model.change_decoding_strategy(decoder_type="ctc") if cfg.use_local_attention: logging.info( "Flag use_local_attention is set to True => will try to use local attention for model if it allows it" ) model.change_attention_model(self_attention_model="rel_pos_local_attn", att_context_size=[64, 64]) if not (isinstance(model, EncDecCTCModel) or isinstance(model, EncDecHybridRNNTCTCModel)): raise NotImplementedError( f"Model is not an instance of NeMo EncDecCTCModel or ENCDecHybridRNNTCTCModel." " Currently only instances of these models are supported" ) if cfg.ctm_file_config.minimum_timestamp_duration > 0: logging.warning( f"cfg.ctm_file_config.minimum_timestamp_duration has been set to {cfg.ctm_file_config.minimum_timestamp_duration} seconds. " "This may cause the alignments for some tokens/words/additional segments to be overlapping." ) buffered_chunk_params = {} if cfg.use_buffered_chunked_streaming: model_cfg = copy.deepcopy(model._cfg) OmegaConf.set_struct(model_cfg.preprocessor, False) # some changes for streaming scenario model_cfg.preprocessor.dither = 0.0 model_cfg.preprocessor.pad_to = 0 if model_cfg.preprocessor.normalize != "per_feature": logging.error( "Only EncDecCTCModelBPE models trained with per_feature normalization are supported currently" ) # Disable config overwriting OmegaConf.set_struct(model_cfg.preprocessor, True) feature_stride = model_cfg.preprocessor['window_stride'] model_stride_in_secs = feature_stride * cfg.model_downsample_factor total_buffer = cfg.total_buffer_in_secs chunk_len = float(cfg.chunk_len_in_secs) tokens_per_chunk = math.ceil(chunk_len / model_stride_in_secs) mid_delay = math.ceil((chunk_len + (total_buffer - chunk_len) / 2) / model_stride_in_secs) logging.info(f"tokens_per_chunk is {tokens_per_chunk}, mid_delay is {mid_delay}") model = FrameBatchASR( asr_model=model, frame_len=chunk_len, total_buffer=cfg.total_buffer_in_secs, batch_size=cfg.chunk_batch_size, ) buffered_chunk_params = { "delay": mid_delay, "model_stride_in_secs": model_stride_in_secs, "tokens_per_chunk": tokens_per_chunk, } # get start and end line IDs of batches starts, ends = get_batch_starts_ends(cfg.manifest_filepath, cfg.batch_size) # init output_timestep_duration = None and we will calculate and update it during the first batch output_timestep_duration = None # init f_manifest_out os.makedirs(cfg.output_dir, exist_ok=True) tgt_manifest_name = str(Path(cfg.manifest_filepath).stem) + "_with_output_file_paths.json" tgt_manifest_filepath = str(Path(cfg.output_dir) / tgt_manifest_name) f_manifest_out = open(tgt_manifest_filepath, 'w') # get alignment and save in CTM batch-by-batch for start, end in zip(starts, ends): manifest_lines_batch = get_manifest_lines_batch(cfg.manifest_filepath, start, end) (log_probs_batch, y_batch, T_batch, U_batch, utt_obj_batch, output_timestep_duration,) = get_batch_variables( manifest_lines_batch, model, cfg.additional_segment_grouping_separator, cfg.align_using_pred_text, cfg.audio_filepath_parts_in_utt_id, output_timestep_duration, cfg.simulate_cache_aware_streaming, cfg.use_buffered_chunked_streaming, buffered_chunk_params, ) alignments_batch = viterbi_decoding(log_probs_batch, y_batch, T_batch, U_batch, viterbi_device) for utt_obj, alignment_utt in zip(utt_obj_batch, alignments_batch): utt_obj = add_t_start_end_to_utt_obj(utt_obj, alignment_utt, output_timestep_duration) if "ctm" in cfg.save_output_file_formats: utt_obj = make_ctm_files(utt_obj, cfg.output_dir, cfg.ctm_file_config,) if "ass" in cfg.save_output_file_formats: utt_obj = make_ass_files(utt_obj, cfg.output_dir, cfg.ass_file_config) write_manifest_out_line( f_manifest_out, utt_obj, ) f_manifest_out.close() return None if __name__ == "__main__": main() [end of tools/nemo_forced_aligner/align.py] </code> I need you to solve this issue by regenerating the full files in the code base that you would like to change. You can change as many files as you like. Please respond with a list of files and their revised contents in the following format. <example> [start of /src/this_file.py] import os def euclidean(a, b): if b == 0: return a return euclidean(b, a % b) [end of /src/this_file.py] [start of /src/another_file.py] def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) x, y = x0, y0 sx = -1 if x0 > x1 else 1 sy = -1 if y0 > y1 else 1 if dx > dy: err = dx / 2.0 while x != x1: points.append((x, y)) err -= dy if err < 0: y += sy err += dx x += sx else: err = dy / 2.0 while y != y1: points.append((x err -= dx if err < 0: x += sx err += dy y += sy points.append((x, y)) return points [end of /src/another_file.py] </example>
NVIDIA/NeMo
15db83ec4a65e649d83b61d7a4a58d911586e853
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-10-03T19:14:38Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,7 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,7 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed @@ -2201,7 +2201,7 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -175,7 +175,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() + method_cfg: ConfidenceMethodConfig = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -118,7 +118,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
slackapi__python-slack-events-api-71
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): </issue> <code> [start of README.rst] Slack Events API adapter for Python =================================== .. image:: https://badge.fury.io/py/slackeventsapi.svg :target: https://pypi.org/project/slackeventsapi/ .. image:: https://travis-ci.org/slackapi/python-slack-events-api.svg?branch=master :target: https://travis-ci.org/slackapi/python-slack-events-api .. image:: https://codecov.io/gh/slackapi/python-slack-events-api/branch/master/graph/badge.svg :target: https://codecov.io/gh/slackapi/python-slack-events-api The Slack Events Adapter is a Python-based solution to receive and parse events from Slackโ€™s Events API. This library uses an event emitter framework to allow you to easily process Slack events by simply attaching functions to event listeners. This adapter enhances and simplifies Slack's Events API by incorporating useful best practices, patterns, and opportunities to abstract out common tasks. ๐Ÿ’ก We wrote a `blog post which explains how`_ the Events API can help you, why we built these tools, and how you can use them to build production-ready Slack apps. .. _blog post which explains how: https://medium.com/@SlackAPI/enhancing-slacks-events-api-7535827829ab ๐Ÿค– Installation ---------------- .. code:: shell pip install slackeventsapi ๐Ÿค– App Setup -------------------- Before you can use the `Events API`_ you must `create a Slack App`_, and turn on `Event Subscriptions`_. ๐Ÿ’ก When you add the Request URL to your app's Event Subscription settings, Slack will send a request containing a `challenge` code to verify that your server is alive. This package handles that URL Verification event for you, so all you need to do is start the example app, start ngrok and configure your URL accordingly. โœ… Once you have your `Request URL` verified, your app is ready to start receiving Team Events. ๐Ÿ”‘ Your server will begin receiving Events from Slack's Events API as soon as a user has authorized your app. ๐Ÿค– Development workflow: =========================== (1) Create a Slack app on https://api.slack.com/apps (2) Add a `bot user` for your app (3) Start the example app on your **Request URL** endpoint (4) Start ngrok and copy the **HTTPS** URL (5) Add your **Request URL** and subscribe your app to events (6) Go to your ngrok URL (e.g. https://myapp12.ngrok.com/) and auth your app **๐ŸŽ‰ Once your app has been authorized, you will begin receiving Slack Events** โš ๏ธ Ngrok is a great tool for developing Slack apps, but we don't recommend using ngrok for production apps. ๐Ÿค– Usage ---------- **โš ๏ธ Keep your app's credentials safe!** - For development, keep them in virtualenv variables. - For production, use a secure data store. - Never post your app's credentials to github. .. code:: python SLACK_SIGNING_SECRET = os.environ["SLACK_SIGNING_SECRET"] Create a Slack Event Adapter for receiving actions via the Events API ----------------------------------------------------------------------- **Using the built-in Flask server:** .. code:: python from slackeventsapi import SlackEventAdapter slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, endpoint="/slack/events") # Create an event listener for "reaction_added" events and print the emoji name @slack_events_adapter.on("reaction_added") def reaction_added(event_data): emoji = event_data["event"]["reaction"] print(emoji) # Start the server on port 3000 slack_events_adapter.start(port=3000) **Using your existing Flask instance:** .. code:: python from flask import Flask from slackeventsapi import SlackEventAdapter # This `app` represents your existing Flask app app = Flask(__name__) # An example of one of your Flask app's routes @app.route("/") def hello(): return "Hello there!" # Bind the Events API route to your existing Flask app by passing the server # instance as the last param, or with `server=app`. slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, "/slack/events", app) # Create an event listener for "reaction_added" events and print the emoji name @slack_events_adapter.on("reaction_added") def reaction_added(event_data): emoji = event_data["event"]["reaction"] print(emoji) # Start the server on port 3000 if __name__ == "__main__": app.run(port=3000) For a comprehensive list of available Slack `Events` and more information on `Scopes`, see https://api.slack.com/events-api ๐Ÿค– Example event listeners ----------------------------- See `example.py`_ for usage examples. This example also utilizes the SlackClient Web API client. .. _example.py: /example/ ๐Ÿค” Support ----------- Need help? Join `Slack Community`_ and talk to us in `#slack-api`_. You can also `create an Issue`_ right here on GitHub. .. _Events API: https://api.slack.com/events-api .. _create a Slack App: https://api.slack.com/apps/new .. _Event Subscriptions: https://api.slack.com/events-api#subscriptions .. _Slack Community: http://slackcommunity.com/ .. _#slack-api: https://dev4slack.slack.com/messages/slack-api/ .. _create an Issue: https://github.com/slackapi/python-slack-events-api/issues/new [end of README.rst] [start of /dev/null] [end of /dev/null] [start of slackeventsapi/server.py] from flask import Flask, request, make_response, Blueprint import json import platform import sys import hmac import hashlib from time import time from .version import __version__ class SlackServer(Flask): def __init__(self, signing_secret, endpoint, emitter, server): self.signing_secret = signing_secret self.emitter = emitter self.endpoint = endpoint self.package_info = self.get_package_info() # If a server is passed in, bind the event handler routes to it, # otherwise create a new Flask instance. if server: if isinstance(server, Flask) or isinstance(server, Blueprint): self.bind_route(server) else: raise TypeError("Server must be an instance of Flask or Blueprint") else: Flask.__init__(self, __name__) self.bind_route(self) def get_package_info(self): client_name = __name__.split('.')[0] client_version = __version__ # Version is returned from version.py # Collect the package info, Python version and OS version. package_info = { "client": "{0}/{1}".format(client_name, client_version), "python": "Python/{v.major}.{v.minor}.{v.micro}".format(v=sys.version_info), "system": "{0}/{1}".format(platform.system(), platform.release()) } # Concatenate and format the user-agent string to be passed into request headers ua_string = [] for key, val in package_info.items(): ua_string.append(val) return " ".join(ua_string) def verify_signature(self, timestamp, signature): # Verify the request signature of the request sent from Slack # Generate a new hash using the app's signing secret and request data # Compare the generated hash and incoming request signature # Python 2.7.6 doesn't support compare_digest # It's recommended to use Python 2.7.7+ # noqa See https://docs.python.org/2/whatsnew/2.7.html#pep-466-network-security-enhancements-for-python-2-7 req = str.encode('v0:' + str(timestamp) + ':') + request.get_data() request_hash = 'v0=' + hmac.new( str.encode(self.signing_secret), req, hashlib.sha256 ).hexdigest() if hasattr(hmac, "compare_digest"): # Compare byte strings for Python 2 if (sys.version_info[0] == 2): return hmac.compare_digest(bytes(request_hash), bytes(signature)) else: return hmac.compare_digest(request_hash, signature) else: if len(request_hash) != len(signature): return False result = 0 if isinstance(request_hash, bytes) and isinstance(signature, bytes): for x, y in zip(request_hash, signature): result |= x ^ y else: for x, y in zip(request_hash, signature): result |= ord(x) ^ ord(y) return result == 0 def bind_route(self, server): @server.route(self.endpoint, methods=['GET', 'POST']) def event(): # If a GET request is made, return 404. if request.method == 'GET': return make_response("These are not the slackbots you're looking for.", 404) # Each request comes with request timestamp and request signature # emit an error if the timestamp is out of range req_timestamp = request.headers.get('X-Slack-Request-Timestamp') if abs(time() - int(req_timestamp)) > 60 * 5: slack_exception = SlackEventAdapterException('Invalid request timestamp') self.emitter.emit('error', slack_exception) return make_response("", 403) # Verify the request signature using the app's signing secret # emit an error if the signature can't be verified req_signature = request.headers.get('X-Slack-Signature') if not self.verify_signature(req_timestamp, req_signature): slack_exception = SlackEventAdapterException('Invalid request signature') self.emitter.emit('error', slack_exception) return make_response("", 403) # Parse the request payload into JSON event_data = json.loads(request.data.decode('utf-8')) # Echo the URL verification challenge code back to Slack if "challenge" in event_data: return make_response( event_data.get("challenge"), 200, {"content_type": "application/json"} ) # Parse the Event payload and emit the event to the event listener if "event" in event_data: event_type = event_data["event"]["type"] self.emitter.emit(event_type, event_data) response = make_response("", 200) response.headers['X-Slack-Powered-By'] = self.package_info return response class SlackEventAdapterException(Exception): """ Base exception for all errors raised by the SlackClient library """ def __init__(self, msg=None): if msg is None: # default error message msg = "An error occurred in the SlackEventsApiAdapter library" super(SlackEventAdapterException, self).__init__(msg) [end of slackeventsapi/server.py] </code> I need you to solve this issue by regenerating the full files in the code base that you would like to change. You can change as many files as you like. Please respond with a list of files and their revised contents in the following format. <example> [start of /src/this_file.py] import os def euclidean(a, b): if b == 0: return a return euclidean(b, a % b) [end of /src/this_file.py] [start of /src/another_file.py] def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) x, y = x0, y0 sx = -1 if x0 > x1 else 1 sy = -1 if y0 > y1 else 1 if dx > dy: err = dx / 2.0 while x != x1: points.append((x, y)) err -= dy if err < 0: y += sy err += dx x += sx else: err = dy / 2.0 while y != y1: points.append((x err -= dx if err < 0: x += sx err += dy y += sy points.append((x, y)) return points [end of /src/another_file.py] </example>
slackapi/python-slack-events-api
0c0ce604b502508622fb14c278a0d64841fa32e3
Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s):
2020-06-12T06:58:10Z
<patch> <patch> diff --git a/example/current_app/main.py b/example/current_app/main.py new file mode 100644 --- /dev/null +++ b/example/current_app/main.py @@ -0,0 +1,49 @@ +# ------------------ +# Only for running this script here +import sys +from os.path import dirname +sys.path.insert(1, f"{dirname(__file__)}/../..") +# ------------------ + +import os +from slack import WebClient +import logging +logging.basicConfig(level=logging.DEBUG) + +from flask import Flask + +app = Flask(__name__) + +with app.app_context(): + from test_module.slack_app import slack_events_adapter + + slack_bot_token = os.environ["SLACK_BOT_TOKEN"] + slack_client = WebClient(slack_bot_token) + + + @slack_events_adapter.on("message") + def handle_message(event_data): + message = event_data["event"] + if message.get("subtype") is None and "hi" in message.get('text'): + channel = message["channel"] + message = "Hi <@%s>! :tada:" % message["user"] + slack_client.chat_postMessage(channel=channel, text=message) + + + @slack_events_adapter.on("error") + def error_handler(err): + print("ERROR: " + str(err)) + +# (Terminal A) +# source env/bin/activate +# (env) $ export SLACK_BOT_TOKEN=xoxb-*** +# (env) $ export SLACK_SIGNING_SECRET=** +# (env) $ cd example/current_app +# (env) $ FLASK_APP=main.py FLASK_ENV=development flask run --port 3000 + +# (Terminal B) +# ngrok http 3000 + +# in Slack +# /invite @{your app's bot user} +# post a message "hi" in the channel diff --git a/slackeventsapi/server.py b/slackeventsapi/server.py --- a/slackeventsapi/server.py +++ b/slackeventsapi/server.py @@ -1,10 +1,13 @@ -from flask import Flask, request, make_response, Blueprint +import hashlib +import hmac import json import platform import sys -import hmac -import hashlib from time import time + +from flask import Flask, request, make_response, Blueprint +from werkzeug.local import LocalProxy + from .version import __version__ @@ -18,10 +21,10 @@ def __init__(self, signing_secret, endpoint, emitter, server): # If a server is passed in, bind the event handler routes to it, # otherwise create a new Flask instance. if server: - if isinstance(server, Flask) or isinstance(server, Blueprint): + if isinstance(server, (Flask, Blueprint, LocalProxy)): self.bind_route(server) else: - raise TypeError("Server must be an instance of Flask or Blueprint") + raise TypeError("Server must be an instance of Flask, Blueprint, or LocalProxy") else: Flask.__init__(self, __name__) self.bind_route(self) </patch> </s> </patch>
diff --git a/example/current_app/test_module/__init__.py b/example/current_app/test_module/__init__.py new file mode 100644 diff --git a/example/current_app/test_module/slack_app.py b/example/current_app/test_module/slack_app.py new file mode 100644 --- /dev/null +++ b/example/current_app/test_module/slack_app.py @@ -0,0 +1,16 @@ +# ------------------ +# Only for running this script here +import logging +import sys +from os.path import dirname + +sys.path.insert(1, f"{dirname(__file__)}/../../..") +logging.basicConfig(level=logging.DEBUG) +# ------------------ + +from flask import current_app as app +from slackeventsapi import SlackEventAdapter +import os + +slack_signing_secret = os.environ["SLACK_SIGNING_SECRET"] +slack_events_adapter = SlackEventAdapter(slack_signing_secret, "/slack/events", app) diff --git a/tests/test_server.py b/tests/test_server.py --- a/tests/test_server.py +++ b/tests/test_server.py @@ -18,7 +18,7 @@ def test_server_not_flask(): with pytest.raises(TypeError) as e: invalid_flask = "I am not a Flask" SlackEventAdapter("SIGNING_SECRET", "/slack/events", invalid_flask) - assert e.value.args[0] == 'Server must be an instance of Flask or Blueprint' + assert e.value.args[0] == 'Server must be an instance of Flask, Blueprint, or LocalProxy' def test_blueprint_server():
1.0
celery__celery-2598
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ``` </issue> <code> [start of README.rst] ================================= celery - Distributed Task Queue ================================= .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png |build-status| |coverage-status| :Version: 3.2.0a1 (Cipater) :Web: http://celeryproject.org/ :Download: http://pypi.python.org/pypi/celery/ :Source: http://github.com/celery/celery/ :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, python, webhooks, queue, distributed -- What is a Task Queue? ===================== Task queues are used as a mechanism to distribute work across threads or machines. A task queue's input is a unit of work, called a task, dedicated worker processes then constantly monitor the queue for new work to perform. Celery communicates via messages, usually using a broker to mediate between clients and workers. To initiate a task a client puts a message on the queue, the broker then delivers the message to a worker. A Celery system can consist of multiple workers and brokers, giving way to high availability and horizontal scaling. Celery is a library written in Python, but the protocol can be implemented in any language. So far there's RCelery_ for the Ruby programming language, and a `PHP client`, but language interoperability can also be achieved by using webhooks. .. _RCelery: https://github.com/leapfrogonline/rcelery .. _`PHP client`: https://github.com/gjedeer/celery-php .. _`using webhooks`: http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html What do I need? =============== Celery version 3.0 runs on, - Python (2.6, 2.7, 3.3, 3.4) - PyPy (1.8, 1.9) - Jython (2.5, 2.7). This is the last version to support Python 2.5, and from Celery 3.1, Python 2.6 or later is required. The last version to support Python 2.4 was Celery series 2.2. *Celery* is usually used with a message broker to send and receive messages. The RabbitMQ, Redis transports are feature complete, but there's also experimental support for a myriad of other solutions, including using SQLite for local development. *Celery* can run on a single machine, on multiple machines, or even across datacenters. Get Started =========== If this is the first time you're trying to use Celery, or you are new to Celery 3.0 coming from previous versions then you should read our getting started tutorials: - `First steps with Celery`_ Tutorial teaching you the bare minimum needed to get started with Celery. - `Next steps`_ A more complete overview, showing more features. .. _`First steps with Celery`: http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html .. _`Next steps`: http://docs.celeryproject.org/en/latest/getting-started/next-steps.html Celery is... ========== - **Simple** Celery is easy to use and maintain, and does *not need configuration files*. It has an active, friendly community you can talk to for support, including a `mailing-list`_ and and an IRC channel. Here's one of the simplest applications you can make:: from celery import Celery app = Celery('hello', broker='amqp://guest@localhost//') @app.task def hello(): return 'hello world' - **Highly Available** Workers and clients will automatically retry in the event of connection loss or failure, and some brokers support HA in way of *Master/Master* or *Master/Slave* replication. - **Fast** A single Celery process can process millions of tasks a minute, with sub-millisecond round-trip latency (using RabbitMQ, py-librabbitmq, and optimized settings). - **Flexible** Almost every part of *Celery* can be extended or used on its own, Custom pool implementations, serializers, compression schemes, logging, schedulers, consumers, producers, autoscalers, broker transports and much more. It supports... ============ - **Message Transports** - RabbitMQ_, Redis_, - MongoDB_ (experimental), Amazon SQS (experimental), - CouchDB_ (experimental), SQLAlchemy_ (experimental), - Django ORM (experimental), `IronMQ`_ - and more... - **Concurrency** - Prefork, Eventlet_, gevent_, threads/single threaded - **Result Stores** - AMQP, Redis - memcached, MongoDB - SQLAlchemy, Django ORM - Apache Cassandra, IronCache - **Serialization** - *pickle*, *json*, *yaml*, *msgpack*. - *zlib*, *bzip2* compression. - Cryptographic message signing. .. _`Eventlet`: http://eventlet.net/ .. _`gevent`: http://gevent.org/ .. _RabbitMQ: http://rabbitmq.com .. _Redis: http://redis.io .. _MongoDB: http://mongodb.org .. _Beanstalk: http://kr.github.com/beanstalkd .. _CouchDB: http://couchdb.apache.org .. _SQLAlchemy: http://sqlalchemy.org .. _`IronMQ`: http://iron.io Framework Integration ===================== Celery is easy to integrate with web frameworks, some of which even have integration packages: +--------------------+------------------------+ | `Django`_ | not needed | +--------------------+------------------------+ | `Pyramid`_ | `pyramid_celery`_ | +--------------------+------------------------+ | `Pylons`_ | `celery-pylons`_ | +--------------------+------------------------+ | `Flask`_ | not needed | +--------------------+------------------------+ | `web2py`_ | `web2py-celery`_ | +--------------------+------------------------+ | `Tornado`_ | `tornado-celery`_ | +--------------------+------------------------+ The integration packages are not strictly necessary, but they can make development easier, and sometimes they add important hooks like closing database connections at ``fork``. .. _`Django`: http://djangoproject.com/ .. _`Pylons`: http://www.pylonsproject.org/ .. _`Flask`: http://flask.pocoo.org/ .. _`web2py`: http://web2py.com/ .. _`Bottle`: http://bottlepy.org/ .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ .. _`django-celery`: http://pypi.python.org/pypi/django-celery .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ .. _`Tornado`: http://www.tornadoweb.org/ .. _`tornado-celery`: http://github.com/mher/tornado-celery/ .. _celery-documentation: Documentation ============= The `latest documentation`_ with user guides, tutorials and API reference is hosted at Read The Docs. .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ .. _celery-installation: Installation ============ You can install Celery either via the Python Package Index (PyPI) or from source. To install using `pip`,:: $ pip install -U Celery To install using `easy_install`,:: $ easy_install -U Celery .. _bundles: Bundles ------- Celery also defines a group of bundles that can be used to install Celery and the dependencies for a given feature. You can specify these in your requirements or on the ``pip`` comand-line by using brackets. Multiple bundles can be specified by separating them by commas. :: $ pip install "celery[librabbitmq]" $ pip install "celery[librabbitmq,redis,auth,msgpack]" The following bundles are available: Serializers ~~~~~~~~~~~ :celery[auth]: for using the auth serializer. :celery[msgpack]: for using the msgpack serializer. :celery[yaml]: for using the yaml serializer. Concurrency ~~~~~~~~~~~ :celery[eventlet]: for using the eventlet pool. :celery[gevent]: for using the gevent pool. :celery[threads]: for using the thread pool. Transports and Backends ~~~~~~~~~~~~~~~~~~~~~~~ :celery[librabbitmq]: for using the librabbitmq C library. :celery[redis]: for using Redis as a message transport or as a result backend. :celery[mongodb]: for using MongoDB as a message transport (*experimental*), or as a result backend (*supported*). :celery[sqs]: for using Amazon SQS as a message transport (*experimental*). :celery[memcache]: for using memcached as a result backend. :celery[cassandra]: for using Apache Cassandra as a result backend. :celery[couchdb]: for using CouchDB as a message transport (*experimental*). :celery[couchbase]: for using CouchBase as a result backend. :celery[beanstalk]: for using Beanstalk as a message transport (*experimental*). :celery[zookeeper]: for using Zookeeper as a message transport. :celery[zeromq]: for using ZeroMQ as a message transport (*experimental*). :celery[sqlalchemy]: for using SQLAlchemy as a message transport (*experimental*), or as a result backend (*supported*). :celery[pyro]: for using the Pyro4 message transport (*experimental*). :celery[slmq]: for using the SoftLayer Message Queue transport (*experimental*). .. _celery-installing-from-source: Downloading and installing from source -------------------------------------- Download the latest version of Celery from http://pypi.python.org/pypi/celery/ You can install it by doing the following,:: $ tar xvfz celery-0.0.0.tar.gz $ cd celery-0.0.0 $ python setup.py build # python setup.py install The last command must be executed as a privileged user if you are not currently using a virtualenv. .. _celery-installing-from-git: Using the development version ----------------------------- With pip ~~~~~~~~ The Celery development version also requires the development versions of ``kombu``, ``amqp`` and ``billiard``. You can install the latest snapshot of these using the following pip commands:: $ pip install https://github.com/celery/celery/zipball/master#egg=celery $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu With git ~~~~~~~~ Please the Contributing section. .. _getting-help: Getting Help ============ .. _mailing-list: Mailing list ------------ For discussions about the usage, development, and future of celery, please join the `celery-users`_ mailing list. .. _`celery-users`: http://groups.google.com/group/celery-users/ .. _irc-channel: IRC --- Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ network. .. _`Freenode`: http://freenode.net .. _bug-tracker: Bug tracker =========== If you have any suggestions, bug reports or annoyances please report them to our issue tracker at http://github.com/celery/celery/issues/ .. _wiki: Wiki ==== http://wiki.github.com/celery/celery/ .. _contributing-short: Contributing ============ Development of `celery` happens at Github: http://github.com/celery/celery You are highly encouraged to participate in the development of `celery`. If you don't like Github (for some reason) you're welcome to send regular patches. Be sure to also read the `Contributing to Celery`_ section in the documentation. .. _`Contributing to Celery`: http://docs.celeryproject.org/en/master/contributing.html .. _license: License ======= This software is licensed under the `New BSD License`. See the ``LICENSE`` file in the top distribution directory for the full license text. .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png :alt: Bitdeli badge :target: https://bitdeli.com/free .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master :target: https://travis-ci.org/celery/celery .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg :target: https://coveralls.io/r/celery/celery [end of README.rst] [start of celery/backends/amqp.py] # -*- coding: utf-8 -*- """ celery.backends.amqp ~~~~~~~~~~~~~~~~~~~~ The AMQP result backend. This backend publishes results as messages. """ from __future__ import absolute_import import socket from collections import deque from operator import itemgetter from kombu import Exchange, Queue, Producer, Consumer from celery import states from celery.exceptions import TimeoutError from celery.five import range, monotonic from celery.utils.functional import dictfilter from celery.utils.log import get_logger from celery.utils.timeutils import maybe_s_to_ms from .base import BaseBackend __all__ = ['BacklogLimitExceeded', 'AMQPBackend'] logger = get_logger(__name__) class BacklogLimitExceeded(Exception): """Too much state history to fast-forward.""" def repair_uuid(s): # Historically the dashes in UUIDS are removed from AMQ entity names, # but there is no known reason to. Hopefully we'll be able to fix # this in v4.0. return '%s-%s-%s-%s-%s' % (s[:8], s[8:12], s[12:16], s[16:20], s[20:]) class NoCacheQueue(Queue): can_cache_declaration = False class AMQPBackend(BaseBackend): """Publishes results by sending messages.""" Exchange = Exchange Queue = NoCacheQueue Consumer = Consumer Producer = Producer BacklogLimitExceeded = BacklogLimitExceeded persistent = True supports_autoexpire = True supports_native_join = True retry_policy = { 'max_retries': 20, 'interval_start': 0, 'interval_step': 1, 'interval_max': 1, } def __init__(self, app, connection=None, exchange=None, exchange_type=None, persistent=None, serializer=None, auto_delete=True, **kwargs): super(AMQPBackend, self).__init__(app, **kwargs) conf = self.app.conf self._connection = connection self.persistent = self.prepare_persistent(persistent) self.delivery_mode = 2 if self.persistent else 1 exchange = exchange or conf.CELERY_RESULT_EXCHANGE exchange_type = exchange_type or conf.CELERY_RESULT_EXCHANGE_TYPE self.exchange = self._create_exchange( exchange, exchange_type, self.delivery_mode, ) self.serializer = serializer or conf.CELERY_RESULT_SERIALIZER self.auto_delete = auto_delete self.queue_arguments = dictfilter({ 'x-expires': maybe_s_to_ms(self.expires), }) def _create_exchange(self, name, type='direct', delivery_mode=2): return self.Exchange(name=name, type=type, delivery_mode=delivery_mode, durable=self.persistent, auto_delete=False) def _create_binding(self, task_id): name = self.rkey(task_id) return self.Queue(name=name, exchange=self.exchange, routing_key=name, durable=self.persistent, auto_delete=self.auto_delete, queue_arguments=self.queue_arguments) def revive(self, channel): pass def rkey(self, task_id): return task_id.replace('-', '') def destination_for(self, task_id, request): if request: return self.rkey(task_id), request.correlation_id or task_id return self.rkey(task_id), task_id def store_result(self, task_id, result, status, traceback=None, request=None, **kwargs): """Send task return value and status.""" routing_key, correlation_id = self.destination_for(task_id, request) if not routing_key: return with self.app.amqp.producer_pool.acquire(block=True) as producer: producer.publish( {'task_id': task_id, 'status': status, 'result': self.encode_result(result, status), 'traceback': traceback, 'children': self.current_task_children(request)}, exchange=self.exchange, routing_key=routing_key, correlation_id=correlation_id, serializer=self.serializer, retry=True, retry_policy=self.retry_policy, declare=self.on_reply_declare(task_id), delivery_mode=self.delivery_mode, ) return result def on_reply_declare(self, task_id): return [self._create_binding(task_id)] def wait_for(self, task_id, timeout=None, cache=True, no_ack=True, on_interval=None, READY_STATES=states.READY_STATES, PROPAGATE_STATES=states.PROPAGATE_STATES, **kwargs): cached_meta = self._cache.get(task_id) if cache and cached_meta and \ cached_meta['status'] in READY_STATES: return cached_meta else: try: return self.consume(task_id, timeout=timeout, no_ack=no_ack, on_interval=on_interval) except socket.timeout: raise TimeoutError('The operation timed out.') def get_task_meta(self, task_id, backlog_limit=1000): # Polling and using basic_get with self.app.pool.acquire_channel(block=True) as (_, channel): binding = self._create_binding(task_id)(channel) binding.declare() prev = latest = acc = None for i in range(backlog_limit): # spool ffwd acc = binding.get( accept=self.accept, no_ack=False, ) if not acc: # no more messages break if acc.payload['task_id'] == task_id: prev, latest = latest, acc if prev: # backends are not expected to keep history, # so we delete everything except the most recent state. prev.ack() prev = None else: raise self.BacklogLimitExceeded(task_id) if latest: payload = self._cache[task_id] = latest.payload latest.requeue() return payload else: # no new state, use previous try: return self._cache[task_id] except KeyError: # result probably pending. return {'status': states.PENDING, 'result': None} poll = get_task_meta # XXX compat def drain_events(self, connection, consumer, timeout=None, on_interval=None, now=monotonic, wait=None): wait = wait or connection.drain_events results = {} def callback(meta, message): if meta['status'] in states.READY_STATES: results[meta['task_id']] = meta consumer.callbacks[:] = [callback] time_start = now() while 1: # Total time spent may exceed a single call to wait() if timeout and now() - time_start >= timeout: raise socket.timeout() try: wait(timeout=1) except socket.timeout: pass if on_interval: on_interval() if results: # got event on the wanted channel. break self._cache.update(results) return results def consume(self, task_id, timeout=None, no_ack=True, on_interval=None): wait = self.drain_events with self.app.pool.acquire_channel(block=True) as (conn, channel): binding = self._create_binding(task_id) with self.Consumer(channel, binding, no_ack=no_ack, accept=self.accept) as consumer: while 1: try: return wait( conn, consumer, timeout, on_interval)[task_id] except KeyError: continue def _many_bindings(self, ids): return [self._create_binding(task_id) for task_id in ids] def get_many(self, task_ids, timeout=None, no_ack=True, on_message=None, now=monotonic, getfields=itemgetter('status', 'task_id'), READY_STATES=states.READY_STATES, PROPAGATE_STATES=states.PROPAGATE_STATES, **kwargs): with self.app.pool.acquire_channel(block=True) as (conn, channel): ids = set(task_ids) cached_ids = set() mark_cached = cached_ids.add for task_id in ids: try: cached = self._cache[task_id] except KeyError: pass else: if cached['status'] in READY_STATES: yield task_id, cached mark_cached(task_id) ids.difference_update(cached_ids) results = deque() push_result = results.append push_cache = self._cache.__setitem__ decode_result = self.meta_from_decoded def _on_message(message): body = decode_result(message.decode()) if on_message is not None: on_message(body) state, uid = getfields(body) if state in READY_STATES: push_result(body) \ if uid in task_ids else push_cache(uid, body) bindings = self._many_bindings(task_ids) with self.Consumer(channel, bindings, on_message=_on_message, accept=self.accept, no_ack=no_ack): wait = conn.drain_events popleft = results.popleft while ids: wait(timeout=timeout) while results: state = popleft() task_id = state['task_id'] ids.discard(task_id) push_cache(task_id, state) yield task_id, state def reload_task_result(self, task_id): raise NotImplementedError( 'reload_task_result is not supported by this backend.') def reload_group_result(self, task_id): """Reload group result, even if it has been previously fetched.""" raise NotImplementedError( 'reload_group_result is not supported by this backend.') def save_group(self, group_id, result): raise NotImplementedError( 'save_group is not supported by this backend.') def restore_group(self, group_id, cache=True): raise NotImplementedError( 'restore_group is not supported by this backend.') def delete_group(self, group_id): raise NotImplementedError( 'delete_group is not supported by this backend.') def __reduce__(self, args=(), kwargs={}): kwargs.update( connection=self._connection, exchange=self.exchange.name, exchange_type=self.exchange.type, persistent=self.persistent, serializer=self.serializer, auto_delete=self.auto_delete, expires=self.expires, ) return super(AMQPBackend, self).__reduce__(args, kwargs) [end of celery/backends/amqp.py] </code> I need you to solve this issue by regenerating the full files in the code base that you would like to change. You can change as many files as you like. Please respond with a list of files and their revised contents in the following format. <example> [start of /src/this_file.py] import os def euclidean(a, b): if b == 0: return a return euclidean(b, a % b) [end of /src/this_file.py] [start of /src/another_file.py] def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) x, y = x0, y0 sx = -1 if x0 > x1 else 1 sy = -1 if y0 > y1 else 1 if dx > dy: err = dx / 2.0 while x != x1: points.append((x, y)) err -= dy if err < 0: y += sy err += dx x += sx else: err = dy / 2.0 while y != y1: points.append((x err -= dx if err < 0: x += sx err += dy y += sy points.append((x, y)) return points [end of /src/another_file.py] </example>
celery/celery
6592ff64b6b024a4b68abcc53b151888fdf0dee3
CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ```
This is biting me as well. Any news?
2015-04-29T14:52:17Z
<patch> <patch> diff --git a/celery/backends/amqp.py b/celery/backends/amqp.py --- a/celery/backends/amqp.py +++ b/celery/backends/amqp.py @@ -195,7 +195,7 @@ def drain_events(self, connection, consumer, def callback(meta, message): if meta['status'] in states.READY_STATES: - results[meta['task_id']] = meta + results[meta['task_id']] = self.meta_from_decoded(meta) consumer.callbacks[:] = [callback] time_start = now() </patch> </s> </patch>
diff --git a/celery/tests/backends/test_amqp.py b/celery/tests/backends/test_amqp.py --- a/celery/tests/backends/test_amqp.py +++ b/celery/tests/backends/test_amqp.py @@ -13,6 +13,7 @@ from celery.backends.amqp import AMQPBackend from celery.exceptions import TimeoutError from celery.five import Empty, Queue, range +from celery.result import AsyncResult from celery.utils import uuid from celery.tests.case import ( @@ -246,10 +247,20 @@ def test_wait_for(self): with self.assertRaises(TimeoutError): b.wait_for(tid, timeout=0.01, cache=False) - def test_drain_events_remaining_timeouts(self): + def test_drain_events_decodes_exceptions_in_meta(self): + tid = uuid() + b = self.create_backend(serializer="json") + b.store_result(tid, RuntimeError("aap"), states.FAILURE) + result = AsyncResult(tid, backend=b) - class Connection(object): + with self.assertRaises(Exception) as cm: + result.get() + self.assertEqual(cm.exception.__class__.__name__, "RuntimeError") + self.assertEqual(str(cm.exception), "aap") + + def test_drain_events_remaining_timeouts(self): + class Connection(object): def drain_events(self, timeout=None): pass
1.0
celery__celery-2840
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost. </issue> <code> [start of README.rst] 1 ================================= 2 celery - Distributed Task Queue 3 ================================= 4 5 .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png 6 7 |build-status| |coverage-status| 8 9 :Version: 3.2.0a1 (Cipater) 10 :Web: http://celeryproject.org/ 11 :Download: http://pypi.python.org/pypi/celery/ 12 :Source: http://github.com/celery/celery/ 13 :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, 14 python, webhooks, queue, distributed 15 16 -- 17 18 What is a Task Queue? 19 ===================== 20 21 Task queues are used as a mechanism to distribute work across threads or 22 machines. 23 24 A task queue's input is a unit of work, called a task, dedicated worker 25 processes then constantly monitor the queue for new work to perform. 26 27 Celery communicates via messages, usually using a broker 28 to mediate between clients and workers. To initiate a task a client puts a 29 message on the queue, the broker then delivers the message to a worker. 30 31 A Celery system can consist of multiple workers and brokers, giving way 32 to high availability and horizontal scaling. 33 34 Celery is a library written in Python, but the protocol can be implemented in 35 any language. So far there's RCelery_ for the Ruby programming language, and a 36 `PHP client`, but language interoperability can also be achieved 37 by using webhooks. 38 39 .. _RCelery: https://github.com/leapfrogonline/rcelery 40 .. _`PHP client`: https://github.com/gjedeer/celery-php 41 .. _`using webhooks`: 42 http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html 43 44 What do I need? 45 =============== 46 47 Celery version 3.0 runs on, 48 49 - Python (2.6, 2.7, 3.3, 3.4) 50 - PyPy (1.8, 1.9) 51 - Jython (2.5, 2.7). 52 53 This is the last version to support Python 2.5, 54 and from Celery 3.1, Python 2.6 or later is required. 55 The last version to support Python 2.4 was Celery series 2.2. 56 57 *Celery* is usually used with a message broker to send and receive messages. 58 The RabbitMQ, Redis transports are feature complete, 59 but there's also experimental support for a myriad of other solutions, including 60 using SQLite for local development. 61 62 *Celery* can run on a single machine, on multiple machines, or even 63 across datacenters. 64 65 Get Started 66 =========== 67 68 If this is the first time you're trying to use Celery, or you are 69 new to Celery 3.0 coming from previous versions then you should read our 70 getting started tutorials: 71 72 - `First steps with Celery`_ 73 74 Tutorial teaching you the bare minimum needed to get started with Celery. 75 76 - `Next steps`_ 77 78 A more complete overview, showing more features. 79 80 .. _`First steps with Celery`: 81 http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html 82 83 .. _`Next steps`: 84 http://docs.celeryproject.org/en/latest/getting-started/next-steps.html 85 86 Celery is... 87 ========== 88 89 - **Simple** 90 91 Celery is easy to use and maintain, and does *not need configuration files*. 92 93 It has an active, friendly community you can talk to for support, 94 including a `mailing-list`_ and and an IRC channel. 95 96 Here's one of the simplest applications you can make:: 97 98 from celery import Celery 99 100 app = Celery('hello', broker='amqp://guest@localhost//') 101 102 @app.task 103 def hello(): 104 return 'hello world' 105 106 - **Highly Available** 107 108 Workers and clients will automatically retry in the event 109 of connection loss or failure, and some brokers support 110 HA in way of *Master/Master* or *Master/Slave* replication. 111 112 - **Fast** 113 114 A single Celery process can process millions of tasks a minute, 115 with sub-millisecond round-trip latency (using RabbitMQ, 116 py-librabbitmq, and optimized settings). 117 118 - **Flexible** 119 120 Almost every part of *Celery* can be extended or used on its own, 121 Custom pool implementations, serializers, compression schemes, logging, 122 schedulers, consumers, producers, autoscalers, broker transports and much more. 123 124 It supports... 125 ============ 126 127 - **Message Transports** 128 129 - RabbitMQ_, Redis_, 130 - MongoDB_ (experimental), Amazon SQS (experimental), 131 - CouchDB_ (experimental), SQLAlchemy_ (experimental), 132 - Django ORM (experimental), `IronMQ`_ 133 - and more... 134 135 - **Concurrency** 136 137 - Prefork, Eventlet_, gevent_, threads/single threaded 138 139 - **Result Stores** 140 141 - AMQP, Redis 142 - memcached, MongoDB 143 - SQLAlchemy, Django ORM 144 - Apache Cassandra, IronCache 145 146 - **Serialization** 147 148 - *pickle*, *json*, *yaml*, *msgpack*. 149 - *zlib*, *bzip2* compression. 150 - Cryptographic message signing. 151 152 .. _`Eventlet`: http://eventlet.net/ 153 .. _`gevent`: http://gevent.org/ 154 155 .. _RabbitMQ: http://rabbitmq.com 156 .. _Redis: http://redis.io 157 .. _MongoDB: http://mongodb.org 158 .. _Beanstalk: http://kr.github.com/beanstalkd 159 .. _CouchDB: http://couchdb.apache.org 160 .. _SQLAlchemy: http://sqlalchemy.org 161 .. _`IronMQ`: http://iron.io 162 163 Framework Integration 164 ===================== 165 166 Celery is easy to integrate with web frameworks, some of which even have 167 integration packages: 168 169 +--------------------+----------------------------------------------------+ 170 | `Django`_ | not needed | 171 +--------------------+----------------------------------------------------+ 172 | `Pyramid`_ | `pyramid_celery`_ | 173 +--------------------+----------------------------------------------------+ 174 | `Pylons`_ | `celery-pylons`_ | 175 +--------------------+----------------------------------------------------+ 176 | `Flask`_ | not needed | 177 +--------------------+----------------------------------------------------+ 178 | `web2py`_ | `web2py-celery`_ | 179 +--------------------+----------------------------------------------------+ 180 | `Tornado`_ | `tornado-celery`_ | `another tornado-celery`_ | 181 +--------------------+----------------------------------------------------+ 182 183 The integration packages are not strictly necessary, but they can make 184 development easier, and sometimes they add important hooks like closing 185 database connections at ``fork``. 186 187 .. _`Django`: http://djangoproject.com/ 188 .. _`Pylons`: http://www.pylonsproject.org/ 189 .. _`Flask`: http://flask.pocoo.org/ 190 .. _`web2py`: http://web2py.com/ 191 .. _`Bottle`: http://bottlepy.org/ 192 .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html 193 .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ 194 .. _`django-celery`: http://pypi.python.org/pypi/django-celery 195 .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons 196 .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ 197 .. _`Tornado`: http://www.tornadoweb.org/ 198 .. _`tornado-celery`: http://github.com/mher/tornado-celery/ 199 .. _`another tornado-celery`: https://github.com/mayflaver/tornado-celery 200 201 .. _celery-documentation: 202 203 Documentation 204 ============= 205 206 The `latest documentation`_ with user guides, tutorials and API reference 207 is hosted at Read The Docs. 208 209 .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ 210 211 .. _celery-installation: 212 213 Installation 214 ============ 215 216 You can install Celery either via the Python Package Index (PyPI) 217 or from source. 218 219 To install using `pip`,:: 220 221 $ pip install -U Celery 222 223 To install using `easy_install`,:: 224 225 $ easy_install -U Celery 226 227 .. _bundles: 228 229 Bundles 230 ------- 231 232 Celery also defines a group of bundles that can be used 233 to install Celery and the dependencies for a given feature. 234 235 You can specify these in your requirements or on the ``pip`` comand-line 236 by using brackets. Multiple bundles can be specified by separating them by 237 commas. 238 :: 239 240 $ pip install "celery[librabbitmq]" 241 242 $ pip install "celery[librabbitmq,redis,auth,msgpack]" 243 244 The following bundles are available: 245 246 Serializers 247 ~~~~~~~~~~~ 248 249 :celery[auth]: 250 for using the auth serializer. 251 252 :celery[msgpack]: 253 for using the msgpack serializer. 254 255 :celery[yaml]: 256 for using the yaml serializer. 257 258 Concurrency 259 ~~~~~~~~~~~ 260 261 :celery[eventlet]: 262 for using the eventlet pool. 263 264 :celery[gevent]: 265 for using the gevent pool. 266 267 :celery[threads]: 268 for using the thread pool. 269 270 Transports and Backends 271 ~~~~~~~~~~~~~~~~~~~~~~~ 272 273 :celery[librabbitmq]: 274 for using the librabbitmq C library. 275 276 :celery[redis]: 277 for using Redis as a message transport or as a result backend. 278 279 :celery[mongodb]: 280 for using MongoDB as a message transport (*experimental*), 281 or as a result backend (*supported*). 282 283 :celery[sqs]: 284 for using Amazon SQS as a message transport (*experimental*). 285 286 :celery[memcache]: 287 for using memcached as a result backend. 288 289 :celery[cassandra]: 290 for using Apache Cassandra as a result backend. 291 292 :celery[couchdb]: 293 for using CouchDB as a message transport (*experimental*). 294 295 :celery[couchbase]: 296 for using CouchBase as a result backend. 297 298 :celery[beanstalk]: 299 for using Beanstalk as a message transport (*experimental*). 300 301 :celery[zookeeper]: 302 for using Zookeeper as a message transport. 303 304 :celery[zeromq]: 305 for using ZeroMQ as a message transport (*experimental*). 306 307 :celery[sqlalchemy]: 308 for using SQLAlchemy as a message transport (*experimental*), 309 or as a result backend (*supported*). 310 311 :celery[pyro]: 312 for using the Pyro4 message transport (*experimental*). 313 314 :celery[slmq]: 315 for using the SoftLayer Message Queue transport (*experimental*). 316 317 .. _celery-installing-from-source: 318 319 Downloading and installing from source 320 -------------------------------------- 321 322 Download the latest version of Celery from 323 http://pypi.python.org/pypi/celery/ 324 325 You can install it by doing the following,:: 326 327 $ tar xvfz celery-0.0.0.tar.gz 328 $ cd celery-0.0.0 329 $ python setup.py build 330 # python setup.py install 331 332 The last command must be executed as a privileged user if 333 you are not currently using a virtualenv. 334 335 .. _celery-installing-from-git: 336 337 Using the development version 338 ----------------------------- 339 340 With pip 341 ~~~~~~~~ 342 343 The Celery development version also requires the development 344 versions of ``kombu``, ``amqp`` and ``billiard``. 345 346 You can install the latest snapshot of these using the following 347 pip commands:: 348 349 $ pip install https://github.com/celery/celery/zipball/master#egg=celery 350 $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard 351 $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp 352 $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu 353 354 With git 355 ~~~~~~~~ 356 357 Please the Contributing section. 358 359 .. _getting-help: 360 361 Getting Help 362 ============ 363 364 .. _mailing-list: 365 366 Mailing list 367 ------------ 368 369 For discussions about the usage, development, and future of celery, 370 please join the `celery-users`_ mailing list. 371 372 .. _`celery-users`: http://groups.google.com/group/celery-users/ 373 374 .. _irc-channel: 375 376 IRC 377 --- 378 379 Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ 380 network. 381 382 .. _`Freenode`: http://freenode.net 383 384 .. _bug-tracker: 385 386 Bug tracker 387 =========== 388 389 If you have any suggestions, bug reports or annoyances please report them 390 to our issue tracker at http://github.com/celery/celery/issues/ 391 392 .. _wiki: 393 394 Wiki 395 ==== 396 397 http://wiki.github.com/celery/celery/ 398 399 400 .. _maintainers: 401 402 Maintainers 403 =========== 404 405 - `@ask`_ (primary maintainer) 406 - `@thedrow`_ 407 - `@chrisgogreen`_ 408 - `@PMickael`_ 409 - `@malinoff`_ 410 - And you? We really need more: https://github.com/celery/celery/issues/2534 411 412 .. _`@ask`: http://github.com/ask 413 .. _`@thedrow`: http://github.com/thedrow 414 .. _`@chrisgogreen`: http://github.com/chrisgogreen 415 .. _`@PMickael`: http://github.com/PMickael 416 .. _`@malinoff`: http://github.com/malinoff 417 418 419 .. _contributing-short: 420 421 Contributing 422 ============ 423 424 Development of `celery` happens at Github: http://github.com/celery/celery 425 426 You are highly encouraged to participate in the development 427 of `celery`. If you don't like Github (for some reason) you're welcome 428 to send regular patches. 429 430 Be sure to also read the `Contributing to Celery`_ section in the 431 documentation. 432 433 .. _`Contributing to Celery`: 434 http://docs.celeryproject.org/en/master/contributing.html 435 436 .. _license: 437 438 License 439 ======= 440 441 This software is licensed under the `New BSD License`. See the ``LICENSE`` 442 file in the top distribution directory for the full license text. 443 444 .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround 445 446 447 .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png 448 :alt: Bitdeli badge 449 :target: https://bitdeli.com/free 450 451 .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master 452 :target: https://travis-ci.org/celery/celery 453 .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg 454 :target: https://coveralls.io/r/celery/celery 455 [end of README.rst] [start of celery/app/defaults.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.app.defaults 4 ~~~~~~~~~~~~~~~~~~~ 5 6 Configuration introspection and defaults. 7 8 """ 9 from __future__ import absolute_import 10 11 import sys 12 13 from collections import deque, namedtuple 14 from datetime import timedelta 15 16 from celery.five import items 17 from celery.utils import strtobool 18 from celery.utils.functional import memoize 19 20 __all__ = ['Option', 'NAMESPACES', 'flatten', 'find'] 21 22 is_jython = sys.platform.startswith('java') 23 is_pypy = hasattr(sys, 'pypy_version_info') 24 25 DEFAULT_POOL = 'prefork' 26 if is_jython: 27 DEFAULT_POOL = 'threads' 28 elif is_pypy: 29 if sys.pypy_version_info[0:3] < (1, 5, 0): 30 DEFAULT_POOL = 'solo' 31 else: 32 DEFAULT_POOL = 'prefork' 33 34 DEFAULT_ACCEPT_CONTENT = ['json', 'pickle', 'msgpack', 'yaml'] 35 DEFAULT_PROCESS_LOG_FMT = """ 36 [%(asctime)s: %(levelname)s/%(processName)s] %(message)s 37 """.strip() 38 DEFAULT_LOG_FMT = '[%(asctime)s: %(levelname)s] %(message)s' 39 DEFAULT_TASK_LOG_FMT = """[%(asctime)s: %(levelname)s/%(processName)s] \ 40 %(task_name)s[%(task_id)s]: %(message)s""" 41 42 _BROKER_OLD = {'deprecate_by': '2.5', 'remove_by': '4.0', 43 'alt': 'BROKER_URL setting'} 44 _REDIS_OLD = {'deprecate_by': '2.5', 'remove_by': '4.0', 45 'alt': 'URL form of CELERY_RESULT_BACKEND'} 46 47 searchresult = namedtuple('searchresult', ('namespace', 'key', 'type')) 48 49 50 class Option(object): 51 alt = None 52 deprecate_by = None 53 remove_by = None 54 typemap = dict(string=str, int=int, float=float, any=lambda v: v, 55 bool=strtobool, dict=dict, tuple=tuple) 56 57 def __init__(self, default=None, *args, **kwargs): 58 self.default = default 59 self.type = kwargs.get('type') or 'string' 60 for attr, value in items(kwargs): 61 setattr(self, attr, value) 62 63 def to_python(self, value): 64 return self.typemap[self.type](value) 65 66 def __repr__(self): 67 return '<Option: type->{0} default->{1!r}>'.format(self.type, 68 self.default) 69 70 NAMESPACES = { 71 'BROKER': { 72 'URL': Option(None, type='string'), 73 'CONNECTION_TIMEOUT': Option(4, type='float'), 74 'CONNECTION_RETRY': Option(True, type='bool'), 75 'CONNECTION_MAX_RETRIES': Option(100, type='int'), 76 'FAILOVER_STRATEGY': Option(None, type='string'), 77 'HEARTBEAT': Option(None, type='int'), 78 'HEARTBEAT_CHECKRATE': Option(3.0, type='int'), 79 'LOGIN_METHOD': Option(None, type='string'), 80 'POOL_LIMIT': Option(10, type='int'), 81 'USE_SSL': Option(False, type='bool'), 82 'TRANSPORT': Option(type='string'), 83 'TRANSPORT_OPTIONS': Option({}, type='dict'), 84 'HOST': Option(type='string', **_BROKER_OLD), 85 'PORT': Option(type='int', **_BROKER_OLD), 86 'USER': Option(type='string', **_BROKER_OLD), 87 'PASSWORD': Option(type='string', **_BROKER_OLD), 88 'VHOST': Option(type='string', **_BROKER_OLD), 89 }, 90 'CASSANDRA': { 91 'COLUMN_FAMILY': Option(type='string'), 92 'DETAILED_MODE': Option(False, type='bool'), 93 'KEYSPACE': Option(type='string'), 94 'READ_CONSISTENCY': Option(type='string'), 95 'SERVERS': Option(type='list'), 96 'WRITE_CONSISTENCY': Option(type='string'), 97 }, 98 'CELERY': { 99 'ACCEPT_CONTENT': Option(DEFAULT_ACCEPT_CONTENT, type='list'), 100 'ACKS_LATE': Option(False, type='bool'), 101 'ALWAYS_EAGER': Option(False, type='bool'), 102 'ANNOTATIONS': Option(type='any'), 103 'BROADCAST_QUEUE': Option('celeryctl'), 104 'BROADCAST_EXCHANGE': Option('celeryctl'), 105 'BROADCAST_EXCHANGE_TYPE': Option('fanout'), 106 'CACHE_BACKEND': Option(), 107 'CACHE_BACKEND_OPTIONS': Option({}, type='dict'), 108 'CHORD_PROPAGATES': Option(True, type='bool'), 109 'COUCHBASE_BACKEND_SETTINGS': Option(None, type='dict'), 110 'CREATE_MISSING_QUEUES': Option(True, type='bool'), 111 'DEFAULT_RATE_LIMIT': Option(type='string'), 112 'DISABLE_RATE_LIMITS': Option(False, type='bool'), 113 'DEFAULT_ROUTING_KEY': Option('celery'), 114 'DEFAULT_QUEUE': Option('celery'), 115 'DEFAULT_EXCHANGE': Option('celery'), 116 'DEFAULT_EXCHANGE_TYPE': Option('direct'), 117 'DEFAULT_DELIVERY_MODE': Option(2, type='string'), 118 'EAGER_PROPAGATES_EXCEPTIONS': Option(False, type='bool'), 119 'ENABLE_UTC': Option(True, type='bool'), 120 'ENABLE_REMOTE_CONTROL': Option(True, type='bool'), 121 'EVENT_SERIALIZER': Option('json'), 122 'EVENT_QUEUE_EXPIRES': Option(60.0, type='float'), 123 'EVENT_QUEUE_TTL': Option(5.0, type='float'), 124 'IMPORTS': Option((), type='tuple'), 125 'INCLUDE': Option((), type='tuple'), 126 'IGNORE_RESULT': Option(False, type='bool'), 127 'MAX_CACHED_RESULTS': Option(100, type='int'), 128 'MESSAGE_COMPRESSION': Option(type='string'), 129 'MONGODB_BACKEND_SETTINGS': Option(type='dict'), 130 'REDIS_HOST': Option(type='string', **_REDIS_OLD), 131 'REDIS_PORT': Option(type='int', **_REDIS_OLD), 132 'REDIS_DB': Option(type='int', **_REDIS_OLD), 133 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 134 'REDIS_MAX_CONNECTIONS': Option(type='int'), 135 'RESULT_BACKEND': Option(type='string'), 136 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 137 'RESULT_DB_TABLENAMES': Option(type='dict'), 138 'RESULT_DBURI': Option(), 139 'RESULT_ENGINE_OPTIONS': Option(type='dict'), 140 'RESULT_EXCHANGE': Option('celeryresults'), 141 'RESULT_EXCHANGE_TYPE': Option('direct'), 142 'RESULT_SERIALIZER': Option('json'), 143 'RESULT_PERSISTENT': Option(None, type='bool'), 144 'RIAK_BACKEND_SETTINGS': Option(type='dict'), 145 'ROUTES': Option(type='any'), 146 'SEND_EVENTS': Option(False, type='bool'), 147 'SEND_TASK_ERROR_EMAILS': Option(False, type='bool'), 148 'SEND_TASK_SENT_EVENT': Option(False, type='bool'), 149 'STORE_ERRORS_EVEN_IF_IGNORED': Option(False, type='bool'), 150 'TASK_PROTOCOL': Option(1, type='int'), 151 'TASK_PUBLISH_RETRY': Option(True, type='bool'), 152 'TASK_PUBLISH_RETRY_POLICY': Option({ 153 'max_retries': 3, 154 'interval_start': 0, 155 'interval_max': 1, 156 'interval_step': 0.2}, type='dict'), 157 'TASK_RESULT_EXPIRES': Option(timedelta(days=1), type='float'), 158 'TASK_SERIALIZER': Option('json'), 159 'TIMEZONE': Option(type='string'), 160 'TRACK_STARTED': Option(False, type='bool'), 161 'REDIRECT_STDOUTS': Option(True, type='bool'), 162 'REDIRECT_STDOUTS_LEVEL': Option('WARNING'), 163 'QUEUES': Option(type='dict'), 164 'QUEUE_HA_POLICY': Option(None, type='string'), 165 'SECURITY_KEY': Option(type='string'), 166 'SECURITY_CERTIFICATE': Option(type='string'), 167 'SECURITY_CERT_STORE': Option(type='string'), 168 'WORKER_DIRECT': Option(False, type='bool'), 169 }, 170 'CELERYD': { 171 'AGENT': Option(None, type='string'), 172 'AUTOSCALER': Option('celery.worker.autoscale:Autoscaler'), 173 'AUTORELOADER': Option('celery.worker.autoreload:Autoreloader'), 174 'CONCURRENCY': Option(0, type='int'), 175 'TIMER': Option(type='string'), 176 'TIMER_PRECISION': Option(1.0, type='float'), 177 'FORCE_EXECV': Option(False, type='bool'), 178 'HIJACK_ROOT_LOGGER': Option(True, type='bool'), 179 'CONSUMER': Option('celery.worker.consumer:Consumer', type='string'), 180 'LOG_FORMAT': Option(DEFAULT_PROCESS_LOG_FMT), 181 'LOG_COLOR': Option(type='bool'), 182 'LOG_LEVEL': Option('WARN', deprecate_by='2.4', remove_by='4.0', 183 alt='--loglevel argument'), 184 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', 185 alt='--logfile argument'), 186 'MAX_TASKS_PER_CHILD': Option(type='int'), 187 'POOL': Option(DEFAULT_POOL), 188 'POOL_PUTLOCKS': Option(True, type='bool'), 189 'POOL_RESTARTS': Option(False, type='bool'), 190 'PREFETCH_MULTIPLIER': Option(4, type='int'), 191 'STATE_DB': Option(), 192 'TASK_LOG_FORMAT': Option(DEFAULT_TASK_LOG_FMT), 193 'TASK_SOFT_TIME_LIMIT': Option(type='float'), 194 'TASK_TIME_LIMIT': Option(type='float'), 195 'WORKER_LOST_WAIT': Option(10.0, type='float') 196 }, 197 'CELERYBEAT': { 198 'SCHEDULE': Option({}, type='dict'), 199 'SCHEDULER': Option('celery.beat:PersistentScheduler'), 200 'SCHEDULE_FILENAME': Option('celerybeat-schedule'), 201 'SYNC_EVERY': Option(0, type='int'), 202 'MAX_LOOP_INTERVAL': Option(0, type='float'), 203 'LOG_LEVEL': Option('INFO', deprecate_by='2.4', remove_by='4.0', 204 alt='--loglevel argument'), 205 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', 206 alt='--logfile argument'), 207 }, 208 'CELERYMON': { 209 'LOG_LEVEL': Option('INFO', deprecate_by='2.4', remove_by='4.0', 210 alt='--loglevel argument'), 211 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', 212 alt='--logfile argument'), 213 'LOG_FORMAT': Option(DEFAULT_LOG_FMT), 214 }, 215 'EMAIL': { 216 'HOST': Option('localhost'), 217 'PORT': Option(25, type='int'), 218 'HOST_USER': Option(), 219 'HOST_PASSWORD': Option(), 220 'TIMEOUT': Option(2, type='float'), 221 'USE_SSL': Option(False, type='bool'), 222 'USE_TLS': Option(False, type='bool'), 223 'CHARSET': Option('us-ascii'), 224 }, 225 'SERVER_EMAIL': Option('celery@localhost'), 226 'ADMINS': Option((), type='tuple'), 227 } 228 229 230 def flatten(d, ns=''): 231 stack = deque([(ns, d)]) 232 while stack: 233 name, space = stack.popleft() 234 for key, value in items(space): 235 if isinstance(value, dict): 236 stack.append((name + key + '_', value)) 237 else: 238 yield name + key, value 239 DEFAULTS = {key: value.default for key, value in flatten(NAMESPACES)} 240 241 242 def find_deprecated_settings(source): 243 from celery.utils import warn_deprecated 244 for name, opt in flatten(NAMESPACES): 245 if (opt.deprecate_by or opt.remove_by) and getattr(source, name, None): 246 warn_deprecated(description='The {0!r} setting'.format(name), 247 deprecation=opt.deprecate_by, 248 removal=opt.remove_by, 249 alternative='Use the {0.alt} instead'.format(opt)) 250 return source 251 252 253 @memoize(maxsize=None) 254 def find(name, namespace='celery'): 255 # - Try specified namespace first. 256 namespace = namespace.upper() 257 try: 258 return searchresult( 259 namespace, name.upper(), NAMESPACES[namespace][name.upper()], 260 ) 261 except KeyError: 262 # - Try all the other namespaces. 263 for ns, keys in items(NAMESPACES): 264 if ns.upper() == name.upper(): 265 return searchresult(None, ns, keys) 266 elif isinstance(keys, dict): 267 try: 268 return searchresult(ns, name.upper(), keys[name.upper()]) 269 except KeyError: 270 pass 271 # - See if name is a qualname last. 272 return searchresult(None, name.upper(), DEFAULTS[name.upper()]) 273 [end of celery/app/defaults.py] [start of celery/app/task.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.app.task 4 ~~~~~~~~~~~~~~~ 5 6 Task Implementation: Task request context, and the base task class. 7 8 """ 9 from __future__ import absolute_import 10 11 import sys 12 13 from billiard.einfo import ExceptionInfo 14 15 from celery import current_app, group 16 from celery import states 17 from celery._state import _task_stack 18 from celery.canvas import signature 19 from celery.exceptions import Ignore, MaxRetriesExceededError, Reject, Retry 20 from celery.five import class_property, items 21 from celery.result import EagerResult 22 from celery.utils import abstract 23 from celery.utils import uuid, maybe_reraise 24 from celery.utils.functional import mattrgetter, maybe_list 25 from celery.utils.imports import instantiate 26 from celery.utils.mail import ErrorMail 27 28 from .annotations import resolve_all as resolve_all_annotations 29 from .registry import _unpickle_task_v2 30 from .utils import appstr 31 32 __all__ = ['Context', 'Task'] 33 34 #: extracts attributes related to publishing a message from an object. 35 extract_exec_options = mattrgetter( 36 'queue', 'routing_key', 'exchange', 'priority', 'expires', 37 'serializer', 'delivery_mode', 'compression', 'time_limit', 38 'soft_time_limit', 'immediate', 'mandatory', # imm+man is deprecated 39 ) 40 41 # We take __repr__ very seriously around here ;) 42 R_BOUND_TASK = '<class {0.__name__} of {app}{flags}>' 43 R_UNBOUND_TASK = '<unbound {0.__name__}{flags}>' 44 R_SELF_TASK = '<@task {0.name} bound to other {0.__self__}>' 45 R_INSTANCE = '<@task: {0.name} of {app}{flags}>' 46 47 #: Here for backwards compatibility as tasks no longer use a custom metaclass. 48 TaskType = type 49 50 51 def _strflags(flags, default=''): 52 if flags: 53 return ' ({0})'.format(', '.join(flags)) 54 return default 55 56 57 def _reprtask(task, fmt=None, flags=None): 58 flags = list(flags) if flags is not None else [] 59 flags.append('v2 compatible') if task.__v2_compat__ else None 60 if not fmt: 61 fmt = R_BOUND_TASK if task._app else R_UNBOUND_TASK 62 return fmt.format( 63 task, flags=_strflags(flags), 64 app=appstr(task._app) if task._app else None, 65 ) 66 67 68 class Context(object): 69 # Default context 70 logfile = None 71 loglevel = None 72 hostname = None 73 id = None 74 args = None 75 kwargs = None 76 retries = 0 77 eta = None 78 expires = None 79 is_eager = False 80 headers = None 81 delivery_info = None 82 reply_to = None 83 root_id = None 84 parent_id = None 85 correlation_id = None 86 taskset = None # compat alias to group 87 group = None 88 chord = None 89 utc = None 90 called_directly = True 91 callbacks = None 92 errbacks = None 93 timelimit = None 94 _children = None # see property 95 _protected = 0 96 97 def __init__(self, *args, **kwargs): 98 self.update(*args, **kwargs) 99 100 def update(self, *args, **kwargs): 101 return self.__dict__.update(*args, **kwargs) 102 103 def clear(self): 104 return self.__dict__.clear() 105 106 def get(self, key, default=None): 107 return getattr(self, key, default) 108 109 def __repr__(self): 110 return '<Context: {0!r}>'.format(vars(self)) 111 112 @property 113 def children(self): 114 # children must be an empy list for every thread 115 if self._children is None: 116 self._children = [] 117 return self._children 118 119 120 class Task(object): 121 """Task base class. 122 123 When called tasks apply the :meth:`run` method. This method must 124 be defined by all tasks (that is unless the :meth:`__call__` method 125 is overridden). 126 127 """ 128 __trace__ = None 129 __v2_compat__ = False # set by old base in celery.task.base 130 131 ErrorMail = ErrorMail 132 MaxRetriesExceededError = MaxRetriesExceededError 133 134 #: Execution strategy used, or the qualified name of one. 135 Strategy = 'celery.worker.strategy:default' 136 137 #: This is the instance bound to if the task is a method of a class. 138 __self__ = None 139 140 #: The application instance associated with this task class. 141 _app = None 142 143 #: Name of the task. 144 name = None 145 146 #: If :const:`True` the task is an abstract base class. 147 abstract = True 148 149 #: Maximum number of retries before giving up. If set to :const:`None`, 150 #: it will **never** stop retrying. 151 max_retries = 3 152 153 #: Default time in seconds before a retry of the task should be 154 #: executed. 3 minutes by default. 155 default_retry_delay = 3 * 60 156 157 #: Rate limit for this task type. Examples: :const:`None` (no rate 158 #: limit), `'100/s'` (hundred tasks a second), `'100/m'` (hundred tasks 159 #: a minute),`'100/h'` (hundred tasks an hour) 160 rate_limit = None 161 162 #: If enabled the worker will not store task state and return values 163 #: for this task. Defaults to the :setting:`CELERY_IGNORE_RESULT` 164 #: setting. 165 ignore_result = None 166 167 #: If enabled the request will keep track of subtasks started by 168 #: this task, and this information will be sent with the result 169 #: (``result.children``). 170 trail = True 171 172 #: When enabled errors will be stored even if the task is otherwise 173 #: configured to ignore results. 174 store_errors_even_if_ignored = None 175 176 #: If enabled an email will be sent to :setting:`ADMINS` whenever a task 177 #: of this type fails. 178 send_error_emails = None 179 180 #: The name of a serializer that are registered with 181 #: :mod:`kombu.serialization.registry`. Default is `'pickle'`. 182 serializer = None 183 184 #: Hard time limit. 185 #: Defaults to the :setting:`CELERYD_TASK_TIME_LIMIT` setting. 186 time_limit = None 187 188 #: Soft time limit. 189 #: Defaults to the :setting:`CELERYD_TASK_SOFT_TIME_LIMIT` setting. 190 soft_time_limit = None 191 192 #: The result store backend used for this task. 193 backend = None 194 195 #: If disabled this task won't be registered automatically. 196 autoregister = True 197 198 #: If enabled the task will report its status as 'started' when the task 199 #: is executed by a worker. Disabled by default as the normal behaviour 200 #: is to not report that level of granularity. Tasks are either pending, 201 #: finished, or waiting to be retried. 202 #: 203 #: Having a 'started' status can be useful for when there are long 204 #: running tasks and there is a need to report which task is currently 205 #: running. 206 #: 207 #: The application default can be overridden using the 208 #: :setting:`CELERY_TRACK_STARTED` setting. 209 track_started = None 210 211 #: When enabled messages for this task will be acknowledged **after** 212 #: the task has been executed, and not *just before* which is the 213 #: default behavior. 214 #: 215 #: Please note that this means the task may be executed twice if the 216 #: worker crashes mid execution (which may be acceptable for some 217 #: applications). 218 #: 219 #: The application default can be overridden with the 220 #: :setting:`CELERY_ACKS_LATE` setting. 221 acks_late = None 222 223 #: Tuple of expected exceptions. 224 #: 225 #: These are errors that are expected in normal operation 226 #: and that should not be regarded as a real error by the worker. 227 #: Currently this means that the state will be updated to an error 228 #: state, but the worker will not log the event as an error. 229 throws = () 230 231 #: Default task expiry time. 232 expires = None 233 234 #: Task request stack, the current request will be the topmost. 235 request_stack = None 236 237 #: Some may expect a request to exist even if the task has not been 238 #: called. This should probably be deprecated. 239 _default_request = None 240 241 _exec_options = None 242 243 __bound__ = False 244 245 from_config = ( 246 ('send_error_emails', 'CELERY_SEND_TASK_ERROR_EMAILS'), 247 ('serializer', 'CELERY_TASK_SERIALIZER'), 248 ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), 249 ('track_started', 'CELERY_TRACK_STARTED'), 250 ('acks_late', 'CELERY_ACKS_LATE'), 251 ('ignore_result', 'CELERY_IGNORE_RESULT'), 252 ('store_errors_even_if_ignored', 253 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), 254 ) 255 256 #: ignored 257 accept_magic_kwargs = False 258 259 _backend = None # set by backend property. 260 261 __bound__ = False 262 263 # - Tasks are lazily bound, so that configuration is not set 264 # - until the task is actually used 265 266 @classmethod 267 def bind(self, app): 268 was_bound, self.__bound__ = self.__bound__, True 269 self._app = app 270 conf = app.conf 271 self._exec_options = None # clear option cache 272 273 for attr_name, config_name in self.from_config: 274 if getattr(self, attr_name, None) is None: 275 setattr(self, attr_name, conf[config_name]) 276 277 # decorate with annotations from config. 278 if not was_bound: 279 self.annotate() 280 281 from celery.utils.threads import LocalStack 282 self.request_stack = LocalStack() 283 284 # PeriodicTask uses this to add itself to the PeriodicTask schedule. 285 self.on_bound(app) 286 287 return app 288 289 @classmethod 290 def on_bound(self, app): 291 """This method can be defined to do additional actions when the 292 task class is bound to an app.""" 293 pass 294 295 @classmethod 296 def _get_app(self): 297 if self._app is None: 298 self._app = current_app 299 if not self.__bound__: 300 # The app property's __set__ method is not called 301 # if Task.app is set (on the class), so must bind on use. 302 self.bind(self._app) 303 return self._app 304 app = class_property(_get_app, bind) 305 306 @classmethod 307 def annotate(self): 308 for d in resolve_all_annotations(self.app.annotations, self): 309 for key, value in items(d): 310 if key.startswith('@'): 311 self.add_around(key[1:], value) 312 else: 313 setattr(self, key, value) 314 315 @classmethod 316 def add_around(self, attr, around): 317 orig = getattr(self, attr) 318 if getattr(orig, '__wrapped__', None): 319 orig = orig.__wrapped__ 320 meth = around(orig) 321 meth.__wrapped__ = orig 322 setattr(self, attr, meth) 323 324 def __call__(self, *args, **kwargs): 325 _task_stack.push(self) 326 self.push_request(args=args, kwargs=kwargs) 327 try: 328 # add self if this is a bound task 329 if self.__self__ is not None: 330 return self.run(self.__self__, *args, **kwargs) 331 return self.run(*args, **kwargs) 332 finally: 333 self.pop_request() 334 _task_stack.pop() 335 336 def __reduce__(self): 337 # - tasks are pickled into the name of the task only, and the reciever 338 # - simply grabs it from the local registry. 339 # - in later versions the module of the task is also included, 340 # - and the receiving side tries to import that module so that 341 # - it will work even if the task has not been registered. 342 mod = type(self).__module__ 343 mod = mod if mod and mod in sys.modules else None 344 return (_unpickle_task_v2, (self.name, mod), None) 345 346 def run(self, *args, **kwargs): 347 """The body of the task executed by workers.""" 348 raise NotImplementedError('Tasks must define the run method.') 349 350 def start_strategy(self, app, consumer, **kwargs): 351 return instantiate(self.Strategy, self, app, consumer, **kwargs) 352 353 def delay(self, *args, **kwargs): 354 """Star argument version of :meth:`apply_async`. 355 356 Does not support the extra options enabled by :meth:`apply_async`. 357 358 :param \*args: positional arguments passed on to the task. 359 :param \*\*kwargs: keyword arguments passed on to the task. 360 361 :returns :class:`celery.result.AsyncResult`: 362 363 """ 364 return self.apply_async(args, kwargs) 365 366 def apply_async(self, args=None, kwargs=None, task_id=None, producer=None, 367 link=None, link_error=None, shadow=None, **options): 368 """Apply tasks asynchronously by sending a message. 369 370 :keyword args: The positional arguments to pass on to the 371 task (a :class:`list` or :class:`tuple`). 372 373 :keyword kwargs: The keyword arguments to pass on to the 374 task (a :class:`dict`) 375 376 :keyword countdown: Number of seconds into the future that the 377 task should execute. Defaults to immediate 378 execution. 379 380 :keyword eta: A :class:`~datetime.datetime` object describing 381 the absolute time and date of when the task should 382 be executed. May not be specified if `countdown` 383 is also supplied. 384 385 :keyword expires: Either a :class:`int`, describing the number of 386 seconds, or a :class:`~datetime.datetime` object 387 that describes the absolute time and date of when 388 the task should expire. The task will not be 389 executed after the expiration time. 390 391 :keyword shadow: Override task name used in logs/monitoring 392 (default from :meth:`shadow_name`). 393 394 :keyword connection: Re-use existing broker connection instead 395 of establishing a new one. 396 397 :keyword retry: If enabled sending of the task message will be retried 398 in the event of connection loss or failure. Default 399 is taken from the :setting:`CELERY_TASK_PUBLISH_RETRY` 400 setting. Note that you need to handle the 401 producer/connection manually for this to work. 402 403 :keyword retry_policy: Override the retry policy used. See the 404 :setting:`CELERY_TASK_PUBLISH_RETRY_POLICY` 405 setting. 406 407 :keyword routing_key: Custom routing key used to route the task to a 408 worker server. If in combination with a 409 ``queue`` argument only used to specify custom 410 routing keys to topic exchanges. 411 412 :keyword queue: The queue to route the task to. This must be a key 413 present in :setting:`CELERY_QUEUES`, or 414 :setting:`CELERY_CREATE_MISSING_QUEUES` must be 415 enabled. See :ref:`guide-routing` for more 416 information. 417 418 :keyword exchange: Named custom exchange to send the task to. 419 Usually not used in combination with the ``queue`` 420 argument. 421 422 :keyword priority: The task priority, a number between 0 and 9. 423 Defaults to the :attr:`priority` attribute. 424 425 :keyword serializer: A string identifying the default 426 serialization method to use. Can be `pickle`, 427 `json`, `yaml`, `msgpack` or any custom 428 serialization method that has been registered 429 with :mod:`kombu.serialization.registry`. 430 Defaults to the :attr:`serializer` attribute. 431 432 :keyword compression: A string identifying the compression method 433 to use. Can be one of ``zlib``, ``bzip2``, 434 or any custom compression methods registered with 435 :func:`kombu.compression.register`. Defaults to 436 the :setting:`CELERY_MESSAGE_COMPRESSION` 437 setting. 438 :keyword link: A single, or a list of tasks to apply if the 439 task exits successfully. 440 :keyword link_error: A single, or a list of tasks to apply 441 if an error occurs while executing the task. 442 443 :keyword producer: :class:`kombu.Producer` instance to use. 444 445 :keyword add_to_parent: If set to True (default) and the task 446 is applied while executing another task, then the result 447 will be appended to the parent tasks ``request.children`` 448 attribute. Trailing can also be disabled by default using the 449 :attr:`trail` attribute 450 451 :keyword publisher: Deprecated alias to ``producer``. 452 453 :keyword headers: Message headers to be sent in the 454 task (a :class:`dict`) 455 456 :rtype :class:`celery.result.AsyncResult`: if 457 :setting:`CELERY_ALWAYS_EAGER` is not set, otherwise 458 :class:`celery.result.EagerResult`: 459 460 Also supports all keyword arguments supported by 461 :meth:`kombu.Producer.publish`. 462 463 .. note:: 464 If the :setting:`CELERY_ALWAYS_EAGER` setting is set, it will 465 be replaced by a local :func:`apply` call instead. 466 467 """ 468 try: 469 check_arguments = self.__header__ 470 except AttributeError: 471 pass 472 else: 473 check_arguments(*(args or ()), **(kwargs or {})) 474 475 app = self._get_app() 476 if app.conf.CELERY_ALWAYS_EAGER: 477 return self.apply(args, kwargs, task_id=task_id or uuid(), 478 link=link, link_error=link_error, **options) 479 # add 'self' if this is a "task_method". 480 if self.__self__ is not None: 481 args = args if isinstance(args, tuple) else tuple(args or ()) 482 args = (self.__self__,) + args 483 shadow = shadow or self.shadow_name(args, kwargs, options) 484 485 preopts = self._get_exec_options() 486 options = dict(preopts, **options) if options else preopts 487 return app.send_task( 488 self.name, args, kwargs, task_id=task_id, producer=producer, 489 link=link, link_error=link_error, result_cls=self.AsyncResult, 490 shadow=shadow, 491 **options 492 ) 493 494 def shadow_name(self, args, kwargs, options): 495 """Override for custom task name in worker logs/monitoring. 496 497 :param args: Task positional arguments. 498 :param kwargs: Task keyword arguments. 499 :param options: Task execution options. 500 501 **Example**: 502 503 .. code-block:: python 504 505 from celery.utils.imports import qualname 506 507 def shadow_name(task, args, kwargs, options): 508 return qualname(args[0]) 509 510 @app.task(shadow_name=shadow_name, serializer='pickle') 511 def apply_function_async(fun, *args, **kwargs): 512 return fun(*args, **kwargs) 513 514 """ 515 pass 516 517 def signature_from_request(self, request=None, args=None, kwargs=None, 518 queue=None, **extra_options): 519 request = self.request if request is None else request 520 args = request.args if args is None else args 521 kwargs = request.kwargs if kwargs is None else kwargs 522 limit_hard, limit_soft = request.timelimit or (None, None) 523 options = { 524 'task_id': request.id, 525 'link': request.callbacks, 526 'link_error': request.errbacks, 527 'group_id': request.group, 528 'chord': request.chord, 529 'soft_time_limit': limit_soft, 530 'time_limit': limit_hard, 531 'reply_to': request.reply_to, 532 'headers': request.headers, 533 } 534 options.update( 535 {'queue': queue} if queue else (request.delivery_info or {}), 536 ) 537 return self.signature( 538 args, kwargs, options, type=self, **extra_options 539 ) 540 subtask_from_request = signature_from_request 541 542 def retry(self, args=None, kwargs=None, exc=None, throw=True, 543 eta=None, countdown=None, max_retries=None, **options): 544 """Retry the task. 545 546 :param args: Positional arguments to retry with. 547 :param kwargs: Keyword arguments to retry with. 548 :keyword exc: Custom exception to report when the max restart 549 limit has been exceeded (default: 550 :exc:`~@MaxRetriesExceededError`). 551 552 If this argument is set and retry is called while 553 an exception was raised (``sys.exc_info()`` is set) 554 it will attempt to reraise the current exception. 555 556 If no exception was raised it will raise the ``exc`` 557 argument provided. 558 :keyword countdown: Time in seconds to delay the retry for. 559 :keyword eta: Explicit time and date to run the retry at 560 (must be a :class:`~datetime.datetime` instance). 561 :keyword max_retries: If set, overrides the default retry limit. 562 A value of :const:`None`, means "use the default", so if you want 563 infinite retries you would have to set the :attr:`max_retries` 564 attribute of the task to :const:`None` first. 565 :keyword time_limit: If set, overrides the default time limit. 566 :keyword soft_time_limit: If set, overrides the default soft 567 time limit. 568 :keyword \*\*options: Any extra options to pass on to 569 meth:`apply_async`. 570 :keyword throw: If this is :const:`False`, do not raise the 571 :exc:`~@Retry` exception, 572 that tells the worker to mark the task as being 573 retried. Note that this means the task will be 574 marked as failed if the task raises an exception, 575 or successful if it returns. 576 577 :raises celery.exceptions.Retry: To tell the worker that 578 the task has been re-sent for retry. This always happens, 579 unless the `throw` keyword argument has been explicitly set 580 to :const:`False`, and is considered normal operation. 581 582 **Example** 583 584 .. code-block:: pycon 585 586 >>> from imaginary_twitter_lib import Twitter 587 >>> from proj.celery import app 588 589 >>> @app.task(bind=True) 590 ... def tweet(self, auth, message): 591 ... twitter = Twitter(oauth=auth) 592 ... try: 593 ... twitter.post_status_update(message) 594 ... except twitter.FailWhale as exc: 595 ... # Retry in 5 minutes. 596 ... raise self.retry(countdown=60 * 5, exc=exc) 597 598 Although the task will never return above as `retry` raises an 599 exception to notify the worker, we use `raise` in front of the retry 600 to convey that the rest of the block will not be executed. 601 602 """ 603 request = self.request 604 retries = request.retries + 1 605 max_retries = self.max_retries if max_retries is None else max_retries 606 607 # Not in worker or emulated by (apply/always_eager), 608 # so just raise the original exception. 609 if request.called_directly: 610 maybe_reraise() # raise orig stack if PyErr_Occurred 611 raise exc or Retry('Task can be retried', None) 612 613 if not eta and countdown is None: 614 countdown = self.default_retry_delay 615 616 is_eager = request.is_eager 617 S = self.signature_from_request( 618 request, args, kwargs, 619 countdown=countdown, eta=eta, retries=retries, 620 **options 621 ) 622 623 if max_retries is not None and retries > max_retries: 624 if exc: 625 # first try to reraise the original exception 626 maybe_reraise() 627 # or if not in an except block then raise the custom exc. 628 raise exc 629 raise self.MaxRetriesExceededError( 630 "Can't retry {0}[{1}] args:{2} kwargs:{3}".format( 631 self.name, request.id, S.args, S.kwargs)) 632 633 ret = Retry(exc=exc, when=eta or countdown) 634 635 if is_eager: 636 # if task was executed eagerly using apply(), 637 # then the retry must also be executed eagerly. 638 S.apply().get() 639 if throw: 640 raise ret 641 return ret 642 643 try: 644 S.apply_async() 645 except Exception as exc: 646 raise Reject(exc, requeue=False) 647 if throw: 648 raise ret 649 return ret 650 651 def apply(self, args=None, kwargs=None, 652 link=None, link_error=None, **options): 653 """Execute this task locally, by blocking until the task returns. 654 655 :param args: positional arguments passed on to the task. 656 :param kwargs: keyword arguments passed on to the task. 657 :keyword throw: Re-raise task exceptions. Defaults to 658 the :setting:`CELERY_EAGER_PROPAGATES_EXCEPTIONS` 659 setting. 660 661 :rtype :class:`celery.result.EagerResult`: 662 663 """ 664 # trace imports Task, so need to import inline. 665 from celery.app.trace import build_tracer 666 667 app = self._get_app() 668 args = args or () 669 # add 'self' if this is a bound method. 670 if self.__self__ is not None: 671 args = (self.__self__,) + tuple(args) 672 kwargs = kwargs or {} 673 task_id = options.get('task_id') or uuid() 674 retries = options.get('retries', 0) 675 throw = app.either('CELERY_EAGER_PROPAGATES_EXCEPTIONS', 676 options.pop('throw', None)) 677 678 # Make sure we get the task instance, not class. 679 task = app._tasks[self.name] 680 681 request = {'id': task_id, 682 'retries': retries, 683 'is_eager': True, 684 'logfile': options.get('logfile'), 685 'loglevel': options.get('loglevel', 0), 686 'callbacks': maybe_list(link), 687 'errbacks': maybe_list(link_error), 688 'headers': options.get('headers'), 689 'delivery_info': {'is_eager': True}} 690 tb = None 691 tracer = build_tracer( 692 task.name, task, eager=True, 693 propagate=throw, app=self._get_app(), 694 ) 695 ret = tracer(task_id, args, kwargs, request) 696 retval = ret.retval 697 if isinstance(retval, ExceptionInfo): 698 retval, tb = retval.exception, retval.traceback 699 state = states.SUCCESS if ret.info is None else ret.info.state 700 return EagerResult(task_id, retval, state, traceback=tb) 701 702 def AsyncResult(self, task_id, **kwargs): 703 """Get AsyncResult instance for this kind of task. 704 705 :param task_id: Task id to get result for. 706 707 """ 708 return self._get_app().AsyncResult(task_id, backend=self.backend, 709 task_name=self.name, **kwargs) 710 711 def signature(self, args=None, *starargs, **starkwargs): 712 """Return :class:`~celery.signature` object for 713 this task, wrapping arguments and execution options 714 for a single task invocation.""" 715 starkwargs.setdefault('app', self.app) 716 return signature(self, args, *starargs, **starkwargs) 717 subtask = signature 718 719 def s(self, *args, **kwargs): 720 """``.s(*a, **k) -> .signature(a, k)``""" 721 return self.signature(args, kwargs) 722 723 def si(self, *args, **kwargs): 724 """``.si(*a, **k) -> .signature(a, k, immutable=True)``""" 725 return self.signature(args, kwargs, immutable=True) 726 727 def chunks(self, it, n): 728 """Creates a :class:`~celery.canvas.chunks` task for this task.""" 729 from celery import chunks 730 return chunks(self.s(), it, n, app=self.app) 731 732 def map(self, it): 733 """Creates a :class:`~celery.canvas.xmap` task from ``it``.""" 734 from celery import xmap 735 return xmap(self.s(), it, app=self.app) 736 737 def starmap(self, it): 738 """Creates a :class:`~celery.canvas.xstarmap` task from ``it``.""" 739 from celery import xstarmap 740 return xstarmap(self.s(), it, app=self.app) 741 742 def send_event(self, type_, **fields): 743 req = self.request 744 with self.app.events.default_dispatcher(hostname=req.hostname) as d: 745 return d.send(type_, uuid=req.id, **fields) 746 747 def replace(self, sig): 748 """Replace the current task, with a new task inheriting the 749 same task id. 750 751 :param sig: :class:`@signature` 752 753 Note: This will raise :exc:`~@Ignore`, so the best practice 754 is to always use ``raise self.replace(...)`` to convey 755 to the reader that the task will not continue after being replaced. 756 757 :param: Signature of new task. 758 759 """ 760 chord = self.request.chord 761 if isinstance(sig, group): 762 sig |= self.app.tasks['celery.accumulate'].s(index=0).set( 763 chord=chord, 764 ) 765 chord = None 766 sig.freeze(self.request.id, 767 group_id=self.request.group, 768 chord=chord, 769 root_id=self.request.root_id) 770 sig.delay() 771 raise Ignore('Chord member replaced by new task') 772 773 def add_to_chord(self, sig, lazy=False): 774 """Add signature to the chord the current task is a member of. 775 776 :param sig: Signature to extend chord with. 777 :param lazy: If enabled the new task will not actually be called, 778 and ``sig.delay()`` must be called manually. 779 780 Currently only supported by the Redis result backend when 781 ``?new_join=1`` is enabled. 782 783 """ 784 if not self.request.chord: 785 raise ValueError('Current task is not member of any chord') 786 result = sig.freeze(group_id=self.request.group, 787 chord=self.request.chord, 788 root_id=self.request.root_id) 789 self.backend.add_to_chord(self.request.group, result) 790 return sig.delay() if not lazy else sig 791 792 def update_state(self, task_id=None, state=None, meta=None): 793 """Update task state. 794 795 :keyword task_id: Id of the task to update, defaults to the 796 id of the current task 797 :keyword state: New state (:class:`str`). 798 :keyword meta: State metadata (:class:`dict`). 799 800 801 802 """ 803 if task_id is None: 804 task_id = self.request.id 805 self.backend.store_result(task_id, meta, state) 806 807 def on_success(self, retval, task_id, args, kwargs): 808 """Success handler. 809 810 Run by the worker if the task executes successfully. 811 812 :param retval: The return value of the task. 813 :param task_id: Unique id of the executed task. 814 :param args: Original arguments for the executed task. 815 :param kwargs: Original keyword arguments for the executed task. 816 817 The return value of this handler is ignored. 818 819 """ 820 pass 821 822 def on_retry(self, exc, task_id, args, kwargs, einfo): 823 """Retry handler. 824 825 This is run by the worker when the task is to be retried. 826 827 :param exc: The exception sent to :meth:`retry`. 828 :param task_id: Unique id of the retried task. 829 :param args: Original arguments for the retried task. 830 :param kwargs: Original keyword arguments for the retried task. 831 832 :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` 833 instance, containing the traceback. 834 835 The return value of this handler is ignored. 836 837 """ 838 pass 839 840 def on_failure(self, exc, task_id, args, kwargs, einfo): 841 """Error handler. 842 843 This is run by the worker when the task fails. 844 845 :param exc: The exception raised by the task. 846 :param task_id: Unique id of the failed task. 847 :param args: Original arguments for the task that failed. 848 :param kwargs: Original keyword arguments for the task 849 that failed. 850 851 :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` 852 instance, containing the traceback. 853 854 The return value of this handler is ignored. 855 856 """ 857 pass 858 859 def after_return(self, status, retval, task_id, args, kwargs, einfo): 860 """Handler called after the task returns. 861 862 :param status: Current task state. 863 :param retval: Task return value/exception. 864 :param task_id: Unique id of the task. 865 :param args: Original arguments for the task. 866 :param kwargs: Original keyword arguments for the task. 867 868 :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` 869 instance, containing the traceback (if any). 870 871 The return value of this handler is ignored. 872 873 """ 874 pass 875 876 def send_error_email(self, context, exc, **kwargs): 877 if self.send_error_emails and \ 878 not getattr(self, 'disable_error_emails', None): 879 self.ErrorMail(self, **kwargs).send(context, exc) 880 881 def add_trail(self, result): 882 if self.trail: 883 self.request.children.append(result) 884 return result 885 886 def push_request(self, *args, **kwargs): 887 self.request_stack.push(Context(*args, **kwargs)) 888 889 def pop_request(self): 890 self.request_stack.pop() 891 892 def __repr__(self): 893 """`repr(task)`""" 894 return _reprtask(self, R_SELF_TASK if self.__self__ else R_INSTANCE) 895 896 def _get_request(self): 897 """Get current request object.""" 898 req = self.request_stack.top 899 if req is None: 900 # task was not called, but some may still expect a request 901 # to be there, perhaps that should be deprecated. 902 if self._default_request is None: 903 self._default_request = Context() 904 return self._default_request 905 return req 906 request = property(_get_request) 907 908 def _get_exec_options(self): 909 if self._exec_options is None: 910 self._exec_options = extract_exec_options(self) 911 return self._exec_options 912 913 @property 914 def backend(self): 915 backend = self._backend 916 if backend is None: 917 return self.app.backend 918 return backend 919 920 @backend.setter 921 def backend(self, value): # noqa 922 self._backend = value 923 924 @property 925 def __name__(self): 926 return self.__class__.__name__ 927 abstract.CallableTask.register(Task) 928 BaseTask = Task # compat alias 929 [end of celery/app/task.py] [start of celery/worker/request.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.worker.request 4 ~~~~~~~~~~~~~~~~~~~~~ 5 6 This module defines the :class:`Request` class, 7 which specifies how tasks are executed. 8 9 """ 10 from __future__ import absolute_import, unicode_literals 11 12 import logging 13 import socket 14 import sys 15 16 from datetime import datetime 17 from weakref import ref 18 19 from kombu.utils.encoding import safe_repr, safe_str 20 21 from celery import signals 22 from celery.app.trace import trace_task, trace_task_ret 23 from celery.exceptions import ( 24 Ignore, TaskRevokedError, InvalidTaskError, 25 SoftTimeLimitExceeded, TimeLimitExceeded, 26 WorkerLostError, Terminated, Retry, Reject, 27 ) 28 from celery.five import string 29 from celery.platforms import signals as _signals 30 from celery.utils.functional import noop 31 from celery.utils.log import get_logger 32 from celery.utils.timeutils import maybe_iso8601, timezone, maybe_make_aware 33 from celery.utils.serialization import get_pickled_exception 34 35 from . import state 36 37 __all__ = ['Request'] 38 39 IS_PYPY = hasattr(sys, 'pypy_version_info') 40 41 logger = get_logger(__name__) 42 debug, info, warn, error = (logger.debug, logger.info, 43 logger.warning, logger.error) 44 _does_info = False 45 _does_debug = False 46 47 48 def __optimize__(): 49 # this is also called by celery.app.trace.setup_worker_optimizations 50 global _does_debug 51 global _does_info 52 _does_debug = logger.isEnabledFor(logging.DEBUG) 53 _does_info = logger.isEnabledFor(logging.INFO) 54 __optimize__() 55 56 # Localize 57 tz_utc = timezone.utc 58 tz_or_local = timezone.tz_or_local 59 send_revoked = signals.task_revoked.send 60 61 task_accepted = state.task_accepted 62 task_ready = state.task_ready 63 revoked_tasks = state.revoked 64 65 66 class Request(object): 67 """A request for task execution.""" 68 acknowledged = False 69 time_start = None 70 worker_pid = None 71 time_limits = (None, None) 72 _already_revoked = False 73 _terminate_on_ack = None 74 _apply_result = None 75 _tzlocal = None 76 77 if not IS_PYPY: # pragma: no cover 78 __slots__ = ( 79 'app', 'type', 'name', 'id', 'on_ack', 'body', 80 'hostname', 'eventer', 'connection_errors', 'task', 'eta', 81 'expires', 'request_dict', 'on_reject', 'utc', 82 'content_type', 'content_encoding', 83 '__weakref__', '__dict__', 84 ) 85 86 def __init__(self, message, on_ack=noop, 87 hostname=None, eventer=None, app=None, 88 connection_errors=None, request_dict=None, 89 task=None, on_reject=noop, body=None, 90 headers=None, decoded=False, utc=True, 91 maybe_make_aware=maybe_make_aware, 92 maybe_iso8601=maybe_iso8601, **opts): 93 if headers is None: 94 headers = message.headers 95 if body is None: 96 body = message.body 97 self.app = app 98 self.message = message 99 self.body = body 100 self.utc = utc 101 if decoded: 102 self.content_type = self.content_encoding = None 103 else: 104 self.content_type, self.content_encoding = ( 105 message.content_type, message.content_encoding, 106 ) 107 108 self.id = headers['id'] 109 type = self.type = self.name = headers['task'] 110 if 'shadow' in headers: 111 self.name = headers['shadow'] 112 if 'timelimit' in headers: 113 self.time_limits = headers['timelimit'] 114 self.on_ack = on_ack 115 self.on_reject = on_reject 116 self.hostname = hostname or socket.gethostname() 117 self.eventer = eventer 118 self.connection_errors = connection_errors or () 119 self.task = task or self.app.tasks[type] 120 121 # timezone means the message is timezone-aware, and the only timezone 122 # supported at this point is UTC. 123 eta = headers.get('eta') 124 if eta is not None: 125 try: 126 eta = maybe_iso8601(eta) 127 except (AttributeError, ValueError, TypeError) as exc: 128 raise InvalidTaskError( 129 'invalid eta value {0!r}: {1}'.format(eta, exc)) 130 self.eta = maybe_make_aware(eta, self.tzlocal) 131 else: 132 self.eta = None 133 134 expires = headers.get('expires') 135 if expires is not None: 136 try: 137 expires = maybe_iso8601(expires) 138 except (AttributeError, ValueError, TypeError) as exc: 139 raise InvalidTaskError( 140 'invalid expires value {0!r}: {1}'.format(expires, exc)) 141 self.expires = maybe_make_aware(expires, self.tzlocal) 142 else: 143 self.expires = None 144 145 delivery_info = message.delivery_info or {} 146 properties = message.properties or {} 147 headers.update({ 148 'reply_to': properties.get('reply_to'), 149 'correlation_id': properties.get('correlation_id'), 150 'delivery_info': { 151 'exchange': delivery_info.get('exchange'), 152 'routing_key': delivery_info.get('routing_key'), 153 'priority': delivery_info.get('priority'), 154 'redelivered': delivery_info.get('redelivered'), 155 } 156 157 }) 158 self.request_dict = headers 159 160 @property 161 def delivery_info(self): 162 return self.request_dict['delivery_info'] 163 164 def execute_using_pool(self, pool, **kwargs): 165 """Used by the worker to send this task to the pool. 166 167 :param pool: A :class:`celery.concurrency.base.TaskPool` instance. 168 169 :raises celery.exceptions.TaskRevokedError: if the task was revoked 170 and ignored. 171 172 """ 173 task_id = self.id 174 task = self.task 175 if self.revoked(): 176 raise TaskRevokedError(task_id) 177 178 time_limit, soft_time_limit = self.time_limits 179 time_limit = time_limit or task.time_limit 180 soft_time_limit = soft_time_limit or task.soft_time_limit 181 result = pool.apply_async( 182 trace_task_ret, 183 args=(self.type, task_id, self.request_dict, self.body, 184 self.content_type, self.content_encoding), 185 accept_callback=self.on_accepted, 186 timeout_callback=self.on_timeout, 187 callback=self.on_success, 188 error_callback=self.on_failure, 189 soft_timeout=soft_time_limit, 190 timeout=time_limit, 191 correlation_id=task_id, 192 ) 193 # cannot create weakref to None 194 self._apply_result = ref(result) if result is not None else result 195 return result 196 197 def execute(self, loglevel=None, logfile=None): 198 """Execute the task in a :func:`~celery.app.trace.trace_task`. 199 200 :keyword loglevel: The loglevel used by the task. 201 :keyword logfile: The logfile used by the task. 202 203 """ 204 if self.revoked(): 205 return 206 207 # acknowledge task as being processed. 208 if not self.task.acks_late: 209 self.acknowledge() 210 211 request = self.request_dict 212 args, kwargs, embed = self.message.payload 213 request.update({'loglevel': loglevel, 'logfile': logfile, 214 'hostname': self.hostname, 'is_eager': False, 215 'args': args, 'kwargs': kwargs}, **embed or {}) 216 retval = trace_task(self.task, self.id, args, kwargs, request, 217 hostname=self.hostname, loader=self.app.loader, 218 app=self.app)[0] 219 self.acknowledge() 220 return retval 221 222 def maybe_expire(self): 223 """If expired, mark the task as revoked.""" 224 if self.expires: 225 now = datetime.now(self.expires.tzinfo) 226 if now > self.expires: 227 revoked_tasks.add(self.id) 228 return True 229 230 def terminate(self, pool, signal=None): 231 signal = _signals.signum(signal or 'TERM') 232 if self.time_start: 233 pool.terminate_job(self.worker_pid, signal) 234 self._announce_revoked('terminated', True, signal, False) 235 else: 236 self._terminate_on_ack = pool, signal 237 if self._apply_result is not None: 238 obj = self._apply_result() # is a weakref 239 if obj is not None: 240 obj.terminate(signal) 241 242 def _announce_revoked(self, reason, terminated, signum, expired): 243 task_ready(self) 244 self.send_event('task-revoked', 245 terminated=terminated, signum=signum, expired=expired) 246 if self.store_errors: 247 self.task.backend.mark_as_revoked(self.id, reason, request=self) 248 self.acknowledge() 249 self._already_revoked = True 250 send_revoked(self.task, request=self, 251 terminated=terminated, signum=signum, expired=expired) 252 253 def revoked(self): 254 """If revoked, skip task and mark state.""" 255 expired = False 256 if self._already_revoked: 257 return True 258 if self.expires: 259 expired = self.maybe_expire() 260 if self.id in revoked_tasks: 261 info('Discarding revoked task: %s[%s]', self.name, self.id) 262 self._announce_revoked( 263 'expired' if expired else 'revoked', False, None, expired, 264 ) 265 return True 266 return False 267 268 def send_event(self, type, **fields): 269 if self.eventer and self.eventer.enabled: 270 self.eventer.send(type, uuid=self.id, **fields) 271 272 def on_accepted(self, pid, time_accepted): 273 """Handler called when task is accepted by worker pool.""" 274 self.worker_pid = pid 275 self.time_start = time_accepted 276 task_accepted(self) 277 if not self.task.acks_late: 278 self.acknowledge() 279 self.send_event('task-started') 280 if _does_debug: 281 debug('Task accepted: %s[%s] pid:%r', self.name, self.id, pid) 282 if self._terminate_on_ack is not None: 283 self.terminate(*self._terminate_on_ack) 284 285 def on_timeout(self, soft, timeout): 286 """Handler called if the task times out.""" 287 task_ready(self) 288 if soft: 289 warn('Soft time limit (%ss) exceeded for %s[%s]', 290 soft, self.name, self.id) 291 exc = SoftTimeLimitExceeded(soft) 292 else: 293 error('Hard time limit (%ss) exceeded for %s[%s]', 294 timeout, self.name, self.id) 295 exc = TimeLimitExceeded(timeout) 296 297 if self.store_errors: 298 self.task.backend.mark_as_failure(self.id, exc, request=self) 299 300 if self.task.acks_late: 301 self.acknowledge() 302 303 def on_success(self, failed__retval__runtime, **kwargs): 304 """Handler called if the task was successfully processed.""" 305 failed, retval, runtime = failed__retval__runtime 306 if failed: 307 if isinstance(retval.exception, (SystemExit, KeyboardInterrupt)): 308 raise retval.exception 309 return self.on_failure(retval, return_ok=True) 310 task_ready(self) 311 312 if self.task.acks_late: 313 self.acknowledge() 314 315 self.send_event('task-succeeded', result=retval, runtime=runtime) 316 317 def on_retry(self, exc_info): 318 """Handler called if the task should be retried.""" 319 if self.task.acks_late: 320 self.acknowledge() 321 322 self.send_event('task-retried', 323 exception=safe_repr(exc_info.exception.exc), 324 traceback=safe_str(exc_info.traceback)) 325 326 def on_failure(self, exc_info, send_failed_event=True, return_ok=False): 327 """Handler called if the task raised an exception.""" 328 task_ready(self) 329 330 if isinstance(exc_info.exception, MemoryError): 331 raise MemoryError('Process got: %s' % (exc_info.exception,)) 332 elif isinstance(exc_info.exception, Reject): 333 return self.reject(requeue=exc_info.exception.requeue) 334 elif isinstance(exc_info.exception, Ignore): 335 return self.acknowledge() 336 337 exc = exc_info.exception 338 339 if isinstance(exc, Retry): 340 return self.on_retry(exc_info) 341 342 # These are special cases where the process would not have had 343 # time to write the result. 344 if self.store_errors: 345 if isinstance(exc, Terminated): 346 self._announce_revoked( 347 'terminated', True, string(exc), False) 348 send_failed_event = False # already sent revoked event 349 elif isinstance(exc, WorkerLostError) or not return_ok: 350 self.task.backend.mark_as_failure( 351 self.id, exc, request=self, 352 ) 353 # (acks_late) acknowledge after result stored. 354 if self.task.acks_late: 355 self.acknowledge() 356 357 if send_failed_event: 358 self.send_event( 359 'task-failed', 360 exception=safe_repr(get_pickled_exception(exc_info.exception)), 361 traceback=exc_info.traceback, 362 ) 363 364 if not return_ok: 365 error('Task handler raised error: %r', exc, 366 exc_info=exc_info.exc_info) 367 368 def acknowledge(self): 369 """Acknowledge task.""" 370 if not self.acknowledged: 371 self.on_ack(logger, self.connection_errors) 372 self.acknowledged = True 373 374 def reject(self, requeue=False): 375 if not self.acknowledged: 376 self.on_reject(logger, self.connection_errors, requeue) 377 self.acknowledged = True 378 379 def info(self, safe=False): 380 return {'id': self.id, 381 'name': self.name, 382 'type': self.type, 383 'body': self.body, 384 'hostname': self.hostname, 385 'time_start': self.time_start, 386 'acknowledged': self.acknowledged, 387 'delivery_info': self.delivery_info, 388 'worker_pid': self.worker_pid} 389 390 def __str__(self): 391 return ' '.join([ 392 self.humaninfo(), 393 ' eta:[{0}]'.format(self.eta) if self.eta else '', 394 ' expires:[{0}]'.format(self.expires) if self.expires else '', 395 ]) 396 shortinfo = __str__ 397 398 def humaninfo(self): 399 return '{0.name}[{0.id}]'.format(self) 400 401 def __repr__(self): 402 return '<{0}: {1}>'.format(type(self).__name__, self.humaninfo()) 403 404 @property 405 def tzlocal(self): 406 if self._tzlocal is None: 407 self._tzlocal = self.app.conf.CELERY_TIMEZONE 408 return self._tzlocal 409 410 @property 411 def store_errors(self): 412 return (not self.task.ignore_result or 413 self.task.store_errors_even_if_ignored) 414 415 @property 416 def task_id(self): 417 # XXX compat 418 return self.id 419 420 @task_id.setter # noqa 421 def task_id(self, value): 422 self.id = value 423 424 @property 425 def task_name(self): 426 # XXX compat 427 return self.name 428 429 @task_name.setter # noqa 430 def task_name(self, value): 431 self.name = value 432 433 @property 434 def reply_to(self): 435 # used by rpc backend when failures reported by parent process 436 return self.request_dict['reply_to'] 437 438 @property 439 def correlation_id(self): 440 # used similarly to reply_to 441 return self.request_dict['correlation_id'] 442 443 444 def create_request_cls(base, task, pool, hostname, eventer, 445 ref=ref, revoked_tasks=revoked_tasks, 446 task_ready=task_ready): 447 from celery.app.trace import trace_task_ret as trace 448 default_time_limit = task.time_limit 449 default_soft_time_limit = task.soft_time_limit 450 apply_async = pool.apply_async 451 acks_late = task.acks_late 452 events = eventer and eventer.enabled 453 454 class Request(base): 455 456 def execute_using_pool(self, pool, **kwargs): 457 task_id = self.id 458 if (self.expires or task_id in revoked_tasks) and self.revoked(): 459 raise TaskRevokedError(task_id) 460 461 time_limit, soft_time_limit = self.time_limits 462 time_limit = time_limit or default_time_limit 463 soft_time_limit = soft_time_limit or default_soft_time_limit 464 result = apply_async( 465 trace, 466 args=(self.type, task_id, self.request_dict, self.body, 467 self.content_type, self.content_encoding), 468 accept_callback=self.on_accepted, 469 timeout_callback=self.on_timeout, 470 callback=self.on_success, 471 error_callback=self.on_failure, 472 soft_timeout=soft_time_limit, 473 timeout=time_limit, 474 correlation_id=task_id, 475 ) 476 # cannot create weakref to None 477 self._apply_result = ref(result) if result is not None else result 478 return result 479 480 def on_success(self, failed__retval__runtime, **kwargs): 481 failed, retval, runtime = failed__retval__runtime 482 if failed: 483 if isinstance(retval.exception, ( 484 SystemExit, KeyboardInterrupt)): 485 raise retval.exception 486 return self.on_failure(retval, return_ok=True) 487 task_ready(self) 488 489 if acks_late: 490 self.acknowledge() 491 492 if events: 493 self.send_event( 494 'task-succeeded', result=retval, runtime=runtime, 495 ) 496 497 return Request 498 [end of celery/worker/request.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
celery/celery
045b52f1450d6d5cc500e0057a4b498250dc5692
Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost.
This is deliberate as if a task is killed it may mean that the next invocation will also cause the same to happen. If the task is redelivered it may cause a loop where the same conditions occur again and again. Also, sadly you cannot distinguish processes killed by OOM from processes killed by other means, and if an administrator kills -9 a task going amok, you usually don't want that task to be called again. There could be a configuration option for not acking terminated tasks, but I'm not sure how useful that would be. A better solution could be to use `basic_reject(requeue=False)` instead of `basic_ack`, that way you can configure a dead letter queue so that the killed tasks will be sent to a queue for manual inspection. I must say, regardless of the status of this feature request, the documentation is misleading. Specifically, [this FAQ makes it seem that process failures would NOT acknowledge messages](http://celery.readthedocs.org/en/latest/faq.html#faq-acks-late-vs-retry). And [this FAQ boldface states](http://celery.readthedocs.org/en/latest/faq.html#id54) that in the event of a kill signal (9), that acks_late will allow the task to re-run (which again, is patently wrong based on this poorly documented behavior). Nowhere in the docs have I found that if the process _dies_, the message will be acknowledged, regardless of acks_late or not. (for instance, I have a set of 10k+ tasks, and some 1% of tasks wind up acknowledged but incomplete when a WorkerLostError is thrown in connection with the worker, although there are no other errors of any kind in any of my logs related to that task). TL;DR at the least, appropriately document the current state when describing the functionality and limitations of acks_late. A work-around would be helpful -- I'm not sure I understand the solution of using `basic_reject`, although I'll keep looking into it. The docs are referring to killing the worker process with KILL, not the child processes. The term worker will always refer to the worker instance, not the pool processes. The section within about acks_late is probably not very helpful and should be removed
2015-10-06T05:34:34Z
<patch> <patch> diff --git a/celery/app/defaults.py b/celery/app/defaults.py --- a/celery/app/defaults.py +++ b/celery/app/defaults.py @@ -132,6 +132,7 @@ def __repr__(self): 'REDIS_DB': Option(type='int', **_REDIS_OLD), 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 'REDIS_MAX_CONNECTIONS': Option(type='int'), + 'REJECT_ON_WORKER_LOST': Option(type='bool'), 'RESULT_BACKEND': Option(type='string'), 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 'RESULT_DB_TABLENAMES': Option(type='dict'), diff --git a/celery/app/task.py b/celery/app/task.py --- a/celery/app/task.py +++ b/celery/app/task.py @@ -220,6 +220,12 @@ class Task(object): #: :setting:`CELERY_ACKS_LATE` setting. acks_late = None + #: When CELERY_ACKS_LATE is set to True, the default behavior to + #: handle worker crash is to acknowledge the message. Setting + #: this to true allows the message to be rejected and requeued so + #: it will be executed again by another worker. + reject_on_worker_lost = None + #: Tuple of expected exceptions. #: #: These are errors that are expected in normal operation @@ -248,6 +254,7 @@ class Task(object): ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), ('track_started', 'CELERY_TRACK_STARTED'), ('acks_late', 'CELERY_ACKS_LATE'), + ('reject_on_worker_lost', 'CELERY_REJECT_ON_WORKER_LOST'), ('ignore_result', 'CELERY_IGNORE_RESULT'), ('store_errors_even_if_ignored', 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), diff --git a/celery/worker/request.py b/celery/worker/request.py --- a/celery/worker/request.py +++ b/celery/worker/request.py @@ -326,7 +326,6 @@ def on_retry(self, exc_info): def on_failure(self, exc_info, send_failed_event=True, return_ok=False): """Handler called if the task raised an exception.""" task_ready(self) - if isinstance(exc_info.exception, MemoryError): raise MemoryError('Process got: %s' % (exc_info.exception,)) elif isinstance(exc_info.exception, Reject): @@ -352,7 +351,13 @@ def on_failure(self, exc_info, send_failed_event=True, return_ok=False): ) # (acks_late) acknowledge after result stored. if self.task.acks_late: - self.acknowledge() + reject_and_requeue = (self.task.reject_on_worker_lost and + isinstance(exc, WorkerLostError) and + self.delivery_info.get('redelivered', False) is False) + if reject_and_requeue: + self.reject(requeue=True) + else: + self.acknowledge() if send_failed_event: self.send_event( </patch> </s> </patch>
diff --git a/celery/tests/worker/test_request.py b/celery/tests/worker/test_request.py --- a/celery/tests/worker/test_request.py +++ b/celery/tests/worker/test_request.py @@ -325,6 +325,20 @@ def test_on_failure_Reject_rejects_with_requeue(self): req_logger, req.connection_errors, True, ) + def test_on_failure_WrokerLostError_rejects_with_requeue(self): + einfo = None + try: + raise WorkerLostError() + except: + einfo = ExceptionInfo(internal=True) + req = self.get_request(self.add.s(2, 2)) + req.task.acks_late = True + req.task.reject_on_worker_lost = True + req.delivery_info['redelivered'] = False + req.on_failure(einfo) + req.on_reject.assert_called_with(req_logger, + req.connection_errors, True) + def test_tzlocal_is_cached(self): req = self.get_request(self.add.s(2, 2)) req._tzlocal = 'foo'
1.0
NVIDIA__NeMo-473
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ``` </issue> <code> [start of README.rst] 1 .. image:: http://www.repostatus.org/badges/latest/active.svg 2 :target: http://www.repostatus.org/#active 3 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 4 5 .. image:: https://img.shields.io/badge/documentation-github.io-blue.svg 6 :target: https://nvidia.github.io/NeMo/ 7 :alt: NeMo documentation on GitHub pages 8 9 .. image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 10 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 11 :alt: NeMo core license and license for collections in this repo 12 13 .. image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 14 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python 15 :alt: Language grade: Python 16 17 .. image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 18 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ 19 :alt: Total alerts 20 21 .. image:: https://img.shields.io/badge/code%20style-black-000000.svg 22 :target: https://github.com/psf/black 23 :alt: Code style: black 24 25 26 27 NVIDIA Neural Modules: NeMo 28 =========================== 29 30 NeMo is a toolkit for defining and building `Conversational AI <https://developer.nvidia.com/conversational-ai#started>`_ applications. 31 32 Goal of the NeMo toolkit is to make it possible for researchers to easily compose complex neural network architectures for conversational AI using reusable components. Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 33 34 **Neural Modules** are conceptual blocks of neural networks that take *typed* inputs and produce *typed* outputs. Such modules typically represent data layers, encoders, decoders, language models, loss functions, or methods of combining activations. 35 36 The toolkit comes with extendable collections of pre-built modules for automatic speech recognition (ASR), natural language processing (NLP) and text synthesis (TTS). 37 38 **Introduction** 39 40 * Watch `this video <https://nvidia.github.io/NeMo/>`_ for a quick walk-through. 41 42 * Documentation (latest released version): https://nvidia.github.io/NeMo/ 43 44 * Read NVIDIA `Developer Blog for example applications <https://devblogs.nvidia.com/how-to-build-domain-specific-automatic-speech-recognition-models-on-gpus/>`_ 45 46 * Read NVIDIA `Developer Blog for Quartznet ASR model <https://devblogs.nvidia.com/develop-smaller-speech-recognition-models-with-nvidias-nemo-framework/>`_ 47 48 * Recommended version to install is **0.9.0** via ``pip install nemo-toolkit`` 49 50 * Recommended NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ 51 52 * Pretrained models are available on NVIDIA `NGC Model repository <https://ngc.nvidia.com/catalog/models?orderBy=modifiedDESC&query=nemo&quickFilter=models&filters=>`_ 53 54 55 Getting started 56 ~~~~~~~~~~~~~~~ 57 58 THE LATEST STABLE VERSION OF NeMo is **0.9.0** (Available via PIP). 59 60 **Requirements** 61 62 1) Python 3.6 or 3.7 63 2) PyTorch 1.4.* with GPU support 64 3) (optional, for best performance) NVIDIA APEX. Install from here: https://github.com/NVIDIA/apex 65 66 **NeMo Docker Container** 67 NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ is now available. 68 69 * Pull the docker: ``docker pull nvcr.io/nvidia/nemo:v0.9`` 70 * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/nemo:v0.9`` 71 72 If you are using the NVIDIA `NGC PyTorch container <https://ngc.nvidia.com/catalog/containers/nvidia:pytorch>`_ follow these instructions 73 74 * Pull the docker: ``docker pull nvcr.io/nvidia/pytorch:20.01-py3`` 75 * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/pytorch:20.01-py3`` 76 * ``apt-get update && apt-get install -y libsndfile1`` 77 * ``pip install nemo_toolkit`` NeMo core 78 * ``pip install nemo_asr`` NeMo ASR (Speech Recognition) collection 79 * ``pip install nemo_nlp`` NeMo NLP (Natural Language Processing) collection 80 * ``pip install nemo_tts`` NeMo TTS (Speech Synthesis) collection 81 82 See `examples/start_here` to get started with the simplest example. The folder `examples` contains several examples to get you started with various tasks in NLP and ASR. 83 84 **Tutorials** 85 86 * `Speech recognition <https://nvidia.github.io/NeMo/asr/intro.html>`_ 87 * `Natural language processing <https://nvidia.github.io/NeMo/nlp/intro.html>`_ 88 * `Speech Synthesis <https://nvidia.github.io/NeMo/tts/intro.html>`_ 89 90 91 DEVELOPMENT 92 ~~~~~~~~~~~ 93 If you'd like to use master branch and/or develop NeMo you can run "reinstall.sh" script. 94 95 `Documentation (master branch) <http://nemo-master-docs.s3-website.us-east-2.amazonaws.com/>`_. 96 97 **Installing From Github** 98 99 If you prefer to use NeMo's latest development version (from GitHub) follow the steps below: 100 101 1) Clone the repository ``git clone https://github.com/NVIDIA/NeMo.git`` 102 2) Go to NeMo folder and re-install the toolkit with collections: 103 104 .. code-block:: bash 105 106 ./reinstall.sh 107 108 **Style tests** 109 110 .. code-block:: bash 111 112 python setup.py style # Checks overall project code style and output issues with diff. 113 python setup.py style --fix # Tries to fix error in-place. 114 python setup.py style --scope=tests # Operates within certain scope (dir of file). 115 116 **Unittests** 117 118 This command runs unittests: 119 120 .. code-block:: bash 121 122 ./reinstall.sh 123 python pytest tests 124 125 126 Citation 127 ~~~~~~~~ 128 129 If you are using NeMo please cite the following publication 130 131 .. code-block:: tex 132 133 @misc{nemo2019, 134 title={NeMo: a toolkit for building AI applications using Neural Modules}, 135 author={Oleksii Kuchaiev and Jason Li and Huyen Nguyen and Oleksii Hrinchuk and Ryan Leary and Boris Ginsburg and Samuel Kriman and Stanislav Beliaev and Vitaly Lavrukhin and Jack Cook and Patrice Castonguay and Mariya Popova and Jocelyn Huang and Jonathan M. Cohen}, 136 year={2019}, 137 eprint={1909.09577}, 138 archivePrefix={arXiv}, 139 primaryClass={cs.LG} 140 } 141 142 [end of README.rst] [start of nemo/backends/pytorch/actions.py] 1 # Copyright (c) 2019 NVIDIA Corporation 2 import copy 3 import importlib 4 import itertools 5 import json 6 import os 7 from collections import defaultdict 8 from contextlib import ExitStack 9 from pathlib import Path 10 from typing import List, Optional 11 12 import torch 13 import torch.distributed as dist 14 import torch.nn as nn 15 import torch.optim as optim 16 from torch.nn.parallel import DistributedDataParallel as DDP 17 18 from nemo import logging 19 from nemo.backends.pytorch.module_wrapper import TrainableNeuralModuleWrapper 20 from nemo.backends.pytorch.nm import DataLayerNM, TrainableNM 21 from nemo.backends.pytorch.optimizers import AdamW, Novograd, master_params 22 from nemo.core import DeploymentFormat, DeviceType, NeuralModule, NmTensor 23 from nemo.core.callbacks import ActionCallback, EvaluatorCallback, SimpleLossLoggerCallback 24 from nemo.core.neural_factory import Actions, ModelMode, Optimization 25 from nemo.core.neural_types import * 26 from nemo.utils.helpers import get_checkpoint_from_dir 27 28 # these imports will happen on as-needed basis 29 amp = None 30 # convert_syncbn = None 31 # create_syncbn_process_group = None 32 LARC = None 33 FusedLAMB = None 34 FusedAdam = None 35 FusedNovoGrad = None 36 37 AmpOptimizations = { 38 Optimization.mxprO0: "O0", 39 Optimization.mxprO1: "O1", 40 Optimization.mxprO2: "O2", 41 Optimization.mxprO3: "O3", 42 } 43 44 _float_2_half_req = { 45 Optimization.mxprO1, 46 Optimization.mxprO2, 47 Optimization.mxprO3, 48 } 49 50 51 class PtActions(Actions): 52 def __init__( 53 self, local_rank=None, global_rank=None, tb_writer=None, optimization_level=Optimization.mxprO0, 54 ): 55 need_apex = local_rank is not None or optimization_level != Optimization.mxprO0 56 if need_apex: 57 try: 58 apex = importlib.import_module('apex') 59 if optimization_level != Optimization.mxprO0: 60 global amp 61 amp = importlib.import_module('apex.amp') 62 if local_rank is not None: 63 # global convert_syncbn 64 # global create_syncbn_process_group 65 global LARC 66 global FusedLAMB 67 global FusedAdam 68 global FusedNovoGrad 69 parallel = importlib.import_module('apex.parallel') 70 apex_optimizer = importlib.import_module('apex.optimizers') 71 # convert_syncbn = parallel.convert_syncbn_model 72 # create_syncbn_process_group = parallel.create_syncbn_process_group 73 LARC = parallel.LARC 74 FusedLAMB = apex_optimizer.FusedLAMB 75 FusedAdam = apex_optimizer.FusedAdam 76 FusedNovoGrad = apex_optimizer.FusedNovoGrad 77 78 except ImportError: 79 raise ImportError( 80 "NVIDIA Apex is necessary for distributed training and" 81 "mixed precision training. It only works on GPUs." 82 "Please install Apex from " 83 "https://www.github.com/nvidia/apex" 84 ) 85 86 super(PtActions, self).__init__( 87 local_rank=local_rank, global_rank=global_rank, optimization_level=optimization_level, 88 ) 89 90 # will be [unique_instance_id -> (NMModule, PTModule)] 91 self.module_reference_table = {} 92 self.step = 0 93 self.epoch_num = 0 94 self.optimizers = [] 95 self.tb_writer = tb_writer 96 self._modules = set() 97 self.cache = None 98 self.amp_initialized = False 99 100 @property 101 def modules(self): 102 return self._modules 103 104 def __get_top_sorted_modules_and_dataloader(self, hook): 105 """ 106 Constructs DAG leading to hook and creates its topological order. 107 It also populates self.module_reference_table. 108 Args: 109 hook: an NmTensor or a list of NmTensors representing leaf nodes 110 in DAG 111 112 Returns: 113 list of modules with their call arguments and outputs, and dataset 114 """ 115 116 def create_node(producer, producer_args): 117 if producer_args is None: 118 return tuple((producer, ())) 119 else: 120 return tuple((producer, tuple([(k, v) for k, v in producer_args.items()]),)) 121 122 def is_in_degree_zero(node, processed_nodes): 123 """A node has in degree of zero""" 124 if node[1] == (): 125 return True 126 for portname, nmtensor in node[1]: 127 nd = create_node(nmtensor.producer, nmtensor.producer_args) 128 if nd not in processed_nodes: 129 return False 130 return True 131 132 hooks = hook if isinstance(hook, list) else [hook] 133 134 # ensures that no tensors are processed twice 135 processed_nmtensors = set() 136 137 indices_to_remove = [] 138 # Check for duplicates in hook 139 for i, nmtensor in enumerate(hook): 140 if nmtensor in processed_nmtensors: 141 indices_to_remove.append(i) 142 else: 143 processed_nmtensors.add(nmtensor) 144 145 for i in reversed(indices_to_remove): 146 hook.pop(i) 147 148 _top_sorted_modules = [] 149 all_nodes = {} 150 151 # extract all nodes to all_nodes set 152 hooks_lst = list(hooks) 153 while len(hooks_lst) > 0: 154 # take nmtensor from the end of the list 155 nmtensor = hooks_lst.pop() 156 node = create_node(nmtensor.producer, nmtensor.producer_args) 157 # Store nmtensor as an output of its producer 158 # first make sure all keys are present per output port 159 # and nm is inside all_nodes 160 if node not in all_nodes: 161 all_nodes[node] = {k: None for k in nmtensor.producer.output_ports} 162 # second, populate output port with current nmtensor 163 # where applicable 164 all_nodes[node][nmtensor.name] = nmtensor 165 processed_nmtensors.add(nmtensor) 166 if nmtensor.producer_args is not None and nmtensor.producer_args != {}: 167 for _, new_nmtensor in nmtensor.producer_args.items(): 168 if new_nmtensor not in processed_nmtensors: 169 # put in the start of list 170 hooks_lst.insert(0, new_nmtensor) 171 172 all_node_with_output = [] 173 # Iterate over all_nodes to create new nodes that include its output 174 # now all nodes have (module, input tensors, output tensors) 175 for node in all_nodes: 176 all_node_with_output.append(tuple((node[0], node[1], all_nodes[node]))) 177 178 processed_nodes = [] 179 while len(all_node_with_output) > 0: 180 for node in all_node_with_output.copy(): 181 # if node's in_degree is zero it can be added to 182 # _top_sorted_modules 183 # this will also reduce in_degree of its children 184 if is_in_degree_zero(node, processed_nodes): 185 _top_sorted_modules.append(node) 186 processed_nodes.append((node[0], node[1])) 187 all_node_with_output.remove(node) 188 189 # Create top_sorted_modules aka callchain 190 top_sorted_modules = [] 191 for i, m in enumerate(_top_sorted_modules): 192 top_sorted_modules.append((m[0], dict(m[1]), m[2])) 193 # Ensure that there is only one dataset in callchain 194 if i > 0 and isinstance(m[0], DataLayerNM): 195 raise ValueError("There were more than one DataLayer NeuralModule inside your DAG.") 196 197 if not isinstance(top_sorted_modules[0][0], DataLayerNM): 198 raise ValueError("The first module in your DAG was not a DataLayer NeuralModule.") 199 200 tdataset = top_sorted_modules[0][0].dataset 201 202 # populate self.module_reference_table 203 for m in top_sorted_modules: 204 if m[0].factory is None and self._local_rank is not None: 205 raise ValueError( 206 "Neural module {0} was created without " 207 "NeuralModuleFactory, but you are trying to" 208 "run in distributed mode. Please instantiate" 209 "NeuralModuleFactory first and pass its " 210 "instance as `factory` parameter to all your" 211 "Neural Module objects." 212 "".format(str(m[0])) 213 ) 214 key = m[0].unique_instance_id 215 if key not in self.module_reference_table: 216 if isinstance(m[0], TrainableNeuralModuleWrapper): 217 self.module_reference_table[key] = (m[0], m[0]._pt_module) 218 else: 219 self.module_reference_table[key] = (m[0], m[0]) 220 221 return top_sorted_modules, tdataset 222 223 def create_optimizer(self, optimizer, things_to_optimize, optimizer_params=None): 224 """ 225 Wrapper function around __setup_optimizer() 226 227 Args: 228 optimizer : A instantiated PyTorch optimizer or string. For 229 currently supported strings, see __setup_optimizer(). 230 things_to_optimize (list): Must be a list of Neural Modules and/or 231 parameters. If a Neural Module is passed, all trainable 232 parameters are extracted and passed to the optimizer. 233 optimizer_params (dict): Optional parameters dictionary. 234 235 Returns: 236 Optimizer 237 """ 238 239 optimizer_instance = None 240 optimizer_class = None 241 if isinstance(optimizer, str): 242 optimizer_class = optimizer 243 elif isinstance(optimizer, torch.optim.Optimizer): 244 optimizer_instance = optimizer 245 else: 246 raise ValueError("`optimizer` must be a string or an instance of torch.optim.Optimizer") 247 248 modules_to_optimize = [] 249 tensors_to_optimize = [] 250 if not isinstance(things_to_optimize, list): 251 things_to_optimize = [things_to_optimize] 252 for thing in things_to_optimize: 253 if isinstance(thing, NeuralModule): 254 modules_to_optimize.append(thing) 255 elif isinstance(thing, NmTensor): 256 tensors_to_optimize.append(thing) 257 else: 258 raise ValueError( 259 "{} passed to create_optimizer() was neither a neural module nor a neural module tensor" 260 ) 261 262 if tensors_to_optimize: 263 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(tensors_to_optimize) 264 265 for module in call_chain: 266 if module[0] not in modules_to_optimize: 267 modules_to_optimize.append(module[0]) 268 269 # Extract trainable weights which will be optimized 270 params_list = [p.parameters() for p in modules_to_optimize if isinstance(p, TrainableNM) or p.is_trainable()] 271 params_to_optimize = itertools.chain(*params_list) 272 273 if optimizer_params is None: 274 optimizer_params = {} 275 # Init amp 276 optimizer = self.__setup_optimizer( 277 optimizer_instance=optimizer_instance, 278 optimizer_class=optimizer_class, 279 optimization_params=optimizer_params, 280 params_to_optimize=params_to_optimize, 281 ) 282 283 self.optimizers.append(optimizer) 284 return optimizer 285 286 @staticmethod 287 def __setup_optimizer( 288 optimizer_instance, optimizer_class, optimization_params, params_to_optimize, 289 ): 290 291 if optimizer_instance is None: 292 # Setup optimizer instance, by default it is SGD 293 lr = optimization_params["lr"] 294 if optimizer_class.lower() == "sgd": 295 optimizer = optim.SGD( 296 params_to_optimize, 297 lr=lr, 298 momentum=optimization_params.get("momentum", 0.9), 299 weight_decay=optimization_params.get("weight_decay", 0.0), 300 ) 301 elif optimizer_class.lower() == "adam": 302 optimizer = optim.Adam( 303 params=params_to_optimize, lr=lr, betas=optimization_params.get("betas", (0.9, 0.999)), 304 ) 305 elif optimizer_class.lower() == "fused_adam": 306 optimizer = FusedAdam(params=params_to_optimize, lr=lr) 307 elif optimizer_class.lower() == "adam_w": 308 optimizer = AdamW( 309 params=params_to_optimize, 310 lr=lr, 311 weight_decay=optimization_params.get("weight_decay", 0.0), 312 betas=optimization_params.get("betas", (0.9, 0.999)), 313 ) 314 elif optimizer_class.lower() == "novograd": 315 optimizer = Novograd( 316 params_to_optimize, 317 lr=lr, 318 weight_decay=optimization_params.get("weight_decay", 0.0), 319 luc=optimization_params.get("luc", False), 320 luc_trust=optimization_params.get("luc_eta", 1e-3), 321 betas=optimization_params.get("betas", (0.95, 0.25)), 322 ) 323 elif optimizer_class.lower() == "fused_novograd": 324 optimizer = FusedNovoGrad( 325 params_to_optimize, 326 lr=lr, 327 weight_decay=optimization_params.get("weight_decay", 0.0), 328 reg_inside_moment=True, 329 grad_averaging=False, 330 betas=optimization_params.get("betas", (0.95, 0.25)), 331 ) 332 elif optimizer_class.lower() == "fused_lamb": 333 optimizer = FusedLAMB(params_to_optimize, lr=lr,) 334 else: 335 raise ValueError("Unknown optimizer class: {0}".format(optimizer_class)) 336 337 if optimization_params.get("larc", False): 338 logging.info("Enabling larc") 339 optimizer = LARC(optimizer, trust_coefficient=optimization_params.get("larc_eta", 2e-2),) 340 else: 341 logging.info("Optimizer instance: {0} is provided.") 342 if optimizer_class is not None and optimizer_class != "": 343 logging.warning("Ignoring `optimizer_class` parameter because `optimizer_instance` is provided") 344 if optimization_params is not None and optimization_params != {}: 345 logging.warning( 346 "Ignoring `optimization_params` parameter for " 347 "optimizer because `optimizer_instance` is provided" 348 ) 349 optimizer = optimizer_instance 350 return optimizer 351 352 def __initialize_amp( 353 self, optimizer, optim_level, amp_max_loss_scale=2.0 ** 24, amp_min_loss_scale=1.0, 354 ): 355 if optim_level not in AmpOptimizations: 356 raise ValueError(f"__initialize_amp() was called with unknown optim_level={optim_level}") 357 # in this case, nothing to do here 358 if optim_level == Optimization.mxprO0: 359 return optimizer 360 361 if len(self.modules) < 1: 362 raise ValueError("There were no modules to initialize") 363 pt_modules = [] 364 for module in self.modules: 365 if isinstance(module, nn.Module): 366 pt_modules.append(module) 367 elif isinstance(module, TrainableNeuralModuleWrapper): 368 pt_modules.append(module._pt_module) 369 370 _, optimizer = amp.initialize( 371 max_loss_scale=amp_max_loss_scale, 372 min_loss_scale=amp_min_loss_scale, 373 models=pt_modules, 374 optimizers=optimizer, 375 opt_level=AmpOptimizations[optim_level], 376 ) 377 self.amp_initialized = True 378 return optimizer 379 380 def __nm_graph_forward_pass( 381 self, call_chain, registered_tensors, mode=ModelMode.train, use_cache=False, 382 ): 383 for ind in range(1, len(call_chain)): 384 if use_cache: 385 in_cache = True 386 for tensor in call_chain[ind][2].values(): 387 if tensor is None: 388 # NM has an output tensor that is not used in the 389 # current call chain, so we don't care if it's not in 390 # cache 391 continue 392 if tensor.unique_name not in registered_tensors: 393 in_cache = False 394 if in_cache: 395 continue 396 call_args = call_chain[ind][1] 397 # module = call_chain[ind][0] 398 m_id = call_chain[ind][0].unique_instance_id 399 pmodule = self.module_reference_table[m_id][1] 400 401 # if self._local_rank is not None: 402 # if isinstance(pmodule, DDP): 403 # if disable_allreduce: 404 # pmodule.disable_allreduce() 405 # else: 406 # pmodule.enable_allreduce() 407 408 if mode == ModelMode.train: 409 # if module.is_trainable(): 410 if isinstance(pmodule, nn.Module): 411 pmodule.train() 412 elif mode == ModelMode.eval: 413 # if module.is_trainable(): 414 if isinstance(pmodule, nn.Module): 415 pmodule.eval() 416 else: 417 raise ValueError("Unknown ModelMode") 418 # prepare call signature for `module` 419 call_set = {} 420 for tensor_name, nmtensor in call_args.items(): 421 # _add_uuid_2_name(nmtensor.name, nmtensor.producer._uuid) 422 key = nmtensor.unique_name 423 call_set[tensor_name] = registered_tensors[key] 424 # actual PyTorch module call with signature 425 if isinstance(self.module_reference_table[m_id][0], TrainableNeuralModuleWrapper,): 426 new_tensors = pmodule(**call_set) 427 else: 428 new_tensors = pmodule(force_pt=True, **call_set) 429 430 if not isinstance(new_tensors, List): 431 if not isinstance(new_tensors, tuple): 432 new_tensors = [new_tensors] 433 else: 434 new_tensors = list(new_tensors) 435 for t_tensor, nm_tensor in zip(new_tensors, call_chain[ind][2].values()): 436 if nm_tensor is None: 437 continue 438 t_name = nm_tensor.unique_name 439 if t_name not in registered_tensors: 440 registered_tensors[t_name] = t_tensor 441 else: 442 raise ValueError("A NMTensor was produced twice in " f"the same DAG. {t_name}") 443 444 @staticmethod 445 def pad_tensor(t: torch.Tensor, target_size: torch.Size): 446 padded_shape = target_size.cpu().data.numpy().tolist() 447 padded_t = torch.zeros(padded_shape).cuda().type_as(t) 448 t_size = t.size() 449 if len(t_size) == 0: 450 padded_t = t 451 elif len(t_size) == 1: 452 padded_t[: t_size[0]] = t 453 elif len(t_size) == 2: 454 padded_t[: t_size[0], : t_size[1]] = t 455 elif len(t_size) == 3: 456 padded_t[: t_size[0], : t_size[1], : t_size[2]] = t 457 elif len(t_size) == 4: 458 padded_t[: t_size[0], : t_size[1], : t_size[2], : t_size[3]] = t 459 else: 460 raise NotImplementedError 461 return padded_t 462 463 @staticmethod 464 def depad_tensor(t: torch.Tensor, original_size: torch.Size): 465 t_size = original_size 466 if len(t_size) == 0: 467 depadded_t = t 468 elif len(t_size) == 1: 469 depadded_t = t[: t_size[0]] 470 elif len(t_size) == 2: 471 depadded_t = t[: t_size[0], : t_size[1]] 472 elif len(t_size) == 3: 473 depadded_t = t[: t_size[0], : t_size[1], : t_size[2]] 474 elif len(t_size) == 4: 475 depadded_t = t[: t_size[0], : t_size[1], : t_size[2], : t_size[3]] 476 else: 477 raise NotImplementedError 478 return depadded_t 479 480 def _eval(self, tensors_2_evaluate, callback, step, verbose=False): 481 """ 482 Evaluation process. 483 WARNING THIS function assumes that all tensors_2_evaluate are based 484 on a single datalayer 485 Args: 486 tensors_2_evaluate: list of NmTensors to evaluate 487 callback: instance of EvaluatorCallback 488 step: current training step, used for logging 489 490 Returns: 491 None 492 """ 493 with torch.no_grad(): 494 # each call chain corresponds to a tensor in tensors_2_evaluate 495 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors_2_evaluate) 496 # "Retrieve" data layer from call chain. 497 dl_nm = call_chain[0][0] 498 499 # Prepare eval_dataloader 500 # For distributed training it should have disjoint subsets of 501 # all data on every worker 502 is_distributed = False 503 world_size = None 504 if dl_nm.placement == DeviceType.AllGpu: 505 assert dist.is_initialized() 506 is_distributed = True 507 world_size = torch.distributed.get_world_size() 508 # logging.info( 509 # "Doing distributed evaluation. Rank {0} of {1}".format( 510 # self.local_rank, world_size 511 # ) 512 # ) 513 if dl_nm.dataset is not None: 514 sampler = torch.utils.data.distributed.DistributedSampler( 515 dataset=dl_nm.dataset, shuffle=dl_nm.shuffle 516 ) 517 eval_dataloader = torch.utils.data.DataLoader( 518 dataset=dl_nm.dataset, 519 sampler=sampler, 520 num_workers=dl_nm.num_workers, 521 batch_size=dl_nm.batch_size, 522 shuffle=False, 523 ) 524 else: 525 eval_dataloader = dl_nm.data_iterator 526 527 if hasattr(eval_dataloader, 'sampler'): 528 eval_dataloader.sampler.set_epoch(0) 529 else: # Not distributed 530 if dl_nm.dataset is not None: 531 # Todo: remove local_parameters 532 eval_dataloader = torch.utils.data.DataLoader( 533 dataset=dl_nm.dataset, 534 sampler=None, # not distributed sampler 535 num_workers=dl_nm.num_workers, 536 batch_size=dl_nm.batch_size, 537 shuffle=dl_nm.shuffle, 538 ) 539 else: 540 eval_dataloader = dl_nm.data_iterator 541 # after this eval_dataloader is ready to be used 542 # reset global_var_dict - results of evaluation will be stored 543 # there 544 545 callback.clear_global_var_dict() 546 dl_device = dl_nm._device 547 548 # Evaluation mini-batch for loop 549 num_batches = None 550 if hasattr(eval_dataloader, "__len__"): 551 num_batches = len(eval_dataloader) 552 for epoch_i, data in enumerate(eval_dataloader, 0): 553 if ( 554 verbose 555 and num_batches is not None 556 and (num_batches < 10 or (epoch_i % int(num_batches / 10) == 0)) 557 ): 558 logging.info(f"Evaluating batch {epoch_i} out of {num_batches}") 559 tensors = [] 560 if isinstance(data, torch.Tensor): 561 data = (data,) 562 for d in data: 563 if isinstance(d, torch.Tensor): 564 tensors.append(d.to(dl_device)) 565 else: 566 tensors.append(d) 567 568 registered_e_tensors = { 569 t.unique_name: d for t, d in zip(call_chain[0][2].values(), tensors) if t is not None 570 } 571 self.__nm_graph_forward_pass( 572 call_chain=call_chain, registered_tensors=registered_e_tensors, mode=ModelMode.eval, 573 ) 574 575 if not is_distributed or self.global_rank == 0: 576 values_dict = {} 577 # If distributed. For the outer loop, we need to ensure that 578 # all processes loop through the elements in the same order 579 for t2e in tensors_2_evaluate: 580 key = t2e.unique_name 581 if key not in registered_e_tensors.keys(): 582 logging.info("WARNING: Tensor {} was not found during eval".format(key)) 583 continue 584 if is_distributed: 585 # where we will all_gather results from all workers 586 tensors_list = [] 587 # where we will all_gather tensor sizes 588 tensor_on_worker = registered_e_tensors[key] 589 if tensor_on_worker.shape != torch.Size([]): 590 tensor_on_worker_size_as_tensor = torch.tensor(tensor_on_worker.shape).cuda() 591 sizes = [] 592 for ind in range(world_size): 593 sizes.append(torch.empty_like(tensor_on_worker_size_as_tensor)) 594 dist.all_gather(sizes, tensor_on_worker_size_as_tensor) 595 mx_dim, _ = torch.max(torch.stack(sizes), dim=0) 596 else: # this is a singleton. For example, loss value 597 sizes = [torch.Size([])] * world_size 598 mx_dim = None 599 for ind in range(world_size): 600 # we have to use max shape for all_gather 601 if mx_dim is None: # singletons 602 tensors_list.append(torch.tensor(2).cuda().type_as(tensor_on_worker)) 603 else: # non-singletons 604 tensors_list.append( 605 torch.empty(mx_dim.cpu().data.numpy().tolist()).cuda().type_as(tensor_on_worker) 606 ) 607 608 if mx_dim is not None: 609 t_to_send = self.pad_tensor(tensor_on_worker, mx_dim) 610 else: 611 t_to_send = tensor_on_worker 612 dist.all_gather(tensors_list, t_to_send) 613 tensors_list = [self.depad_tensor(t, size) for t, size in zip(tensors_list, sizes)] 614 if self.global_rank == 0: 615 values_dict["IS_FROM_DIST_EVAL"] = True 616 values_dict[key] = tensors_list 617 else: # NON-DISTRIBUTED TRAINING 618 values_dict["IS_FROM_DIST_EVAL"] = False 619 values_dict[key] = [registered_e_tensors[key]] 620 if callback.user_iter_callback and (self.global_rank is None or self.global_rank == 0): 621 # values_dict will contain results from all workers 622 callback.user_iter_callback(values_dict, callback._global_var_dict) 623 624 # final aggregation (over minibatches) and logging of results 625 # should happend only on one worker 626 if callback.user_done_callback and (self.global_rank is None or self.global_rank == 0): 627 vals_to_log = callback.user_done_callback(callback._global_var_dict) 628 # log results to Tensorboard or Weights & Biases 629 if vals_to_log is not None: 630 if hasattr(callback, 'swriter') and callback.swriter is not None: 631 if hasattr(callback, 'tb_writer_func') and callback.tb_writer_func is not None: 632 callback.tb_writer_func(callback.swriter, vals_to_log, step) 633 else: 634 for key, val in vals_to_log.items(): 635 callback.swriter.add_scalar(key, val, step) 636 if hasattr(callback, 'wandb_log'): 637 callback.wandb_log(vals_to_log) 638 639 def _infer( 640 self, tensors_to_return, verbose=False, cache=False, use_cache=False, offload_to_cpu=True, 641 ): 642 """ 643 Does the same as _eval() just with tensors instead of eval callback. 644 """ 645 # Checking that cache is used properly 646 if cache and use_cache: 647 raise ValueError( 648 "cache and use_cache were both set. However cache must first be created prior to using it." 649 ) 650 if cache: 651 if self.cache is not None: 652 raise ValueError("cache was set but was not empty") 653 self.cache = [] 654 if use_cache: 655 if not self.cache: 656 raise ValueError("use_cache was set, but cache was empty") 657 658 with torch.no_grad(): 659 # each call chain corresponds to a tensor in tensors_2_evaluate 660 dl_nm = None 661 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors_to_return) 662 dl_nm = call_chain[0][0] 663 664 # Prepare eval_dataloader 665 # For distributed training it should have disjoint subsets of 666 # all data on every worker 667 is_distributed = False 668 world_size = None 669 if dl_nm.placement == DeviceType.AllGpu: 670 if self.cache or use_cache: 671 raise NotImplementedError("Caching is not available for distributed training.") 672 assert dist.is_initialized() 673 is_distributed = True 674 world_size = torch.distributed.get_world_size() 675 # logging.info( 676 # "Doing distributed evaluation. Rank {0} of {1}".format( 677 # self.local_rank, world_size 678 # ) 679 # ) 680 if dl_nm.dataset is not None: 681 sampler = torch.utils.data.distributed.DistributedSampler( 682 dataset=dl_nm.dataset, shuffle=dl_nm.shuffle 683 ) 684 eval_dataloader = torch.utils.data.DataLoader( 685 dataset=dl_nm.dataset, 686 sampler=sampler, 687 num_workers=dl_nm.num_workers, 688 batch_size=dl_nm.batch_size, 689 shuffle=False, 690 ) 691 else: 692 eval_dataloader = dl_nm.data_iterator 693 eval_dataloader.sampler.set_epoch(0) 694 elif not use_cache: # Not distributed and not using cache 695 # Dataloaders are only used if use_cache is False 696 # When caching, the DAG must cache all outputs from dataloader 697 if dl_nm.dataset is not None: 698 # Todo: remove local_parameters 699 eval_dataloader = torch.utils.data.DataLoader( 700 dataset=dl_nm.dataset, 701 sampler=None, # not distributed sampler 702 num_workers=dl_nm.num_workers, 703 batch_size=dl_nm.batch_size, 704 shuffle=dl_nm.shuffle, 705 ) 706 else: 707 eval_dataloader = dl_nm.data_iterator 708 # after this eval_dataloader is ready to be used 709 # reset global_var_dict - results of evaluation will be stored 710 # there 711 712 if not is_distributed or self.global_rank == 0: 713 values_dict = {} 714 for t in tensors_to_return: 715 values_dict[t.unique_name] = [] 716 dl_device = dl_nm._device 717 718 # Evaluation mini-batch for loop 719 if use_cache: 720 num_batches = len(self.cache) 721 loop_iterator = self.cache 722 else: 723 num_batches = len(eval_dataloader) 724 loop_iterator = eval_dataloader 725 726 for epoch_i, data in enumerate(loop_iterator, 0): 727 logging.debug(torch.cuda.memory_allocated()) 728 if verbose and (num_batches < 10 or (epoch_i % int(num_batches / 10) == 0)): 729 logging.info(f"Evaluating batch {epoch_i} out of {num_batches}") 730 tensors = [] 731 if use_cache: 732 registered_e_tensors = data 733 # delete tensors_to_return 734 for t in tensors_to_return: 735 if t.unique_name in registered_e_tensors: 736 del registered_e_tensors[t.unique_name] 737 # Need to check for device type mismatch 738 for t in registered_e_tensors: 739 registered_e_tensors[t].to(dl_device) 740 else: 741 if isinstance(data, torch.Tensor): 742 data = (data,) 743 for d in data: 744 if isinstance(d, torch.Tensor): 745 tensors.append(d.to(dl_device)) 746 else: 747 tensors.append(d) 748 749 registered_e_tensors = { 750 t.unique_name: d for t, d in zip(call_chain[0][2].values(), tensors) if t is not None 751 } 752 self.__nm_graph_forward_pass( 753 call_chain=call_chain, 754 registered_tensors=registered_e_tensors, 755 mode=ModelMode.eval, 756 use_cache=use_cache, 757 ) 758 759 # if offload_to_cpu: 760 # # Take all cuda tensors and save them to value_dict as 761 # # cpu tensors to save GPU memory 762 # for name, tensor in registered_e_tensors.items(): 763 # if isinstance(tensor, torch.Tensor): 764 # registered_e_tensors[name] = tensor.cpu() 765 if cache: 766 self.append_to_cache(registered_e_tensors, offload_to_cpu) 767 768 # If distributed. For the outer loop, we need to ensure that 769 # all processes loop through the elements in the same order 770 for t2e in tensors_to_return: 771 key = t2e.unique_name 772 if key not in registered_e_tensors.keys(): 773 logging.info("WARNING: Tensor {} was not found during eval".format(key)) 774 continue 775 if is_distributed: 776 # where we will all_gather results from all workers 777 tensors_list = [] 778 # where we will all_gather tensor sizes 779 tensor_on_worker = registered_e_tensors[key] 780 if tensor_on_worker.shape != torch.Size([]): 781 tensor_on_worker_size_as_tensor = torch.tensor(tensor_on_worker.shape).cuda() 782 sizes = [] 783 for ind in range(world_size): 784 sizes.append(torch.empty_like(tensor_on_worker_size_as_tensor)) 785 dist.all_gather(sizes, tensor_on_worker_size_as_tensor) 786 mx_dim, _ = torch.max(torch.stack(sizes), dim=0) 787 else: # this is a singleton. For example, loss value 788 sizes = [torch.Size([])] * world_size 789 mx_dim = None 790 for ind in range(world_size): 791 # we have to use max shape for all_gather 792 if mx_dim is None: # singletons 793 tensors_list.append(torch.tensor(2).cuda().type_as(tensor_on_worker)) 794 else: # non-singletons 795 tensors_list.append( 796 torch.empty(mx_dim.cpu().data.numpy().tolist()).cuda().type_as(tensor_on_worker) 797 ) 798 799 if mx_dim is not None: 800 t_to_send = self.pad_tensor(tensor_on_worker, mx_dim) 801 else: 802 t_to_send = tensor_on_worker 803 dist.all_gather(tensors_list, t_to_send) 804 tensors_list = [self.depad_tensor(t, size) for t, size in zip(tensors_list, sizes)] 805 if offload_to_cpu: 806 tensors_list = [t.cpu() for t in tensors_list] 807 if self.global_rank == 0: 808 values_dict[key] += tensors_list 809 else: # NON-DISTRIBUTED TRAINING 810 tensor = registered_e_tensors[key] 811 if offload_to_cpu and isinstance(tensor, torch.Tensor): 812 tensor = tensor.cpu() 813 values_dict[key] += [tensor] 814 815 if not is_distributed or self.global_rank == 0: 816 inferred_tensors = [] 817 for t in tensors_to_return: 818 inferred_tensors.append(values_dict[t.unique_name]) 819 return inferred_tensors 820 821 # For all other ranks 822 return None 823 824 def append_to_cache(self, registered_tensors: dict, offload_to_cpu): 825 """Simpler helper function to add results of __nm_graph_forward_pass to 826 current cache. 827 """ 828 if offload_to_cpu: 829 for t in registered_tensors: 830 registered_tensors[t] = registered_tensors[t].cpu() 831 self.cache.append(registered_tensors) 832 833 def clear_cache(self): 834 """ Simple helpful function to clear cache by setting self.cache to 835 None 836 """ 837 self.cache = None 838 839 def save_state_to(self, path: str): 840 """ 841 Saves current state such as step, epoch and optimizer parameters 842 Args: 843 path: 844 845 Returns: 846 847 """ 848 state = { 849 "step": self.step, 850 "epoch_num": self.epoch_num, 851 "optimizer_state": [opt.state_dict() for opt in self.optimizers], 852 } 853 torch.save(state, path) 854 855 def restore_state_from(self, path: str): 856 """ 857 Restores state such as step, epoch and optimizer parameters 858 Args: 859 path: 860 861 Returns: 862 863 """ 864 if os.path.isfile(path): 865 # map_location could be cuda:<device_id> but cpu seems to be more 866 # general since we are also saving step and epoch_num 867 # load_state_dict should move the variables to the relevant device 868 checkpoint = torch.load(path, map_location="cpu") 869 self.step = checkpoint["step"] 870 self.epoch_num = checkpoint["epoch_num"] 871 if checkpoint["optimizer_state"]: 872 for opt, opt_chkpt in zip(self.optimizers, checkpoint["optimizer_state"]): 873 opt.load_state_dict(opt_chkpt) 874 else: 875 raise FileNotFoundError("Could not find checkpoint file: {0}".format(path)) 876 877 @staticmethod 878 def _check_all_tensors(list_of_tensors): 879 """Method that checks if the passed list contains all NmTensors 880 """ 881 if not isinstance(list_of_tensors, list): 882 return False 883 for tensor in list_of_tensors: 884 if not isinstance(tensor, NmTensor): 885 return False 886 return True 887 888 @staticmethod 889 def _check_tuples(list_of_tuples): 890 """Method that checks if the passed tuple contains an optimizer in the 891 first element, and a list of NmTensors in the second. 892 """ 893 for tup in list_of_tuples: 894 if not (isinstance(tup[0], torch.optim.Optimizer) and PtActions._check_all_tensors(tup[1])): 895 return False 896 return True 897 898 def _get_all_modules(self, training_loop, callbacks, logging_callchain=None): 899 """Gets all neural modules that will be used by train() and eval() via 900 EvaluatorCallbacks. Saves all modules to self.modules 901 """ 902 # If there is a SimpleLossLoggerCallback, create an logger_callchain 903 # with all callchains from training_loop and 904 # SimpleLossLoggerCallback.tensors 905 if logging_callchain: 906 for module in logging_callchain: 907 self.modules.add(module[0]) 908 909 # Else grab all callchains from training_loop 910 else: 911 for step in training_loop: 912 for module in step[2]: 913 self.modules.add(module[0]) 914 915 # Lastly, grab all eval modules 916 if callbacks is not None: 917 for callback in callbacks: 918 if isinstance(callback, EvaluatorCallback): 919 (callchain, _,) = self.__get_top_sorted_modules_and_dataloader(hook=callback.eval_tensors) 920 for module in callchain: 921 self.modules.add(module[0]) 922 923 @staticmethod 924 def __module_export(module, output, d_format: DeploymentFormat, input_example=None, output_example=None): 925 # Check if output already exists 926 destination = Path(output) 927 if destination.exists(): 928 raise FileExistsError(f"Destination {output} already exists. " f"Aborting export.") 929 930 input_names = list(module.input_ports.keys()) 931 output_names = list(module.output_ports.keys()) 932 dynamic_axes = defaultdict(list) 933 934 def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defaultdict): 935 if ntype.axes: 936 for ind, axis in enumerate(ntype.axes): 937 if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: 938 dynamic_axes[port_name].append(ind) 939 940 # This is a hack for Jasper to Jarvis export -- need re-design for this 941 inputs_to_drop = set() 942 outputs_to_drop = set() 943 if type(module).__name__ == "JasperEncoder": 944 logging.info( 945 "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " 946 "deployment" 947 ) 948 inputs_to_drop.add("length") 949 outputs_to_drop.add("encoded_lengths") 950 951 # for input_ports 952 for port_name, ntype in module.input_ports.items(): 953 if port_name in inputs_to_drop: 954 input_names.remove(port_name) 955 continue 956 __extract_dynamic_axes(port_name, ntype, dynamic_axes) 957 # for output_ports 958 for port_name, ntype in module.output_ports.items(): 959 if port_name in outputs_to_drop: 960 output_names.remove(port_name) 961 continue 962 __extract_dynamic_axes(port_name, ntype, dynamic_axes) 963 964 if len(dynamic_axes) == 0: 965 dynamic_axes = None 966 967 # Make a deep copy of init parameters. 968 init_params_copy = copy.deepcopy(module._init_params) 969 970 # Remove NeMo-related things from the module 971 # We need to change __call__ method. Note that this will change the 972 # whole class, not just this object! Which is why we need to repair it 973 # in the finally block 974 type(module).__call__ = torch.nn.Module.__call__ 975 976 # Reset standard instance field - making the file (probably) lighter. 977 module._init_params = None 978 module._placement = None 979 module._factory = None 980 module._device = None 981 982 module.eval() 983 try: 984 if d_format == DeploymentFormat.TORCHSCRIPT: 985 if input_example is None: 986 # Route 1 - via torch.jit.script 987 traced_m = torch.jit.script(module) 988 traced_m.save(output) 989 else: 990 # Route 2 - via tracing 991 traced_m = torch.jit.trace(module, input_example) 992 traced_m.save(output) 993 elif d_format == DeploymentFormat.ONNX or d_format == DeploymentFormat.TRTONNX: 994 if input_example is None: 995 raise ValueError(f'Example input is None, but ONNX tracing was' f' attempted') 996 if output_example is None: 997 if isinstance(input_example, tuple): 998 output_example = module.forward(*input_example) 999 else: 1000 output_example = module.forward(input_example) 1001 with torch.jit.optimized_execution(True): 1002 jitted_model = torch.jit.trace(module, input_example) 1003 1004 torch.onnx.export( 1005 jitted_model, 1006 input_example, 1007 output, 1008 input_names=input_names, 1009 output_names=output_names, 1010 verbose=False, 1011 export_params=True, 1012 do_constant_folding=True, 1013 dynamic_axes=dynamic_axes, 1014 opset_version=11, 1015 example_outputs=output_example, 1016 ) 1017 # fn = output + ".readable" 1018 # with open(fn, 'w') as f: 1019 # tempModel = onnx.load(output) 1020 # onnx.save(tempModel, output + ".copy") 1021 # onnx.checker.check_model(tempModel) 1022 # pgraph = onnx.helper.printable_graph(tempModel.graph) 1023 # f.write(pgraph) 1024 1025 elif d_format == DeploymentFormat.PYTORCH: 1026 torch.save(module.state_dict(), output) 1027 with open(output + ".json", 'w') as outfile: 1028 json.dump(init_params_copy, outfile) 1029 1030 else: 1031 raise NotImplementedError(f"Not supported deployment format: {d_format}") 1032 except Exception as e: # nopep8 1033 logging.error(f'module export failed for {module} ' f'with exception {e}') 1034 finally: 1035 1036 def __old_call__(self, force_pt=False, *input, **kwargs): 1037 pt_call = len(input) > 0 or force_pt 1038 if pt_call: 1039 return nn.Module.__call__(self, *input, **kwargs) 1040 else: 1041 return NeuralModule.__call__(self, **kwargs) 1042 1043 type(module).__call__ = __old_call__ 1044 1045 @staticmethod 1046 def deployment_export(module, output: str, d_format: DeploymentFormat, input_example=None, output_example=None): 1047 """Exports Neural Module instance for deployment. 1048 1049 Args: 1050 module: neural module to export 1051 output (str): where export results should be saved 1052 d_format (DeploymentFormat): which deployment format to use 1053 input_example: sometimes tracing will require input examples 1054 output_example: Should match inference on input_example 1055 amp_max_loss_scale (float): Max value for amp loss scaling. 1056 Defaults to 2.0**24. 1057 """ 1058 1059 with torch.no_grad(): 1060 PtActions.__module_export( 1061 module=module, 1062 output=output, 1063 d_format=d_format, 1064 input_example=input_example, 1065 output_example=output_example, 1066 ) 1067 1068 def train( 1069 self, 1070 tensors_to_optimize, 1071 optimizer=None, 1072 optimization_params=None, 1073 callbacks: Optional[List[ActionCallback]] = None, 1074 lr_policy=None, 1075 batches_per_step=None, 1076 stop_on_nan_loss=False, 1077 synced_batchnorm=False, 1078 synced_batchnorm_groupsize=0, 1079 gradient_predivide=False, 1080 amp_max_loss_scale=2.0 ** 24, 1081 ): 1082 if gradient_predivide: 1083 logging.error( 1084 "gradient_predivide is currently disabled, and is under consideration for removal in future versions. " 1085 "If this functionality is needed, please raise a github issue." 1086 ) 1087 if not optimization_params: 1088 optimization_params = {} 1089 num_epochs = optimization_params.get("num_epochs", None) 1090 max_steps = optimization_params.get("max_steps", None) 1091 if num_epochs is None and max_steps is None: 1092 raise ValueError("You must specify either max_steps or num_epochs") 1093 grad_norm_clip = optimization_params.get('grad_norm_clip', None) 1094 1095 if batches_per_step is None: 1096 batches_per_step = 1 1097 # this is necessary because we average gradients over batch 1098 bps_scale = torch.FloatTensor([1.0 / batches_per_step]).squeeze() 1099 1100 if tensors_to_optimize is None: 1101 # This is Evaluation Mode 1102 self._init_callbacks(callbacks) 1103 # Do action start callbacks 1104 self._perform_on_action_end(callbacks=callbacks) 1105 return 1106 # Check if tensors_to_optimize is just a list of NmTensors 1107 elif tensors_to_optimize is not None and ( 1108 isinstance(tensors_to_optimize[0], NmTensor) and PtActions._check_all_tensors(tensors_to_optimize) 1109 ): 1110 # Parse graph into a topologically sorted sequence of neural 1111 # modules' calls 1112 (opt_call_chain, t_dataset,) = self.__get_top_sorted_modules_and_dataloader(hook=tensors_to_optimize) 1113 # Extract trainable weights which will be optimized 1114 params_list = [ 1115 p[0].parameters() for p in opt_call_chain if isinstance(p[0], TrainableNM) or p[0].is_trainable() 1116 ] 1117 params_to_optimize = itertools.chain(*params_list) 1118 1119 # Setup optimizer instance. By default it is SGD 1120 optimizer_instance = None 1121 optimizer_class = None 1122 if isinstance(optimizer, str): 1123 optimizer_class = optimizer 1124 elif isinstance(optimizer, torch.optim.Optimizer): 1125 optimizer_instance = optimizer 1126 else: 1127 raise ValueError("optimizer was not understood") 1128 optimizer = self.__setup_optimizer( 1129 optimizer_instance=optimizer_instance, 1130 optimizer_class=optimizer_class, 1131 optimization_params=optimization_params, 1132 params_to_optimize=params_to_optimize, 1133 ) 1134 1135 training_loop = [(optimizer, tensors_to_optimize, opt_call_chain)] 1136 1137 self.optimizers.append(optimizer) 1138 assert ( 1139 len(self.optimizers) == 1 1140 ), "There was more than one optimizer, was create_optimizer() called before train()?" 1141 1142 elif PtActions._check_tuples(tensors_to_optimize): 1143 if batches_per_step != 1: 1144 raise ValueError("Gradient accumlation with multiple optimizers is not supported") 1145 datasets = [] 1146 training_loop = [] 1147 for step in tensors_to_optimize: 1148 (step_call_chain, dataset,) = self.__get_top_sorted_modules_and_dataloader(hook=step[1]) 1149 datasets.append(dataset) 1150 training_loop.append((step[0], step[1], step_call_chain)) 1151 1152 t_dataset = datasets[0] 1153 for dataset in datasets: 1154 if type(dataset) is not type(t_dataset): 1155 raise ValueError("There were two training datasets, we only support 1.") 1156 else: 1157 raise ValueError("tensors_to_optimize was not understood") 1158 1159 logging_callchain = None 1160 # callbacks setup 1161 if callbacks is not None: 1162 for callback in callbacks: 1163 if not isinstance(callback, ActionCallback): 1164 raise ValueError("A callback was received that was not a child of ActionCallback") 1165 elif isinstance(callback, SimpleLossLoggerCallback): 1166 if logging_callchain: 1167 raise ValueError("We only support one logger callback but more than one were found") 1168 logger_step_freq = callback._step_freq 1169 logging_tensors = callback.tensors 1170 all_tensors = logging_tensors 1171 for step in training_loop: 1172 all_tensors = all_tensors + step[1] 1173 (logging_callchain, _,) = self.__get_top_sorted_modules_and_dataloader(hook=all_tensors) 1174 1175 self._get_all_modules(training_loop, callbacks, logging_callchain) 1176 1177 # Intialize Amp if needed 1178 if self._optim_level in AmpOptimizations: 1179 # Store mapping of self.optimizers to optimizer in callchain 1180 training_loop_opts = [] 1181 for opt in training_loop: 1182 training_loop_opts.append(self.optimizers.index(opt[0])) 1183 self.optimizers = self.__initialize_amp( 1184 optimizer=self.optimizers, 1185 optim_level=self._optim_level, 1186 amp_max_loss_scale=amp_max_loss_scale, 1187 amp_min_loss_scale=optimization_params.get('amp_min_loss_scale', 1.0), 1188 ) 1189 # Use stored mapping to map amp_init opts to training loop 1190 for i, step in enumerate(training_loop): 1191 training_loop[i] = ( 1192 self.optimizers[training_loop_opts[i]], 1193 step[1], 1194 step[2], 1195 ) 1196 1197 dataNM = training_loop[0][2][0][0] 1198 if dataNM.placement == DeviceType.AllGpu: 1199 # if len(training_loop) > 1: 1200 # raise NotImplementedError( 1201 # "Distributed training does nor work with multiple " 1202 # "optimizers") 1203 logging.info("Doing distributed training") 1204 if t_dataset is not None: 1205 train_sampler = torch.utils.data.distributed.DistributedSampler( 1206 dataset=t_dataset, shuffle=dataNM.shuffle 1207 ) 1208 train_dataloader = torch.utils.data.DataLoader( 1209 dataset=t_dataset, 1210 sampler=train_sampler, 1211 num_workers=dataNM.num_workers, 1212 batch_size=dataNM.batch_size, 1213 shuffle=False, 1214 ) 1215 else: 1216 train_dataloader = dataNM.data_iterator 1217 if hasattr(train_dataloader, 'sampler'): 1218 train_sampler = train_dataloader.sampler 1219 else: 1220 train_sampler = None 1221 1222 for train_iter in training_loop: 1223 call_chain = train_iter[2] 1224 for i in range(1, len(call_chain) - 1): 1225 key = call_chain[i][0].unique_instance_id 1226 pmodule = self.module_reference_table[key][1] 1227 if not isinstance(pmodule, DDP) and isinstance(pmodule, torch.nn.Module): 1228 # gpf = 1 1229 # if gradient_predivide: 1230 # gpf = dist.get_world_size() 1231 # pmodule = DDP(pmodule, gradient_predivide_factor=gpf) # Old Apex Method 1232 1233 # Per pytorch docs, convert sync bn prior to DDP 1234 if synced_batchnorm: 1235 world_size = dist.get_world_size() 1236 sync_batchnorm_group = None 1237 if synced_batchnorm_groupsize > 0: 1238 if world_size % synced_batchnorm_groupsize != 0: 1239 raise ValueError( 1240 f"Synchronized batch norm group size ({synced_batchnorm_groupsize}) must be 0" 1241 f" or divide total number of GPUs ({world_size})." 1242 ) 1243 # Find ranks of other nodes in the same batchnorm group 1244 rank = torch.distributed.get_rank() 1245 group = rank // synced_batchnorm_groupsize 1246 group_rank_ids = range( 1247 group * synced_batchnorm_groupsize, (group + 1) * synced_batchnorm_groupsize 1248 ) 1249 sync_batchnorm_group = torch.distributed.new_group(group_rank_ids) 1250 1251 pmodule = nn.SyncBatchNorm.convert_sync_batchnorm( 1252 pmodule, process_group=sync_batchnorm_group 1253 ) 1254 1255 # By default, disable broadcast_buffers. This disables batch norm synchronization on forward 1256 # pass 1257 pmodule = DDP( 1258 pmodule, device_ids=[self.local_rank], broadcast_buffers=False, find_unused_parameters=True 1259 ) 1260 1261 # # Convert batchnorm modules to synced if applicable 1262 # if synced_batchnorm and isinstance(pmodule, torch.nn.Module): 1263 # world_size = dist.get_world_size() 1264 # if synced_batchnorm_groupsize > 0 and world_size % synced_batchnorm_groupsize != 0: 1265 # raise ValueError( 1266 # f"Synchronized batch norm group size" 1267 # f" ({synced_batchnorm_groupsize}) must be 0" 1268 # f" or divide total number of GPUs" 1269 # f" ({world_size})." 1270 # ) 1271 # process_group = create_syncbn_process_group(synced_batchnorm_groupsize) 1272 # pmodule = convert_syncbn(pmodule, process_group=process_group) 1273 1274 self.module_reference_table[key] = ( 1275 self.module_reference_table[key][0], 1276 pmodule, 1277 ) 1278 # single GPU/CPU training 1279 else: 1280 if t_dataset is not None: 1281 train_sampler = None 1282 train_dataloader = torch.utils.data.DataLoader( 1283 dataset=t_dataset, 1284 sampler=None, 1285 num_workers=dataNM.num_workers, 1286 batch_size=dataNM.batch_size, 1287 shuffle=dataNM.shuffle, 1288 ) 1289 else: 1290 train_dataloader = dataNM.data_iterator 1291 train_sampler = None 1292 1293 self._init_callbacks(callbacks) 1294 # Do action start callbacks 1295 self._perform_on_action_start(callbacks=callbacks) 1296 1297 # MAIN TRAINING LOOP 1298 # iteration over epochs 1299 while num_epochs is None or self.epoch_num < num_epochs: 1300 if train_sampler is not None: 1301 train_sampler.set_epoch(self.epoch_num) 1302 if max_steps is not None and self.step >= max_steps: 1303 break 1304 1305 # Register epochs start with callbacks 1306 self._perform_on_epoch_start(callbacks=callbacks) 1307 1308 # iteration over batches in epoch 1309 batch_counter = 0 1310 for _, data in enumerate(train_dataloader, 0): 1311 if max_steps is not None and self.step >= max_steps: 1312 break 1313 1314 if batch_counter == 0: 1315 # Started step, zero gradients 1316 curr_optimizer = training_loop[self.step % len(training_loop)][0] 1317 curr_optimizer.zero_grad() 1318 # Register iteration start with callbacks 1319 self._perform_on_iteration_start(callbacks=callbacks) 1320 1321 # set learning rate policy 1322 if lr_policy is not None: 1323 adjusted_lr = lr_policy(optimization_params["lr"], self.step, self.epoch_num) 1324 for param_group in curr_optimizer.param_groups: 1325 param_group["lr"] = adjusted_lr 1326 if self.tb_writer is not None: 1327 value = curr_optimizer.param_groups[0]['lr'] 1328 self.tb_writer.add_scalar('param/lr', value, self.step) 1329 if callbacks is not None: 1330 for callback in callbacks: 1331 callback.learning_rate = curr_optimizer.param_groups[0]['lr'] 1332 1333 # registered_tensors will contain created tensors 1334 # named by output port and uuid of module which created them 1335 # Get and properly name tensors returned by data layer 1336 curr_call_chain = training_loop[self.step % len(training_loop)][2] 1337 dl_device = curr_call_chain[0][0]._device 1338 if logging_callchain and self.step % logger_step_freq == 0: 1339 curr_call_chain = logging_callchain 1340 tensors = [] 1341 if isinstance(data, torch.Tensor): 1342 data = (data,) 1343 for d in data: 1344 if isinstance(d, torch.Tensor): 1345 tensors.append(d.to(dl_device)) 1346 else: 1347 tensors.append(d) 1348 1349 registered_tensors = { 1350 t.unique_name: d for t, d in zip(curr_call_chain[0][2].values(), tensors) if t is not None 1351 } 1352 disable_allreduce = batch_counter < (batches_per_step - 1) 1353 self.__nm_graph_forward_pass( 1354 call_chain=curr_call_chain, registered_tensors=registered_tensors, 1355 ) 1356 1357 curr_tensors_to_optimize = training_loop[self.step % len(training_loop)][1] 1358 final_loss = 0 1359 nan = False 1360 for tensor in curr_tensors_to_optimize: 1361 if ( 1362 torch.isnan(registered_tensors[tensor.unique_name]).any() 1363 or torch.isinf(registered_tensors[tensor.unique_name]).any() 1364 ): 1365 if stop_on_nan_loss: 1366 raise ValueError('Loss is NaN or inf - exiting') 1367 logging.warning('Loss is NaN or inf') 1368 curr_optimizer.zero_grad() 1369 nan = True 1370 break 1371 final_loss += registered_tensors[tensor.unique_name] 1372 if nan: 1373 continue 1374 if self._optim_level in AmpOptimizations and self._optim_level != Optimization.mxprO0: 1375 with amp.scale_loss(final_loss, curr_optimizer, delay_unscale=disable_allreduce) as scaled_loss: 1376 if torch.isnan(scaled_loss).any() or torch.isinf(scaled_loss).any(): 1377 if stop_on_nan_loss: 1378 raise ValueError('Loss is NaN or inf -' ' exiting') 1379 logging.warning('WARNING: Loss is NaN or inf') 1380 curr_optimizer.zero_grad() 1381 continue 1382 if disable_allreduce: 1383 with ExitStack() as stack: 1384 for mod in self.get_DDP_modules(curr_call_chain): 1385 stack.enter_context(mod.no_sync()) 1386 scaled_loss.backward(bps_scale.to(scaled_loss.get_device())) 1387 else: 1388 scaled_loss.backward(bps_scale.to(scaled_loss.get_device())) 1389 # no AMP optimizations needed 1390 else: 1391 # multi-GPU, float32 1392 if self._local_rank is not None: 1393 if disable_allreduce: 1394 with ExitStack() as stack: 1395 for mod in self.get_DDP_modules(curr_call_chain): 1396 stack.enter_context(mod.no_sync()) 1397 final_loss.backward(bps_scale.to(final_loss.get_device())) 1398 else: 1399 final_loss.backward(bps_scale.to(final_loss.get_device())) 1400 # single device (CPU or GPU) 1401 else: 1402 # Fix (workaround?) enabling to backpropagate gradiens on CPUs. 1403 if final_loss.get_device() < 0: 1404 final_loss.backward(bps_scale) 1405 else: 1406 final_loss.backward(bps_scale.to(final_loss.get_device())) 1407 1408 batch_counter += 1 1409 1410 if batch_counter == batches_per_step: 1411 # Ended step. Do optimizer update 1412 if grad_norm_clip is not None: 1413 torch.nn.utils.clip_grad_norm_(master_params(curr_optimizer), grad_norm_clip) 1414 curr_optimizer.step() 1415 batch_counter = 0 1416 # Register iteration end with callbacks 1417 self._update_callbacks( 1418 callbacks=callbacks, registered_tensors=registered_tensors, 1419 ) 1420 self._perform_on_iteration_end(callbacks=callbacks) 1421 self.step += 1 1422 # End of epoch for loop 1423 # Register epochs end with callbacks 1424 self._perform_on_epoch_end(callbacks=callbacks) 1425 self.epoch_num += 1 1426 self._perform_on_action_end(callbacks=callbacks) 1427 1428 def infer( 1429 self, 1430 tensors, 1431 checkpoint_dir=None, 1432 ckpt_pattern='', 1433 verbose=True, 1434 cache=False, 1435 use_cache=False, 1436 offload_to_cpu=True, 1437 modules_to_restore=None, 1438 ): 1439 """See NeuralModuleFactory.infer() 1440 """ 1441 1442 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors) 1443 if checkpoint_dir: 1444 # Find all modules that need to be restored 1445 if modules_to_restore is None: 1446 modules_to_restore = [] 1447 modules_to_restore_name = [] 1448 for op in call_chain: 1449 if op[0].num_weights > 0: 1450 modules_to_restore.append(op[0]) 1451 1452 if not isinstance(modules_to_restore, list): 1453 modules_to_restore = [modules_to_restore] 1454 modules_to_restore_name = [] 1455 for mod in modules_to_restore: 1456 if not isinstance(mod, NeuralModule): 1457 raise ValueError("Found something that was not a Neural Module inside modules_to_restore") 1458 elif mod.num_weights == 0: 1459 raise ValueError("Found a Neural Module with 0 weights inside modules_to_restore") 1460 modules_to_restore_name.append(str(mod)) 1461 1462 module_checkpoints = get_checkpoint_from_dir(modules_to_restore_name, checkpoint_dir, ckpt_pattern) 1463 1464 for mod, checkpoint in zip(modules_to_restore, module_checkpoints): 1465 logging.info(f"Restoring {mod} from {checkpoint}") 1466 mod.restore_from(checkpoint, self._local_rank) 1467 1468 # Init Amp 1469 if ( 1470 self._optim_level in AmpOptimizations 1471 and self._optim_level != Optimization.mxprO0 1472 and not self.amp_initialized 1473 ): 1474 pt_modules = [] 1475 for i in range(len(call_chain)): 1476 if isinstance(call_chain[i][0], nn.Module): 1477 pt_modules.append(call_chain[i][0]) 1478 elif isinstance(call_chain[i][0], TrainableNeuralModuleWrapper): 1479 pt_modules.append(call_chain[i][0]._pt_module) 1480 1481 amp.initialize( 1482 min_loss_scale=1.0, models=pt_modules, optimizers=None, opt_level=AmpOptimizations[self._optim_level], 1483 ) 1484 self.amp_initialized = True 1485 1486 # Run infer 1487 return self._infer( 1488 tensors_to_return=tensors, 1489 verbose=verbose, 1490 cache=cache, 1491 use_cache=use_cache, 1492 offload_to_cpu=offload_to_cpu, 1493 ) 1494 1495 def get_DDP_modules(self, call_chain): 1496 modules = [] 1497 for ind in range(1, len(call_chain)): 1498 m_id = call_chain[ind][0].unique_instance_id 1499 module = self.module_reference_table[m_id][1] 1500 if isinstance(module, DDP): 1501 modules.append(module) 1502 1503 return modules 1504 [end of nemo/backends/pytorch/actions.py] [start of nemo/collections/asr/jasper.py] 1 # Copyright (c) 2019 NVIDIA Corporation 2 from typing import Optional 3 4 import torch 5 import torch.nn as nn 6 import torch.nn.functional as F 7 8 import nemo 9 from .parts.jasper import JasperBlock, init_weights, jasper_activations 10 from nemo.backends.pytorch.nm import TrainableNM 11 from nemo.core.neural_types import * 12 from nemo.utils.decorators import add_port_docs 13 14 logging = nemo.logging 15 16 17 class JasperEncoder(TrainableNM): 18 """ 19 Jasper Encoder creates the pre-processing (prologue), Jasper convolution 20 block, and the first 3 post-processing (epilogue) layers as described in 21 Jasper (https://arxiv.org/abs/1904.03288) 22 23 Args: 24 jasper (list): A list of dictionaries. Each element in the list 25 represents the configuration of one Jasper Block. Each element 26 should contain:: 27 28 { 29 # Required parameters 30 'filters' (int) # Number of output channels, 31 'repeat' (int) # Number of sub-blocks, 32 'kernel' (int) # Size of conv kernel, 33 'stride' (int) # Conv stride 34 'dilation' (int) # Conv dilation 35 'dropout' (float) # Dropout probability 36 'residual' (bool) # Whether to use residual or not. 37 # Optional parameters 38 'residual_dense' (bool) # Whether to use Dense Residuals 39 # or not. 'residual' must be True for 'residual_dense' 40 # to be enabled. 41 # Defaults to False. 42 'separable' (bool) # Whether to use separable convolutions. 43 # Defaults to False 44 'groups' (int) # Number of groups in each conv layer. 45 # Defaults to 1 46 'heads' (int) # Sharing of separable filters 47 # Defaults to -1 48 'tied' (bool) # Whether to use the same weights for all 49 # sub-blocks. 50 # Defaults to False 51 'se' (bool) # Whether to add Squeeze and Excitation 52 # sub-blocks. 53 # Defaults to False 54 'se_reduction_ratio' (int) # The reduction ratio of the Squeeze 55 # sub-module. 56 # Must be an integer > 1. 57 # Defaults to 16 58 'kernel_size_factor' (float) # Conv kernel size multiplier 59 # Can be either an int or float 60 # Kernel size is recomputed as below: 61 # new_kernel_size = int(max(1, (kernel_size * kernel_width))) 62 # to prevent kernel sizes than 1. 63 # Note: If rescaled kernel size is an even integer, 64 # adds 1 to the rescaled kernel size to allow "same" 65 # padding. 66 } 67 68 activation (str): Activation function used for each sub-blocks. Can be 69 one of ["hardtanh", "relu", "selu"]. 70 feat_in (int): Number of channels being input to this module 71 normalization_mode (str): Normalization to be used in each sub-block. 72 Can be one of ["batch", "layer", "instance", "group"] 73 Defaults to "batch". 74 residual_mode (str): Type of residual connection. 75 Can be "add" or "max". 76 Defaults to "add". 77 norm_groups (int): Number of groups for "group" normalization type. 78 If set to -1, number of channels is used. 79 Defaults to -1. 80 conv_mask (bool): Controls the use of sequence length masking prior 81 to convolutions. 82 Defaults to True. 83 frame_splicing (int): Defaults to 1. 84 init_mode (str): Describes how neural network parameters are 85 initialized. Options are ['xavier_uniform', 'xavier_normal', 86 'kaiming_uniform','kaiming_normal']. 87 Defaults to "xavier_uniform". 88 """ 89 90 length: Optional[torch.Tensor] 91 92 @property 93 @add_port_docs() 94 def input_ports(self): 95 """Returns definitions of module input ports. 96 """ 97 return { 98 # "audio_signal": NeuralType( 99 # {0: AxisType(BatchTag), 1: AxisType(SpectrogramSignalTag), 2: AxisType(ProcessedTimeTag),} 100 # ), 101 # "length": NeuralType({0: AxisType(BatchTag)}), 102 "audio_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 103 "length": NeuralType(tuple('B'), LengthsType()), 104 } 105 106 @property 107 @add_port_docs() 108 def output_ports(self): 109 """Returns definitions of module output ports. 110 """ 111 return { 112 # "outputs": NeuralType( 113 # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag),} 114 # ), 115 # "encoded_lengths": NeuralType({0: AxisType(BatchTag)}), 116 "outputs": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), 117 "encoded_lengths": NeuralType(tuple('B'), LengthsType()), 118 } 119 120 @property 121 def disabled_deployment_input_ports(self): 122 return set(["length"]) 123 124 @property 125 def disabled_deployment_output_ports(self): 126 return set(["encoded_lengths"]) 127 128 def prepare_for_deployment(self): 129 m_count = 0 130 for m in self.modules(): 131 if type(m).__name__ == "MaskedConv1d": 132 m.use_mask = False 133 m_count += 1 134 logging.warning(f"Turned off {m_count} masked convolutions") 135 136 def __init__( 137 self, 138 jasper, 139 activation, 140 feat_in, 141 normalization_mode="batch", 142 residual_mode="add", 143 norm_groups=-1, 144 conv_mask=True, 145 frame_splicing=1, 146 init_mode='xavier_uniform', 147 ): 148 super().__init__() 149 150 activation = jasper_activations[activation]() 151 feat_in = feat_in * frame_splicing 152 153 residual_panes = [] 154 encoder_layers = [] 155 self.dense_residual = False 156 for lcfg in jasper: 157 dense_res = [] 158 if lcfg.get('residual_dense', False): 159 residual_panes.append(feat_in) 160 dense_res = residual_panes 161 self.dense_residual = True 162 groups = lcfg.get('groups', 1) 163 separable = lcfg.get('separable', False) 164 heads = lcfg.get('heads', -1) 165 se = lcfg.get('se', False) 166 se_reduction_ratio = lcfg.get('se_reduction_ratio', 16) 167 kernel_size_factor = lcfg.get('kernel_size_factor', 1.0) 168 encoder_layers.append( 169 JasperBlock( 170 feat_in, 171 lcfg['filters'], 172 repeat=lcfg['repeat'], 173 kernel_size=lcfg['kernel'], 174 stride=lcfg['stride'], 175 dilation=lcfg['dilation'], 176 dropout=lcfg['dropout'], 177 residual=lcfg['residual'], 178 groups=groups, 179 separable=separable, 180 heads=heads, 181 residual_mode=residual_mode, 182 normalization=normalization_mode, 183 norm_groups=norm_groups, 184 activation=activation, 185 residual_panes=dense_res, 186 conv_mask=conv_mask, 187 se=se, 188 se_reduction_ratio=se_reduction_ratio, 189 kernel_size_factor=kernel_size_factor, 190 ) 191 ) 192 feat_in = lcfg['filters'] 193 194 self.encoder = nn.Sequential(*encoder_layers) 195 self.apply(lambda x: init_weights(x, mode=init_mode)) 196 self.to(self._device) 197 198 def forward(self, audio_signal, length=None): 199 # type: (Tensor, Optional[Tensor]) -> Tensor, Optional[Tensor] 200 201 s_input, length = self.encoder(([audio_signal], length)) 202 if length is None: 203 return s_input[-1] 204 return s_input[-1], length 205 206 207 class JasperDecoderForCTC(TrainableNM): 208 """ 209 Jasper Decoder creates the final layer in Jasper that maps from the outputs 210 of Jasper Encoder to the vocabulary of interest. 211 212 Args: 213 feat_in (int): Number of channels being input to this module 214 num_classes (int): Number of characters in ASR model's vocab/labels. 215 This count should not include the CTC blank symbol. 216 init_mode (str): Describes how neural network parameters are 217 initialized. Options are ['xavier_uniform', 'xavier_normal', 218 'kaiming_uniform','kaiming_normal']. 219 Defaults to "xavier_uniform". 220 """ 221 222 @property 223 @add_port_docs() 224 def input_ports(self): 225 """Returns definitions of module input ports. 226 """ 227 return { 228 # "encoder_output": NeuralType( 229 # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag),} 230 # ) 231 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()) 232 } 233 234 @property 235 @add_port_docs() 236 def output_ports(self): 237 """Returns definitions of module output ports. 238 """ 239 # return {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag), 2: AxisType(ChannelTag),})} 240 return {"output": NeuralType(('B', 'T', 'D'), LogprobsType())} 241 242 def __init__(self, feat_in, num_classes, init_mode="xavier_uniform"): 243 super().__init__() 244 245 self._feat_in = feat_in 246 # Add 1 for blank char 247 self._num_classes = num_classes + 1 248 249 self.decoder_layers = nn.Sequential(nn.Conv1d(self._feat_in, self._num_classes, kernel_size=1, bias=True)) 250 self.apply(lambda x: init_weights(x, mode=init_mode)) 251 self.to(self._device) 252 253 def forward(self, encoder_output): 254 return F.log_softmax(self.decoder_layers(encoder_output).transpose(1, 2), dim=-1) 255 256 257 class JasperDecoderForClassification(TrainableNM): 258 """ 259 Jasper Decoder creates the final layer in Jasper that maps from the outputs 260 of Jasper Encoder to one class label. 261 262 Args: 263 feat_in (int): Number of channels being input to this module 264 num_classes (int): Number of characters in ASR model's vocab/labels. 265 This count should not include the CTC blank symbol. 266 init_mode (str): Describes how neural network parameters are 267 initialized. Options are ['xavier_uniform', 'xavier_normal', 268 'kaiming_uniform','kaiming_normal']. 269 Defaults to "xavier_uniform". 270 """ 271 272 @property 273 def input_ports(self): 274 """Returns definitions of module input ports. 275 """ 276 return { 277 # "encoder_output": NeuralType( 278 # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag)} 279 # ) 280 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()) 281 } 282 283 @property 284 def output_ports(self): 285 """Returns definitions of module output ports. 286 """ 287 # return {"logits": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} 288 return {"logits": NeuralType(('B', 'D'), LogitsType())} 289 290 def __init__( 291 self, *, feat_in, num_classes, init_mode="xavier_uniform", return_logits=True, pooling_type='avg', **kwargs 292 ): 293 TrainableNM.__init__(self, **kwargs) 294 295 self._feat_in = feat_in 296 self._return_logits = return_logits 297 self._num_classes = num_classes 298 299 if pooling_type == 'avg': 300 self.pooling = nn.AdaptiveAvgPool1d(1) 301 elif pooling_type == 'max': 302 self.pooling = nn.AdaptiveMaxPool1d(1) 303 else: 304 raise ValueError('Pooling type chosen is not valid. Must be either `avg` or `max`') 305 306 self.decoder_layers = nn.Sequential(nn.Linear(self._feat_in, self._num_classes, bias=True)) 307 self.apply(lambda x: init_weights(x, mode=init_mode)) 308 self.to(self._device) 309 310 def forward(self, encoder_output): 311 batch, in_channels, timesteps = encoder_output.size() 312 313 encoder_output = self.pooling(encoder_output).view(batch, in_channels) # [B, C] 314 logits = self.decoder_layers(encoder_output) # [B, num_classes] 315 316 if self._return_logits: 317 return logits 318 319 return F.softmax(logits, dim=-1) 320 [end of nemo/collections/asr/jasper.py] [start of nemo/core/neural_factory.py] 1 # ! /usr/bin/python 2 # -*- coding: utf-8 -*- 3 4 # Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. 5 # 6 # Licensed under the Apache License, Version 2.0 (the "License"); 7 # you may not use this file except in compliance with the License. 8 # You may obtain a copy of the License at 9 # 10 # http://www.apache.org/licenses/LICENSE-2.0 11 # 12 # Unless required by applicable law or agreed to in writing, software 13 # distributed under the License is distributed on an "AS IS" BASIS, 14 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 15 # See the License for the specific language governing permissions and 16 # limitations under the License. 17 18 __all__ = [ 19 'Backend', 20 'ModelMode', 21 'Optimization', 22 'DeviceType', 23 'Actions', 24 'NeuralModuleFactory', 25 'DeploymentFormat', 26 ] 27 28 import random 29 from abc import ABC, abstractmethod 30 from enum import Enum 31 from typing import List, Optional 32 33 import numpy as np 34 35 import nemo 36 from ..utils import ExpManager 37 from .callbacks import ActionCallback, EvaluatorCallback 38 from .neural_types import * 39 from nemo.utils.decorators import deprecated 40 41 logging = nemo.logging 42 43 44 class DeploymentFormat(Enum): 45 """Which format to use when exporting a Neural Module for deployment""" 46 47 AUTO = 0 48 PYTORCH = 1 49 TORCHSCRIPT = 2 50 ONNX = 3 51 TRTONNX = 4 52 53 54 class Backend(Enum): 55 """Supported backends. For now, it is only PyTorch.""" 56 57 PyTorch = 1 58 NotSupported = 2 59 60 61 class ModelMode(Enum): 62 """Training Mode or Evaluation/Inference""" 63 64 train = 0 65 eval = 1 66 67 68 class Optimization(Enum): 69 """Various levels of Apex/amp Optimization. 70 WARNING: This might have effect on model accuracy.""" 71 72 mxprO0 = 0 73 mxprO1 = 1 74 mxprO2 = 2 75 mxprO3 = 3 76 77 78 class DeviceType(Enum): 79 """Device types where Neural Modules can be placed.""" 80 81 GPU = 1 82 CPU = 2 83 AllGpu = 3 84 85 86 class Actions(ABC): 87 """Basic actions allowed on graphs of Neural Modules""" 88 89 def __init__(self, local_rank, global_rank, optimization_level=Optimization.mxprO0): 90 self._local_rank = local_rank 91 self._global_rank = global_rank 92 self._optim_level = optimization_level 93 self.step = None 94 self.epoch_num = None 95 96 @property 97 def local_rank(self): 98 """Local rank during distributed execution. None if single GPU/CPU 99 100 Returns: 101 (int) rank or worker or None if not in distributed model 102 """ 103 return self._local_rank 104 105 @property 106 def global_rank(self): 107 """Global rank during distributed execution. None if single GPU/CPU 108 109 Returns: 110 (int) rank or worker or None if not in distributed model 111 """ 112 return self._global_rank 113 114 @abstractmethod 115 def train( 116 self, 117 tensors_to_optimize: List[NmTensor], 118 callbacks: Optional[List[ActionCallback]], 119 lr_policy=None, 120 batches_per_step=None, 121 stop_on_nan_loss=False, 122 ): 123 """This action executes training and (optionally) evaluation. 124 125 Args: 126 tensors_to_optimize: which tensors to optimize. Typically this is 127 single loss tesnor. 128 callbacks: list of callback objects 129 lr_policy: function which should take (initial_lr, step, epoch) and 130 return learning rate 131 batches_per_step: number of mini-batches to process before one 132 optimizer step. (default: None, same as 1). Use this 133 to simulate larger batch sizes on hardware which could not fit 134 larger batch in memory otherwise. Effectively, this will make 135 "algorithmic" batch size per GPU/worker = batches_per_step* 136 batch_size 137 stop_on_nan_loss: (default: False) If set to True, the training 138 will stop if loss=nan. If set to False, the training will 139 continue, but the gradients will be zeroed before next 140 mini-batch. 141 142 Returns: 143 None 144 """ 145 pass 146 147 @abstractmethod 148 def infer(self, tensors: List[NmTensor]): 149 """This action executes inference. Nothing is optimized. 150 Args: 151 tensors: which tensors to evaluate. 152 153 Returns: 154 None 155 """ 156 pass 157 158 @abstractmethod 159 def save_state_to(self, path: str): 160 """ 161 Saves current state such as step, epoch and optimizer parameters 162 Args: 163 path: 164 165 Returns: 166 167 """ 168 pass 169 170 @abstractmethod 171 def restore_state_from(self, path: str): 172 """ 173 Restores state such as step, epoch and optimizer parameters 174 Args: 175 path: 176 177 Returns: 178 179 """ 180 pass 181 182 @abstractmethod 183 def create_optimizer(self, optimizer, things_to_optimize, optimizer_params): 184 """ 185 Creates an optimizer object to be use in the train() method. 186 187 Args: 188 optimizer: Specifies which optimizer to use. 189 things_to_optimize: A list of neural modules or tensors to be 190 optimized. 191 optimizer_params: Specifies the parameters of the optimizer 192 193 Returns: 194 Optimizer 195 """ 196 pass 197 198 def _perform_on_iteration_start(self, callbacks): 199 # TODO: Most of these checks can be relaxed since we enforce callbacks 200 # to be a list of ActionCallback objects 201 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 202 for callback in callbacks: 203 callback.on_iteration_start() 204 205 def _perform_on_iteration_end(self, callbacks): 206 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 207 for callback in callbacks: 208 callback.on_iteration_end() 209 210 def _perform_on_action_start(self, callbacks): 211 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 212 for callback in callbacks: 213 callback.on_action_start() 214 215 def _perform_on_action_end(self, callbacks): 216 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 217 for callback in callbacks: 218 callback.on_action_end() 219 220 def _perform_on_epoch_start(self, callbacks): 221 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 222 for callback in callbacks: 223 callback.on_epoch_start() 224 225 def _perform_on_epoch_end(self, callbacks): 226 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 227 for callback in callbacks: 228 callback.on_epoch_end() 229 230 def _init_callbacks(self, callbacks): 231 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 232 for callback in callbacks: 233 callback.action = self 234 235 def _update_callbacks( 236 self, callbacks=None, registered_tensors=None, 237 ): 238 # if self.local_rank is None or self.local_rank == 0: 239 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 240 for callback in callbacks: 241 callback._registered_tensors = registered_tensors 242 243 244 def _str_to_opt_level(opt_str: str) -> Optimization: 245 number = int(opt_str[1:]) 246 if number not in Optimization._value2member_map_: 247 raise ValueError(f"Unknown optimization value {opt_str}") 248 return Optimization(number) 249 250 251 class NeuralModuleFactory(object): 252 _DEFAULT = None 253 254 """ 255 Neural Module Factory instance is used to create neural modules and 256 trainers 257 258 Args: 259 backend (Backend): Currently only Backend.PyTorch is supported 260 local_rank (int): Process rank. Should be set by distributed runner 261 optimization_level (Optimization): Level of optimization to use. Will 262 be passed to neural modules and actions created by this factory. 263 placement (DeviceType: where to place NeuralModule instances by default 264 cudnn_benchmark (bool): (default False) If set to True it will use 265 cudnnFind method to find the best kernels instead of using 266 heuristics. If the shapes of your inputs are constant this 267 should help, for various shapes it can slow things down. Give it 268 few iterations to warmup if set to True. Currently only supported 269 by PyTorch backend. 270 random_seed (int): (default None) Sets random seed to control for 271 randomness. This should be used for debugging purposes as it might 272 have negative impact on performance. Can't be used when 273 `cudnn_benchmark=True`. 274 master_process (bool): (default True) Flag for master process 275 indication 276 set_default (bool): (default True) True if should set this instance as 277 default factory for modules instantiating. 278 """ 279 280 def __init__( 281 self, 282 backend=Backend.PyTorch, 283 local_rank=None, 284 optimization_level=Optimization.mxprO0, 285 placement=None, 286 cudnn_benchmark=False, 287 random_seed=None, 288 set_default=True, 289 log_dir=None, 290 checkpoint_dir=None, 291 tensorboard_dir=None, 292 create_tb_writer=False, 293 files_to_copy=None, 294 add_time_to_log_dir=False, 295 ): 296 self._local_rank = local_rank 297 self._global_rank = None 298 299 if isinstance(optimization_level, str): 300 optimization_level = _str_to_opt_level(optimization_level) 301 self._optim_level = optimization_level 302 303 if placement is None: 304 if local_rank is not None: 305 device = DeviceType.AllGpu 306 else: 307 device = DeviceType.GPU 308 309 self._placement = device 310 else: 311 self._placement = placement 312 313 self._backend = backend 314 self._world_size = 1 315 broadcast_func = None 316 if backend == Backend.PyTorch: 317 # TODO: Move all framework specific code from this file 318 import torch 319 320 if self._placement != DeviceType.CPU: 321 if not torch.cuda.is_available(): 322 raise ValueError( 323 "You requested to use GPUs but CUDA is " 324 "not installed. You can try running using" 325 " CPU-only. To do this, instantiate your" 326 " factory with placement=DeviceType.CPU" 327 "\n" 328 "Note that this is slow and is not " 329 "well supported." 330 ) 331 332 torch.backends.cudnn.benchmark = cudnn_benchmark 333 if random_seed is not None and cudnn_benchmark: 334 raise ValueError("cudnn_benchmark can not be set to True when random_seed is not None.") 335 if random_seed is not None: 336 torch.backends.cudnn.deterministic = True 337 torch.backends.cudnn.benchmark = False 338 torch.manual_seed(random_seed) 339 np.random.seed(random_seed) 340 random.seed(random_seed) 341 342 if self._local_rank is not None: 343 torch.distributed.init_process_group(backend="nccl", init_method="env://") 344 345 cuda_set = True 346 # Try to set cuda device. This should fail if self._local_rank 347 # is greater than the number of available GPUs 348 try: 349 torch.cuda.set_device(self._local_rank) 350 except RuntimeError: 351 # Note in this case, all tensors are now sent to GPU 0 352 # who could crash because of OOM. Thus init_process_group() 353 # must be done before any cuda tensors are allocated 354 cuda_set = False 355 cuda_set_t = torch.cuda.IntTensor([cuda_set]) 356 357 # Do an all_reduce to ensure all workers obtained a GPU 358 # For the strangest reason, BAND doesn't work so I am resorting 359 # to MIN. 360 torch.distributed.all_reduce(cuda_set_t, op=torch.distributed.ReduceOp.MIN) 361 if cuda_set_t.item() == 0: 362 raise RuntimeError( 363 "There was an error initializing distributed training." 364 " Perhaps you specified more gpus than you have " 365 "available" 366 ) 367 368 del cuda_set_t 369 torch.cuda.empty_cache() 370 # Remove test tensor from memory 371 372 self._world_size = torch.distributed.get_world_size() 373 self._global_rank = torch.distributed.get_rank() 374 375 def torch_broadcast_wrapper(str_len=None, string=None, src=0): 376 """Wrapper function to broadcast string values across all 377 workers 378 """ 379 # Create byte cuda torch tensor 380 if string is not None: 381 string_tensor = torch.tensor(list(string.encode()), dtype=torch.uint8).cuda() 382 else: 383 string_tensor = torch.tensor([0] * str_len, dtype=torch.uint8).cuda() 384 # Run broadcast 385 torch.distributed.broadcast(string_tensor, src) 386 # turn byte tensor back to string 387 return_string = string_tensor.cpu().numpy() 388 return_string = b''.join(return_string).decode() 389 return return_string 390 391 broadcast_func = torch_broadcast_wrapper 392 else: 393 raise NotImplementedError("Only Pytorch backend is currently supported.") 394 395 # Create ExpManager 396 # if log_dir is None, only create logger 397 self._exp_manager = ExpManager( 398 work_dir=log_dir, 399 ckpt_dir=checkpoint_dir, 400 use_tb=create_tb_writer, 401 tb_dir=tensorboard_dir, 402 local_rank=local_rank, 403 global_rank=self._global_rank, 404 files_to_copy=files_to_copy, 405 add_time=add_time_to_log_dir, 406 exist_ok=True, 407 broadcast_func=broadcast_func, 408 ) 409 self._tb_writer = self._exp_manager.tb_writer 410 411 # Create trainer 412 self._trainer = self._get_trainer(tb_writer=self._tb_writer) 413 414 if set_default: 415 NeuralModuleFactory.set_default_factory(self) 416 417 @classmethod 418 def get_default_factory(cls): 419 return cls._DEFAULT 420 421 @classmethod 422 def set_default_factory(cls, factory): 423 cls._DEFAULT = factory 424 425 @classmethod 426 def reset_default_factory(cls): 427 cls._DEFAULT = None 428 429 @staticmethod 430 def __name_import(name): 431 components = name.split(".") 432 mod = __import__(components[0]) 433 for comp in components[1:]: 434 mod = getattr(mod, comp) 435 return mod 436 437 @deprecated(version=0.11) 438 def __get_pytorch_module(self, name, collection, params, pretrained): 439 # TK: "factory" is not passed as parameter anymore. 440 # params["factory"] = self 441 442 if collection == "toys" or collection == "tutorials" or collection == "other": 443 constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.tutorials." + name) 444 elif collection == "nemo_nlp": 445 constructor = NeuralModuleFactory.__name_import("nemo_nlp." + name) 446 if name == "BERT" and pretrained is True: 447 params["pretrained"] = True 448 elif collection == "nemo_asr": 449 constructor = NeuralModuleFactory.__name_import("nemo_asr." + name) 450 elif collection == "nemo_lpr": 451 constructor = NeuralModuleFactory.__name_import("nemo_lpr." + name) 452 elif collection == 'common': 453 constructor = NeuralModuleFactory.__name_import('nemo.backends.pytorch.common.' + name) 454 elif collection == "torchvision": 455 import torchvision.models as tv_models 456 import nemo.backends.pytorch.module_wrapper as mw 457 import torch.nn as nn 458 459 if name == "ImageFolderDataLayer": 460 constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.torchvision.data." + name) 461 instance = constructor(**params) 462 return instance 463 else: 464 _nm_name = name.lower() 465 if _nm_name == "resnet18": 466 input_ports = { 467 "x": NeuralType( 468 { 469 0: AxisType(BatchTag), 470 1: AxisType(ChannelTag), 471 2: AxisType(HeightTag, 224), 472 3: AxisType(WidthTag, 224), 473 } 474 ) 475 } 476 output_ports = {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} 477 478 pt_model = tv_models.resnet18(pretrained=pretrained) 479 num_classes = params.get("num_classes", None) 480 if num_classes is not None: 481 pt_model.fc = nn.Linear(512, params["num_classes"]) 482 return mw.TrainableNeuralModuleWrapper( 483 pt_nn_module=pt_model, input_ports_dict=input_ports, output_ports_dict=output_ports, 484 ) 485 elif _nm_name == "resnet50": 486 input_ports = { 487 "x": NeuralType( 488 { 489 0: AxisType(BatchTag), 490 1: AxisType(ChannelTag), 491 2: AxisType(HeightTag, 224), 492 3: AxisType(WidthTag, 224), 493 } 494 ) 495 } 496 output_ports = {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} 497 498 pt_model = tv_models.resnet50(pretrained=pretrained) 499 num_classes = params.get("num_classes", None) 500 if num_classes is not None: 501 pt_model.fc = nn.Linear(2048, params["num_classes"]) 502 return mw.TrainableNeuralModuleWrapper( 503 pt_nn_module=pt_model, input_ports_dict=input_ports, output_ports_dict=output_ports, 504 ) 505 else: 506 collection_path = "nemo.collections." + collection + "." + name 507 constructor = NeuralModuleFactory.__name_import(collection_path) 508 if name == "BERT" and pretrained is True: 509 params["pretrained"] = True 510 511 # TK: "placement" is not passed as parameter anymore. 512 # if "placement" not in params: 513 # params["placement"] = self._placement 514 instance = constructor(**params) 515 return instance 516 517 @deprecated(version=0.11) 518 def get_module(self, name, collection, params, pretrained=False): 519 """ 520 Creates NeuralModule instance 521 522 Args: 523 name (str): name of NeuralModule which instance should be returned. 524 params (dict): local parameters which should be passed to 525 NeuralModule's constructor. 526 collection (str): in which collection to look for 527 `neural_module_name` 528 pretrained (bool): return pre-trained instance or randomly 529 initialized (default) 530 531 Returns: 532 NeuralModule instance 533 """ 534 535 # TK: "optimization_level" is not passed as parameter anymore. 536 # if params is not None and "optimization_level" in params: 537 # if params["optimization_level"] != self._optim_level: 538 # logging.warning( 539 # "Module's {0} requested optimization level {1} is" 540 # "different from the one specified by factory - {2}." 541 # "Using: {3} for this module".format( 542 # name, params["optimization_level"], self._optim_level, params["optimization_level"], 543 # ) 544 # ) 545 # else: 546 # if params is None: 547 # params = {} 548 # params["optimization_level"] = self._optim_level 549 550 if self._backend == Backend.PyTorch: 551 return self.__get_pytorch_module(name=name, collection=collection, params=params, pretrained=pretrained,) 552 else: 553 return None 554 555 def create_optimizer(self, optimizer, things_to_optimize, optimizer_params): 556 return self._trainer.create_optimizer( 557 optimizer=optimizer, things_to_optimize=things_to_optimize, optimizer_params=optimizer_params, 558 ) 559 560 def train( 561 self, 562 tensors_to_optimize, 563 optimizer=None, 564 optimization_params=None, 565 callbacks: Optional[List[ActionCallback]] = None, 566 lr_policy=None, 567 batches_per_step=None, 568 stop_on_nan_loss=False, 569 synced_batchnorm=False, 570 synced_batchnorm_groupsize=0, 571 gradient_predivide=False, 572 amp_max_loss_scale=2.0 ** 24, 573 reset=False, 574 ): 575 if reset: 576 self.reset_trainer() 577 return self._trainer.train( 578 tensors_to_optimize=tensors_to_optimize, 579 optimizer=optimizer, 580 optimization_params=optimization_params, 581 callbacks=callbacks, 582 lr_policy=lr_policy, 583 batches_per_step=batches_per_step, 584 stop_on_nan_loss=stop_on_nan_loss, 585 synced_batchnorm=synced_batchnorm, 586 synced_batchnorm_groupsize=synced_batchnorm_groupsize, 587 gradient_predivide=gradient_predivide, 588 amp_max_loss_scale=amp_max_loss_scale, 589 ) 590 591 def eval(self, callbacks: List[EvaluatorCallback]): 592 if callbacks is None or len(callbacks) == 0: 593 raise ValueError(f"You need to provide at lease one evaluation" f"callback to eval") 594 for callback in callbacks: 595 if not isinstance(callback, EvaluatorCallback): 596 raise TypeError(f"All callbacks passed to the eval action must" f"be inherited from EvaluatorCallback") 597 self.train( 598 tensors_to_optimize=None, optimizer='sgd', callbacks=callbacks, optimization_params={'num_epochs': 1}, 599 ) 600 601 def deployment_export( 602 self, module, output: str, d_format: DeploymentFormat, input_example=None, output_example=None 603 ): 604 """Exports Neural Module instance for deployment. 605 606 Args: 607 module: neural module to export 608 output (str): where export results should be saved 609 d_format (DeploymentFormat): which deployment format to use 610 input_example: sometimes tracing will require input examples 611 output_example: Should match inference on input_example 612 """ 613 module.prepare_for_deployment() 614 615 return self._trainer.deployment_export( 616 module=module, 617 output=output, 618 d_format=d_format, 619 input_example=input_example, 620 output_example=output_example, 621 ) 622 623 def infer( 624 self, 625 tensors: List[NmTensor], 626 checkpoint_dir=None, 627 ckpt_pattern='', 628 verbose=True, 629 cache=False, 630 use_cache=False, 631 offload_to_cpu=True, 632 modules_to_restore=None, 633 ): 634 """Runs inference to obtain values for tensors 635 636 Args: 637 tensors (list[NmTensor]): List of NeMo tensors that we want to get 638 values of. 639 checkpoint_dir (str): Path to checkpoint directory. Default is None 640 which does not load checkpoints. 641 ckpt_pattern (str): Pattern used to check for checkpoints inside 642 checkpoint_dir. Default is '' which matches any checkpoints 643 inside checkpoint_dir. 644 verbose (bool): Controls printing. Defaults to True. 645 cache (bool): If True, cache all `tensors` and intermediate tensors 646 so that future calls that have use_cache set will avoid 647 computation. Defaults to False. 648 use_cache (bool): Values from `tensors` will be always re-computed. 649 It will re-use intermediate tensors from the DAG leading to 650 `tensors`. If you want something to be re-computed, put it into 651 `tensors` list. Defaults to False. 652 offload_to_cpu (bool): If True, all evaluated tensors are moved to 653 cpu memory after each inference batch. Defaults to True. 654 modules_to_restore (list): Defaults to None, in which case all 655 NMs inside callchain with weights will be restored. If 656 specified only the modules inside this list will be restored. 657 658 Returns: 659 List of evaluated tensors. Each element in the list is also a list 660 where each element is now a batch of tensor values. 661 """ 662 return self._trainer.infer( 663 tensors=tensors, 664 checkpoint_dir=checkpoint_dir, 665 ckpt_pattern=ckpt_pattern, 666 verbose=verbose, 667 cache=cache, 668 use_cache=use_cache, 669 offload_to_cpu=offload_to_cpu, 670 modules_to_restore=modules_to_restore, 671 ) 672 673 def clear_cache(self): 674 """Helper function to clean inference cache.""" 675 self._trainer.clear_cache() 676 677 @deprecated(version="future") 678 def _get_trainer(self, tb_writer=None): 679 if self._backend == Backend.PyTorch: 680 constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.PtActions") 681 instance = constructor( 682 local_rank=self._local_rank, 683 global_rank=self._global_rank, 684 tb_writer=tb_writer, 685 optimization_level=self._optim_level, 686 ) 687 return instance 688 else: 689 raise ValueError("Only PyTorch backend is currently supported.") 690 691 @deprecated( 692 version="future", 693 explanation="Please use .train(...), .eval(...), .infer(...) and " 694 f".create_optimizer(...) of the NeuralModuleFactory instance directly.", 695 ) 696 def get_trainer(self, tb_writer=None): 697 if self._trainer: 698 logging.warning( 699 "The trainer instance was created during initialization of " 700 "Neural factory, using the already created instance." 701 ) 702 return self._trainer 703 return self._get_trainer(tb_writer) 704 705 def reset_trainer(self): 706 del self._trainer 707 self._trainer = self._get_trainer(tb_writer=self._tb_writer) 708 709 def sync_all_processes(self, status=True): 710 """ Helper function for testing that allows proccess 0 to inform all 711 other processes of failures. Does nothing if not using distributed 712 training. Usage example can be seen in examples/asr/jasper_an4.py 713 714 Args: 715 status (bool): Defaults to True. If any proccess passes False, it 716 will trigger a graceful exit on all other processes. It is 717 assumed that the process that passed False will print an error 718 message on its own and exit 719 """ 720 if self._world_size == 1: 721 logging.info("sync_all_processes does nothing if there is one process") 722 return 723 if self._backend == Backend.PyTorch: 724 import torch 725 726 status_tensor = torch.cuda.IntTensor([status]) 727 torch.distributed.all_reduce(status_tensor, op=torch.distributed.ReduceOp.MIN) 728 if status_tensor.item() == 0: 729 logging.error("At least one process had a failure") 730 if status: 731 raise ValueError( 732 f"Process with global rank {self._global_rank} entered" 733 " sync_all_processes with a passing status, but " 734 "another process indicated a failure" 735 ) 736 737 @property 738 def world_size(self): 739 return self._world_size 740 741 @property 742 def tb_writer(self): 743 return self._tb_writer 744 745 @property 746 def placement(self): 747 return self._placement 748 749 @property 750 def optim_level(self): 751 return self._optim_level 752 753 @property 754 @deprecated(version=0.11, explanation="Please use ``nemo.logging instead``") 755 def logger(self): 756 return nemo.logging 757 758 @property 759 def checkpoint_dir(self): 760 return self._exp_manager.ckpt_dir 761 762 @property 763 def work_dir(self): 764 return self._exp_manager.work_dir 765 766 @property 767 def global_rank(self): 768 return self._global_rank 769 [end of nemo/core/neural_factory.py] [start of nemo/core/neural_modules.py] 1 # ! /usr/bin/python 2 # -*- coding: utf-8 -*- 3 4 # Copyright (c) 2019-, NVIDIA CORPORATION. All rights reserved. 5 # 6 # Licensed under the Apache License, Version 2.0 (the "License"); 7 # you may not use this file except in compliance with the License. 8 # You may obtain a copy of the License at 9 # 10 # http://www.apache.org/licenses/LICENSE-2.0 11 # 12 # Unless required by applicable law or agreed to in writing, software 13 # distributed under the License is distributed on an "AS IS" BASIS, 14 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 15 # See the License for the specific language governing permissions and 16 # limitations under the License. 17 18 """This file contains NeuralModule and NmTensor classes.""" 19 __all__ = ['WeightShareTransform', 'NeuralModule'] 20 21 import collections 22 import uuid 23 from abc import ABC, abstractmethod 24 from collections import namedtuple 25 from enum import Enum 26 from inspect import getargvalues, getfullargspec, stack 27 from os import path 28 from typing import Dict, List, Optional, Set, Tuple 29 30 from ruamel.yaml import YAML 31 32 from .neural_types import ( 33 CanNotInferResultNeuralType, 34 NeuralPortNameMismatchError, 35 NeuralPortNmTensorMismatchError, 36 NeuralType, 37 NeuralTypeComparisonResult, 38 NmTensor, 39 ) 40 from nemo import logging 41 from nemo.core import NeuralModuleFactory 42 from nemo.package_info import __version__ as nemo_version 43 from nemo.utils.decorators.deprecated import deprecated 44 45 YAML = YAML(typ='safe') 46 47 48 class WeightShareTransform(Enum): 49 """When sharing parameters, what kind of transform to apply.""" 50 51 SAME = 0 52 TRANSPOSE = 1 53 54 55 PretrainedModelInfo = namedtuple( 56 "PretrainedModleInfo", ("pretrained_model_name", "description", "parameters", "location"), 57 ) 58 59 60 class NeuralModule(ABC): 61 """Abstract class that every Neural Module must inherit from. 62 """ 63 64 def __init__(self): 65 66 # Get default factory. 67 self._factory = NeuralModuleFactory.get_default_factory() 68 69 # Set module properties from factory else use defaults 70 self._placement = self._factory.placement 71 # If one needs to change that should override it manually. 72 73 # Optimization level. 74 self._opt_level = self._factory.optim_level 75 76 # Get object UUID. 77 self._uuid = str(uuid.uuid4()) 78 79 # Retrieve dictionary of parameters (keys, values) passed to init. 80 self._init_params = self.__extract_init_params() 81 82 # Pint the types of the values. 83 # for key, value in self._init_params.items(): 84 # print("{}: {} ({})".format(key, value, type(value))) 85 86 # Validate the parameters. 87 # self._validate_params(self._init_params) 88 89 @property 90 def init_params(self) -> Optional[Dict]: 91 """ 92 Property returning parameters used to instantiate the module. 93 94 Returns: 95 Dictionary containing parameters used to instantiate the module. 96 """ 97 return self._init_params 98 99 def __extract_init_params(self): 100 """ 101 Retrieves the dictionary of of parameters (keys, values) passed to constructor of a class derived 102 (also indirectly) from the Neural Module class. 103 104 Returns: 105 Dictionary containing parameters passed to init(). 106 """ 107 # Get names of arguments of the original module init method. 108 init_keys = getfullargspec(type(self).__init__).args 109 110 # Remove self. 111 if "self" in init_keys: 112 init_keys.remove("self") 113 114 # Create list of params. 115 init_params = {}.fromkeys(init_keys) 116 117 # Retrieve values of those params from the call list. 118 for frame in stack()[1:]: 119 localvars = getargvalues(frame[0]).locals 120 # print("localvars: ", localvars) 121 for key in init_keys: 122 # Found the variable! 123 if key in localvars.keys(): 124 # Save the value. 125 init_params[key] = localvars[key] 126 127 # Return parameters. 128 return init_params 129 130 def __validate_params(self, params): 131 """ 132 Checks whether dictionary contains parameters being primitive types (string, int, float etc.) 133 or (lists of)+ primitive types. 134 135 Args: 136 params: dictionary of parameters. 137 138 Returns: 139 True if all parameters were ok, False otherwise. 140 """ 141 ok = True 142 143 # Iterate over parameters and check them one by one. 144 for key, variable in params.items(): 145 if not self.__is_of_allowed_type(variable): 146 logging.warning( 147 "Parameter '{}' contains a variable '{}' of type '{}' which is not allowed.".format( 148 key, variable, type(variable) 149 ) 150 ) 151 ok = False 152 153 # Return the result. 154 return ok 155 156 def __is_of_allowed_type(self, var): 157 """ 158 A recursive function that checks if a given variable is of allowed type. 159 160 Args: 161 pretrained_model_name (str): name of pretrained model to use in order. 162 163 Returns: 164 True if all parameters were ok, False otherwise. 165 """ 166 # Special case: None is also allowed. 167 if var is None: 168 return True 169 170 var_type = type(var) 171 172 # If this is list - check its elements. 173 if var_type == list: 174 for list_var in var: 175 if not self.__is_of_allowed_type(list_var): 176 return False 177 178 # If this is dict - check its elements. 179 elif var_type == dict: 180 for _, dict_var in var.items(): 181 if not self.__is_of_allowed_type(dict_var): 182 return False 183 184 elif var_type not in (str, int, float, bool): 185 return False 186 187 # Well, seems that everything is ok. 188 return True 189 190 def _create_config_header(self): 191 """ A protected method that create a header stored later in the configuration file. """ 192 193 # Get module "full specification". 194 module_full_spec = str(self.__module__) + "." + str(self.__class__.__qualname__) 195 module_class_name = type(self).__name__ 196 # print(module_full_spec) 197 198 # Check whether module belongs to a collection. 199 spec_list = module_full_spec.split(".") 200 201 # Do not check Neural Modules from unit tests. 202 if spec_list[0] == "tests": 203 # Set collection variables. 204 collection_type = "tests" 205 collection_version = None 206 else: 207 # Check if component belongs to any collection 208 if len(spec_list) < 3 or (spec_list[0] != "nemo" and spec_list[1] != "collection"): 209 logging.warning( 210 "Module `{}` does not belong to any collection. This won't be allowed in the next release.".format( 211 module_class_name 212 ) 213 ) 214 collection_type = "unknown" 215 collection_version = None 216 else: 217 # Ok, set collection. 218 collection_type = spec_list[2] 219 collection_version = None 220 # TODO: to be SET! 221 # print(getattr("nemo.collections.nlp", __version__)) 222 223 # Create a "header" with module "specification". 224 header = { 225 "nemo_core_version": nemo_version, 226 "collection_type": collection_type, 227 "collection_version": collection_version, 228 # "class": module_class_name, # Operating only on full_spec now. 229 "full_spec": module_full_spec, 230 } 231 return header 232 233 def export_to_config(self, config_file): 234 """ 235 A function that exports module "configuration" (i.e. init parameters) to a YAML file. 236 Raises a ValueError exception in case then parameters coudn't be exported. 237 238 Args: 239 config_file: path (absolute or relative) and name of the config file (YML) 240 """ 241 # Check if generic export will work. 242 if not self.__validate_params(self._init_params): 243 raise ValueError( 244 "Generic configuration export enables to use of parameters of primitive types (string, int, float) " 245 F"or (lists of/dicts of) primitive types. Please implement your own custom `export_to_config()` and " 246 F"`import_from_config()` methods for your custom Module class." 247 ) 248 249 # Greate an absolute path. 250 abs_path_file = path.expanduser(config_file) 251 252 # Create the dictionary to be exported. 253 to_export = {} 254 255 # Add "header" with module "specification". 256 to_export["header"] = self._create_config_header() 257 258 # Add init parameters. 259 to_export["init_params"] = self._init_params 260 # print(to_export) 261 262 # All parameters are ok, let's export. 263 with open(abs_path_file, 'w') as outfile: 264 YAML.dump(to_export, outfile) 265 266 logging.info( 267 "Configuration of module {} ({}) exported to {}".format(self._uuid, type(self).__name__, abs_path_file) 268 ) 269 270 @classmethod 271 def _validate_config_file(cls, config_file, section_name=None): 272 """ 273 Class method validating whether the config file has a proper content (sections, specification etc.). 274 Raises an ImportError exception when config file is invalid or 275 incompatible (when called from a particular class). 276 277 Args: 278 config_file: path (absolute or relative) and name of the config file (YML) 279 280 section_name: section in the configuration file storing module configuration (optional, DEFAULT: None) 281 282 Returns: 283 A loaded configuration file (dictionary). 284 """ 285 # Greate an absolute path. 286 abs_path_file = path.expanduser(config_file) 287 288 # Open the config file. 289 with open(abs_path_file, 'r') as stream: 290 loaded_config = YAML.load(stream) 291 292 # Check section. 293 if section_name is not None: 294 if section_name not in loaded_config: 295 raise ImportError( 296 "The loaded config `{}` doesn't contain the indicated `{}` section".format( 297 config_file, section_name 298 ) 299 ) 300 # Section exists - use only it for configuration. 301 loaded_config = loaded_config[section_name] 302 303 # Make sure that the config is valid. 304 if "header" not in loaded_config: 305 raise ImportError("The loaded config `{}` doesn't contain the `header` section".format(config_file)) 306 307 if "init_params" not in loaded_config: 308 raise ImportError("The loaded config `{}` doesn't contain the `init_params` section".format(config_file)) 309 310 # Parse the "full specification". 311 spec_list = loaded_config["header"]["full_spec"].split(".") 312 313 # Check if config contains data of a compatible class. 314 if cls.__name__ != "NeuralModule" and spec_list[-1] != cls.__name__: 315 txt = "The loaded file `{}` contains configuration of ".format(config_file) 316 txt = txt + "`{}` thus cannot be used for instantiation of an object of type `{}`".format( 317 spec_list[-1], cls.__name__ 318 ) 319 raise ImportError(txt) 320 321 # Success - return configuration. 322 return loaded_config 323 324 @classmethod 325 def import_from_config(cls, config_file, section_name=None, overwrite_params={}): 326 """ 327 Class method importing the configuration file. 328 Raises an ImportError exception when config file is invalid or 329 incompatible (when called from a particular class). 330 331 Args: 332 config_file: path (absolute or relative) and name of the config file (YML) 333 334 section_name: section in the configuration file storing module configuration (optional, DEFAULT: None) 335 336 overwrite_params: Dictionary containing parameters that will be added to or overwrite (!) the default 337 parameters loaded from the configuration file 338 339 Returns: 340 Instance of the created NeuralModule object. 341 """ 342 # Validate the content of the configuration file (its header). 343 loaded_config = cls._validate_config_file(config_file, section_name) 344 345 # Parse the "full specification". 346 spec_list = loaded_config["header"]["full_spec"].split(".") 347 348 # Get object class from "full specification". 349 mod_obj = __import__(spec_list[0]) 350 for spec in spec_list[1:]: 351 mod_obj = getattr(mod_obj, spec) 352 # print(mod_obj) 353 354 # Get init parameters. 355 init_params = loaded_config["init_params"] 356 # Update parameters with additional ones. 357 init_params.update(overwrite_params) 358 359 # Create and return the object. 360 obj = mod_obj(**init_params) 361 logging.info( 362 "Instantiated a new Neural Module of type `{}` using configuration loaded from the `{}` file".format( 363 spec_list[-1], config_file 364 ) 365 ) 366 return obj 367 368 @deprecated(version=0.11) 369 @staticmethod 370 def create_ports(**kwargs): 371 """ Deprecated method, to be remoted in the next release.""" 372 raise Exception( 373 'Deprecated method. Please implement ``inputs`` and ``outputs`` \ 374 properties to define module ports instead' 375 ) 376 377 @property 378 @abstractmethod 379 def input_ports(self) -> Optional[Dict[str, NeuralType]]: 380 """Returns definitions of module input ports 381 382 Returns: 383 A (dict) of module's input ports names to NeuralTypes mapping 384 """ 385 386 @property 387 @abstractmethod 388 def output_ports(self) -> Optional[Dict[str, NeuralType]]: 389 """Returns definitions of module output ports 390 391 Returns: 392 A (dict) of module's output ports names to NeuralTypes mapping 393 """ 394 395 @property 396 def disabled_deployment_input_ports(self) -> Optional[Set[str]]: 397 """Returns names of input ports that will not be included in an export 398 399 Returns: 400 A (set) of module's input port names that are not exportable 401 """ 402 return set([]) 403 404 @property 405 def disabled_deployment_output_ports(self) -> Optional[Set[str]]: 406 """Returns names of output ports that will not be included in an export 407 408 Returns: 409 A (set) of module's output port names that are not exportable 410 """ 411 return set([]) 412 413 def prepare_for_deployment(self) -> None: 414 """Patch the module if required to prepare for deployment 415 416 """ 417 return 418 419 @staticmethod 420 def pretrained_storage(): 421 return '' 422 423 def __call__(self, **kwargs): 424 """This method allows objects to be called with their port names 425 426 Args: 427 kwargs: Input ports and their values. For example: 428 ... 429 mymodule1 = Subclass1_of_NeuralModule(...) 430 mymodule2 = Subclass2_of_NeuralModule(...) 431 ... 432 out_port1, out_port2 = mymodule1(input_port1=value1, 433 input_port2=value2, 434 input_port3=value3) 435 out_port11 = mymodule2(input_port1=out_port2) 436 ... 437 438 Returns: 439 NmTensor object or tuple of NmTensor objects 440 """ 441 # Get input and output ports definitions. 442 input_port_defs = self.input_ports 443 output_port_defs = self.output_ports 444 445 first_input_nmtensor_type = None 446 input_nmtensors_are_of_same_type = True 447 for port_name, tgv in kwargs.items(): 448 # make sure that passed arguments correspond to input port names 449 if port_name not in input_port_defs.keys(): 450 raise NeuralPortNameMismatchError("Wrong input port name: {0}".format(port_name)) 451 452 input_port = input_port_defs[port_name] 453 type_comatibility = input_port.compare(tgv) 454 if ( 455 type_comatibility != NeuralTypeComparisonResult.SAME 456 and type_comatibility != NeuralTypeComparisonResult.GREATER 457 ): 458 raise NeuralPortNmTensorMismatchError( 459 "\n\nIn {0}. \n" 460 "Port: {1} and a NmTensor it was fed are \n" 461 "of incompatible neural types:\n\n{2} \n\n and \n\n{3}" 462 "\n\nType comparison result: {4}".format( 463 self.__class__.__name__, port_name, input_port_defs[port_name], tgv, type_comatibility, 464 ) 465 ) 466 467 # if first_input_nmtensor_type is None: 468 # first_input_nmtensor_type = NeuralType(tgv._axis2type) 469 # else: 470 # if first_input_nmtensor_type._axis2type is None: 471 # input_nmtensors_are_of_same_type = True 472 # else: 473 # input_nmtensors_are_of_same_type = first_input_nmtensor_type.compare( 474 # tgv 475 # ) == NeuralTypeComparisonResult.SAME and len(first_input_nmtensor_type._axis2type) 476 # if not ( 477 # type_comatibility == NeuralTypeComparisonResult.SAME 478 # or type_comatibility == NeuralTypeComparisonResult.GREATER 479 # ): 480 # raise NeuralPortNmTensorMismatchError( 481 # "\n\nIn {0}. \n" 482 # "Port: {1} and a NmTensor it was fed are \n" 483 # "of incompatible neural types:\n\n{2} \n\n and \n\n{3}" 484 # "\n\nType comparison result: {4}".format( 485 # self.__class__.__name__, port_name, input_port_defs[port_name], tgv, type_comatibility, 486 # ) 487 # ) 488 # if type_comatibility == NeuralTypeComparisonResult.LESS: 489 # print('Types were raised') 490 491 if len(output_port_defs) == 1: 492 out_name = list(output_port_defs)[0] 493 out_type = output_port_defs[out_name] 494 if out_type is None: 495 if input_nmtensors_are_of_same_type: 496 out_type = first_input_nmtensor_type 497 else: 498 raise CanNotInferResultNeuralType( 499 "Can't infer output neural type. Likely your inputs are of different type." 500 ) 501 return NmTensor(producer=self, producer_args=kwargs, name=out_name, ntype=out_type,) 502 else: 503 result = [] 504 for out_port, n_type in output_port_defs.items(): 505 out_type = n_type 506 if out_type is None: 507 if input_nmtensors_are_of_same_type: 508 out_type = first_input_nmtensor_type 509 else: 510 raise CanNotInferResultNeuralType( 511 "Can't infer output neural type. Likely your inputs are of different type." 512 ) 513 result.append(NmTensor(producer=self, producer_args=kwargs, name=out_port, ntype=out_type,)) 514 515 # Creating ad-hoc class for returning from module's forward pass. 516 output_class_name = f'{self.__class__.__name__}Output' 517 field_names = list(output_port_defs) 518 result_type = collections.namedtuple(typename=output_class_name, field_names=field_names,) 519 520 # Tie tuple of output tensors with corresponding names. 521 result = result_type(*result) 522 523 return result 524 525 def __str__(self): 526 return self.__class__.__name__ 527 528 @abstractmethod 529 def get_weights(self) -> Optional[Dict[(str, bool)]]: 530 """Returns NeuralModule's weights copy. 531 532 Returns: 533 Dictionary of name -> (weights, trainable)""" 534 pass 535 536 @abstractmethod 537 def set_weights( 538 self, 539 name2weight: Dict[(str, Tuple[str, bool])], 540 name2name_and_transform: Dict[(str, Tuple[str, WeightShareTransform])] = None, 541 ): 542 """Sets weight from given values. For every named weight in 543 name2weight, 544 if weight with the same name is found in the model, it will be set to 545 found value. 546 547 WARNING: This will NOT tie weights. It will copy values. 548 549 If ``name2name_and_transform`` is provided then if will set weights 550 using 551 name mapping and transform. For example, suppose ``objec1.X = 3x5 552 weight``. 553 Then, if ``name2name_and_transform['X']=('Y', 554 WeightShareTransform.TRANSPOSE)`` 555 and ``Y`` is 5x3 weight and ``name2weight['Y']=Y. Then: 556 ``object1.set_weights(name2weight, name2name_and_transform)`` will 557 set object1.X=transpose(Y). 558 559 Args: 560 name2weight (dict): dictionary of name to (weight, trainable). 561 Typically this is output of get_weights method. 562 name2name_and_transform: mapping from name -> (name, transform) 563 """ 564 pass 565 566 @staticmethod 567 def list_pretrained_models() -> Optional[List[PretrainedModelInfo]]: 568 """List all available pre-trained models (e.g. weights) for this NM. 569 570 Returns: 571 A list of PretrainedModelInfo tuples. 572 The pretrained_model_name field of the tuple can be used to 573 retrieve pre-trained model's weights (pass it as 574 pretrained_model_name argument to the module's constructor) 575 """ 576 return None 577 578 def get_config_dict_and_checkpoint(self, pretrained_model_name): 579 """WARNING: This part is work in progress""" 580 return None 581 582 @abstractmethod 583 def tie_weights_with( 584 self, 585 module, 586 weight_names=List[str], 587 name2name_and_transform: Dict[(str, Tuple[str, WeightShareTransform])] = None, 588 ): 589 """Ties weights between self and module. For every weight name in 590 weight_names, if weight with the same name is found in self, it will 591 be tied 592 with a same weight from ``module``. 593 594 WARNING: Once weights are tied, updates to one weights's weights 595 will affect 596 other module's weights. 597 598 599 If ``name2name_and_transform`` is provided then if will set weights 600 using 601 name mapping and transform. For example, suppose ``objec1.X = 3x5 602 weights`` 603 and ``object2.Y = 5x3 weights``. Then these weights can be tied like 604 this: 605 606 .. code-block:: python 607 608 object1.tie_weights_with(object2, weight_names=['X'], 609 name2name_and_transform = 610 { 'X': ('Y', WeightShareTransform.TRANSPOSE)}) 611 612 613 Args: 614 module: with which module to tie weights 615 weight_names (List[str]): list of self weights' names 616 name2name_and_transform: mapping from name -> (name, transform) 617 """ 618 pass 619 620 def is_trainable(self) -> bool: 621 """ 622 Checks if NeuralModule is trainable. 623 A NeuralModule is trainable IFF it contains at least one trainable 624 weight 625 626 Returns: 627 True if module has trainable weights, False otherwise 628 """ 629 weights = self.get_weights() 630 if weights is None: 631 return False 632 for name, w in weights.items(): 633 if w[1]: 634 return True 635 return False 636 637 @abstractmethod 638 def save_to(self, path: str): 639 """Save module state to file. 640 641 Args: 642 path (string): path to while where to save. 643 """ 644 pass 645 646 @abstractmethod 647 def restore_from(self, path: str): 648 """Restore module's state from file. 649 650 Args: 651 path (string): path to where to restore from. 652 """ 653 pass 654 655 @abstractmethod 656 def freeze(self, weights: Set[str] = None): 657 """Freeze weights 658 659 Args: 660 weights (set): set of weight names to freeze 661 If None, all weights are freezed. 662 """ 663 pass 664 665 @abstractmethod 666 def unfreeze(self, weights: Set[str] = None): 667 """Unfreeze weights 668 669 Args: 670 weights (set): set of weight names to unfreeze 671 If None, all weights are unfreezed. 672 """ 673 pass 674 675 @property 676 def placement(self): 677 """Module's placement. Currently CPU or GPU. 678 DataParallel and ModelParallel will come later. 679 680 Returns: 681 (DeviceType) Device where NM's weights are located 682 """ 683 return self._placement 684 685 @property 686 @deprecated(version=0.11) 687 def local_parameters(self) -> Optional[Dict]: 688 """Get module's parameters 689 690 Returns: 691 module's parameters 692 """ 693 return self._init_params 694 # return self._local_parameters 695 696 @property 697 def unique_instance_id(self): 698 """A unique instance id for this object 699 700 Returns: 701 A uniq uuid which can be used to identify this object 702 """ 703 return self._uuid 704 705 @property 706 def factory(self): 707 """ Neural module factory which created this module 708 Returns: NeuralModuleFactory instance or None 709 """ 710 return self._factory 711 712 @property 713 @abstractmethod 714 def num_weights(self): 715 """Number of module's weights 716 """ 717 pass 718 [end of nemo/core/neural_modules.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
ba4616f1f011d599de87f0cb3315605e715d402a
Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ```
2020-03-10T03:03:23Z
<patch> <patch> diff --git a/nemo/backends/pytorch/actions.py b/nemo/backends/pytorch/actions.py --- a/nemo/backends/pytorch/actions.py +++ b/nemo/backends/pytorch/actions.py @@ -937,26 +937,16 @@ def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defa if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: dynamic_axes[port_name].append(ind) - # This is a hack for Jasper to Jarvis export -- need re-design for this - inputs_to_drop = set() - outputs_to_drop = set() - if type(module).__name__ == "JasperEncoder": - logging.info( - "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " - "deployment" - ) - inputs_to_drop.add("length") - outputs_to_drop.add("encoded_lengths") - + # extract dynamic axes and remove unnecessary inputs/outputs # for input_ports for port_name, ntype in module.input_ports.items(): - if port_name in inputs_to_drop: + if port_name in module._disabled_deployment_input_ports: input_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) # for output_ports for port_name, ntype in module.output_ports.items(): - if port_name in outputs_to_drop: + if port_name in module._disabled_deployment_output_ports: output_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) diff --git a/nemo/collections/asr/jasper.py b/nemo/collections/asr/jasper.py --- a/nemo/collections/asr/jasper.py +++ b/nemo/collections/asr/jasper.py @@ -118,14 +118,14 @@ def output_ports(self): } @property - def disabled_deployment_input_ports(self): + def _disabled_deployment_input_ports(self): return set(["length"]) @property - def disabled_deployment_output_ports(self): + def _disabled_deployment_output_ports(self): return set(["encoded_lengths"]) - def prepare_for_deployment(self): + def _prepare_for_deployment(self): m_count = 0 for m in self.modules(): if type(m).__name__ == "MaskedConv1d": diff --git a/nemo/core/neural_factory.py b/nemo/core/neural_factory.py --- a/nemo/core/neural_factory.py +++ b/nemo/core/neural_factory.py @@ -610,7 +610,7 @@ def deployment_export( input_example: sometimes tracing will require input examples output_example: Should match inference on input_example """ - module.prepare_for_deployment() + module._prepare_for_deployment() return self._trainer.deployment_export( module=module, diff --git a/nemo/core/neural_modules.py b/nemo/core/neural_modules.py --- a/nemo/core/neural_modules.py +++ b/nemo/core/neural_modules.py @@ -393,7 +393,7 @@ def output_ports(self) -> Optional[Dict[str, NeuralType]]: """ @property - def disabled_deployment_input_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_input_ports(self) -> Optional[Set[str]]: """Returns names of input ports that will not be included in an export Returns: @@ -402,7 +402,7 @@ def disabled_deployment_input_ports(self) -> Optional[Set[str]]: return set([]) @property - def disabled_deployment_output_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_output_ports(self) -> Optional[Set[str]]: """Returns names of output ports that will not be included in an export Returns: @@ -410,7 +410,7 @@ def disabled_deployment_output_ports(self) -> Optional[Set[str]]: """ return set([]) - def prepare_for_deployment(self) -> None: + def _prepare_for_deployment(self) -> None: """Patch the module if required to prepare for deployment """ </patch> </s> </patch>
diff --git a/tests/unit/core/test_deploy_export.py b/tests/unit/core/test_deploy_export.py --- a/tests/unit/core/test_deploy_export.py +++ b/tests/unit/core/test_deploy_export.py @@ -46,9 +46,11 @@ import nemo.collections.nlp.nm.trainables.common.token_classification_nm from nemo import logging +TRT_ONNX_DISABLED = False + # Check if the required libraries and runtimes are installed. +# Only initialize GPU after this runner is activated. try: - # Only initialize GPU after this runner is activated. import pycuda.autoinit # This import causes pycuda to automatically manage CUDA context creation and cleanup. @@ -63,16 +65,17 @@ ) from .tensorrt_runner import TensorRTRunnerV2 except: - # Skip tests. - pytestmark = pytest.mark.skip + TRT_ONNX_DISABLED = True @pytest.mark.usefixtures("neural_factory") class TestDeployExport(TestCase): - def setUp(self): - logging.setLevel(logging.WARNING) - device = nemo.core.DeviceType.GPU - self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) + # def setUp(self): + # super().setUp() + + # logging.setLevel(logging.WARNING) + # device = nemo.core.DeviceType.GPU + # self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) def __test_export_route(self, module, out_name, mode, input_example=None): out = Path(out_name) @@ -112,7 +115,13 @@ def __test_export_route(self, module, out_name, mode, input_example=None): loader_cache = DataLoaderCache(data_loader) profile_shapes = OrderedDict() names = list(module.input_ports) + list(module.output_ports) - + names = list( + filter( + lambda x: x + not in (module._disabled_deployment_input_ports | module._disabled_deployment_output_ports), + names, + ) + ) if isinstance(input_example, tuple): si = [tuple(input_example[i].shape) for i in range(len(input_example))] elif isinstance(input_example, OrderedDict): @@ -152,7 +161,7 @@ def __test_export_route(self, module, out_name, mode, input_example=None): input_names = list(input_metadata.keys()) for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: + if input_name in module._disabled_deployment_input_ports: continue inputs[input_name] = ( input_example[input_name].cpu().numpy() @@ -209,8 +218,8 @@ def __test_export_route(self, module, out_name, mode, input_example=None): ort_inputs = ort_session.get_inputs() for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: - input_name = ort_inputs[i].name + if input_name in module._disabled_deployment_input_ports: + continue inputs[input_name] = ( input_example[input_name].cpu().numpy() if isinstance(input_example, OrderedDict) @@ -263,9 +272,10 @@ def __test_export_route(self, module, out_name, mode, input_example=None): def __test_export_route_all(self, module, out_name, input_example=None): if input_example is not None: - self.__test_export_route( - module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example - ) + if not TRT_ONNX_DISABLED: + self.__test_export_route( + module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example + ) self.__test_export_route(module, out_name + '.onnx', nemo.core.DeploymentFormat.ONNX, input_example) self.__test_export_route(module, out_name + '.pt', nemo.core.DeploymentFormat.PYTORCH, input_example) self.__test_export_route(module, out_name + '.ts', nemo.core.DeploymentFormat.TORCHSCRIPT, input_example) @@ -323,9 +333,7 @@ def test_jasper_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="jasper_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randn(256).cuda()), + module=jasper_encoder, out_name="jasper_encoder", input_example=torch.randn(16, 64, 256).cuda(), ) @pytest.mark.unit @@ -343,7 +351,5 @@ def test_quartz_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="quartz_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randint(20, (16,)).cuda()), + module=jasper_encoder, out_name="quartz_encoder", input_example=torch.randn(16, 64, 256).cuda(), )
1.0
NVIDIA__NeMo-3632
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> ./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`. </issue> <code> [start of README.rst] 1 2 |status| |documentation| |license| |lgtm_grade| |lgtm_alerts| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |lgtm_grade| image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 17 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python 18 :alt: Language grade: Python 19 20 .. |lgtm_alerts| image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 21 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ 22 :alt: Total alerts 23 24 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 25 :target: https://github.com/psf/black 26 :alt: Code style: black 27 28 .. _main-readme: 29 30 **NVIDIA NeMo** 31 =============== 32 33 Introduction 34 ------------ 35 36 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), natural language processing (NLP), and text-to-speech synthesis (TTS). 37 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 38 39 `Pre-trained NeMo models. <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_ 40 41 `Introductory video. <https://www.youtube.com/embed/wBgpMf_KQVw>`_ 42 43 Key Features 44 ------------ 45 46 * Speech processing 47 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 48 * Supported models: Jasper, QuartzNet, CitriNet, Conformer-CTC, Conformer-Transducer, ContextNet, ... 49 * Supports CTC and Transducer/RNNT losses/decoders 50 * Beam Search decoding 51 * `Language Modelling for ASR <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 52 * Streaming and Buffered ASR (CTC/Transdcer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/main/examples/asr/asr_chunked_inference>`_ 53 * `Speech Classification and Speech Command Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition) 54 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 55 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 56 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 57 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 58 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 59 * Natural Language Processing 60 * `Compatible with Hugging Face Transformers and NVIDIA Megatron <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html>`_ 61 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation.html>`_ 62 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 63 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 64 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 65 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 66 * `BERT pre-training <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/bert_pretraining.html>`_ 67 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 68 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 69 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 70 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 71 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 72 * `Neural Duplex Text Normalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization.html>`_ 73 * `Prompt Tuning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html#prompt-tuning>`_ 74 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 75 * `Speech synthesis (TTS) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 76 * Spectrogram generation: Tacotron2, GlowTTS, TalkNet, FastPitch, FastSpeech2, Mixer-TTS, Mixer-TTS-X 77 * Vocoders: WaveGlow, SqueezeWave, UniGlow, MelGAN, HiFiGAN, UnivNet 78 * End-to-end speech generation: FastPitch_HifiGan_E2E, FastSpeech2_HifiGan_E2E 79 * `NGC collection of pre-trained TTS models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 80 * `Tools <https://github.com/NVIDIA/NeMo/tree/main/tools>`_ 81 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/text_processing_deployment.html>`_ 82 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 83 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 84 85 86 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 87 88 Requirements 89 ------------ 90 91 1) Python 3.6, 3.7 or 3.8 92 2) Pytorch 1.10.0 or above 93 3) NVIDIA GPU for training 94 95 Documentation 96 ------------- 97 98 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 99 :alt: Documentation Status 100 :scale: 100% 101 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 102 103 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 104 :alt: Documentation Status 105 :scale: 100% 106 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 107 108 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 109 | Version | Status | Description | 110 +=========+=============+==========================================================================================================================================+ 111 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 112 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 113 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 114 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 115 116 Tutorials 117 --------- 118 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 119 120 Getting help with NeMo 121 ---------------------- 122 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 123 124 125 Installation 126 ------------ 127 128 Pip 129 ~~~ 130 Use this installation mode if you want the latest released version. 131 132 .. code-block:: bash 133 134 apt-get update && apt-get install -y libsndfile1 ffmpeg 135 pip install Cython 136 pip install nemo_toolkit['all'] 137 138 .. note:: 139 140 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 141 142 Pip from source 143 ~~~~~~~~~~~~~~~ 144 Use this installation mode if you want the a version from particular GitHub branch (e.g main). 145 146 .. code-block:: bash 147 148 apt-get update && apt-get install -y libsndfile1 ffmpeg 149 pip install Cython 150 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 151 152 153 From source 154 ~~~~~~~~~~~ 155 Use this installation mode if you are contributing to NeMo. 156 157 .. code-block:: bash 158 159 apt-get update && apt-get install -y libsndfile1 ffmpeg 160 git clone https://github.com/NVIDIA/NeMo 161 cd NeMo 162 ./reinstall.sh 163 164 .. note:: 165 166 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 167 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 168 169 RNNT 170 ~~~~ 171 Note that RNNT requires numba to be installed from conda. 172 173 .. code-block:: bash 174 175 conda remove numba 176 pip uninstall numba 177 conda install -c conda-forge numba 178 179 Megatron GPT 180 ~~~~~~~~~~~~ 181 Megatron GPT training requires NVIDIA Apex to be installed. 182 183 .. code-block:: bash 184 185 git clone https://github.com/NVIDIA/apex 186 cd apex 187 git checkout c8bcc98176ad8c3a0717082600c70c907891f9cb 188 pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" ./ 189 190 Docker containers: 191 ~~~~~~~~~~~~~~~~~~ 192 To build a nemo container with Dockerfile from a branch, please run 193 194 .. code-block:: bash 195 196 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 197 198 199 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 22.01-py3 and then installing from GitHub. 200 201 .. code-block:: bash 202 203 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 204 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 205 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:22.01-py3 206 207 Examples 208 -------- 209 210 Many examples can be found under `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 211 212 213 Contributing 214 ------------ 215 216 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 217 218 Publications 219 ------------ 220 221 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/blob/main/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 222 223 Citation 224 -------- 225 226 .. code-block:: bash 227 228 @article{kuchaiev2019nemo, 229 title={Nemo: a toolkit for building ai applications using neural modules}, 230 author={Kuchaiev, Oleksii and Li, Jason and Nguyen, Huyen and Hrinchuk, Oleksii and Leary, Ryan and Ginsburg, Boris and Kriman, Samuel and Beliaev, Stanislav and Lavrukhin, Vitaly and Cook, Jack and others}, 231 journal={arXiv preprint arXiv:1909.09577}, 232 year={2019} 233 } 234 235 License 236 ------- 237 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 238 [end of README.rst] [start of /dev/null] 1 [end of /dev/null] [start of nemo_text_processing/text_normalization/__init__.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from nemo.utils import logging 16 17 try: 18 import pynini 19 20 PYNINI_AVAILABLE = True 21 except (ModuleNotFoundError, ImportError): 22 logging.warning( 23 "`pynini` is not installed ! \n" 24 "Please run the `nemo_text_processing/setup.sh` script" 25 "prior to usage of this toolkit." 26 ) 27 28 PYNINI_AVAILABLE = False 29 [end of nemo_text_processing/text_normalization/__init__.py] [start of nemo_text_processing/text_normalization/en/graph_utils.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 import os 17 import string 18 from pathlib import Path 19 from typing import Dict 20 21 from nemo_text_processing.text_normalization.en.utils import get_abs_path 22 23 try: 24 import pynini 25 from pynini import Far 26 from pynini.export import export 27 from pynini.examples import plurals 28 from pynini.lib import byte, pynutil, utf8 29 30 NEMO_CHAR = utf8.VALID_UTF8_CHAR 31 32 NEMO_DIGIT = byte.DIGIT 33 NEMO_LOWER = pynini.union(*string.ascii_lowercase).optimize() 34 NEMO_UPPER = pynini.union(*string.ascii_uppercase).optimize() 35 NEMO_ALPHA = pynini.union(NEMO_LOWER, NEMO_UPPER).optimize() 36 NEMO_ALNUM = pynini.union(NEMO_DIGIT, NEMO_ALPHA).optimize() 37 NEMO_HEX = pynini.union(*string.hexdigits).optimize() 38 NEMO_NON_BREAKING_SPACE = u"\u00A0" 39 NEMO_SPACE = " " 40 NEMO_WHITE_SPACE = pynini.union(" ", "\t", "\n", "\r", u"\u00A0").optimize() 41 NEMO_NOT_SPACE = pynini.difference(NEMO_CHAR, NEMO_WHITE_SPACE).optimize() 42 NEMO_NOT_QUOTE = pynini.difference(NEMO_CHAR, r'"').optimize() 43 44 NEMO_PUNCT = pynini.union(*map(pynini.escape, string.punctuation)).optimize() 45 NEMO_GRAPH = pynini.union(NEMO_ALNUM, NEMO_PUNCT).optimize() 46 47 NEMO_SIGMA = pynini.closure(NEMO_CHAR) 48 49 delete_space = pynutil.delete(pynini.closure(NEMO_WHITE_SPACE)) 50 insert_space = pynutil.insert(" ") 51 delete_extra_space = pynini.cross(pynini.closure(NEMO_WHITE_SPACE, 1), " ") 52 delete_preserve_order = pynini.closure( 53 pynutil.delete(" preserve_order: true") 54 | (pynutil.delete(" field_order: \"") + NEMO_NOT_QUOTE + pynutil.delete("\"")) 55 ) 56 57 suppletive = pynini.string_file(get_abs_path("data/suppletive.tsv")) 58 # _v = pynini.union("a", "e", "i", "o", "u") 59 _c = pynini.union( 60 "b", "c", "d", "f", "g", "h", "j", "k", "l", "m", "n", "p", "q", "r", "s", "t", "v", "w", "x", "y", "z" 61 ) 62 _ies = NEMO_SIGMA + _c + pynini.cross("y", "ies") 63 _es = NEMO_SIGMA + pynini.union("s", "sh", "ch", "x", "z") + pynutil.insert("es") 64 _s = NEMO_SIGMA + pynutil.insert("s") 65 66 graph_plural = plurals._priority_union( 67 suppletive, plurals._priority_union(_ies, plurals._priority_union(_es, _s, NEMO_SIGMA), NEMO_SIGMA), NEMO_SIGMA 68 ).optimize() 69 70 SINGULAR_TO_PLURAL = graph_plural 71 PLURAL_TO_SINGULAR = pynini.invert(graph_plural) 72 TO_LOWER = pynini.union(*[pynini.cross(x, y) for x, y in zip(string.ascii_uppercase, string.ascii_lowercase)]) 73 TO_UPPER = pynini.invert(TO_LOWER) 74 75 PYNINI_AVAILABLE = True 76 except (ModuleNotFoundError, ImportError): 77 # Create placeholders 78 NEMO_CHAR = None 79 80 NEMO_DIGIT = None 81 NEMO_LOWER = None 82 NEMO_UPPER = None 83 NEMO_ALPHA = None 84 NEMO_ALNUM = None 85 NEMO_HEX = None 86 NEMO_NON_BREAKING_SPACE = u"\u00A0" 87 NEMO_SPACE = " " 88 NEMO_WHITE_SPACE = None 89 NEMO_NOT_SPACE = None 90 NEMO_NOT_QUOTE = None 91 92 NEMO_PUNCT = None 93 NEMO_GRAPH = None 94 95 NEMO_SIGMA = None 96 97 delete_space = None 98 insert_space = None 99 delete_extra_space = None 100 delete_preserve_order = None 101 102 suppletive = None 103 # _v = pynini.union("a", "e", "i", "o", "u") 104 _c = None 105 _ies = None 106 _es = None 107 _s = None 108 109 graph_plural = None 110 111 SINGULAR_TO_PLURAL = None 112 PLURAL_TO_SINGULAR = None 113 TO_LOWER = None 114 TO_UPPER = None 115 116 PYNINI_AVAILABLE = False 117 118 119 def generator_main(file_name: str, graphs: Dict[str, 'pynini.FstLike']): 120 """ 121 Exports graph as OpenFst finite state archive (FAR) file with given file name and rule name. 122 123 Args: 124 file_name: exported file name 125 graphs: Mapping of a rule name and Pynini WFST graph to be exported 126 """ 127 exporter = export.Exporter(file_name) 128 for rule, graph in graphs.items(): 129 exporter[rule] = graph.optimize() 130 exporter.close() 131 print(f'Created {file_name}') 132 133 134 def get_plurals(fst): 135 """ 136 Given singular returns plurals 137 138 Args: 139 fst: Fst 140 141 Returns plurals to given singular forms 142 """ 143 return SINGULAR_TO_PLURAL @ fst 144 145 146 def get_singulars(fst): 147 """ 148 Given plural returns singulars 149 150 Args: 151 fst: Fst 152 153 Returns singulars to given plural forms 154 """ 155 return PLURAL_TO_SINGULAR @ fst 156 157 158 def convert_space(fst) -> 'pynini.FstLike': 159 """ 160 Converts space to nonbreaking space. 161 Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" 162 This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. 163 164 Args: 165 fst: input fst 166 167 Returns output fst where breaking spaces are converted to non breaking spaces 168 """ 169 return fst @ pynini.cdrewrite(pynini.cross(NEMO_SPACE, NEMO_NON_BREAKING_SPACE), "", "", NEMO_SIGMA) 170 171 172 class GraphFst: 173 """ 174 Base class for all grammar fsts. 175 176 Args: 177 name: name of grammar class 178 kind: either 'classify' or 'verbalize' 179 deterministic: if True will provide a single transduction option, 180 for False multiple transduction are generated (used for audio-based normalization) 181 """ 182 183 def __init__(self, name: str, kind: str, deterministic: bool = True): 184 self.name = name 185 self.kind = str 186 self._fst = None 187 self.deterministic = deterministic 188 189 self.far_path = Path(os.path.dirname(__file__) + '/grammars/' + kind + '/' + name + '.far') 190 if self.far_exist(): 191 self._fst = Far(self.far_path, mode="r", arc_type="standard", far_type="default").get_fst() 192 193 def far_exist(self) -> bool: 194 """ 195 Returns true if FAR can be loaded 196 """ 197 return self.far_path.exists() 198 199 @property 200 def fst(self) -> 'pynini.FstLike': 201 return self._fst 202 203 @fst.setter 204 def fst(self, fst): 205 self._fst = fst 206 207 def add_tokens(self, fst) -> 'pynini.FstLike': 208 """ 209 Wraps class name around to given fst 210 211 Args: 212 fst: input fst 213 214 Returns: 215 Fst: fst 216 """ 217 return pynutil.insert(f"{self.name} {{ ") + fst + pynutil.insert(" }") 218 219 def delete_tokens(self, fst) -> 'pynini.FstLike': 220 """ 221 Deletes class name wrap around output of given fst 222 223 Args: 224 fst: input fst 225 226 Returns: 227 Fst: fst 228 """ 229 res = ( 230 pynutil.delete(f"{self.name}") 231 + delete_space 232 + pynutil.delete("{") 233 + delete_space 234 + fst 235 + delete_space 236 + pynutil.delete("}") 237 ) 238 return res @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) 239 [end of nemo_text_processing/text_normalization/en/graph_utils.py] [start of nemo_text_processing/text_normalization/en/taggers/punctuation.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 import sys 17 from unicodedata import category 18 19 from nemo_text_processing.text_normalization.en.graph_utils import GraphFst 20 21 try: 22 import pynini 23 from pynini.lib import pynutil 24 25 PYNINI_AVAILABLE = False 26 except (ModuleNotFoundError, ImportError): 27 PYNINI_AVAILABLE = False 28 29 30 class PunctuationFst(GraphFst): 31 """ 32 Finite state transducer for classifying punctuation 33 e.g. a, -> tokens { name: "a" } tokens { name: "," } 34 35 Args: 36 deterministic: if True will provide a single transduction option, 37 for False multiple transduction are generated (used for audio-based normalization) 38 39 """ 40 41 def __init__(self, deterministic: bool = True): 42 super().__init__(name="punctuation", kind="classify", deterministic=deterministic) 43 44 s = "!#%&\'()*+,-./:;<=>?@^_`{|}~\"" 45 46 punct_unicode = [chr(i) for i in range(sys.maxunicode) if category(chr(i)).startswith("P")] 47 punct_unicode.remove('[') 48 punct_unicode.remove(']') 49 punct = pynini.union(*s) | pynini.union(*punct_unicode) 50 51 self.graph = punct 52 self.fst = (pynutil.insert("name: \"") + self.graph + pynutil.insert("\"")).optimize() 53 [end of nemo_text_processing/text_normalization/en/taggers/punctuation.py] [start of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 17 from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space 18 19 try: 20 import pynini 21 from pynini.lib import pynutil 22 23 PYNINI_AVAILABLE = True 24 except (ModuleNotFoundError, ImportError): 25 PYNINI_AVAILABLE = False 26 27 28 class WhiteListFst(GraphFst): 29 """ 30 Finite state transducer for verbalizing whitelist 31 e.g. tokens { name: "misses" } } -> misses 32 33 Args: 34 deterministic: if True will provide a single transduction option, 35 for False multiple transduction are generated (used for audio-based normalization) 36 """ 37 38 def __init__(self, deterministic: bool = True): 39 super().__init__(name="whitelist", kind="verbalize", deterministic=deterministic) 40 graph = ( 41 pynutil.delete("name:") 42 + delete_space 43 + pynutil.delete("\"") 44 + pynini.closure(NEMO_CHAR - " ", 1) 45 + pynutil.delete("\"") 46 ) 47 graph = graph @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) 48 self.fst = graph.optimize() 49 [end of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] [start of nemo_text_processing/text_normalization/en/verbalizers/word.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space 17 18 try: 19 import pynini 20 from pynini.lib import pynutil 21 22 PYNINI_AVAILABLE = True 23 except (ModuleNotFoundError, ImportError): 24 PYNINI_AVAILABLE = False 25 26 27 class WordFst(GraphFst): 28 """ 29 Finite state transducer for verbalizing word 30 e.g. tokens { name: "sleep" } -> sleep 31 32 Args: 33 deterministic: if True will provide a single transduction option, 34 for False multiple transduction are generated (used for audio-based normalization) 35 """ 36 37 def __init__(self, deterministic: bool = True): 38 super().__init__(name="word", kind="verbalize", deterministic=deterministic) 39 chars = pynini.closure(NEMO_CHAR - " ", 1) 40 char = pynutil.delete("name:") + delete_space + pynutil.delete("\"") + chars + pynutil.delete("\"") 41 graph = char @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) 42 43 self.fst = graph.optimize() 44 [end of nemo_text_processing/text_normalization/en/verbalizers/word.py] [start of nemo_text_processing/text_normalization/normalize.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import itertools 16 import os 17 import re 18 from argparse import ArgumentParser 19 from collections import OrderedDict 20 from math import factorial 21 from typing import Dict, List, Union 22 23 from nemo_text_processing.text_normalization.data_loader_utils import get_installation_msg, pre_process 24 from nemo_text_processing.text_normalization.token_parser import PRESERVE_ORDER_KEY, TokenParser 25 from tqdm import tqdm 26 27 try: 28 import pynini 29 30 PYNINI_AVAILABLE = True 31 32 except (ModuleNotFoundError, ImportError): 33 PYNINI_AVAILABLE = False 34 35 try: 36 from nemo.collections.common.tokenizers.moses_tokenizers import MosesProcessor 37 from nemo.collections.nlp.data.text_normalization.utils import post_process_punct 38 39 NLP_AVAILABLE = True 40 except (ModuleNotFoundError, ImportError): 41 NLP_AVAILABLE = False 42 43 44 SPACE_DUP = re.compile(' {2,}') 45 46 47 class Normalizer: 48 """ 49 Normalizer class that converts text from written to spoken form. 50 Useful for TTS preprocessing. 51 52 Args: 53 input_case: expected input capitalization 54 lang: language specifying the TN rules, by default: English 55 cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. 56 overwrite_cache: set to True to overwrite .far files 57 whitelist: path to a file with whitelist replacements 58 """ 59 60 def __init__( 61 self, 62 input_case: str, 63 lang: str = 'en', 64 deterministic: bool = True, 65 cache_dir: str = None, 66 overwrite_cache: bool = False, 67 whitelist: str = None, 68 ): 69 assert input_case in ["lower_cased", "cased"] 70 71 if not PYNINI_AVAILABLE: 72 raise ImportError(get_installation_msg()) 73 74 if lang == 'en' and deterministic: 75 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify import ClassifyFst 76 from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst 77 elif lang == 'en' and not deterministic: 78 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify_with_audio import ClassifyFst 79 from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst 80 elif lang == 'ru': 81 # Ru TN only support non-deterministic cases and produces multiple normalization options 82 # use normalize_with_audio.py 83 from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst 84 from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst 85 elif lang == 'de': 86 # Ru TN only support non-deterministic cases and produces multiple normalization options 87 # use normalize_with_audio.py 88 from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst 89 from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst 90 self.tagger = ClassifyFst( 91 input_case=input_case, 92 deterministic=deterministic, 93 cache_dir=cache_dir, 94 overwrite_cache=overwrite_cache, 95 whitelist=whitelist, 96 ) 97 self.verbalizer = VerbalizeFinalFst(deterministic=deterministic) 98 self.parser = TokenParser() 99 self.lang = lang 100 101 if NLP_AVAILABLE: 102 self.processor = MosesProcessor(lang_id=lang) 103 else: 104 self.processor = None 105 print("NeMo NLP is not available. Moses de-tokenization will be skipped.") 106 107 def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: 108 """ 109 NeMo text normalizer 110 111 Args: 112 texts: list of input strings 113 verbose: whether to print intermediate meta information 114 115 Returns converted list input strings 116 """ 117 res = [] 118 for input in tqdm(texts): 119 try: 120 text = self.normalize(input, verbose=verbose, punct_post_process=punct_post_process) 121 except: 122 print(input) 123 raise Exception 124 res.append(text) 125 return res 126 127 def _estimate_number_of_permutations_in_nested_dict( 128 self, token_group: Dict[str, Union[OrderedDict, str, bool]] 129 ) -> int: 130 num_perms = 1 131 for k, inner in token_group.items(): 132 if isinstance(inner, dict): 133 num_perms *= self._estimate_number_of_permutations_in_nested_dict(inner) 134 num_perms *= factorial(len(token_group)) 135 return num_perms 136 137 def _split_tokens_to_reduce_number_of_permutations( 138 self, tokens: List[dict], max_number_of_permutations_per_split: int = 729 139 ) -> List[List[dict]]: 140 """ 141 Splits a sequence of tokens in a smaller sequences of tokens in a way that maximum number of composite 142 tokens permutations does not exceed ``max_number_of_permutations_per_split``. 143 144 For example, 145 146 .. code-block:: python 147 tokens = [ 148 {"tokens": {"date": {"year": "twenty eighteen", "month": "december", "day": "thirty one"}}}, 149 {"tokens": {"date": {"year": "twenty eighteen", "month": "january", "day": "eight"}}}, 150 ] 151 split = normalizer._split_tokens_to_reduce_number_of_permutations( 152 tokens, max_number_of_permutations_per_split=6 153 ) 154 assert split == [ 155 [{"tokens": {"date": {"year": "twenty eighteen", "month": "december", "day": "thirty one"}}}], 156 [{"tokens": {"date": {"year": "twenty eighteen", "month": "january", "day": "eight"}}}], 157 ] 158 159 Date tokens contain 3 items each which gives 6 permutations for every date. Since there are 2 dates, total 160 number of permutations would be ``6 * 6 == 36``. Parameter ``max_number_of_permutations_per_split`` equals 6, 161 so input sequence of tokens is split into 2 smaller sequences. 162 163 Args: 164 tokens (:obj:`List[dict]`): a list of dictionaries, possibly nested. 165 max_number_of_permutations_per_split (:obj:`int`, `optional`, defaults to :obj:`243`): a maximum number 166 of permutations which can be generated from input sequence of tokens. 167 168 Returns: 169 :obj:`List[List[dict]]`: a list of smaller sequences of tokens resulting from ``tokens`` split. 170 """ 171 splits = [] 172 prev_end_of_split = 0 173 current_number_of_permutations = 1 174 for i, token_group in enumerate(tokens): 175 n = self._estimate_number_of_permutations_in_nested_dict(token_group) 176 if n * current_number_of_permutations > max_number_of_permutations_per_split: 177 splits.append(tokens[prev_end_of_split:i]) 178 prev_end_of_split = i 179 current_number_of_permutations = 1 180 if n > max_number_of_permutations_per_split: 181 raise ValueError( 182 f"Could not split token list with respect to condition that every split can generate number of " 183 f"permutations less or equal to " 184 f"`max_number_of_permutations_per_split={max_number_of_permutations_per_split}`. " 185 f"There is an unsplittable token group that generates more than " 186 f"{max_number_of_permutations_per_split} permutations. Try to increase " 187 f"`max_number_of_permutations_per_split` parameter." 188 ) 189 current_number_of_permutations *= n 190 splits.append(tokens[prev_end_of_split:]) 191 assert sum([len(s) for s in splits]) == len(tokens) 192 return splits 193 194 def normalize( 195 self, text: str, verbose: bool = False, punct_pre_process: bool = False, punct_post_process: bool = False 196 ) -> str: 197 """ 198 Main function. Normalizes tokens from written to spoken form 199 e.g. 12 kg -> twelve kilograms 200 201 Args: 202 text: string that may include semiotic classes 203 verbose: whether to print intermediate meta information 204 punct_pre_process: whether to perform punctuation pre-processing, for example, [25] -> [ 25 ] 205 punct_post_process: whether to normalize punctuation 206 207 Returns: spoken form 208 """ 209 original_text = text 210 if punct_pre_process: 211 text = pre_process(text) 212 text = text.strip() 213 if not text: 214 if verbose: 215 print(text) 216 return text 217 text = pynini.escape(text) 218 tagged_lattice = self.find_tags(text) 219 tagged_text = self.select_tag(tagged_lattice) 220 if verbose: 221 print(tagged_text) 222 self.parser(tagged_text) 223 tokens = self.parser.parse() 224 split_tokens = self._split_tokens_to_reduce_number_of_permutations(tokens) 225 output = "" 226 for s in split_tokens: 227 tags_reordered = self.generate_permutations(s) 228 verbalizer_lattice = None 229 for tagged_text in tags_reordered: 230 tagged_text = pynini.escape(tagged_text) 231 232 verbalizer_lattice = self.find_verbalizer(tagged_text) 233 if verbalizer_lattice.num_states() != 0: 234 break 235 if verbalizer_lattice is None: 236 raise ValueError(f"No permutations were generated from tokens {s}") 237 output += ' ' + self.select_verbalizer(verbalizer_lattice) 238 output = SPACE_DUP.sub(' ', output[1:]) 239 if punct_post_process: 240 # do post-processing based on Moses detokenizer 241 if self.processor: 242 output = self.processor.moses_detokenizer.detokenize([output], unescape=False) 243 output = post_process_punct(input=original_text, normalized_text=output) 244 else: 245 print("NEMO_NLP collection is not available: skipping punctuation post_processing") 246 return output 247 248 def _permute(self, d: OrderedDict) -> List[str]: 249 """ 250 Creates reorderings of dictionary elements and serializes as strings 251 252 Args: 253 d: (nested) dictionary of key value pairs 254 255 Return permutations of different string serializations of key value pairs 256 """ 257 l = [] 258 if PRESERVE_ORDER_KEY in d.keys(): 259 d_permutations = [d.items()] 260 else: 261 d_permutations = itertools.permutations(d.items()) 262 for perm in d_permutations: 263 subl = [""] 264 for k, v in perm: 265 if isinstance(v, str): 266 subl = ["".join(x) for x in itertools.product(subl, [f"{k}: \"{v}\" "])] 267 elif isinstance(v, OrderedDict): 268 rec = self._permute(v) 269 subl = ["".join(x) for x in itertools.product(subl, [f" {k} {{ "], rec, [f" }} "])] 270 elif isinstance(v, bool): 271 subl = ["".join(x) for x in itertools.product(subl, [f"{k}: true "])] 272 else: 273 raise ValueError() 274 l.extend(subl) 275 return l 276 277 def generate_permutations(self, tokens: List[dict]): 278 """ 279 Generates permutations of string serializations of list of dictionaries 280 281 Args: 282 tokens: list of dictionaries 283 284 Returns string serialization of list of dictionaries 285 """ 286 287 def _helper(prefix: str, tokens: List[dict], idx: int): 288 """ 289 Generates permutations of string serializations of given dictionary 290 291 Args: 292 tokens: list of dictionaries 293 prefix: prefix string 294 idx: index of next dictionary 295 296 Returns string serialization of dictionary 297 """ 298 if idx == len(tokens): 299 yield prefix 300 return 301 token_options = self._permute(tokens[idx]) 302 for token_option in token_options: 303 yield from _helper(prefix + token_option, tokens, idx + 1) 304 305 return _helper("", tokens, 0) 306 307 def find_tags(self, text: str) -> 'pynini.FstLike': 308 """ 309 Given text use tagger Fst to tag text 310 311 Args: 312 text: sentence 313 314 Returns: tagged lattice 315 """ 316 lattice = text @ self.tagger.fst 317 return lattice 318 319 def select_tag(self, lattice: 'pynini.FstLike') -> str: 320 """ 321 Given tagged lattice return shortest path 322 323 Args: 324 tagged_text: tagged text 325 326 Returns: shortest path 327 """ 328 tagged_text = pynini.shortestpath(lattice, nshortest=1, unique=True).string() 329 return tagged_text 330 331 def find_verbalizer(self, tagged_text: str) -> 'pynini.FstLike': 332 """ 333 Given tagged text creates verbalization lattice 334 This is context-independent. 335 336 Args: 337 tagged_text: input text 338 339 Returns: verbalized lattice 340 """ 341 lattice = tagged_text @ self.verbalizer.fst 342 return lattice 343 344 def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: 345 """ 346 Given verbalized lattice return shortest path 347 348 Args: 349 lattice: verbalization lattice 350 351 Returns: shortest path 352 """ 353 output = pynini.shortestpath(lattice, nshortest=1, unique=True).string() 354 return output 355 356 357 def parse_args(): 358 parser = ArgumentParser() 359 parser.add_argument("input_string", help="input string", type=str) 360 parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) 361 parser.add_argument( 362 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 363 ) 364 parser.add_argument("--verbose", help="print info for debugging", action='store_true') 365 parser.add_argument( 366 "--punct_post_process", help="set to True to enable punctuation post processing", action="store_true" 367 ) 368 parser.add_argument( 369 "--punct_pre_process", help="set to True to enable punctuation pre processing", action="store_true" 370 ) 371 parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") 372 parser.add_argument("--whitelist", help="path to a file with with whitelist", default=None, type=str) 373 parser.add_argument( 374 "--cache_dir", 375 help="path to a dir with .far grammar file. Set to None to avoid using cache", 376 default=None, 377 type=str, 378 ) 379 return parser.parse_args() 380 381 382 if __name__ == "__main__": 383 args = parse_args() 384 whitelist = os.path.abspath(args.whitelist) if args.whitelist else None 385 normalizer = Normalizer( 386 input_case=args.input_case, 387 cache_dir=args.cache_dir, 388 overwrite_cache=args.overwrite_cache, 389 whitelist=whitelist, 390 lang=args.language, 391 ) 392 print( 393 normalizer.normalize( 394 args.input_string, 395 verbose=args.verbose, 396 punct_pre_process=args.punct_pre_process, 397 punct_post_process=args.punct_post_process, 398 ) 399 ) 400 [end of nemo_text_processing/text_normalization/normalize.py] [start of nemo_text_processing/text_normalization/normalize_with_audio.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import json 16 import os 17 import time 18 from argparse import ArgumentParser 19 from glob import glob 20 from typing import List, Tuple 21 22 from joblib import Parallel, delayed 23 from nemo_text_processing.text_normalization.normalize import Normalizer 24 from tqdm import tqdm 25 26 try: 27 from nemo.collections.asr.metrics.wer import word_error_rate 28 from nemo.collections.asr.models import ASRModel 29 30 ASR_AVAILABLE = True 31 except (ModuleNotFoundError, ImportError): 32 ASR_AVAILABLE = False 33 34 try: 35 import pynini 36 from pynini.lib import rewrite 37 38 PYNINI_AVAILABLE = True 39 except (ModuleNotFoundError, ImportError): 40 PYNINI_AVAILABLE = False 41 42 try: 43 from nemo.collections.nlp.data.text_normalization.utils import post_process_punct 44 from nemo_text_processing.text_normalization.data_loader_utils import pre_process 45 46 NLP_AVAILABLE = True 47 except (ModuleNotFoundError, ImportError): 48 NLP_AVAILABLE = False 49 50 """ 51 The script provides multiple normalization options and chooses the best one that minimizes CER of the ASR output 52 (most of the semiotic classes use deterministic=False flag). 53 54 To run this script with a .json manifest file, the manifest file should contain the following fields: 55 "audio_data" - path to the audio file 56 "text" - raw text 57 "pred_text" - ASR model prediction 58 59 See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions 60 61 When the manifest is ready, run: 62 python normalize_with_audio.py \ 63 --audio_data PATH/TO/MANIFEST.JSON \ 64 --language en 65 66 67 To run with a single audio file, specify path to audio and text with: 68 python normalize_with_audio.py \ 69 --audio_data PATH/TO/AUDIO.WAV \ 70 --language en \ 71 --text raw text OR PATH/TO/.TXT/FILE 72 --model QuartzNet15x5Base-En \ 73 --verbose 74 75 To see possible normalization options for a text input without an audio file (could be used for debugging), run: 76 python python normalize_with_audio.py --text "RAW TEXT" 77 78 Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference 79 """ 80 81 82 class NormalizerWithAudio(Normalizer): 83 """ 84 Normalizer class that converts text from written to spoken form. 85 Useful for TTS preprocessing. 86 87 Args: 88 input_case: expected input capitalization 89 lang: language 90 cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. 91 overwrite_cache: set to True to overwrite .far files 92 whitelist: path to a file with whitelist replacements 93 """ 94 95 def __init__( 96 self, 97 input_case: str, 98 lang: str = 'en', 99 cache_dir: str = None, 100 overwrite_cache: bool = False, 101 whitelist: str = None, 102 ): 103 104 super().__init__( 105 input_case=input_case, 106 lang=lang, 107 deterministic=False, 108 cache_dir=cache_dir, 109 overwrite_cache=overwrite_cache, 110 whitelist=whitelist, 111 ) 112 113 def normalize(self, text: str, n_tagged: int, punct_post_process: bool = True, verbose: bool = False,) -> str: 114 """ 115 Main function. Normalizes tokens from written to spoken form 116 e.g. 12 kg -> twelve kilograms 117 118 Args: 119 text: string that may include semiotic classes 120 n_tagged: number of tagged options to consider, -1 - to get all possible tagged options 121 punct_post_process: whether to normalize punctuation 122 verbose: whether to print intermediate meta information 123 124 Returns: 125 normalized text options (usually there are multiple ways of normalizing a given semiotic class) 126 """ 127 original_text = text 128 129 if self.lang == "en": 130 text = pre_process(text) 131 text = text.strip() 132 if not text: 133 if verbose: 134 print(text) 135 return text 136 text = pynini.escape(text) 137 138 if n_tagged == -1: 139 if self.lang == "en": 140 try: 141 tagged_texts = rewrite.rewrites(text, self.tagger.fst_no_digits) 142 except pynini.lib.rewrite.Error: 143 tagged_texts = rewrite.rewrites(text, self.tagger.fst) 144 else: 145 tagged_texts = rewrite.rewrites(text, self.tagger.fst) 146 else: 147 if self.lang == "en": 148 try: 149 tagged_texts = rewrite.top_rewrites(text, self.tagger.fst_no_digits, nshortest=n_tagged) 150 except pynini.lib.rewrite.Error: 151 tagged_texts = rewrite.top_rewrites(text, self.tagger.fst, nshortest=n_tagged) 152 else: 153 tagged_texts = rewrite.top_rewrites(text, self.tagger.fst, nshortest=n_tagged) 154 155 # non-deterministic Eng normalization uses tagger composed with verbalizer, no permutation in between 156 if self.lang == "en": 157 normalized_texts = tagged_texts 158 else: 159 normalized_texts = [] 160 for tagged_text in tagged_texts: 161 self._verbalize(tagged_text, normalized_texts, verbose=verbose) 162 163 if len(normalized_texts) == 0: 164 raise ValueError() 165 166 if punct_post_process: 167 # do post-processing based on Moses detokenizer 168 if self.processor: 169 normalized_texts = [self.processor.detokenize([t]) for t in normalized_texts] 170 normalized_texts = [ 171 post_process_punct(input=original_text, normalized_text=t) for t in normalized_texts 172 ] 173 174 normalized_texts = set(normalized_texts) 175 return normalized_texts 176 177 def _verbalize(self, tagged_text: str, normalized_texts: List[str], verbose: bool = False): 178 """ 179 Verbalizes tagged text 180 181 Args: 182 tagged_text: text with tags 183 normalized_texts: list of possible normalization options 184 verbose: if true prints intermediate classification results 185 """ 186 187 def get_verbalized_text(tagged_text): 188 return rewrite.rewrites(tagged_text, self.verbalizer.fst) 189 190 self.parser(tagged_text) 191 tokens = self.parser.parse() 192 tags_reordered = self.generate_permutations(tokens) 193 for tagged_text_reordered in tags_reordered: 194 try: 195 tagged_text_reordered = pynini.escape(tagged_text_reordered) 196 normalized_texts.extend(get_verbalized_text(tagged_text_reordered)) 197 if verbose: 198 print(tagged_text_reordered) 199 200 except pynini.lib.rewrite.Error: 201 continue 202 203 def select_best_match( 204 self, 205 normalized_texts: List[str], 206 input_text: str, 207 pred_text: str, 208 verbose: bool = False, 209 remove_punct: bool = False, 210 ): 211 """ 212 Selects the best normalization option based on the lowest CER 213 214 Args: 215 normalized_texts: normalized text options 216 input_text: input text 217 pred_text: ASR model transcript of the audio file corresponding to the normalized text 218 verbose: whether to print intermediate meta information 219 remove_punct: whether to remove punctuation before calculating CER 220 221 Returns: 222 normalized text with the lowest CER and CER value 223 """ 224 if pred_text == "": 225 return input_text, 1000 226 227 normalized_texts_cer = calculate_cer(normalized_texts, pred_text, remove_punct) 228 normalized_texts_cer = sorted(normalized_texts_cer, key=lambda x: x[1]) 229 normalized_text, cer = normalized_texts_cer[0] 230 231 if verbose: 232 print('-' * 30) 233 for option in normalized_texts: 234 print(option) 235 print('-' * 30) 236 return normalized_text, cer 237 238 239 def calculate_cer(normalized_texts: List[str], pred_text: str, remove_punct=False) -> List[Tuple[str, float]]: 240 """ 241 Calculates character error rate (CER) 242 243 Args: 244 normalized_texts: normalized text options 245 pred_text: ASR model output 246 247 Returns: normalized options with corresponding CER 248 """ 249 normalized_options = [] 250 for text in normalized_texts: 251 text_clean = text.replace('-', ' ').lower() 252 if remove_punct: 253 for punct in "!?:;,.-()*+-/<=>@^_": 254 text_clean = text_clean.replace(punct, "") 255 cer = round(word_error_rate([pred_text], [text_clean], use_cer=True) * 100, 2) 256 normalized_options.append((text, cer)) 257 return normalized_options 258 259 260 def get_asr_model(asr_model): 261 """ 262 Returns ASR Model 263 264 Args: 265 asr_model: NeMo ASR model 266 """ 267 if os.path.exists(args.model): 268 asr_model = ASRModel.restore_from(asr_model) 269 elif args.model in ASRModel.get_available_model_names(): 270 asr_model = ASRModel.from_pretrained(asr_model) 271 else: 272 raise ValueError( 273 f'Provide path to the pretrained checkpoint or choose from {ASRModel.get_available_model_names()}' 274 ) 275 return asr_model 276 277 278 def parse_args(): 279 parser = ArgumentParser() 280 parser.add_argument("--text", help="input string or path to a .txt file", default=None, type=str) 281 parser.add_argument( 282 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 283 ) 284 parser.add_argument( 285 "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str 286 ) 287 parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") 288 parser.add_argument( 289 '--model', type=str, default='QuartzNet15x5Base-En', help='Pre-trained model name or path to model checkpoint' 290 ) 291 parser.add_argument( 292 "--n_tagged", 293 type=int, 294 default=30, 295 help="number of tagged options to consider, -1 - return all possible tagged options", 296 ) 297 parser.add_argument("--verbose", help="print info for debugging", action="store_true") 298 parser.add_argument( 299 "--no_remove_punct_for_cer", 300 help="Set to True to NOT remove punctuation before calculating CER", 301 action="store_true", 302 ) 303 parser.add_argument( 304 "--no_punct_post_process", help="set to True to disable punctuation post processing", action="store_true" 305 ) 306 parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") 307 parser.add_argument("--whitelist", help="path to a file with with whitelist", default=None, type=str) 308 parser.add_argument( 309 "--cache_dir", 310 help="path to a dir with .far grammar file. Set to None to avoid using cache", 311 default=None, 312 type=str, 313 ) 314 parser.add_argument("--n_jobs", default=-2, type=int, help="The maximum number of concurrently running jobs") 315 parser.add_argument("--batch_size", default=200, type=int, help="Number of examples for each process") 316 return parser.parse_args() 317 318 319 def _normalize_line(normalizer: NormalizerWithAudio, n_tagged, verbose, line: str, remove_punct, punct_post_process): 320 line = json.loads(line) 321 pred_text = line["pred_text"] 322 323 normalized_texts = normalizer.normalize( 324 text=line["text"], verbose=verbose, n_tagged=n_tagged, punct_post_process=punct_post_process, 325 ) 326 327 normalized_text, cer = normalizer.select_best_match( 328 normalized_texts=normalized_texts, 329 input_text=line["text"], 330 pred_text=pred_text, 331 verbose=verbose, 332 remove_punct=remove_punct, 333 ) 334 line["nemo_normalized"] = normalized_text 335 line["CER_nemo_normalized"] = cer 336 return line 337 338 339 def normalize_manifest( 340 normalizer, 341 audio_data: str, 342 n_jobs: int, 343 n_tagged: int, 344 remove_punct: bool, 345 punct_post_process: bool, 346 batch_size: int, 347 ): 348 """ 349 Args: 350 args.audio_data: path to .json manifest file. 351 """ 352 353 def __process_batch(batch_idx, batch, dir_name): 354 normalized_lines = [ 355 _normalize_line( 356 normalizer, 357 n_tagged, 358 verbose=False, 359 line=line, 360 remove_punct=remove_punct, 361 punct_post_process=punct_post_process, 362 ) 363 for line in tqdm(batch) 364 ] 365 366 with open(f"{dir_name}/{batch_idx}.json", "w") as f_out: 367 for line in normalized_lines: 368 f_out.write(json.dumps(line, ensure_ascii=False) + '\n') 369 370 print(f"Batch -- {batch_idx} -- is complete") 371 return normalized_lines 372 373 manifest_out = audio_data.replace('.json', '_normalized.json') 374 with open(audio_data, 'r') as f: 375 lines = f.readlines() 376 377 print(f'Normalizing {len(lines)} lines of {audio_data}...') 378 379 # to save intermediate results to a file 380 batch = min(len(lines), batch_size) 381 382 tmp_dir = manifest_out.replace(".json", "_parts") 383 os.makedirs(tmp_dir, exist_ok=True) 384 385 Parallel(n_jobs=n_jobs)( 386 delayed(__process_batch)(idx, lines[i : i + batch], tmp_dir) 387 for idx, i in enumerate(range(0, len(lines), batch)) 388 ) 389 390 # aggregate all intermediate files 391 with open(manifest_out, "w") as f_out: 392 for batch_f in sorted(glob(f"{tmp_dir}/*.json")): 393 with open(batch_f, "r") as f_in: 394 lines = f_in.read() 395 f_out.write(lines) 396 397 print(f'Normalized version saved at {manifest_out}') 398 399 400 if __name__ == "__main__": 401 args = parse_args() 402 403 if not ASR_AVAILABLE and args.audio_data: 404 raise ValueError("NeMo ASR collection is not installed.") 405 start = time.time() 406 args.whitelist = os.path.abspath(args.whitelist) if args.whitelist else None 407 if args.text is not None: 408 normalizer = NormalizerWithAudio( 409 input_case=args.input_case, 410 lang=args.language, 411 cache_dir=args.cache_dir, 412 overwrite_cache=args.overwrite_cache, 413 whitelist=args.whitelist, 414 ) 415 416 if os.path.exists(args.text): 417 with open(args.text, 'r') as f: 418 args.text = f.read().strip() 419 normalized_texts = normalizer.normalize( 420 text=args.text, 421 verbose=args.verbose, 422 n_tagged=args.n_tagged, 423 punct_post_process=not args.no_punct_post_process, 424 ) 425 426 if args.audio_data: 427 asr_model = get_asr_model(args.model) 428 pred_text = asr_model.transcribe([args.audio_data])[0] 429 normalized_text, cer = normalizer.select_best_match( 430 normalized_texts=normalized_texts, 431 pred_text=pred_text, 432 input_text=args.text, 433 verbose=args.verbose, 434 remove_punct=not args.no_remove_punct_for_cer, 435 ) 436 print(f"Transcript: {pred_text}") 437 print(f"Normalized: {normalized_text}") 438 else: 439 print("Normalization options:") 440 for norm_text in normalized_texts: 441 print(norm_text) 442 elif not os.path.exists(args.audio_data): 443 raise ValueError(f"{args.audio_data} not found.") 444 elif args.audio_data.endswith('.json'): 445 normalizer = NormalizerWithAudio( 446 input_case=args.input_case, 447 lang=args.language, 448 cache_dir=args.cache_dir, 449 overwrite_cache=args.overwrite_cache, 450 whitelist=args.whitelist, 451 ) 452 normalize_manifest( 453 normalizer=normalizer, 454 audio_data=args.audio_data, 455 n_jobs=args.n_jobs, 456 n_tagged=args.n_tagged, 457 remove_punct=not args.no_remove_punct_for_cer, 458 punct_post_process=not args.no_punct_post_process, 459 batch_size=args.batch_size, 460 ) 461 else: 462 raise ValueError( 463 "Provide either path to .json manifest in '--audio_data' OR " 464 + "'--audio_data' path to audio file and '--text' path to a text file OR" 465 "'--text' string text (for debugging without audio)" 466 ) 467 print(f'Execution time: {round((time.time() - start)/60, 2)} min.') 468 [end of nemo_text_processing/text_normalization/normalize_with_audio.py] [start of tools/text_processing_deployment/pynini_export.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 17 import os 18 import time 19 from argparse import ArgumentParser 20 21 from nemo.utils import logging 22 23 try: 24 import pynini 25 from nemo_text_processing.text_normalization.en.graph_utils import generator_main 26 27 PYNINI_AVAILABLE = True 28 except (ModuleNotFoundError, ImportError): 29 30 logging.warning( 31 "`pynini` is not installed ! \n" 32 "Please run the `nemo_text_processing/setup.sh` script" 33 "prior to usage of this toolkit." 34 ) 35 36 PYNINI_AVAILABLE = False 37 38 39 # This script exports compiled grammars inside nemo_text_processing into OpenFst finite state archive files 40 # tokenize_and_classify.far and verbalize.far for production purposes 41 42 43 def itn_grammars(**kwargs): 44 d = {} 45 d['classify'] = { 46 'TOKENIZE_AND_CLASSIFY': ITNClassifyFst( 47 cache_dir=kwargs["cache_dir"], overwrite_cache=kwargs["overwrite_cache"] 48 ).fst 49 } 50 d['verbalize'] = {'ALL': ITNVerbalizeFst().fst, 'REDUP': pynini.accep("REDUP")} 51 return d 52 53 54 def tn_grammars(**kwargs): 55 d = {} 56 d['classify'] = { 57 'TOKENIZE_AND_CLASSIFY': TNClassifyFst( 58 input_case=kwargs["input_case"], 59 deterministic=True, 60 cache_dir=kwargs["cache_dir"], 61 overwrite_cache=kwargs["overwrite_cache"], 62 ).fst 63 } 64 d['verbalize'] = {'ALL': TNVerbalizeFst(deterministic=True).fst, 'REDUP': pynini.accep("REDUP")} 65 return d 66 67 68 def export_grammars(output_dir, grammars): 69 """ 70 Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. 71 72 Args: 73 output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. 74 grammars: grammars to be exported 75 """ 76 77 for category, graphs in grammars.items(): 78 out_dir = os.path.join(output_dir, category) 79 if not os.path.exists(out_dir): 80 os.makedirs(out_dir) 81 time.sleep(1) 82 if category == "classify": 83 category = "tokenize_and_classify" 84 generator_main(f"{out_dir}/{category}.far", graphs) 85 86 87 def parse_args(): 88 parser = ArgumentParser() 89 parser.add_argument("--output_dir", help="output directory for grammars", required=True, type=str) 90 parser.add_argument( 91 "--language", help="language", choices=["en", "de", "es", "ru", 'fr', 'vi'], type=str, default='en' 92 ) 93 parser.add_argument( 94 "--grammars", help="grammars to be exported", choices=["tn_grammars", "itn_grammars"], type=str, required=True 95 ) 96 parser.add_argument( 97 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 98 ) 99 parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") 100 parser.add_argument( 101 "--cache_dir", 102 help="path to a dir with .far grammar file. Set to None to avoid using cache", 103 default=None, 104 type=str, 105 ) 106 return parser.parse_args() 107 108 109 if __name__ == '__main__': 110 args = parse_args() 111 112 if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': 113 raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') 114 115 if args.language == 'en': 116 from nemo_text_processing.inverse_text_normalization.en.taggers.tokenize_and_classify import ( 117 ClassifyFst as ITNClassifyFst, 118 ) 119 from nemo_text_processing.inverse_text_normalization.en.verbalizers.verbalize import ( 120 VerbalizeFst as ITNVerbalizeFst, 121 ) 122 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify import ( 123 ClassifyFst as TNClassifyFst, 124 ) 125 from nemo_text_processing.text_normalization.en.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst 126 elif args.language == 'de': 127 from nemo_text_processing.inverse_text_normalization.de.taggers.tokenize_and_classify import ( 128 ClassifyFst as ITNClassifyFst, 129 ) 130 from nemo_text_processing.inverse_text_normalization.de.verbalizers.verbalize import ( 131 VerbalizeFst as ITNVerbalizeFst, 132 ) 133 from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ( 134 ClassifyFst as TNClassifyFst, 135 ) 136 from nemo_text_processing.text_normalization.de.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst 137 elif args.language == 'ru': 138 from nemo_text_processing.inverse_text_normalization.ru.taggers.tokenize_and_classify import ( 139 ClassifyFst as ITNClassifyFst, 140 ) 141 from nemo_text_processing.inverse_text_normalization.ru.verbalizers.verbalize import ( 142 VerbalizeFst as ITNVerbalizeFst, 143 ) 144 elif args.language == 'es': 145 from nemo_text_processing.inverse_text_normalization.es.taggers.tokenize_and_classify import ( 146 ClassifyFst as ITNClassifyFst, 147 ) 148 from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( 149 VerbalizeFst as ITNVerbalizeFst, 150 ) 151 elif args.language == 'fr': 152 from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( 153 ClassifyFst as ITNClassifyFst, 154 ) 155 from nemo_text_processing.inverse_text_normalization.fr.verbalizers.verbalize import ( 156 VerbalizeFst as ITNVerbalizeFst, 157 ) 158 elif args.language == 'vi': 159 from nemo_text_processing.inverse_text_normalization.vi.taggers.tokenize_and_classify import ( 160 ClassifyFst as ITNClassifyFst, 161 ) 162 from nemo_text_processing.inverse_text_normalization.vi.verbalizers.verbalize import ( 163 VerbalizeFst as ITNVerbalizeFst, 164 ) 165 166 output_dir = os.path.join(args.output_dir, args.language) 167 export_grammars( 168 output_dir=output_dir, 169 grammars=locals()[args.grammars]( 170 input_case=args.input_case, cache_dir=args.cache_dir, overwrite_cache=args.overwrite_cache 171 ), 172 ) 173 [end of tools/text_processing_deployment/pynini_export.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
022f0292aecbc98d591d49423d5045235394f793
./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`.
2022-02-09T05:12:31Z
<patch> <patch> diff --git a/nemo_text_processing/text_normalization/__init__.py b/nemo_text_processing/text_normalization/__init__.py --- a/nemo_text_processing/text_normalization/__init__.py +++ b/nemo_text_processing/text_normalization/__init__.py @@ -21,7 +21,7 @@ except (ModuleNotFoundError, ImportError): logging.warning( "`pynini` is not installed ! \n" - "Please run the `nemo_text_processing/setup.sh` script" + "Please run the `nemo_text_processing/setup.sh` script " "prior to usage of this toolkit." ) diff --git a/nemo_text_processing/text_normalization/en/graph_utils.py b/nemo_text_processing/text_normalization/en/graph_utils.py --- a/nemo_text_processing/text_normalization/en/graph_utils.py +++ b/nemo_text_processing/text_normalization/en/graph_utils.py @@ -159,7 +159,7 @@ def convert_space(fst) -> 'pynini.FstLike': """ Converts space to nonbreaking space. Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" - This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. + This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. Args: fst: input fst @@ -208,9 +208,9 @@ def add_tokens(self, fst) -> 'pynini.FstLike': """ Wraps class name around to given fst - Args: + Args: fst: input fst - + Returns: Fst: fst """ diff --git a/nemo_text_processing/text_normalization/en/taggers/punctuation.py b/nemo_text_processing/text_normalization/en/taggers/punctuation.py --- a/nemo_text_processing/text_normalization/en/taggers/punctuation.py +++ b/nemo_text_processing/text_normalization/en/taggers/punctuation.py @@ -22,7 +22,7 @@ import pynini from pynini.lib import pynutil - PYNINI_AVAILABLE = False + PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py --- a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py @@ -12,8 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -21,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/word.py b/nemo_text_processing/text_normalization/en/verbalizers/word.py --- a/nemo_text_processing/text_normalization/en/verbalizers/word.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/word.py @@ -12,7 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -20,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/es/__init__.py b/nemo_text_processing/text_normalization/es/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/__init__.py @@ -0,0 +1,15 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +LOCALIZATION = "eu" # Set to am for alternate formatting diff --git a/nemo_text_processing/text_normalization/es/data/__init__.py b/nemo_text_processing/text_normalization/es/data/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/dates/__init__.py b/nemo_text_processing/text_normalization/es/data/dates/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/dates/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/electronic/__init__.py b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/fractions/__init__.py b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/measures/__init__.py b/nemo_text_processing/text_normalization/es/data/measures/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/measures/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/money/__init__.py b/nemo_text_processing/text_normalization/es/data/money/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/money/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/numbers/__init__.py b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/roman/__init__.py b/nemo_text_processing/text_normalization/es/data/roman/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/roman/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/time/__init__.py b/nemo_text_processing/text_normalization/es/data/time/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/time/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/graph_utils.py b/nemo_text_processing/text_normalization/es/graph_utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/graph_utils.py @@ -0,0 +1,179 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, NEMO_SPACE +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + digits = pynini.project(pynini.string_file(get_abs_path("data/numbers/digit.tsv")), "input") + tens = pynini.project(pynini.string_file(get_abs_path("data/numbers/ties.tsv")), "input") + teens = pynini.project(pynini.string_file(get_abs_path("data/numbers/teen.tsv")), "input") + twenties = pynini.project(pynini.string_file(get_abs_path("data/numbers/twenties.tsv")), "input") + hundreds = pynini.project(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv")), "input") + + accents = pynini.string_map([("รก", "a"), ("รฉ", "e"), ("รญ", "i"), ("รณ", "o"), ("รบ", "u")]) + + if LOCALIZATION == "am": # Setting localization for central and northern america formatting + cardinal_separator = pynini.string_map([",", NEMO_SPACE]) + decimal_separator = pynini.accep(".") + else: + cardinal_separator = pynini.string_map([".", NEMO_SPACE]) + decimal_separator = pynini.accep(",") + + ones = pynini.union("un", "รบn") + fem_ones = pynini.union(pynini.cross("un", "una"), pynini.cross("รบn", "una"), pynini.cross("uno", "una")) + one_to_one_hundred = pynini.union(digits, tens, teens, twenties, tens + pynini.accep(" y ") + digits) + fem_hundreds = hundreds @ pynini.cdrewrite(pynini.cross("ientos", "ientas"), "", "", NEMO_SIGMA) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digits = None + tens = None + teens = None + twenties = None + hundreds = None + + accents = None + + cardinal_separator = None + decimal_separator = None + + ones = None + fem_ones = None + one_to_one_hundred = None + fem_hundreds = None + + PYNINI_AVAILABLE = False + + +def strip_accent(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Converts all accented vowels to non-accented equivalents + + Args: + fst: Any fst. Composes vowel conversion onto fst's output strings + """ + return fst @ pynini.cdrewrite(accents, "", "", NEMO_SIGMA) + + +def shift_cardinal_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Applies gender conversion rules to a cardinal string. These include: rendering all masculine forms of "uno" (including apocopated forms) as "una" and + Converting all gendered numbers in the hundreds series (200,300,400...) to feminine equivalent (e.g. "doscientos" -> "doscientas"). Converssion only applies + to value place for <1000 and multiple of 1000. (e.g. "doscientos mil doscientos" -> "doscientas mil doscientas".) For place values greater than the thousands, there + is no gender shift as the higher powers of ten ("millones", "billones") are masculine nouns and any conversion would be formally + ungrammatical. + e.g. + "doscientos" -> "doscientas" + "doscientos mil" -> "doscientas mil" + "doscientos millones" -> "doscientos millones" + "doscientos mil millones" -> "doscientos mil millones" + "doscientos millones doscientos mil doscientos" -> "doscientos millones doscientas mil doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + before_mil = ( + NEMO_SPACE + + (pynini.accep("mil") | pynini.accep("milรฉsimo")) + + pynini.closure(NEMO_SPACE + hundreds, 0, 1) + + pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + + pynini.union(pynini.accep("[EOS]"), pynini.accep("\""), decimal_separator) + ) + before_double_digits = pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + pynini.union( + pynini.accep("[EOS]"), pynini.accep("\"") + ) + + fem_allign = pynini.cdrewrite(fem_hundreds, "", before_mil, NEMO_SIGMA) # doscientas mil dosciento + fem_allign @= pynini.cdrewrite(fem_hundreds, "", before_double_digits, NEMO_SIGMA) # doscientas mil doscienta + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union("[EOS]", "\"", decimal_separator), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def shift_number_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Performs gender conversion on all verbalized numbers in output. All values in the hundreds series (200,300,400) are changed to + feminine gender (e.g. "doscientos" -> "doscientas") and all forms of "uno" (including apocopated forms) are converted to "una". + This has no boundary restriction and will perform shift across all values in output string. + e.g. + "doscientos" -> "doscientas" + "doscientos millones" -> "doscientas millones" + "doscientos millones doscientos" -> "doscientas millones doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + fem_allign = pynini.cdrewrite(fem_hundreds, "", "", NEMO_SIGMA) + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union(NEMO_SPACE, pynini.accep("[EOS]"), pynini.accep("\"")), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def strip_cardinal_apocope(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Reverts apocope on cardinal strings in line with formation rules. e.g. "un" -> "uno". Due to cardinal formation rules, this in effect only + affects strings where the final value is a variation of "un". + e.g. + "un" -> "uno" + "veintiรบn" -> "veintiuno" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + # Since cardinals use apocope by default for large values (e.g. "millรณn"), this only needs to act on the last instance of one + strip = pynini.cross("un", "uno") | pynini.cross("รบn", "uno") + strip = pynini.cdrewrite(strip, "", pynini.union("[EOS]", "\""), NEMO_SIGMA) + return fst @ strip + + +def roman_to_int(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Alters given fst to convert Roman integers (lower and upper cased) into Arabic numerals. Valid for values up to 1000. + e.g. + "V" -> "5" + "i" -> "1" + + Args: + fst: Any fst. Composes fst onto Roman conversion outputs. + """ + + def _load_roman(file: str): + roman = load_labels(get_abs_path(file)) + roman_numerals = [(x, y) for x, y in roman] + [(x.upper(), y) for x, y in roman] + return pynini.string_map(roman_numerals) + + digit = _load_roman("data/roman/digit.tsv") + ties = _load_roman("data/roman/ties.tsv") + hundreds = _load_roman("data/roman/hundreds.tsv") + + graph = ( + digit + | ties + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + | ( + hundreds + + (ties | pynutil.add_weight(pynutil.insert("0"), 0.01)) + + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + ) + ).optimize() + + return graph @ fst diff --git a/nemo_text_processing/text_normalization/es/taggers/__init__.py b/nemo_text_processing/text_normalization/es/taggers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/taggers/cardinal.py b/nemo_text_processing/text_normalization/es/taggers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/cardinal.py @@ -0,0 +1,190 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import cardinal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + teen = pynini.invert(pynini.string_file(get_abs_path("data/numbers/teen.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/ties.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/twenties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + zero = None + digit = None + teen = None + ties = None + twenties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def filter_punctuation(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Helper function for parsing number strings. Converts common cardinal strings (groups of three digits delineated by 'cardinal_separator' - see graph_utils) + and converts to a string of digits: + "1 000" -> "1000" + "1.000.000" -> "1000000" + Args: + fst: Any pynini.FstLike object. Function composes fst onto string parser fst + + Returns: + fst: A pynini.FstLike object + """ + exactly_three_digits = NEMO_DIGIT ** 3 # for blocks of three + up_to_three_digits = pynini.closure(NEMO_DIGIT, 1, 3) # for start of string + + cardinal_string = pynini.closure( + NEMO_DIGIT, 1 + ) # For string w/o punctuation (used for page numbers, thousand series) + + cardinal_string |= ( + up_to_three_digits + + pynutil.delete(cardinal_separator) + + pynini.closure(exactly_three_digits + pynutil.delete(cardinal_separator)) + + exactly_three_digits + ) + + return cardinal_string @ fst + + +class CardinalFst(GraphFst): + """ + Finite state transducer for classifying cardinals, e.g. + "1000" -> cardinal { integer: "mil" } + "2.000.000" -> cardinal { integer: "dos millones" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="classify", deterministic=deterministic) + + # Any single digit + graph_digit = digit + digits_no_one = (NEMO_DIGIT - "1") @ graph_digit + + # Any double digit + graph_tens = teen + graph_tens |= ties + (pynutil.delete('0') | (pynutil.insert(" y ") + graph_digit)) + graph_tens |= twenties + + self.tens = graph_tens.optimize() + + self.two_digit_non_zero = pynini.union( + graph_digit, graph_tens, (pynini.cross("0", NEMO_SPACE) + graph_digit) + ).optimize() + + # Three digit strings + graph_hundreds = hundreds + pynini.union( + pynutil.delete("00"), (insert_space + graph_tens), (pynini.cross("0", NEMO_SPACE) + graph_digit) + ) + graph_hundreds |= pynini.cross("100", "cien") + graph_hundreds |= ( + pynini.cross("1", "ciento") + insert_space + pynini.union(graph_tens, pynutil.delete("0") + graph_digit) + ) + + self.hundreds = graph_hundreds.optimize() + + # For all three digit strings with leading zeroes (graph appends '0's to manage place in string) + graph_hundreds_component = pynini.union(graph_hundreds, pynutil.delete("0") + graph_tens) + + graph_hundreds_component_at_least_one_none_zero_digit = graph_hundreds_component | ( + pynutil.delete("00") + graph_digit + ) + graph_hundreds_component_at_least_one_none_zero_digit_no_one = graph_hundreds_component | ( + pynutil.delete("00") + digits_no_one + ) + + graph_thousands_component_at_least_one_none_zero_digit = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_thousands_component_at_least_one_none_zero_digit_no_one = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit_no_one, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_million = pynutil.add_weight(pynini.cross("000001", "un millรณn"), -0.001) + graph_million |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" millones") + graph_million |= pynutil.delete("000000") + graph_million += insert_space + + graph_billion = pynutil.add_weight(pynini.cross("000001", "un billรณn"), -0.001) + graph_billion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" billones") + graph_billion |= pynutil.delete("000000") + graph_billion += insert_space + + graph_trillion = pynutil.add_weight(pynini.cross("000001", "un trillรณn"), -0.001) + graph_trillion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" trillones") + graph_trillion |= pynutil.delete("000000") + graph_trillion += insert_space + + graph = ( + graph_trillion + + graph_billion + + graph_million + + (graph_thousands_component_at_least_one_none_zero_digit | pynutil.delete("000000")) + ) + + self.graph = ( + ((NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 0)) + @ pynini.cdrewrite(pynini.closure(pynutil.insert("0")), "[BOS]", "", NEMO_SIGMA) + @ NEMO_DIGIT ** 24 + @ graph + @ pynini.cdrewrite(delete_space, "[BOS]", "", NEMO_SIGMA) + @ pynini.cdrewrite(delete_space, "", "[EOS]", NEMO_SIGMA) + @ pynini.cdrewrite( + pynini.cross(pynini.closure(NEMO_WHITE_SPACE, 2), NEMO_SPACE), NEMO_ALPHA, NEMO_ALPHA, NEMO_SIGMA + ) + ) + self.graph |= zero + + self.graph = filter_punctuation(self.graph).optimize() + + optional_minus_graph = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + final_graph = optional_minus_graph + pynutil.insert("integer: \"") + self.graph + pynutil.insert("\"") + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/date.py b/nemo_text_processing/text_normalization/es/taggers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/date.py @@ -0,0 +1,107 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_DIGIT, NEMO_SPACE, GraphFst, delete_extra_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + articles = pynini.union("de", "del", "el", "del aรฑo") + delete_leading_zero = (pynutil.delete("0") | (NEMO_DIGIT - "0")) + NEMO_DIGIT + month_numbers = pynini.string_file(get_abs_path("data/dates/months.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + articles = None + delete_leading_zero = None + month_numbers = None + + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for classifying date, e.g. + "01.04.2010" -> date { day: "un" month: "enero" year: "dos mil diez" preserve_order: true } + "marzo 4 2000" -> date { month: "marzo" day: "cuatro" year: "dos mil" } + "1990-20-01" -> date { year: "mil novecientos noventa" day: "veinte" month: "enero" } + + Args: + cardinal: cardinal GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool): + super().__init__(name="date", kind="classify", deterministic=deterministic) + + number_to_month = month_numbers.optimize() + month_graph = pynini.project(number_to_month, "output") + + numbers = cardinal.graph + optional_leading_zero = delete_leading_zero | NEMO_DIGIT + + # 01, 31, 1 + digit_day = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 32)]) @ numbers + day = (pynutil.insert("day: \"") + digit_day + pynutil.insert("\"")).optimize() + + digit_month = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 13)]) + number_to_month = digit_month @ number_to_month + + month_name = (pynutil.insert("month: \"") + month_graph + pynutil.insert("\"")).optimize() + month_number = (pynutil.insert("month: \"") + number_to_month + pynutil.insert("\"")).optimize() + + # prefer cardinal over year + year = (NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 1, 3) # 90, 990, 1990 + year @= numbers + self.year = year + + year_only = pynutil.insert("year: \"") + year + pynutil.insert("\"") + year_with_articles = ( + pynutil.insert("year: \"") + pynini.closure(articles + NEMO_SPACE, 0, 1) + year + pynutil.insert("\"") + ) + + graph_dmy = ( + day + + pynini.closure(pynutil.delete(" de")) + + NEMO_SPACE + + month_name + + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + graph_mdy = ( # English influences on language + month_name + delete_extra_space + day + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + separators = [".", "-", "/"] + for sep in separators: + year_optional = pynini.closure(pynini.cross(sep, NEMO_SPACE) + year_only, 0, 1) + new_graph = day + pynini.cross(sep, NEMO_SPACE) + month_number + year_optional + graph_dmy |= new_graph + if not deterministic: + new_graph = month_number + pynini.cross(sep, NEMO_SPACE) + day + year_optional + graph_mdy |= new_graph + + dash = "-" + day_optional = pynini.closure(pynini.cross(dash, NEMO_SPACE) + day, 0, 1) + graph_ymd = NEMO_DIGIT ** 4 @ year_only + pynini.cross(dash, NEMO_SPACE) + month_number + day_optional + + final_graph = graph_dmy + pynutil.insert(" preserve_order: true") + final_graph |= graph_ymd + final_graph |= graph_mdy + + self.final_graph = final_graph.optimize() + self.fst = self.add_tokens(self.final_graph).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/decimals.py b/nemo_text_processing/text_normalization/es/taggers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/decimals.py @@ -0,0 +1,138 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + cardinal_separator, + decimal_separator, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + quantities = pynini.string_file(get_abs_path("data/numbers/quantities.tsv")) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + quantities = None + digit = None + zero = None + + PYNINI_AVAILABLE = False + + +def get_quantity(decimal_graph: 'pynini.FstLike', cardinal_graph: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Returns FST that transforms either a cardinal or decimal followed by a quantity into a numeral, + e.g. 2 millones -> integer_part: "dos" quantity: "millones" + e.g. 2,4 millones -> integer_part: "dos" fractional_part: "quatro" quantity: "millones" + e.g. 2,400 millones -> integer_part: "dos mil cuatrocientos" fractional_part: "quatro" quantity: "millones" + + Args: + decimal_graph: DecimalFST + cardinal_graph: CardinalFST + """ + numbers = pynini.closure(NEMO_DIGIT, 1, 6) @ cardinal_graph + numbers = pynini.cdrewrite(pynutil.delete(cardinal_separator), "", "", NEMO_SIGMA) @ numbers + + res = ( + pynutil.insert("integer_part: \"") + + numbers # The cardinal we're passing only produces 'un' for one, so gender agreement is safe (all quantities are masculine). Limit to 10^6 power. + + pynutil.insert("\"") + + NEMO_SPACE + + pynutil.insert("quantity: \"") + + quantities + + pynutil.insert("\"") + ) + res |= decimal_graph + NEMO_SPACE + pynutil.insert("quantity: \"") + quantities + pynutil.insert("\"") + return res + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + -11,4006 billones -> decimal { negative: "true" integer_part: "once" fractional_part: "cuatro cero cero seis" quantity: "billones" preserve_order: true } + 1 billรณn -> decimal { integer_part: "un" quantity: "billรณn" preserve_order: true } + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + graph_digit = digit | zero + + if not deterministic: + graph = pynini.union(graph_digit, cardinal.hundreds, cardinal.tens) + graph += pynini.closure(insert_space + graph) + + else: + # General pattern seems to be 1-3 digits: map as cardinal, default to digits otherwise \ + graph = pynini.union( + graph_digit, + cardinal.tens, + cardinal.hundreds, + graph_digit + pynini.closure(insert_space + graph_digit, 3), + zero + + pynini.closure(insert_space + zero) + + pynini.closure(insert_space + graph_digit), # For cases such as "1,010" + ) + + # Need to strip apocope everywhere BUT end of string + reverse_apocope = pynini.string_map([("un", "uno"), ("รบn", "uno")]) + apply_reverse_apocope = pynini.cdrewrite(reverse_apocope, "", NEMO_SPACE, NEMO_SIGMA) + graph @= apply_reverse_apocope + + # Technically decimals should be space delineated groups of three, e.g. (1,333 333). This removes any possible spaces + strip_formatting = pynini.cdrewrite(delete_space, "", "", NEMO_SIGMA) + graph = strip_formatting @ graph + + self.graph = graph.optimize() + + graph_separator = pynutil.delete(decimal_separator) + optional_graph_negative = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + self.graph_fractional = pynutil.insert("fractional_part: \"") + self.graph + pynutil.insert("\"") + + # Integer graph maintains apocope except for ones place + graph_integer = ( + strip_cardinal_apocope(cardinal.graph) + if deterministic + else pynini.union(cardinal.graph, strip_cardinal_apocope(cardinal.graph)) + ) # Gives us forms w/ and w/o apocope + self.graph_integer = pynutil.insert("integer_part: \"") + graph_integer + pynutil.insert("\"") + final_graph_wo_sign = self.graph_integer + graph_separator + insert_space + self.graph_fractional + + self.final_graph_wo_negative = ( + final_graph_wo_sign | get_quantity(final_graph_wo_sign, cardinal.graph).optimize() + ) + final_graph = optional_graph_negative + self.final_graph_wo_negative + + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/electronic.py b/nemo_text_processing/text_normalization/es/taggers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/electronic.py @@ -0,0 +1,84 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_ALPHA, NEMO_DIGIT, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + common_domains = [x[0] for x in load_labels(get_abs_path("data/electronic/domain.tsv"))] + symbols = [x[0] for x in load_labels(get_abs_path("data/electronic/symbols.tsv"))] + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + common_domains = None + symbols = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for classifying electronic: email addresses + e.g. "abc@hotmail.com" -> electronic { username: "abc" domain: "hotmail.com" preserve_order: true } + e.g. "www.abc.com/123" -> electronic { protocol: "www." domain: "abc.com/123" preserve_order: true } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="classify", deterministic=deterministic) + + dot = pynini.accep(".") + accepted_common_domains = pynini.union(*common_domains) + accepted_symbols = pynini.union(*symbols) - dot + accepted_characters = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols) + acceepted_characters_with_dot = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols | dot) + + # email + username = ( + pynutil.insert("username: \"") + + acceepted_characters_with_dot + + pynutil.insert("\"") + + pynini.cross('@', ' ') + ) + domain_graph = accepted_characters + dot + accepted_characters + domain_graph = pynutil.insert("domain: \"") + domain_graph + pynutil.insert("\"") + domain_common_graph = ( + pynutil.insert("domain: \"") + + accepted_characters + + accepted_common_domains + + pynini.closure((accepted_symbols | dot) + pynini.closure(accepted_characters, 1), 0, 1) + + pynutil.insert("\"") + ) + graph = (username + domain_graph) | domain_common_graph + + # url + protocol_start = pynini.accep("https://") | pynini.accep("http://") + protocol_end = ( + pynini.accep("www.") + if deterministic + else pynini.accep("www.") | pynini.cross("www.", "doble ve doble ve doble ve.") + ) + protocol = protocol_start | protocol_end | (protocol_start + protocol_end) + protocol = pynutil.insert("protocol: \"") + protocol + pynutil.insert("\"") + graph |= protocol + insert_space + (domain_graph | domain_common_graph) + self.graph = graph + + final_graph = self.add_tokens(self.graph + pynutil.insert(" preserve_order: true")) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/fraction.py b/nemo_text_processing/text_normalization/es/taggers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/fraction.py @@ -0,0 +1,124 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + ordinal_exceptions = pynini.string_file(get_abs_path("data/fractions/ordinal_exceptions.tsv")) + higher_powers_of_ten = pynini.string_file(get_abs_path("data/fractions/powers_of_ten.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + ordinal_exceptions = None + higher_powers_of_ten = None + + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for classifying fraction + "23 4/5" -> + tokens { fraction { integer: "veintitrรฉs" numerator: "cuatro" denominator: "quinto" mophosyntactic_features: "ordinal" } } + + Args: + cardinal: CardinalFst + ordinal: OrdinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, ordinal: GraphFst, deterministic: bool = True): + super().__init__(name="fraction", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + ordinal_graph = ordinal.graph + + # 2-10 are all ordinals + three_to_ten = pynini.string_map(["2", "3", "4", "5", "6", "7", "8", "9", "10",]) + block_three_to_ten = pynutil.delete(three_to_ten) # To block cardinal productions + if not deterministic: # Multiples of tens are sometimes rendered as ordinals + three_to_ten |= pynini.string_map(["20", "30", "40", "50", "60", "70", "80", "90",]) + graph_three_to_ten = three_to_ten @ ordinal_graph + graph_three_to_ten @= pynini.cdrewrite(ordinal_exceptions, "", "", NEMO_SIGMA) + + # Higher powers of tens (and multiples) are converted to ordinals. + hundreds = pynini.string_map(["100", "200", "300", "400", "500", "600", "700", "800", "900",]) + graph_hundreds = hundreds @ ordinal_graph + + multiples_of_thousand = ordinal.multiples_of_thousand # So we can have X milรฉsimos + + graph_higher_powers_of_ten = ( + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + + pynini.closure("mil ", 0, 1) + + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + ) # x millones / x mil millones / x mil z millones + graph_higher_powers_of_ten += higher_powers_of_ten + graph_higher_powers_of_ten = cardinal_graph @ graph_higher_powers_of_ten + graph_higher_powers_of_ten @= pynini.cdrewrite( + pynutil.delete("un "), pynini.accep("[BOS]"), pynini.project(higher_powers_of_ten, "output"), NEMO_SIGMA + ) # we drop 'un' from these ordinals (millionths, not one-millionths) + + graph_higher_powers_of_ten = multiples_of_thousand | graph_hundreds | graph_higher_powers_of_ten + block_higher_powers_of_ten = pynutil.delete( + pynini.project(graph_higher_powers_of_ten, "input") + ) # For cardinal graph + + graph_fractions_ordinals = graph_higher_powers_of_ten | graph_three_to_ten + graph_fractions_ordinals += pynutil.insert( + "\" morphosyntactic_features: \"ordinal\"" + ) # We note the root for processing later + + # Blocking the digits and hundreds from Cardinal graph + graph_fractions_cardinals = pynini.cdrewrite( + block_three_to_ten | block_higher_powers_of_ten, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fractions_cardinals @= NEMO_CHAR.plus @ pynini.cdrewrite( + pynutil.delete("0"), pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) # Empty characters become '0' for NEMO_CHAR fst, so ned to block + graph_fractions_cardinals @= cardinal_graph + graph_fractions_cardinals += pynutil.insert( + "\" morphosyntactic_features: \"add_root\"" + ) # blocking these entries to reduce erroneous possibilities in debugging + + if deterministic: + graph_fractions_cardinals = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ graph_fractions_cardinals + ) # Past hundreds the conventional scheme can be hard to read. For determinism we stop here + + graph_denominator = pynini.union( + graph_fractions_ordinals, + graph_fractions_cardinals, + pynutil.add_weight(cardinal_graph + pynutil.insert("\""), 0.001), + ) # Last form is simply recording the cardinal. Weighting so last resort + + integer = pynutil.insert("integer_part: \"") + cardinal_graph + pynutil.insert("\"") + NEMO_SPACE + numerator = ( + pynutil.insert("numerator: \"") + cardinal_graph + (pynini.cross("/", "\" ") | pynini.cross(" / ", "\" ")) + ) + denominator = pynutil.insert("denominator: \"") + graph_denominator + + self.graph = pynini.closure(integer, 0, 1) + numerator + denominator + + final_graph = self.add_tokens(self.graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/measure.py b/nemo_text_processing/text_normalization/es/taggers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/measure.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_NON_BREAKING_SPACE, + NEMO_SPACE, + GraphFst, + convert_space, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit = pynini.string_file(get_abs_path("data/measures/measurements.tsv")) + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit = None + unit_plural_fem = None + unit_plural_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for classifying measure, e.g. + "2,4 g" -> measure { cardinal { integer_part: "dos" fractional_part: "cuatro" units: "gramos" preserve_order: true } } + "1 g" -> measure { cardinal { integer: "un" units: "gramo" preserve_order: true } } + "1 millรณn g" -> measure { cardinal { integer: "un quantity: "millรณn" units: "gramos" preserve_order: true } } + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + This class also converts words containing numbers and letters + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + + + Args: + cardinal: CardinalFst + decimal: DecimalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, fraction: GraphFst, deterministic: bool = True): + super().__init__(name="measure", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + + unit_singular = unit + unit_plural = unit_singular @ (unit_plural_fem | unit_plural_masc) + + graph_unit_singular = convert_space(unit_singular) + graph_unit_plural = convert_space(unit_plural) + + optional_graph_negative = pynini.closure("-", 0, 1) + + graph_unit_denominator = ( + pynini.cross("/", "por") + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_singular + ) + + optional_unit_denominator = pynini.closure( + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_denominator, 0, 1, + ) + + unit_plural = ( + pynutil.insert("units: \"") + + ((graph_unit_plural + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + unit_singular_graph = ( + pynutil.insert("units: \"") + + ((graph_unit_singular + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + subgraph_decimal = decimal.fst + insert_space + pynini.closure(NEMO_SPACE, 0, 1) + unit_plural + + subgraph_cardinal = ( + (optional_graph_negative + (pynini.closure(NEMO_DIGIT) - "1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_plural + ) + + subgraph_cardinal |= ( + (optional_graph_negative + pynini.accep("1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_singular_graph + ) + + subgraph_fraction = fraction.fst + insert_space + pynini.closure(delete_space, 0, 1) + unit_plural + + decimal_times = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_times = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.insert("\" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_dash_alpha = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.delete('-') + + pynutil.insert("\" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + decimal_dash_alpha = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.delete('-') + + pynutil.insert(" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + alpha_dash_cardinal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" cardinal { integer: \"") + + cardinal_graph + + pynutil.insert("\" } preserve_order: true") + ) + + alpha_dash_decimal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } preserve_order: true") + ) + + final_graph = ( + subgraph_decimal + | subgraph_cardinal + | subgraph_fraction + | cardinal_dash_alpha + | alpha_dash_cardinal + | decimal_dash_alpha + | decimal_times + | cardinal_times + | alpha_dash_decimal + ) + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/money.py b/nemo_text_processing/text_normalization/es/taggers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/money.py @@ -0,0 +1,194 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import decimal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + maj_singular_labels = load_labels(get_abs_path("data/money/currency_major.tsv")) + maj_singular = pynini.string_file((get_abs_path("data/money/currency_major.tsv"))) + min_singular = pynini.string_file(get_abs_path("data/money/currency_minor.tsv")) + fem_plural = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc_plural = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + maj_singular_labels = None + min_singular = None + maj_singular = None + fem_plural = None + masc_plural = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for classifying money, e.g. + "โ‚ฌ1" -> money { currency_maj: "euro" integer_part: "un"} + "โ‚ฌ1,000" -> money { currency_maj: "euro" integer_part: "un" } + "โ‚ฌ1,001" -> money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un" } + "ยฃ1,4" -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true } + -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "penique" preserve_order: true } + "0,01 ยฃ" -> money { fractional_part: "un" currency_min: "penique" preserve_order: true } + "0,02 ยฃ" -> money { fractional_part: "dos" currency_min: "peniques" preserve_order: true } + "ยฃ0,01 million" -> money { currency_maj: "libra" integer_part: "cero" fractional_part: "cero un" quantity: "million" } + + Args: + cardinal: CardinalFst + decimal: DecimalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + graph_decimal_final = decimal.final_graph_wo_negative + + maj_singular_graph = maj_singular + min_singular_graph = min_singular + maj_plural_graph = maj_singular @ (fem_plural | masc_plural) + min_plural_graph = min_singular @ (fem_plural | masc_plural) + + graph_maj_singular = pynutil.insert("currency_maj: \"") + maj_singular_graph + pynutil.insert("\"") + graph_maj_plural = pynutil.insert("currency_maj: \"") + maj_plural_graph + pynutil.insert("\"") + + graph_integer_one = pynutil.insert("integer_part: \"") + pynini.cross("1", "un") + pynutil.insert("\"") + + decimal_with_quantity = (NEMO_SIGMA + NEMO_ALPHA) @ graph_decimal_final + + graph_decimal_plural = pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, # 1,05 $ + ) + graph_decimal_plural = ( + (NEMO_SIGMA - "1") + decimal_separator + NEMO_SIGMA + ) @ graph_decimal_plural # Can't have "un euros" + + graph_decimal_singular = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, # 1,05 $ + ) + graph_decimal_singular = (pynini.accep("1") + decimal_separator + NEMO_SIGMA) @ graph_decimal_singular + + graph_decimal = pynini.union( + graph_decimal_singular, + graph_decimal_plural, + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + decimal_with_quantity, + ) + + graph_integer = ( + pynutil.insert("integer_part: \"") + ((NEMO_SIGMA - "1") @ cardinal_graph) + pynutil.insert("\"") + ) + + graph_integer_only = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer_one, + graph_integer_one + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, + ) + graph_integer_only |= pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer, + graph_integer + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, + ) + + graph = graph_integer_only | graph_decimal + + # remove trailing zeros of non zero number in the first 2 digits and fill up to 2 digits + # e.g. 2000 -> 20, 0200->02, 01 -> 01, 10 -> 10 + # not accepted: 002, 00, 0, + two_digits_fractional_part = ( + pynini.closure(NEMO_DIGIT) + (NEMO_DIGIT - "0") + pynini.closure(pynutil.delete("0")) + ) @ ( + (pynutil.delete("0") + (NEMO_DIGIT - "0")) + | ((NEMO_DIGIT - "0") + pynutil.insert("0")) + | ((NEMO_DIGIT - "0") + NEMO_DIGIT) + ) + + graph_min_singular = pynutil.insert("currency_min: \"") + min_singular_graph + pynutil.insert("\"") + graph_min_plural = pynutil.insert("currency_min: \"") + min_plural_graph + pynutil.insert("\"") + + # format ** euro ** cent + decimal_graph_with_minor = None + for curr_symbol, _ in maj_singular_labels: + preserve_order = pynutil.insert(" preserve_order: true") + + integer_plus_maj = pynini.union( + graph_integer + insert_space + pynutil.insert(curr_symbol) @ graph_maj_plural, + graph_integer_one + insert_space + pynutil.insert(curr_symbol) @ graph_maj_singular, + ) + # non zero integer part + integer_plus_maj = (pynini.closure(NEMO_DIGIT) - "0") @ integer_plus_maj + + graph_fractional_one = ( + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ pynini.cross("1", "un") + + pynutil.insert("\"") + ) + + graph_fractional = ( + two_digits_fractional_part @ (pynini.closure(NEMO_DIGIT, 1, 2) - "1") @ cardinal.two_digit_non_zero + ) + graph_fractional = pynutil.insert("fractional_part: \"") + graph_fractional + pynutil.insert("\"") + + fractional_plus_min = pynini.union( + graph_fractional + insert_space + pynutil.insert(curr_symbol) @ graph_min_plural, + graph_fractional_one + insert_space + pynutil.insert(curr_symbol) @ graph_min_singular, + ) + + decimal_graph_with_minor_curr = ( + integer_plus_maj + pynini.cross(decimal_separator, NEMO_SPACE) + fractional_plus_min + ) + decimal_graph_with_minor_curr |= pynutil.add_weight( + integer_plus_maj + + pynini.cross(decimal_separator, NEMO_SPACE) + + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ cardinal.two_digit_non_zero + + pynutil.insert("\""), + weight=0.0001, + ) + + decimal_graph_with_minor_curr |= pynutil.delete("0,") + fractional_plus_min + decimal_graph_with_minor_curr = pynini.union( + pynutil.delete(curr_symbol) + + pynini.closure(delete_space, 0, 1) + + decimal_graph_with_minor_curr + + preserve_order, + decimal_graph_with_minor_curr + + preserve_order + + pynini.closure(delete_space, 0, 1) + + pynutil.delete(curr_symbol), + ) + + decimal_graph_with_minor = ( + decimal_graph_with_minor_curr + if decimal_graph_with_minor is None + else pynini.union(decimal_graph_with_minor, decimal_graph_with_minor_curr) + ) + + final_graph = graph | pynutil.add_weight(decimal_graph_with_minor, -0.001) + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/ordinal.py b/nemo_text_processing/text_normalization/es/taggers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/ordinal.py @@ -0,0 +1,186 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import roman_to_int, strip_accent +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/digit.tsv"))) + teens = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/teen.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/twenties.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/ties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ImportError, ModuleNotFoundError): + digit = None + teens = None + twenties = None + ties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def get_one_to_one_thousand(cardinal: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Produces an acceptor for verbalizations of all numbers from 1 to 1000. Needed for ordinals and fractions. + + Args: + cardinal: CardinalFst + + Returns: + fst: A pynini.FstLike object + """ + numbers = pynini.string_map([str(_) for _ in range(1, 1000)]) @ cardinal + return pynini.project(numbers, "output").optimize() + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for classifying ordinal + "21.ยบ" -> ordinal { integer: "vigรฉsimo primero" morphosyntactic_features: "gender_masc" } + This class converts ordinal up to the millionth (millonรฉsimo) order (exclusive). + + This FST also records the ending of the ordinal (called "morphosyntactic_features"): + either as gender_masc, gender_fem, or apocope. Also introduces plural feature for non-deterministic graphs. + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="ordinal", kind="classify") + cardinal_graph = cardinal.graph + + graph_digit = digit.optimize() + graph_teens = teens.optimize() + graph_ties = ties.optimize() + graph_twenties = twenties.optimize() + graph_hundreds = hundreds.optimize() + + if not deterministic: + # Some alternative derivations + graph_ties = graph_ties | pynini.cross("sesenta", "setuagรฉsimo") + + graph_teens = graph_teens | pynini.cross("once", "decimoprimero") + graph_teens |= pynini.cross("doce", "decimosegundo") + + graph_digit = graph_digit | pynini.cross("nueve", "nono") + graph_digit |= pynini.cross("siete", "sรฉtimo") + + graph_tens_component = ( + graph_teens + | (graph_ties + pynini.closure(pynini.cross(" y ", NEMO_SPACE) + graph_digit, 0, 1)) + | graph_twenties + ) + + graph_hundred_component = pynini.union( + graph_hundreds + pynini.closure(NEMO_SPACE + pynini.union(graph_tens_component, graph_digit), 0, 1), + graph_tens_component, + graph_digit, + ) + + # Need to go up to thousands for fractions + self.one_to_one_thousand = get_one_to_one_thousand(cardinal_graph) + + thousands = pynini.cross("mil", "milรฉsimo") + + graph_thousands = ( + strip_accent(self.one_to_one_thousand) + NEMO_SPACE + thousands + ) # Cardinals become prefix for thousands series. Snce accent on the powers of ten we strip accent from leading words + graph_thousands @= pynini.cdrewrite(delete_space, "", "milรฉsimo", NEMO_SIGMA) # merge as a prefix + graph_thousands |= thousands + + self.multiples_of_thousand = (cardinal_graph @ graph_thousands).optimize() + + if ( + not deterministic + ): # Formally the words preceding the power of ten should be a prefix, but some maintain word boundaries. + graph_thousands |= (self.one_to_one_thousand @ graph_hundred_component) + NEMO_SPACE + thousands + + graph_thousands += pynini.closure(NEMO_SPACE + graph_hundred_component, 0, 1) + + ordinal_graph = graph_thousands | graph_hundred_component + ordinal_graph = cardinal_graph @ ordinal_graph + + if not deterministic: + # The 10's and 20's series can also be two words + split_words = pynini.cross("decimo", "dรฉcimo ") | pynini.cross("vigesimo", "vigรฉsimo ") + split_words = pynini.cdrewrite(split_words, "", NEMO_CHAR, NEMO_SIGMA) + ordinal_graph |= ordinal_graph @ split_words + + # If "octavo" is preceeded by the "o" within string, it needs deletion + ordinal_graph @= pynini.cdrewrite(pynutil.delete("o"), "", "octavo", NEMO_SIGMA) + + self.graph = ordinal_graph.optimize() + + masc = pynini.accep("gender_masc") + fem = pynini.accep("gender_fem") + apocope = pynini.accep("apocope") + + delete_period = pynini.closure(pynutil.delete("."), 0, 1) # Sometimes the period is omitted f + + accept_masc = delete_period + pynini.cross("ยบ", masc) + accep_fem = delete_period + pynini.cross("ยช", fem) + accep_apocope = delete_period + pynini.cross("แต‰สณ", apocope) + + # Managing Romanization + graph_roman = pynutil.insert("integer: \"") + roman_to_int(ordinal_graph) + pynutil.insert("\"") + if not deterministic: + # Introduce plural + plural = pynini.closure(pynutil.insert("/plural"), 0, 1) + accept_masc += plural + accep_fem += plural + + # Romanizations have no morphology marker, so in non-deterministic case we provide option for all + insert_morphology = pynutil.insert(pynini.union(masc, fem)) + plural + insert_morphology |= pynutil.insert(apocope) + insert_morphology = ( + pynutil.insert(" morphosyntactic_features: \"") + insert_morphology + pynutil.insert("\"") + ) + + graph_roman += insert_morphology + + else: + # We assume masculine gender as default + graph_roman += pynutil.insert(" morphosyntactic_features: \"gender_masc\"") + + # Rest of graph + convert_abbreviation = accept_masc | accep_fem | accep_apocope + + graph = ( + pynutil.insert("integer: \"") + + ordinal_graph + + pynutil.insert("\"") + + pynutil.insert(" morphosyntactic_features: \"") + + convert_abbreviation + + pynutil.insert("\"") + ) + graph = pynini.union(graph, graph_roman) + + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/telephone.py b/nemo_text_processing/text_normalization/es/taggers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/telephone.py @@ -0,0 +1,156 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.graph_utils import ones +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + graph_digit = pynini.string_file(get_abs_path("data/numbers/digit.tsv")) + graph_ties = pynini.string_file(get_abs_path("data/numbers/ties.tsv")) + graph_teen = pynini.string_file(get_abs_path("data/numbers/teen.tsv")) + graph_twenties = pynini.string_file(get_abs_path("data/numbers/twenties.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + graph_digit = None + graph_ties = None + graph_teen = None + graph_twenties = None + + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for classifying telephone numbers, e.g. + 123-123-5678 -> { number_part: "uno dos tres uno dos tres cinco seis siete ocho" }. + In Spanish, digits are generally read individually, or as 2-digit numbers, + eg. "123" = "uno dos tres", + "1234" = "doce treinta y cuatro". + This will verbalize sequences of 10 (3+3+4 e.g. 123-456-7890). + 9 (3+3+3 e.g. 123-456-789) and 8 (4+4 e.g. 1234-5678) digits. + + (we ignore more complicated cases such as "doscientos y dos" or "tres nueves"). + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="classify") + + # create `single_digits` and `double_digits` graphs as these will be + # the building blocks of possible telephone numbers + single_digits = pynini.invert(graph_digit).optimize() | pynini.cross("0", "cero") + + double_digits = pynini.union( + graph_twenties, + graph_teen, + (graph_ties + pynutil.delete("0")), + (graph_ties + insert_space + pynutil.insert("y") + insert_space + graph_digit), + ) + double_digits = pynini.invert(double_digits) + + # define `ten_digit_graph`, `nine_digit_graph`, `eight_digit_graph` + # which produces telephone numbers spoken (1) only with single digits, + # or (2) spoken with double digits (and sometimes single digits) + + # 10-digit option (1): all single digits + ten_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + # 9-digit option (1): all single digits + nine_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 2, 2) + + single_digits + ) + + # 8-digit option (1): all single digits + eight_digit_graph = ( + pynini.closure(single_digits + insert_space, 4, 4) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + if not deterministic: + # 10-digit option (2): (1+2) + (1+2) + (2+2) digits + ten_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + # 9-digit option (2): (1+2) + (1+2) + (1+2) digits + nine_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + ) + + # 8-digit option (2): (2+2) + (2+2) digits + eight_digit_graph |= ( + double_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + number_part = pynini.union(ten_digit_graph, nine_digit_graph, eight_digit_graph) + number_part @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", "", NEMO_SIGMA) + + number_part = pynutil.insert("number_part: \"") + number_part + pynutil.insert("\"") + + graph = number_part + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/time.py b/nemo_text_processing/text_normalization/es/taggers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/time.py @@ -0,0 +1,218 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + time_zone_graph = pynini.string_file(get_abs_path("data/time/time_zone.tsv")) + suffix = pynini.string_file(get_abs_path("data/time/time_suffix.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + time_zone_graph = None + suffix = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for classifying time, e.g. + "02:15 est" -> time { hours: "dos" minutes: "quince" zone: "e s t"} + "2 h" -> time { hours: "dos" } + "9 h" -> time { hours: "nueve" } + "02:15:10 h" -> time { hours: "dos" minutes: "quince" seconds: "diez"} + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="time", kind="classify", deterministic=deterministic) + + delete_time_delimiter = pynutil.delete(pynini.union(".", ":")) + + one = pynini.string_map([("un", "una"), ("รบn", "una")]) + change_one = pynini.cdrewrite(one, "", "", NEMO_SIGMA) + cardinal_graph = cardinal.graph @ change_one + + day_suffix = pynutil.insert("suffix: \"") + suffix + pynutil.insert("\"") + day_suffix = delete_space + insert_space + day_suffix + + delete_hora_suffix = delete_space + insert_space + pynutil.delete("h") + delete_minute_suffix = delete_space + insert_space + pynutil.delete("min") + delete_second_suffix = delete_space + insert_space + pynutil.delete("s") + + labels_hour_24 = [ + str(x) for x in range(0, 25) + ] # Can see both systems. Twelve hour requires am/pm for ambiguity resolution + labels_hour_12 = [str(x) for x in range(1, 13)] + labels_minute_single = [str(x) for x in range(1, 10)] + labels_minute_double = [str(x) for x in range(10, 60)] + + delete_leading_zero_to_double_digit = ( + pynini.closure(pynutil.delete("0") | (NEMO_DIGIT - "0"), 0, 1) + NEMO_DIGIT + ) + + graph_24 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_24) + ) + graph_12 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_12) + ) + + graph_hour_24 = graph_24 @ cardinal_graph + graph_hour_12 = graph_12 @ cardinal_graph + + graph_minute_single = delete_leading_zero_to_double_digit @ pynini.union(*labels_minute_single) + graph_minute_double = pynini.union(*labels_minute_double) + + graph_minute = pynini.union(graph_minute_single, graph_minute_double) @ cardinal_graph + + final_graph_hour_only_24 = ( + pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + delete_hora_suffix + ) + final_graph_hour_only_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + day_suffix + + final_graph_hour_24 = pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + final_graph_hour_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + + final_graph_minute = pynutil.insert("minutes: \"") + graph_minute + pynutil.insert("\"") + final_graph_second = pynutil.insert("seconds: \"") + graph_minute + pynutil.insert("\"") + final_time_zone_optional = pynini.closure( + delete_space + insert_space + pynutil.insert("zone: \"") + time_zone_graph + pynutil.insert("\""), 0, 1, + ) + + # 02.30 h + graph_hm = ( + final_graph_hour_24 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 h + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_24 + + delete_hora_suffix + + delete_space + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + delete_minute_suffix + + pynini.closure( + delete_space + + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)) + + delete_second_suffix, + 0, + 1, + ) # For seconds + + final_time_zone_optional + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_12 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 a. m. + + day_suffix + + final_time_zone_optional + ) + + graph_h = ( + pynini.union(final_graph_hour_only_24, final_graph_hour_only_12) + final_time_zone_optional + ) # Should always have a time indicator, else we'll pass to cardinals + + if not deterministic: + # This includes alternate vocalization (hour menos min, min para hour), here we shift the times and indicate a `style` tag + hour_shift_24 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_24.tsv"))) + hour_shift_12 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_12.tsv"))) + minute_shift = pynini.string_file(get_abs_path("data/time/minute_to.tsv")) + + graph_hour_to_24 = graph_24 @ hour_shift_24 @ cardinal_graph + graph_hour_to_12 = graph_12 @ hour_shift_12 @ cardinal_graph + + graph_minute_to = pynini.union(graph_minute_single, graph_minute_double) @ minute_shift @ cardinal_graph + + final_graph_hour_to_24 = pynutil.insert("hours: \"") + graph_hour_to_24 + pynutil.insert("\"") + final_graph_hour_to_12 = pynutil.insert("hours: \"") + graph_hour_to_12 + pynutil.insert("\"") + + final_graph_minute_to = pynutil.insert("minutes: \"") + graph_minute_to + pynutil.insert("\"") + + graph_menos = pynutil.insert(" style: \"1\"") + graph_para = pynutil.insert(" style: \"2\"") + + final_graph_style = graph_menos | graph_para + + # 02.30 h (omitting seconds since a bit awkward) + graph_hm |= ( + final_graph_hour_to_24 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + + final_graph_style + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_to_24 + + delete_hora_suffix + + delete_space + + insert_space + + final_graph_minute_to + + delete_minute_suffix + + final_time_zone_optional + + final_graph_style + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_to_12 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + day_suffix + + final_time_zone_optional + + final_graph_style + ) + + final_graph = graph_hm | graph_h + if deterministic: + final_graph = final_graph + pynutil.insert(" preserve_order: true") + final_graph = final_graph.optimize() + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py @@ -0,0 +1,157 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os + +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_space, + generator_main, +) +from nemo_text_processing.text_normalization.en.taggers.punctuation import PunctuationFst +from nemo_text_processing.text_normalization.es.taggers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.taggers.date import DateFst +from nemo_text_processing.text_normalization.es.taggers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.taggers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.taggers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.taggers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.taggers.money import MoneyFst +from nemo_text_processing.text_normalization.es.taggers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.taggers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.taggers.time import TimeFst +from nemo_text_processing.text_normalization.es.taggers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.taggers.word import WordFst + +from nemo.utils import logging + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class ClassifyFst(GraphFst): + """ + Final class that composes all other classification grammars. This class can process an entire sentence, that is lower cased. + For deployment, this grammar will be compiled and exported to OpenFst Finate State aRchive (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. + overwrite_cache: set to True to overwrite .far files + whitelist: path to a file with whitelist replacements + """ + + def __init__( + self, + input_case: str, + deterministic: bool = False, + cache_dir: str = None, + overwrite_cache: bool = False, + whitelist: str = None, + ): + super().__init__(name="tokenize_and_classify", kind="classify", deterministic=deterministic) + far_file = None + if cache_dir is not None and cache_dir != "None": + os.makedirs(cache_dir, exist_ok=True) + whitelist_file = os.path.basename(whitelist) if whitelist else "" + far_file = os.path.join( + cache_dir, f"_{input_case}_es_tn_{deterministic}_deterministic{whitelist_file}.far" + ) + if not overwrite_cache and far_file and os.path.exists(far_file): + self.fst = pynini.Far(far_file, mode="r")["tokenize_and_classify"] + logging.info(f"ClassifyFst.fst was restored from {far_file}.") + else: + logging.info(f"Creating ClassifyFst grammars. This might take some time...") + + self.cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = self.cardinal.fst + + self.ordinal = OrdinalFst(cardinal=self.cardinal, deterministic=deterministic) + ordinal_graph = self.ordinal.fst + + self.decimal = DecimalFst(cardinal=self.cardinal, deterministic=deterministic) + decimal_graph = self.decimal.fst + + self.fraction = FractionFst(cardinal=self.cardinal, ordinal=self.ordinal, deterministic=deterministic) + fraction_graph = self.fraction.fst + self.measure = MeasureFst( + cardinal=self.cardinal, decimal=self.decimal, fraction=self.fraction, deterministic=deterministic + ) + measure_graph = self.measure.fst + self.date = DateFst(cardinal=self.cardinal, deterministic=deterministic) + date_graph = self.date.fst + word_graph = WordFst(deterministic=deterministic).fst + self.time = TimeFst(self.cardinal, deterministic=deterministic) + time_graph = self.time.fst + self.telephone = TelephoneFst(deterministic=deterministic) + telephone_graph = self.telephone.fst + self.electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = self.electronic.fst + self.money = MoneyFst(cardinal=self.cardinal, decimal=self.decimal, deterministic=deterministic) + money_graph = self.money.fst + self.whitelist = WhiteListFst(input_case=input_case, deterministic=deterministic, input_file=whitelist) + whitelist_graph = self.whitelist.fst + punct_graph = PunctuationFst(deterministic=deterministic).fst + + classify = ( + pynutil.add_weight(whitelist_graph, 1.01) + | pynutil.add_weight(time_graph, 1.09) + | pynutil.add_weight(measure_graph, 1.08) + | pynutil.add_weight(cardinal_graph, 1.1) + | pynutil.add_weight(fraction_graph, 1.09) + | pynutil.add_weight(date_graph, 1.1) + | pynutil.add_weight(ordinal_graph, 1.1) + | pynutil.add_weight(decimal_graph, 1.1) + | pynutil.add_weight(money_graph, 1.1) + | pynutil.add_weight(telephone_graph, 1.1) + | pynutil.add_weight(electronic_graph, 1.1) + | pynutil.add_weight(word_graph, 200) + ) + punct = pynutil.insert("tokens { ") + pynutil.add_weight(punct_graph, weight=2.1) + pynutil.insert(" }") + punct = pynini.closure( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct), + 1, + ) + token = pynutil.insert("tokens { ") + classify + pynutil.insert(" }") + token_plus_punct = ( + pynini.closure(punct + pynutil.insert(" ")) + token + pynini.closure(pynutil.insert(" ") + punct) + ) + + graph = token_plus_punct + pynini.closure( + ( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct + pynutil.insert(" ")) + ) + + token_plus_punct + ) + + graph = delete_space + graph + delete_space + graph |= punct + + self.fst = graph.optimize() + + if far_file: + generator_main(far_file, {"tokenize_and_classify": self.fst}) + logging.info(f"ClassifyFst grammars are saved to {far_file}.") diff --git a/nemo_text_processing/text_normalization/es/taggers/whitelist.py b/nemo_text_processing/text_normalization/es/taggers/whitelist.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/whitelist.py @@ -0,0 +1,69 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, convert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WhiteListFst(GraphFst): + """ + Finite state transducer for classifying whitelist, e.g. + "sr." -> tokens { name: "seรฑor" } + This class has highest priority among all classifier grammars. Whitelisted tokens are defined and loaded from "data/whitelist.tsv". + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + input_file: path to a file with whitelist replacements + """ + + def __init__(self, input_case: str, deterministic: bool = True, input_file: str = None): + super().__init__(name="whitelist", kind="classify", deterministic=deterministic) + + def _get_whitelist_graph(input_case, file): + whitelist = load_labels(file) + if input_case == "lower_cased": + whitelist = [[x[0].lower()] + x[1:] for x in whitelist] + graph = pynini.string_map(whitelist) + return graph + + graph = _get_whitelist_graph(input_case, get_abs_path("data/whitelist.tsv")) + if not deterministic and input_case != "lower_cased": + graph |= pynutil.add_weight( + _get_whitelist_graph("lower_cased", get_abs_path("data/whitelist.tsv")), weight=0.0001 + ) + + if input_file: + whitelist_provided = _get_whitelist_graph(input_case, input_file) + if not deterministic: + graph |= whitelist_provided + else: + graph = whitelist_provided + + if not deterministic: + units_graph = _get_whitelist_graph(input_case, file=get_abs_path("data/measures/measurements.tsv")) + graph |= units_graph + + self.graph = graph + self.final_graph = convert_space(self.graph).optimize() + self.fst = (pynutil.insert("name: \"") + self.final_graph + pynutil.insert("\"")).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/word.py b/nemo_text_processing/text_normalization/es/taggers/word.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/word.py @@ -0,0 +1,39 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_SPACE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WordFst(GraphFst): + """ + Finite state transducer for classifying word. + e.g. dormir -> tokens { name: "dormir" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="word", kind="classify") + word = pynutil.insert("name: \"") + pynini.closure(NEMO_NOT_SPACE, 1) + pynutil.insert("\"") + self.fst = word.optimize() diff --git a/nemo_text_processing/text_normalization/es/utils.py b/nemo_text_processing/text_normalization/es/utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/utils.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import csv +import os + + +def get_abs_path(rel_path): + """ + Get absolute path + + Args: + rel_path: relative path to this file + + Returns absolute path + """ + return os.path.dirname(os.path.abspath(__file__)) + '/' + rel_path + + +def load_labels(abs_path): + """ + loads relative path file as dictionary + + Args: + abs_path: absolute path + + Returns dictionary of mappings + """ + label_tsv = open(abs_path) + labels = list(csv.reader(label_tsv, delimiter="\t")) + return labels diff --git a/nemo_text_processing/text_normalization/es/verbalizers/__init__.py b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py @@ -0,0 +1,57 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_cardinal_gender, strip_cardinal_apocope + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class CardinalFst(GraphFst): + """ + Finite state transducer for verbalizing cardinals + e.g. cardinal { integer: "dos" } -> "dos" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="verbalize", deterministic=deterministic) + optional_sign = pynini.closure(pynini.cross("negative: \"true\" ", "menos "), 0, 1) + self.optional_sign = optional_sign + + integer = pynini.closure(NEMO_NOT_QUOTE, 1) + self.integer = pynutil.delete(" \"") + integer + pynutil.delete("\"") + + integer = pynutil.delete("integer:") + self.integer + self.numbers = integer + graph = optional_sign + self.numbers + + if not deterministic: + # For alternate renderings + no_adjust = graph + fem_adjust = shift_cardinal_gender(graph) + apocope_adjust = strip_cardinal_apocope(graph) + graph = no_adjust | fem_adjust | apocope_adjust + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/date.py b/nemo_text_processing/text_normalization/es/verbalizers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/date.py @@ -0,0 +1,86 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.taggers.date import articles + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for verbalizing date, e.g. + date { day: "treinta y uno" month: "marzo" year: "dos mil" } -> "treinta y uno de marzo de dos mil" + date { day: "uno" month: "mayo" year: "del mil novecientos noventa" } -> "primero de mayo del mil novecientos noventa" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="date", kind="verbalize", deterministic=deterministic) + + day_cardinal = pynutil.delete("day: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + day = strip_cardinal_apocope(day_cardinal) + + primero = pynini.cdrewrite(pynini.cross("uno", "primero"), "[BOS]", "[EOS]", NEMO_SIGMA) + day = ( + (day @ primero) if deterministic else pynini.union(day, day @ primero) + ) # Primero for first day is traditional, but will vary depending on region + + month = pynutil.delete("month: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + year = ( + pynutil.delete("year: \"") + + articles + + NEMO_SPACE + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # Insert preposition if wasn't originally with the year. This would mean a space was present + year = pynutil.add_weight(year, -0.001) + year |= ( + pynutil.delete("year: \"") + + pynutil.insert("de ") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # day month year + graph_dmy = day + pynini.cross(NEMO_SPACE, " de ") + month + pynini.closure(pynini.accep(" ") + year, 0, 1) + + graph_mdy = month + NEMO_SPACE + day + pynini.closure(NEMO_SPACE + year, 0, 1) + if deterministic: + graph_mdy += pynutil.delete(" preserve_order: true") # Only accepts this if was explicitly passed + + self.graph = graph_dmy | graph_mdy + final_graph = self.graph + delete_preserve_order + + delete_tokens = self.delete_tokens(final_graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/decimals.py b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py @@ -0,0 +1,87 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + decimal { negative: "true" integer_part: "dos" fractional_part: "cuatro cero" quantity: "billones" } -> menos dos coma quatro cero billones + decimal { integer_part: "un" quantity: "billรณn" } -> un billรณn + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + + self.optional_sign = pynini.closure(pynini.cross("negative: \"true\"", "menos ") + delete_space, 0, 1) + self.integer = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + self.fractional_default = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + conjunction = pynutil.insert(" punto ") if LOCALIZATION == "am" else pynutil.insert(" coma ") + if not deterministic: + conjunction |= pynutil.insert(pynini.union(" con ", " y ")) + self.fractional_default |= strip_cardinal_apocope(self.fractional_default) + self.fractional = conjunction + self.fractional_default + + self.quantity = ( + delete_space + + insert_space + + pynutil.delete("quantity: \"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + self.optional_quantity = pynini.closure(self.quantity, 0, 1) + + graph = self.optional_sign + pynini.union( + (self.integer + self.quantity), (self.integer + delete_space + self.fractional + self.optional_quantity) + ) + + self.numbers = graph.optimize() + self.numbers_no_quantity = self.integer + delete_space + self.fractional + self.optional_quantity + + if not deterministic: + graph |= self.optional_sign + ( + shift_cardinal_gender(self.integer + delete_space) + shift_number_gender(self.fractional) + ) + + graph += delete_preserve_order + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/electronic.py b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py @@ -0,0 +1,91 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit_no_zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + graph_symbols = pynini.string_file(get_abs_path("data/electronic/symbols.tsv")) + server_common = pynini.string_file(get_abs_path("data/electronic/server_name.tsv")) + domain_common = pynini.string_file(get_abs_path("data/electronic/domain.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digit_no_zero = None + zero = None + + graph_symbols = None + server_common = None + domain_common = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for verbalizing electronic + e.g. electronic { username: "abc" domain: "hotmail.com" } -> "a b c arroba hotmail punto com" + -> "a b c arroba h o t m a i l punto c o m" + -> "a b c arroba hotmail punto c o m" + -> "a b c at h o t m a i l punto com" + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="verbalize", deterministic=deterministic) + + graph_digit_no_zero = ( + digit_no_zero @ pynini.cdrewrite(pynini.cross("un", "uno"), "", "", NEMO_SIGMA).optimize() + ) + graph_digit = graph_digit_no_zero | zero + + def add_space_after_char(): + return pynini.closure(NEMO_NOT_QUOTE - pynini.accep(" ") + insert_space) + ( + NEMO_NOT_QUOTE - pynini.accep(" ") + ) + + verbalize_characters = pynini.cdrewrite(graph_symbols | graph_digit, "", "", NEMO_SIGMA) + + user_name = pynutil.delete("username: \"") + add_space_after_char() + pynutil.delete("\"") + user_name @= verbalize_characters + + convert_defaults = pynutil.add_weight(NEMO_NOT_QUOTE, weight=0.0001) | domain_common | server_common + domain = convert_defaults + pynini.closure(insert_space + convert_defaults) + domain @= verbalize_characters + + domain = pynutil.delete("domain: \"") + domain + pynutil.delete("\"") + protocol = ( + pynutil.delete("protocol: \"") + + add_space_after_char() @ pynini.cdrewrite(graph_symbols, "", "", NEMO_SIGMA) + + pynutil.delete("\"") + ) + self.graph = (pynini.closure(protocol + pynini.accep(" "), 0, 1) + domain) | ( + user_name + pynini.accep(" ") + pynutil.insert("arroba ") + domain + ) + delete_tokens = self.delete_tokens(self.graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/fraction.py b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_NOT_QUOTE, + NEMO_NOT_SPACE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + accents, + shift_cardinal_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for verbalizing fraction + e.g. tokens { fraction { integer: "treinta y tres" numerator: "cuatro" denominator: "quinto" } } -> + treinta y tres y cuatro quintos + + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="fraction", kind="verbalize", deterministic=deterministic) + + # Derivational strings append 'avo' as a suffix. Adding space for processing aid + fraction_stem = pynutil.insert(" avo") + plural = pynutil.insert("s") + + integer = ( + pynutil.delete("integer_part: \"") + + strip_cardinal_apocope(pynini.closure(NEMO_NOT_QUOTE)) + + pynutil.delete("\"") + ) + + numerator_one = pynutil.delete("numerator: \"") + pynini.accep("un") + pynutil.delete("\" ") + numerator = ( + pynutil.delete("numerator: \"") + + pynini.difference(pynini.closure(NEMO_NOT_QUOTE), "un") + + pynutil.delete("\" ") + ) + + denominator_add_stem = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + + fraction_stem + + pynutil.delete("\" morphosyntactic_features: \"add_root\"") + ) + denominator_ordinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\" morphosyntactic_features: \"ordinal\"") + ) + denominator_cardinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + ) + + denominator_singular = pynini.union(denominator_add_stem, denominator_ordinal) + denominator_plural = denominator_singular + plural + + if not deterministic: + # Occasional exceptions + denominator_singular |= denominator_add_stem @ pynini.string_map( + [("once avo", "undรฉcimo"), ("doce avo", "duodรฉcimo")] + ) + + # Merging operations + merge = pynini.cdrewrite( + pynini.cross(" y ", "i"), "", "", NEMO_SIGMA + ) # The denominator must be a single word, with the conjunction "y" replaced by i + merge @= pynini.cdrewrite(delete_space, "", pynini.difference(NEMO_CHAR, "parte"), NEMO_SIGMA) + + # The merger can produce duplicate vowels. This is not allowed in orthography + delete_duplicates = pynini.string_map([("aa", "a"), ("oo", "o")]) # Removes vowels + delete_duplicates = pynini.cdrewrite(delete_duplicates, "", "", NEMO_SIGMA) + + remove_accents = pynini.cdrewrite( + accents, + pynini.union(NEMO_SPACE, pynini.accep("[BOS]")) + pynini.closure(NEMO_NOT_SPACE), + pynini.closure(NEMO_NOT_SPACE) + pynini.union("avo", "ava", "รฉsimo", "รฉsima"), + NEMO_SIGMA, + ) + merge_into_single_word = merge @ remove_accents @ delete_duplicates + + fraction_default = numerator + delete_space + insert_space + (denominator_plural @ merge_into_single_word) + fraction_with_one = ( + numerator_one + delete_space + insert_space + (denominator_singular @ merge_into_single_word) + ) + + fraction_with_cardinal = strip_cardinal_apocope(numerator | numerator_one) + fraction_with_cardinal += ( + delete_space + pynutil.insert(" sobre ") + strip_cardinal_apocope(denominator_cardinal) + ) + + conjunction = pynutil.insert(" y ") + + if not deterministic: + # There is an alternative rendering where ordinals act as adjectives for 'parte'. This requires use of the feminine + # Other rules will manage use of "un" at end, so just worry about endings + exceptions = pynini.string_map([("tercia", "tercera")]) + apply_exceptions = pynini.cdrewrite(exceptions, "", "", NEMO_SIGMA) + vowel_change = pynini.cdrewrite(pynini.cross("o", "a"), "", pynini.accep("[EOS]"), NEMO_SIGMA) + + denominator_singular_fem = shift_cardinal_gender(denominator_singular) @ vowel_change @ apply_exceptions + denominator_plural_fem = denominator_singular_fem + plural + + numerator_one_fem = shift_cardinal_gender(numerator_one) + numerator_fem = shift_cardinal_gender(numerator) + + fraction_with_cardinal |= ( + (numerator_one_fem | numerator_fem) + + delete_space + + pynutil.insert(" sobre ") + + shift_cardinal_gender(denominator_cardinal) + ) + + # Still need to manage stems + merge_stem = pynini.cdrewrite( + delete_space, "", pynini.union("avo", "ava", "avos", "avas"), NEMO_SIGMA + ) # For managing alternative spacing + merge_stem @= remove_accents @ delete_duplicates + + fraction_with_one_fem = numerator_one_fem + delete_space + insert_space + fraction_with_one_fem += pynini.union( + denominator_singular_fem @ merge_stem, denominator_singular_fem @ merge_into_single_word + ) # Both forms exists + fraction_with_one_fem @= pynini.cdrewrite(pynini.cross("una media", "media"), "", "", NEMO_SIGMA) + fraction_with_one_fem += pynutil.insert(" parte") + + fraction_default_fem = numerator_fem + delete_space + insert_space + fraction_default_fem += pynini.union( + denominator_plural_fem @ merge_stem, denominator_plural_fem @ merge_into_single_word + ) + fraction_default_fem += pynutil.insert(" partes") + + fraction_default |= ( + numerator + delete_space + insert_space + denominator_plural @ merge_stem + ) # Case of no merger + fraction_default |= fraction_default_fem + + fraction_with_one |= numerator_one + delete_space + insert_space + denominator_singular @ merge_stem + fraction_with_one |= fraction_with_one_fem + + # Integers are influenced by dominant noun, need to allow feminine forms as well + integer |= shift_cardinal_gender(integer) + + # Remove 'un medio' + fraction_with_one @= pynini.cdrewrite(pynini.cross("un medio", "medio"), "", "", NEMO_SIGMA) + + integer = pynini.closure(integer + delete_space + conjunction, 0, 1) + + fraction = fraction_with_one | fraction_default | fraction_with_cardinal + + graph = integer + fraction + + self.graph = graph + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/measure.py b/nemo_text_processing/text_normalization/es/verbalizers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/measure.py @@ -0,0 +1,110 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ones, shift_cardinal_gender +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + unit_singular_fem = pynini.project(unit_plural_fem, "input") + unit_singular_masc = pynini.project(unit_plural_masc, "input") + + unit_plural_fem = pynini.project(unit_plural_fem, "output") + unit_plural_masc = pynini.project(unit_plural_masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit_plural_fem = None + unit_plural_masc = None + + unit_singular_fem = None + unit_singular_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for verbalizing measure, e.g. + measure { cardinal { integer: "dos" units: "gramos" } } -> "dos gramos" + measure { cardinal { integer_part: "dos" quantity: "millones" units: "gramos" } } -> "dos millones de gramos" + + Args: + decimal: DecimalFst + cardinal: CardinalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, cardinal: GraphFst, fraction: GraphFst, deterministic: bool): + super().__init__(name="measure", kind="verbalize", deterministic=deterministic) + + graph_decimal = decimal.fst + graph_cardinal = cardinal.fst + graph_fraction = fraction.fst + + unit_masc = (unit_plural_masc | unit_singular_masc) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_masc |= "por" + pynini.closure(NEMO_NOT_QUOTE, 1) + unit_masc = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_masc) + pynutil.delete("\"") + + unit_fem = (unit_plural_fem | unit_singular_fem) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_fem = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_fem) + pynutil.delete("\"") + + graph_masc = (graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_masc + graph_fem = ( + shift_cardinal_gender(graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_fem + ) + graph = graph_masc | graph_fem + + graph = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph + ) # billones de xyz + + graph @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", NEMO_WHITE_SPACE + "por", NEMO_SIGMA) + + # To manage alphanumeric combonations ("a-8, 5x"), we let them use a weighted default path. + alpha_num_unit = pynutil.delete("units: \"") + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + graph_alpha_num = pynini.union( + (graph_cardinal | graph_decimal) + NEMO_SPACE + alpha_num_unit, + alpha_num_unit + delete_extra_space + (graph_cardinal | graph_decimal), + ) + + graph |= pynutil.add_weight(graph_alpha_num, 0.01) + + graph += delete_preserve_order + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/money.py b/nemo_text_processing/text_normalization/es/verbalizers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/money.py @@ -0,0 +1,195 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + fem = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + fem_singular = pynini.project(fem, "input") + masc_singular = pynini.project(masc, "input") + + fem_plural = pynini.project(fem, "output") + masc_plural = pynini.project(masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + fem_plural = None + masc_plural = None + + fem_singular = None + masc_singular = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for verbalizing money, e.g. + money { currency_maj: "euro" integer_part: "un"} -> "un euro" + money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un"} -> "uno coma cero cero uno euros" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true} -> "una libra cuarenta" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "peniques" preserve_order: true} -> "una libra con cuarenta peniques" + money { fractional_part: "un" currency_min: "penique" preserve_order: true} -> "un penique" + + Args: + decimal: GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="verbalize", deterministic=deterministic) + + maj_singular_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + maj_singular_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + maj_plural_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + maj_plural_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + maj_masc = maj_plural_masc | maj_singular_masc # Tagger kept quantity resolution stable + maj_fem = maj_plural_fem | maj_singular_fem + + min_singular_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + min_singular_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + min_plural_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + min_plural_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + min_masc = min_plural_masc | min_singular_masc + min_fem = min_plural_fem | min_singular_fem + + fractional_part = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + integer_part = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + optional_add_and = pynini.closure(pynutil.insert(pynini.union("con ", "y ")), 0, 1) + + # *** currency_maj + graph_integer_masc = integer_part + NEMO_SPACE + maj_masc + graph_integer_fem = shift_cardinal_gender(integer_part) + NEMO_SPACE + maj_fem + graph_integer = graph_integer_fem | graph_integer_masc + + # *** currency_maj + (***) | ((con) *** current_min) + graph_integer_with_minor_masc = ( + integer_part + + NEMO_SPACE + + maj_masc + + NEMO_SPACE + + pynini.union( + optional_add_and + strip_cardinal_apocope(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor_fem = ( + shift_cardinal_gender(integer_part) + + NEMO_SPACE + + maj_fem + + NEMO_SPACE + + pynini.union( + optional_add_and + shift_cardinal_gender(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor = graph_integer_with_minor_fem | graph_integer_with_minor_masc + + # *** coma *** currency_maj + graph_decimal_masc = decimal.numbers + NEMO_SPACE + maj_masc + + # Need to fix some of the inner parts, so don't use decimal here (note: quantities covered by masc) + graph_decimal_fem = ( + pynini.accep("integer_part: \"") + + shift_cardinal_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SPACE + + pynini.accep("fractional_part: \"") + + shift_number_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SIGMA + ) + graph_decimal_fem @= decimal.numbers_no_quantity + graph_decimal_fem += NEMO_SPACE + maj_fem + + graph_decimal = graph_decimal_fem | graph_decimal_masc + graph_decimal = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph_decimal + ) # formally it's millones/billones de *** + + # *** current_min + graph_minor_masc = fractional_part + NEMO_SPACE + min_masc + delete_preserve_order + graph_minor_fem = shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem + delete_preserve_order + graph_minor = graph_minor_fem | graph_minor_masc + + graph = graph_integer | graph_integer_with_minor | graph_decimal | graph_minor + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py @@ -0,0 +1,76 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, NEMO_SIGMA, NEMO_SPACE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_number_gender + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for verbalizing ordinals + e.g. ordinal { integer: "tercer" } } -> "tercero" + -> "tercera" + -> "tercer" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="ordinal", kind="verbalize", deterministic=deterministic) + + graph = pynutil.delete("integer: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + # masculne gender we leave as is + graph_masc = graph + pynutil.delete(" morphosyntactic_features: \"gender_masc") + + # shift gender + graph_fem_ending = graph @ pynini.cdrewrite( + pynini.cross("o", "a"), "", NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fem = shift_number_gender(graph_fem_ending) + pynutil.delete(" morphosyntactic_features: \"gender_fem") + + # Apocope just changes tercero and primero. May occur if someone wrote 11.er (uncommon) + graph_apocope = ( + pynini.cross("tercero", "tercer") + | pynini.cross("primero", "primer") + | pynini.cross("undรฉcimo", "decimoprimer") + ) # In case someone wrote 11.er with deterministic + graph_apocope = (graph @ pynini.cdrewrite(graph_apocope, "", "", NEMO_SIGMA)) + pynutil.delete( + " morphosyntactic_features: \"apocope" + ) + + graph = graph_apocope | graph_masc | graph_fem + + if not deterministic: + # Plural graph + graph_plural = pynini.cdrewrite( + pynutil.insert("s"), pynini.union("o", "a"), NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + + graph |= (graph @ graph_plural) + pynutil.delete("/plural") + + self.graph = graph + pynutil.delete("\"") + + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/telephone.py b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for verbalizing telephone, e.g. + telephone { number_part: "uno dos tres uno dos tres cinco seis siete ocho" } + -> uno dos tres uno dos tres cinco seis siete ocho + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="verbalize") + + number_part = pynutil.delete("number_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + delete_tokens = self.delete_tokens(number_part) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/time.py b/nemo_text_processing/text_normalization/es/verbalizers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/time.py @@ -0,0 +1,269 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + alt_minutes = pynini.string_file(get_abs_path("data/time/alt_minutes.tsv")) + + morning_times = pynini.string_file(get_abs_path("data/time/morning_times.tsv")) + afternoon_times = pynini.string_file(get_abs_path("data/time/afternoon_times.tsv")) + evening_times = pynini.string_file(get_abs_path("data/time/evening_times.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + alt_minutes = None + + morning_times = None + afternoon_times = None + evening_times = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for verbalizing time, e.g. + time { hours: "doce" minutes: "media" suffix: "a m" } -> doce y media de la noche + time { hours: "doce" } -> twelve o'clock + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="time", kind="verbalize", deterministic=deterministic) + + change_minutes = pynini.cdrewrite(alt_minutes, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA) + + morning_phrases = pynini.cross("am", "de la maรฑana") + afternoon_phrases = pynini.cross("pm", "de la tarde") + evening_phrases = pynini.cross("pm", "de la noche") + + # For the 12's + mid_times = pynini.accep("doce") + mid_phrases = ( + pynini.string_map([("pm", "del mediodรญa"), ("am", "de la noche")]) + if deterministic + else pynini.string_map( + [ + ("pm", "de la maรฑana"), + ("pm", "del dรญa"), + ("pm", "del mediodรญa"), + ("am", "de la noche"), + ("am", "de la medianoche"), + ] + ) + ) + + hour = ( + pynutil.delete("hours:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = ( + pynutil.delete("minutes:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = (minute @ change_minutes) if deterministic else pynini.union(minute, minute @ change_minutes) + + suffix = ( + pynutil.delete("suffix:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + zone = ( + pynutil.delete("zone:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + optional_zone = pynini.closure(delete_space + insert_space + zone, 0, 1) + second = ( + pynutil.delete("seconds:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + graph_hms = ( + hour + + pynutil.insert(" horas ") + + delete_space + + minute + + pynutil.insert(" minutos y ") + + delete_space + + second + + pynutil.insert(" segundos") + ) + + graph_hm = hour + delete_space + pynutil.insert(" y ") + minute + graph_hm |= pynini.union( + (hour @ morning_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases), + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases), + (hour @ evening_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases), + (hour @ mid_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases), + ) + + graph_h = pynini.union( + hour, + (hour @ morning_times) + delete_space + insert_space + (suffix @ morning_phrases), + (hour @ afternoon_times) + delete_space + insert_space + (suffix @ afternoon_phrases), + (hour @ evening_times) + delete_space + insert_space + (suffix @ evening_phrases), + (hour @ mid_times) + delete_space + insert_space + (suffix @ mid_phrases), + ) + + graph = (graph_hms | graph_hm | graph_h) + optional_zone + + if not deterministic: + graph_style_1 = pynutil.delete(" style: \"1\"") + graph_style_2 = pynutil.delete(" style: \"2\"") + + graph_menos = hour + delete_space + pynutil.insert(" menos ") + minute + graph_style_1 + graph_menos |= ( + (hour @ morning_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ evening_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ mid_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_1 + ) + graph_menos += optional_zone + + graph_para = minute + pynutil.insert(" para las ") + delete_space + hour + graph_style_2 + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ morning_times) + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ afternoon_times) + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ evening_times) + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ mid_times) + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_2 + ) + graph_para += optional_zone + graph_para @= pynini.cdrewrite( + pynini.cross(" las ", " la "), "para", "una", NEMO_SIGMA + ) # Need agreement with one + + graph |= graph_menos | graph_para + delete_tokens = self.delete_tokens(graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py @@ -0,0 +1,73 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst +from nemo_text_processing.text_normalization.en.verbalizers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.verbalizers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.verbalizers.date import DateFst +from nemo_text_processing.text_normalization.es.verbalizers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.verbalizers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.verbalizers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.verbalizers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.verbalizers.money import MoneyFst +from nemo_text_processing.text_normalization.es.verbalizers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.verbalizers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.verbalizers.time import TimeFst + + +class VerbalizeFst(GraphFst): + """ + Composes other verbalizer grammars. + For deployment, this grammar will be compiled and exported to OpenFst Finate State Archiv (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize", kind="verbalize", deterministic=deterministic) + cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = cardinal.fst + ordinal = OrdinalFst(deterministic=deterministic) + ordinal_graph = ordinal.fst + decimal = DecimalFst(deterministic=deterministic) + decimal_graph = decimal.fst + fraction = FractionFst(deterministic=deterministic) + fraction_graph = fraction.fst + date = DateFst(deterministic=deterministic) + date_graph = date.fst + measure = MeasureFst(cardinal=cardinal, decimal=decimal, fraction=fraction, deterministic=deterministic) + measure_graph = measure.fst + electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = electronic.fst + whitelist_graph = WhiteListFst(deterministic=deterministic).fst + money_graph = MoneyFst(decimal=decimal, deterministic=deterministic).fst + telephone_graph = TelephoneFst(deterministic=deterministic).fst + time_graph = TimeFst(deterministic=deterministic).fst + + graph = ( + cardinal_graph + | measure_graph + | decimal_graph + | ordinal_graph + | date_graph + | electronic_graph + | money_graph + | fraction_graph + | whitelist_graph + | telephone_graph + | time_graph + ) + self.fst = graph diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py @@ -0,0 +1,52 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, delete_extra_space, delete_space +from nemo_text_processing.text_normalization.en.verbalizers.word import WordFst +from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class VerbalizeFinalFst(GraphFst): + """ + Finite state transducer that verbalizes an entire sentence + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize_final", kind="verbalize", deterministic=deterministic) + verbalize = VerbalizeFst(deterministic=deterministic).fst + word = WordFst(deterministic=deterministic).fst + types = verbalize | word + graph = ( + pynutil.delete("tokens") + + delete_space + + pynutil.delete("{") + + delete_space + + types + + delete_space + + pynutil.delete("}") + ) + graph = delete_space + pynini.closure(graph + delete_extra_space) + graph + delete_space + self.fst = graph diff --git a/nemo_text_processing/text_normalization/normalize.py b/nemo_text_processing/text_normalization/normalize.py --- a/nemo_text_processing/text_normalization/normalize.py +++ b/nemo_text_processing/text_normalization/normalize.py @@ -46,8 +46,8 @@ class Normalizer: """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -83,10 +83,11 @@ def __init__( from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst elif lang == 'de': - # Ru TN only support non-deterministic cases and produces multiple normalization options - # use normalize_with_audio.py from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst + elif lang == 'es': + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ClassifyFst + from nemo_text_processing.text_normalization.es.verbalizers.verbalize_final import VerbalizeFinalFst self.tagger = ClassifyFst( input_case=input_case, deterministic=deterministic, @@ -106,7 +107,7 @@ def __init__( def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: """ - NeMo text normalizer + NeMo text normalizer Args: texts: list of input strings @@ -357,7 +358,7 @@ def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: def parse_args(): parser = ArgumentParser() parser.add_argument("input_string", help="input string", type=str) - parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) + parser.add_argument("--language", help="language", choices=["en", "de", "es"], default="en", type=str) parser.add_argument( "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) diff --git a/nemo_text_processing/text_normalization/normalize_with_audio.py b/nemo_text_processing/text_normalization/normalize_with_audio.py --- a/nemo_text_processing/text_normalization/normalize_with_audio.py +++ b/nemo_text_processing/text_normalization/normalize_with_audio.py @@ -55,15 +55,15 @@ "audio_data" - path to the audio file "text" - raw text "pred_text" - ASR model prediction - + See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions - + When the manifest is ready, run: python normalize_with_audio.py \ --audio_data PATH/TO/MANIFEST.JSON \ - --language en - - + --language en + + To run with a single audio file, specify path to audio and text with: python normalize_with_audio.py \ --audio_data PATH/TO/AUDIO.WAV \ @@ -71,18 +71,18 @@ --text raw text OR PATH/TO/.TXT/FILE --model QuartzNet15x5Base-En \ --verbose - + To see possible normalization options for a text input without an audio file (could be used for debugging), run: python python normalize_with_audio.py --text "RAW TEXT" - + Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference """ class NormalizerWithAudio(Normalizer): """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -282,7 +282,7 @@ def parse_args(): "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) parser.add_argument( - "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str + "--language", help="Select target language", choices=["en", "ru", "de", "es"], default="en", type=str ) parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") parser.add_argument( diff --git a/tools/text_processing_deployment/pynini_export.py b/tools/text_processing_deployment/pynini_export.py --- a/tools/text_processing_deployment/pynini_export.py +++ b/tools/text_processing_deployment/pynini_export.py @@ -67,7 +67,7 @@ def tn_grammars(**kwargs): def export_grammars(output_dir, grammars): """ - Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. + Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. Args: output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. @@ -109,7 +109,7 @@ def parse_args(): if __name__ == '__main__': args = parse_args() - if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': + if args.language in ['ru', 'fr', 'vi'] and args.grammars == 'tn_grammars': raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') if args.language == 'en': @@ -148,6 +148,10 @@ def parse_args(): from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ( + ClassifyFst as TNClassifyFst, + ) + from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst elif args.language == 'fr': from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, </patch> </s> </patch>
diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt @@ -0,0 +1,86 @@ +1~un +2~dos +3~tres +4~cuatro +5~cinco +6~seis +7~siete +8~ocho +9~nueve +10~diez +11~once +12~doce +13~trece +14~catorce +15~quince +16~diecisรฉis +17~diecisiete +18~dieciocho +19~diecinueve +20~veinte +21~veintiรบn +22~veintidรณs +23~veintitrรฉs +24~veinticuatro +25~veinticinco +26~veintisรฉis +27~veintisiete +28~veintiocho +29~veintinueve +30~treinta +31~treinta y un +40~cuarenta +41~cuarenta y un +50~cincuenta +51~cincuenta y un +60~sesenta +70~setenta +80~ochenta +90~noventa +100~cien +101~ciento un +120~ciento veinte +121~ciento veintiรบn +130~ciento treinta +131~ciento treinta y un +200~doscientos +201~doscientos un +300~trescientos +301~trescientos un +1000~mil +1 000~mil +1.000~mil +1001~mil un +1010~mil diez +1020~mil veinte +1021~mil veintiรบn +1100~mil cien +1101~mil ciento un +1110~mil ciento diez +1111~mil ciento once +1234~mil doscientos treinta y cuatro +2000~dos mil +2001~dos mil un +2010~dos mil diez +2020~dos mil veinte +2100~dos mil cien +2101~dos mil ciento un +2110~dos mil ciento diez +2111~dos mil ciento once +2222~dos mil doscientos veintidรณs +10000~diez mil +10 000~diez mil +10.000~diez mil +100000~cien mil +100 000~cien mil +100.000~cien mil +1 000 000~un millรณn +1.000.000~un millรณn +1 234 568~un millรณn doscientos treinta y cuatro mil quinientos sesenta y ocho +2.000.000~dos millones +1.000.000.000~mil millones +2.000.000.000~dos mil millones +3 000 000 000 000~tres billones +3.000.000.000.000~tres billones +100 000 000 000 000 000 000 000~cien mil trillones +100 000 000 000 000 000 000 001~cien mil trillones un diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt @@ -0,0 +1,13 @@ +1 enero~primero de enero +5 febrero~cinco de febrero +20 de marzo~veinte de marzo +abril 30~treinta de abril +31 marzo~treinta y uno de marzo +10 mayo 1990~diez de mayo de mil novecientos noventa +junio 11 2000~once de junio de dos mil +30 julio del 2020~treinta de julio del dos mil veinte +30-2-1990~treinta de febrero de mil novecientos noventa +30/2/1990~treinta de febrero de mil novecientos noventa +30.2.1990~treinta de febrero de mil novecientos noventa +1990-2-30~treinta de febrero de mil novecientos noventa +1990-02-30~treinta de febrero de mil novecientos noventa \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt @@ -0,0 +1,27 @@ +0,1~cero coma un +0,01~cero coma cero un +0,010~cero coma cero uno cero +1,0101~uno coma cero uno cero un +0,0~cero coma cero +1,0~uno coma cero +1,00~uno coma cero cero +1,1~uno coma un +233,32~doscientos treinta y tres coma treinta y dos +32,22 millones~treinta y dos coma veintidรณs millones +320 320,22 millones~trescientos veinte mil trescientos veinte coma veintidรณs millones +5.002,232~cinco mil dos coma doscientos treinta y dos +3,2 trillones~tres coma dos trillones +3 millones~tres millones +3 000 millones~tres mil millones +3000 millones~tres mil millones +3.000 millones~tres mil millones +3.001 millones~tres mil un millones +1 millรณn~un millรณn +1 000 millones~mil millones +1000 millones~mil millones +1.000 millones~mil millones +2,33302 millones~dos coma tres tres tres cero dos millones +1,5332 millรณn~uno coma cinco tres tres dos millรณn +1,53322 millรณn~uno coma cinco tres tres dos dos millรณn +1,53321 millรณn~uno coma cinco tres tres dos un millรณn +101,010101 millones~ciento uno coma cero uno cero uno cero un millones \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt @@ -0,0 +1,12 @@ +a.bc@gmail.com~a punto b c arroba gmail punto com +cdf@abc.edu~c d f arroba a b c punto e d u +abc@gmail.abc~a b c arroba gmail punto a b c +abc@abc.com~a b c arroba a b c punto com +asdf123@abc.com~a s d f uno dos tres arroba a b c punto com +a1b2@abc.com~a uno b dos arroba a b c punto com +ab3.sdd.3@gmail.com~a b tres punto s d d punto tres arroba gmail punto com +https://www.nvidia.com~h t t p s dos puntos barra barra w w w punto nvidia punto com +www.nvidia.com~w w w punto nvidia punto com +www.abc.es/efg~w w w punto a b c punto es barra e f g +www.abc.es~w w w punto a b c punto es +http://www.ourdailynews.com.sm~h t t p dos puntos barra barra w w w punto o u r d a i l y n e w s punto com punto s m \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt @@ -0,0 +1,76 @@ +1/2~medio +1 1/2~uno y medio +3/2~tres medios +1 3/2~uno y tres medios +1/3~un tercio +2/3~dos tercios +1/4~un cuarto +2/4~dos cuartos +1/5~un quinto +2/5~dos quintos +1/6~un sexto +2/6~dos sextos +1/7~un sรฉptimo +2/7~dos sรฉptimos +1/8~un octavo +2/8~dos octavos +1/9~un noveno +2/9~dos novenos +1/10~un dรฉcimo +2/10~dos dรฉcimos +1/11~un onceavo +1/12~un doceavo +1/13~un treceavo +1/14~un catorceavo +1/15~un quinceavo +1/16~un dieciseisavo +1/17~un diecisieteavo +1/18~un dieciochoavo +1/19~un diecinueveavo +1/20~un veinteavo +1/21~un veintiunavo +1/22~un veintidosavo +1/30~un treintavo +1/31~un treintaiunavo +1/40~un cuarentavo +1/41~un cuarentaiunavo +1/50~un cincuentavo +1/60~un sesentavo +1/70~un setentavo +1/80~un ochentavo +1/90~un noventavo +1/100~un centรฉsimo +2/100~dos centรฉsimos +1 2/100~uno y dos centรฉsimos +1/101~uno sobre ciento uno +1/110~uno sobre ciento diez +1/111~uno sobre ciento once +1/112~uno sobre ciento doce +1/123~uno sobre ciento veintitrรฉs +1/134~uno sobre ciento treinta y cuatro +1/200~un ducentรฉsimo +1/201~uno sobre doscientos uno +1/234~uno sobre doscientos treinta y cuatro +1/300~un tricentรฉsimo +1/345~uno sobre trescientos cuarenta y cinco +1/400~un cuadringentรฉsimo +1/456~uno sobre cuatrocientos cincuenta y seis +1/500~un quingentรฉsimo +1/600~un sexcentรฉsimo +1/700~un septingentรฉsimo +1/800~un octingentรฉsimo +1/900~un noningentรฉsimo +1/1000~un milรฉsimo +2/1000~dos milรฉsimos +1 2/1000~uno y dos milรฉsimos +1/1001~uno sobre mil uno +1/1100~uno sobre mil cien +1/1200~uno sobre mil doscientos +1/1234~uno sobre mil doscientos treinta y cuatro +1/2000~un dosmilรฉsimo +1/5000~un cincomilรฉsimo +1/10000~un diezmilรฉsimo +1/100.000~un cienmilรฉsimo +1/1.000.000~un millonรฉsimo +1/100.000.000~un cienmillonรฉsimo +1/1.200.000.000~un mildoscientosmillonรฉsimo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt @@ -0,0 +1,17 @@ +1,2-a~uno coma dos a +a-5~a cinco +200 m~doscientos metros +3 h~tres horas +1 h~una hora +245 mph~doscientas cuarenta y cinco millas por hora +2 kg~dos kilogramos +60,2400 kg~sesenta coma dos cuatro cero cero kilogramos +-60,2400 kg~menos sesenta coma dos cuatro cero cero kilogramos +8,52 %~ocho coma cincuenta y dos por ciento +-8,52 %~menos ocho coma cincuenta y dos por ciento +1 %~uno por ciento +3 cm~tres centรญmetros +4 s~cuatro segundos +5 l~cinco litros +4,51/s~cuatro coma cincuenta y uno por segundo +0,0101 s~cero coma cero uno cero un segundos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt @@ -0,0 +1,24 @@ +$1~un dรณlar +1 $~un dรณlar +$1,50~un dรณlar cincuenta centavos +1,50 $~un dรณlar cincuenta centavos +ยฃ200.000.001~doscientos millones una libras +200.000.001 ยฃ~doscientos millones una libras +2 billones de euros~dos billones de euros +โ‚ฌ2 billones~dos billones de euros +โ‚ฌ 2 billones~dos billones de euros +โ‚ฌ 2,3 billones~dos coma tres billones de euros +2,3 billones de euros~dos coma tres billones de euros +โ‚ฌ5,50~cinco euros cincuenta cรฉntimos +5,50 โ‚ฌ~cinco euros cincuenta cรฉntimos +5,01 โ‚ฌ~cinco euros un cรฉntimo +5,01 ยฃ~cinco libras un penique +21 czk~veintiuna coronas checas +czk21~veintiuna coronas checas +czk21,1 millones~veintiuna coma una millones de coronas checas +czk 5,50 billones~cinco coma cincuenta billones de coronas checas +rs 5,50 billones~cinco coma cincuenta billones de rupias +czk5,50 billones~cinco coma cincuenta billones de coronas checas +0,55 $~cincuenta y cinco centavos +1,01 $~un dรณlar un centavo +ยฅ12,05~doce yenes cinco centavos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt @@ -0,0 +1,120 @@ +~121 +ciento veintiรบn +ciento veintiuno +ciento veintiuna +121 +~200 +doscientos +doscientas +200 +~201 +doscientos un +doscientos uno +doscientas una +201 +~1 +un +uno +una +1 +~550.000.001 +quinientos cincuenta millones un +quinientos cincuenta millones una +quinientos cincuenta millones uno +550.000.001 +~500.501 +quinientos mil quinientos un +quinientos mil quinientos uno +quinientas mil quinientas una +500.501 +~500.001.ยบ +quinientosmilรฉsimo primero +quingentรฉsimo milรฉsimo primero +quinientosmilรฉsimos primeros +quingentรฉsimos milรฉsimos primeros +500.001.ยบ +~500.001.ยช +quinientasmilรฉsima primera +quingentรฉsima milรฉsima primera +quinientasmilรฉsimas primeras +quingentรฉsimas milรฉsimas primeras +500.001.ยช +~11.ยช +dรฉcima primera +decimoprimera +dรฉcimas primeras +decimoprimeras +undรฉcima +undรฉcimas +11.ยช +~11.ยบ +dรฉcimo primero +decimoprimero +dรฉcimos primeros +decimoprimeros +undรฉcimo +undรฉcimos +11.ยบ +~12.ยบ +dรฉcimo segundo +decimosegundo +dรฉcimos segundos +decimosegundos +duodรฉcimo +duodรฉcimos +12.ยบ +~200,0101 +doscientos coma cero uno cero un +doscientos coma cero uno cero uno +doscientas coma cero una cero una +200,0101 +~1.000.200,21 +un millรณn doscientos coma veintiรบn +un millรณn doscientos coma veintiuno +un millรณn doscientas coma veintiuna +un millรณn doscientos coma dos un +un millรณn doscientos coma dos uno +un millรณn doscientas coma dos una +1.000.200,21 +~1/12 +un doceavo +una doceava parte +un duodรฉcimo +una duodรฉcima parte +uno sobre doce +1/12 +~5/200 +cinco ducentรฉsimos +cinco ducentรฉsimas partes +cinco sobre doscientos +5/200 +~1 5/3 +uno y cinco tercios +una y cinco terceras partes +uno y cinco sobre tres +una y cinco sobre tres +~1/5/2020 +primero de mayo de dos mil veinte +uno de mayo de dos mil veinte +cinco de enero de dos mil veinte +~$5,50 +cinco dรณlares con cincuenta +cinco dรณlares y cincuenta +cinco dรณlares cincuenta +cinco dรณlares con cincuenta centavos +cinco dรณlares y cincuenta centavos +cinco dรณlares cincuenta centavos +~2.30 h +dos y treinta +dos y media +tres menos treinta +tres menos media +treinta para las tres +~12.30 a.m. +doce y treinta de la medianoche +doce y treinta de la noche +doce y media de la medianoche +doce y media de la noche +una menos treinta de la maรฑana +una menos media de la maรฑana +treinta para la una de la maรฑana \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt @@ -0,0 +1,137 @@ +1.แต‰สณ~primer +1.ยบ~primero +1.ยช~primera +2.ยบ~segundo +2.ยช~segunda +ii~segundo +II~segundo +3.แต‰สณ~tercer +3.ยบ~tercero +3.ยช~tercera +4.ยบ~cuarto +4.ยช~cuarta +5.ยบ~quinto +5.ยช~quinta +6.ยบ~sexto +6.ยช~sexta +7.ยบ~sรฉptimo +7.ยช~sรฉptima +8.ยบ~octavo +8.ยช~octava +9.ยบ~noveno +9.ยช~novena +10.ยบ~dรฉcimo +10.ยช~dรฉcima +11.แต‰สณ~decimoprimer +11.ยบ~undรฉcimo +11.ยช~undรฉcima +12.ยบ~duodรฉcimo +12.ยช~duodรฉcima +13.แต‰สณ~decimotercer +13.ยบ~decimotercero +13.ยช~decimotercera +14.ยบ~decimocuarto +14.ยช~decimocuarta +15.ยบ~decimoquinto +15.ยช~decimoquinta +16.ยบ~decimosexto +16.ยช~decimosexta +17.ยบ~decimosรฉptimo +17.ยช~decimosรฉptima +18.ยบ~decimoctavo +18.ยช~decimoctava +19.ยบ~decimonoveno +19.ยช~decimonovena +20.ยบ~vigรฉsimo +20.ยช~vigรฉsima +21.แต‰สณ~vigesimoprimer +21.ยบ~vigesimoprimero +21.ยช~vigesimoprimera +30.ยบ~trigรฉsimo +30.ยช~trigรฉsima +31.แต‰สณ~trigรฉsimo primer +31.ยบ~trigรฉsimo primero +31.ยช~trigรฉsima primera +40.ยบ~cuadragรฉsimo +40.ยช~cuadragรฉsima +41.แต‰สณ~cuadragรฉsimo primer +41.ยบ~cuadragรฉsimo primero +41.ยช~cuadragรฉsima primera +50.ยบ~quincuagรฉsimo +50.ยช~quincuagรฉsima +51.แต‰สณ~quincuagรฉsimo primer +51.ยบ~quincuagรฉsimo primero +51.ยช~quincuagรฉsima primera +60.ยบ~sexagรฉsimo +60.ยช~sexagรฉsima +70.ยบ~septuagรฉsimo +70.ยช~septuagรฉsima +80.ยบ~octogรฉsimo +80.ยช~octogรฉsima +90.ยบ~nonagรฉsimo +90.ยช~nonagรฉsima +100.ยบ~centรฉsimo +100.ยช~centรฉsima +101.แต‰สณ~centรฉsimo primer +101.ยบ~centรฉsimo primero +101.ยช~centรฉsima primera +134.ยบ~centรฉsimo trigรฉsimo cuarto +134.ยช~centรฉsima trigรฉsima cuarta +200.ยบ~ducentรฉsimo +200.ยช~ducentรฉsima +300.ยบ~tricentรฉsimo +300.ยช~tricentรฉsima +400.ยบ~cuadringentรฉsimo +400.ยช~cuadringentรฉsima +500.ยบ~quingentรฉsimo +500.ยช~quingentรฉsima +600.ยบ~sexcentรฉsimo +600.ยช~sexcentรฉsima +700.ยบ~septingentรฉsimo +700.ยช~septingentรฉsima +800.ยบ~octingentรฉsimo +800.ยช~octingentรฉsima +900.ยบ~noningentรฉsimo +900.ยช~noningentรฉsima +1000.ยบ~milรฉsimo +1000.ยช~milรฉsima +1001.แต‰สณ~milรฉsimo primer +1 000.ยบ~milรฉsimo +1 000.ยช~milรฉsima +1 001.แต‰สณ~milรฉsimo primer +1.000.ยบ~milรฉsimo +1.000.ยช~milรฉsima +1.001.แต‰สณ~milรฉsimo primer +1248.ยบ~milรฉsimo ducentรฉsimo cuadragรฉsimo octavo +1248.ยช~milรฉsima ducentรฉsima cuadragรฉsima octava +2000.ยบ~dosmilรฉsimo +100 000.ยบ~cienmilรฉsimo +i~primero +I~primero +ii~segundo +II~segundo +iii~tercero +III~tercero +iv~cuarto +IV~cuarto +V~quinto +VI~sexto +VII~sรฉptimo +VIII~octavo +IX~noveno +X~dรฉcimo +XI~undรฉcimo +XII~duodรฉcimo +XIII~decimotercero +XX~vigรฉsimo +XXI~vigesimoprimero +XXX~trigรฉsimo +XL~cuadragรฉsimo +L~quincuagรฉsimo +XC~nonagรฉsimo +C~centรฉsimo +CD~cuadringentรฉsimo +D~quingentรฉsimo +CM~noningentรฉsimo +999.ยบ~noningentรฉsimo nonagรฉsimo noveno +cmxcix~noningentรฉsimo nonagรฉsimo noveno \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt @@ -0,0 +1,3 @@ +123-123-5678~uno dos tres uno dos tres cinco seis siete ocho +123-456-789~uno dos tres cuatro cinco seis siete ocho nueve +1234-5678~uno dos tres cuatro cinco seis siete ocho \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt @@ -0,0 +1,26 @@ +1.00~una +1:00~una +01:00~una +01 h~una +3 h~tres horas +1 h~una hora +1.05 h~una y cinco +01.05 h~una y cinco +1.00 h~una +1.00 a.m.~una de la maรฑana +1.00 a.m~una de la maรฑana +1.00 p.m.~una de la tarde +1.00 p.m est~una de la tarde e s t +1.00 est~una e s t +5:02 est~cinco y dos e s t +5:02 p.m pst~cinco y dos de la noche p s t +5:02 p.m.~cinco y dos de la noche +12.15~doce y cuarto +12.15 a.m.~doce y cuarto de la noche +12.15 p.m.~doce y cuarto del mediodรญa +13.30~trece y media +14.05~catorce y cinco +24:50~veinticuatro y cincuenta +3:02:32 pst~tres horas dos minutos y treinta y dos segundos p s t +00:52~cero y cincuenta y dos +0:52~cero y cincuenta y dos diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt @@ -0,0 +1,3 @@ +el dr.~el doctor +sr. rodriguez~seรฑor rodriguez +182 esq. toledo~ciento ochenta y dos esquina toledo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt @@ -0,0 +1,48 @@ +~ +yahoo!~yahoo! +veinte!~veinte! +โ€”~โ€” +aaa~aaa +aabach~aabach +aabenraa~aabenraa +aabye~aabye +aaccessed~aaccessed +aach~aach +aachen's~aachen's +aadri~aadri +aafia~aafia +aagaard~aagaard +aagadu~aagadu +aagard~aagard +aagathadi~aagathadi +aaghart's~aaghart's +aagnes~aagnes +aagomoni~aagomoni +aagon~aagon +aagoo~aagoo +aagot~aagot +aahar~aahar +aahh~aahh +aahperd~aahperd +aaibinterstate~aaibinterstate +aajab~aajab +aakasa~aakasa +aakervik~aakervik +aakirkeby~aakirkeby +aalam~aalam +aalbaek~aalbaek +aaldiu~aaldiu +aalem~aalem +a'ali~a'ali +aalilaassamthey~aalilaassamthey +aalin~aalin +aaliyan~aaliyan +aaliyan's~aaliyan's +aamadu~aamadu +aamara~aamara +aambala~aambala +aamera~aamera +aamer's~aamer's +aamina~aamina +aaminah~aaminah +aamjiwnaang~aamjiwnaang diff --git a/tests/nemo_text_processing/es/test_cardinal.py b/tests/nemo_text_processing/es/test_cardinal.py --- a/tests/nemo_text_processing/es/test_cardinal.py +++ b/tests/nemo_text_processing/es/test_cardinal.py @@ -22,7 +22,8 @@ class TestCardinal: - inverse_normalizer_es = ( + + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +33,34 @@ class TestCardinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_cardinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_date.py b/tests/nemo_text_processing/es/test_date.py --- a/tests/nemo_text_processing/es/test_date.py +++ b/tests/nemo_text_processing/es/test_date.py @@ -22,7 +22,7 @@ class TestDate: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDate: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_date.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_decimal.py b/tests/nemo_text_processing/es/test_decimal.py --- a/tests/nemo_text_processing/es/test_decimal.py +++ b/tests/nemo_text_processing/es/test_decimal.py @@ -22,7 +22,7 @@ class TestDecimal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDecimal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_decimal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_electronic.py b/tests/nemo_text_processing/es/test_electronic.py --- a/tests/nemo_text_processing/es/test_electronic.py +++ b/tests/nemo_text_processing/es/test_electronic.py @@ -35,3 +35,31 @@ class TestElectronic: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_electronic.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_fraction.py b/tests/nemo_text_processing/es/test_fraction.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_fraction.py @@ -0,0 +1,51 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import pytest +from nemo_text_processing.text_normalization.normalize import Normalizer +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, parse_test_case_file + + +class TestFraction: + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_fraction.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_measure.py b/tests/nemo_text_processing/es/test_measure.py --- a/tests/nemo_text_processing/es/test_measure.py +++ b/tests/nemo_text_processing/es/test_measure.py @@ -36,3 +36,31 @@ class TestMeasure: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_measure.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_money.py b/tests/nemo_text_processing/es/test_money.py --- a/tests/nemo_text_processing/es/test_money.py +++ b/tests/nemo_text_processing/es/test_money.py @@ -23,7 +23,7 @@ class TestMoney: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,34 @@ class TestMoney: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_money.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_normalization_with_audio.py b/tests/nemo_text_processing/es/test_normalization_with_audio.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_normalization_with_audio.py @@ -0,0 +1,43 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import pytest +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, get_test_cases_multiple + + +class TestNormalizeWithAudio: + + normalizer_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + @parameterized.expand(get_test_cases_multiple('es/data_text_normalization/test_cases_normalize_with_audio.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, n_tagged=1000, punct_post_process=False) + print(expected) + print("pred") + print(pred) + assert len(set(pred).intersection(set(expected))) == len( + expected + ), f'missing: {set(expected).difference(set(pred))}' diff --git a/tests/nemo_text_processing/es/test_ordinal.py b/tests/nemo_text_processing/es/test_ordinal.py --- a/tests/nemo_text_processing/es/test_ordinal.py +++ b/tests/nemo_text_processing/es/test_ordinal.py @@ -23,7 +23,7 @@ class TestOrdinal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,33 @@ class TestOrdinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_ordinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=30, punct_post_process=False, + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh @@ -0,0 +1,84 @@ +#! /bin/sh + +PROJECT_DIR=/workspace/tests + +runtest () { + input=$1 + cd /workspace/sparrowhawk/documentation/grammars + + # read test file + while read testcase; do + IFS='~' read written spoken <<< $testcase + denorm_pred=$(echo $written | normalizer_main --config=sparrowhawk_configuration.ascii_proto 2>&1 | tail -n 1) + + # trim white space + spoken="$(echo -e "${spoken}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + denorm_pred="$(echo -e "${denorm_pred}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + + # input expected actual + assertEquals "$written" "$spoken" "$denorm_pred" + done < "$input" +} + +testTNCardinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_cardinal.txt + runtest $input +} + +testTNDate() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_date.txt + runtest $input +} + +testTNDecimal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_decimal.txt + runtest $input +} + +testTNElectronic() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_electronic.txt + runtest $input +} + +testTNFraction() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_fraction.txt + runtest $input +} + +testTNMoney() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_money.txt + runtest $input +} + +testTNOrdinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTelephone() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTime() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_time.txt + runtest $input +} + +testTNMeasure() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_measure.txt + runtest $input +} + +testTNWhitelist() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_whitelist.txt + runtest $input +} + +testTNWord() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_word.txt + runtest $input +} + +# Load shUnit2 +. $PROJECT_DIR/../shunit2/shunit2 diff --git a/tests/nemo_text_processing/es/test_telephone.py b/tests/nemo_text_processing/es/test_telephone.py --- a/tests/nemo_text_processing/es/test_telephone.py +++ b/tests/nemo_text_processing/es/test_telephone.py @@ -36,3 +36,31 @@ class TestTelephone: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_telephone.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_time.py b/tests/nemo_text_processing/es/test_time.py --- a/tests/nemo_text_processing/es/test_time.py +++ b/tests/nemo_text_processing/es/test_time.py @@ -35,3 +35,31 @@ class TestTime: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_time.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_whitelist.py b/tests/nemo_text_processing/es/test_whitelist.py --- a/tests/nemo_text_processing/es/test_whitelist.py +++ b/tests/nemo_text_processing/es/test_whitelist.py @@ -35,3 +35,30 @@ class TestWhitelist: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_whitelist.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=10, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_word.py b/tests/nemo_text_processing/es/test_word.py --- a/tests/nemo_text_processing/es/test_word.py +++ b/tests/nemo_text_processing/es/test_word.py @@ -35,3 +35,30 @@ class TestWord: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer_es = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_word.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, verbose=False) + assert pred == expected, f"input: {test_input}" + + if self.normalizer_with_audio_es: + pred_non_deterministic = self.normalizer_with_audio_es.normalize( + test_input, n_tagged=150, punct_post_process=False + ) + assert expected in pred_non_deterministic, f"input: {test_input}"
1.0
NVIDIA__NeMo-7582
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] 1 2 |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg 17 :target: https://badge.fury.io/py/nemo-toolkit 18 :alt: Release version 19 20 .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg 21 :target: https://badge.fury.io/py/nemo-toolkit 22 :alt: Python version 23 24 .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads 25 :target: https://pepy.tech/project/nemo-toolkit 26 :alt: PyPi total downloads 27 28 .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push 29 :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml 30 :alt: CodeQL 31 32 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 33 :target: https://github.com/psf/black 34 :alt: Code style: black 35 36 .. _main-readme: 37 38 **NVIDIA NeMo** 39 =============== 40 41 Introduction 42 ------------ 43 44 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), 45 text-to-speech synthesis (TTS), large language models (LLMs), and 46 natural language processing (NLP). 47 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) 48 and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 49 50 All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and 51 training is automatically scalable to 1000s of GPUs. 52 Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. 53 NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. 54 55 Getting started with NeMo is simple. 56 State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and 57 `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. 58 These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. 59 60 We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that 61 can all be run on `Google Colab <https://colab.research.google.com>`_. 62 63 For advanced users that want to train NeMo models from scratch or finetune existing NeMo models 64 we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. 65 66 For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. 67 The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ 68 which can be used to find the optimal model parallel configuration for training on a specific cluster. 69 70 Also see our `introductory video <https://www.youtube.com/embed/wBgpMf_KQVw>`_ for a high level overview of NeMo. 71 72 Key Features 73 ------------ 74 75 * Speech processing 76 * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ 77 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 78 * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ 79 * Jasper, QuartzNet, CitriNet, ContextNet 80 * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer 81 * Squeezeformer-CTC and Squeezeformer-Transducer 82 * LSTM-Transducer (RNNT) and LSTM-CTC 83 * Supports the following decoders/losses: 84 * CTC 85 * Transducer/RNNT 86 * Hybrid Transducer/CTC 87 * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ 88 * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ 89 * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ 90 * Beam Search decoding 91 * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 92 * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ 93 * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) 94 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 95 * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ 96 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 97 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ 98 * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet 99 * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) 100 * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer 101 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 102 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 103 * Natural Language Processing 104 * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ 105 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ 106 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 107 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 108 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 109 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 110 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 111 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 112 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 113 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 114 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 115 * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ 116 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 117 * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ 118 * Text-to-Speech Synthesis (TTS): 119 * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 120 * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 121 * Vocoders: HiFiGAN, UnivNet, WaveGlow 122 * End-to-End Models: VITS 123 * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 124 * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ 125 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ 126 * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ 127 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 128 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 129 * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ 130 131 132 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 133 134 Requirements 135 ------------ 136 137 1) Python 3.10 or above 138 2) Pytorch 1.13.1 or above 139 3) NVIDIA GPU for training 140 141 Documentation 142 ------------- 143 144 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 145 :alt: Documentation Status 146 :scale: 100% 147 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 148 149 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 150 :alt: Documentation Status 151 :scale: 100% 152 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 153 154 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 155 | Version | Status | Description | 156 +=========+=============+==========================================================================================================================================+ 157 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 158 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 159 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 160 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 161 162 Tutorials 163 --------- 164 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 165 166 Getting help with NeMo 167 ---------------------- 168 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 169 170 171 Installation 172 ------------ 173 174 Conda 175 ~~~~~ 176 177 We recommend installing NeMo in a fresh Conda environment. 178 179 .. code-block:: bash 180 181 conda create --name nemo python==3.10.12 182 conda activate nemo 183 184 Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. 185 186 .. code-block:: bash 187 188 conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia 189 190 The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. 191 192 Pip 193 ~~~ 194 Use this installation mode if you want the latest released version. 195 196 .. code-block:: bash 197 198 apt-get update && apt-get install -y libsndfile1 ffmpeg 199 pip install Cython 200 pip install nemo_toolkit['all'] 201 202 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 203 204 Pip from source 205 ~~~~~~~~~~~~~~~ 206 Use this installation mode if you want the version from a particular GitHub branch (e.g main). 207 208 .. code-block:: bash 209 210 apt-get update && apt-get install -y libsndfile1 ffmpeg 211 pip install Cython 212 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 213 214 215 From source 216 ~~~~~~~~~~~ 217 Use this installation mode if you are contributing to NeMo. 218 219 .. code-block:: bash 220 221 apt-get update && apt-get install -y libsndfile1 ffmpeg 222 git clone https://github.com/NVIDIA/NeMo 223 cd NeMo 224 ./reinstall.sh 225 226 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 227 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 228 229 RNNT 230 ~~~~ 231 Note that RNNT requires numba to be installed from conda. 232 233 .. code-block:: bash 234 235 conda remove numba 236 pip uninstall numba 237 conda install -c conda-forge numba 238 239 NeMo Megatron 240 ~~~~~~~~~~~~~ 241 NeMo Megatron training requires NVIDIA Apex to be installed. 242 Install it manually if not using the NVIDIA PyTorch container. 243 244 To install Apex, run 245 246 .. code-block:: bash 247 248 git clone https://github.com/NVIDIA/apex.git 249 cd apex 250 git checkout 52e18c894223800cb611682dce27d88050edf1de 251 pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ 252 253 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. 254 255 While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. 256 This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 257 258 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: 259 260 .. code-block:: bash 261 262 conda install -c nvidia cuda-nvprof=11.8 263 264 packaging is also needed: 265 266 .. code-block:: bash 267 268 pip install packaging 269 270 With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. 271 272 273 Transformer Engine 274 ~~~~~~~~~~~~~~~~~~ 275 NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ 276 Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. 277 `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. 278 279 .. code-block:: bash 280 281 pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable 282 283 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. 284 285 Transformer Engine requires PyTorch to be built with CUDA 11.8. 286 287 288 Flash Attention 289 ~~~~~~~~~~~~~~~~~~~~ 290 Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. 291 292 .. code-block:: bash 293 294 pip install flash-attn 295 pip install triton==2.0.0.dev20221202 296 297 NLP inference UI 298 ~~~~~~~~~~~~~~~~~~~~ 299 To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. 300 301 .. code-block:: bash 302 303 pip install gradio==3.34.0 304 305 NeMo Text Processing 306 ~~~~~~~~~~~~~~~~~~~~ 307 NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. 308 309 Docker containers: 310 ~~~~~~~~~~~~~~~~~~ 311 We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. 312 313 To use built container, please run 314 315 .. code-block:: bash 316 317 docker pull nvcr.io/nvidia/nemo:23.06 318 319 To build a nemo container with Dockerfile from a branch, please run 320 321 .. code-block:: bash 322 323 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 324 325 326 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. 327 328 .. code-block:: bash 329 330 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 331 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 332 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 333 334 Examples 335 -------- 336 337 Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 338 339 340 Contributing 341 ------------ 342 343 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 344 345 Publications 346 ------------ 347 348 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 349 350 License 351 ------- 352 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 353 [end of README.rst] [start of examples/asr/experimental/k2/align_speech_parallel.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 # Based on examples/asr/transcribe_speech_parallel.py 17 # ASR alignment with multi-GPU/multi-node support for large datasets 18 # It supports both tarred and non-tarred datasets 19 # Arguments 20 # model: path to a nemo/PTL checkpoint file or name of a pretrained model 21 # predict_ds: config of the dataset/dataloader 22 # aligner_args: aligner config 23 # output_path: path to store the predictions 24 # model_stride: model downsampling factor, 8 for Citrinet models and 4 for Conformer models 25 # 26 # Results of each GPU/worker is written into a file named 'predictions_{rank}.json, and aggregated results of all workers are written into 'predictions_all.json' 27 28 Example for non-tarred datasets: 29 30 python align_speech_parallel.py \ 31 model=stt_en_conformer_ctc_large \ 32 predict_ds.manifest_filepath=/dataset/manifest_file.json \ 33 predict_ds.batch_size=16 \ 34 output_path=/tmp/ 35 36 Example for tarred datasets: 37 38 python align_speech_parallel.py \ 39 predict_ds.is_tarred=true \ 40 predict_ds.manifest_filepath=/tarred_dataset/tarred_audio_manifest.json \ 41 predict_ds.tarred_audio_filepaths=/tarred_dataset/audio__OP_0..127_CL_.tar \ 42 ... 43 44 By default the trainer uses all the GPUs available and default precision is FP32. 45 By setting the trainer config you may control these configs. For example to do the predictions with AMP on just two GPUs: 46 47 python align_speech_parallel.py \ 48 trainer.precision=16 \ 49 trainer.gpus=2 \ 50 ... 51 52 You may control the dataloader's config by setting the predict_ds: 53 54 python align_speech_parallel.py \ 55 predict_ds.num_workers=8 \ 56 predict_ds.min_duration=2.0 \ 57 predict_ds.sample_rate=16000 \ 58 model=stt_en_conformer_ctc_small \ 59 ... 60 61 You may control the aligner's config by setting the aligner_args: 62 aligner_args.alignment_type=argmax \ 63 aligner_args.word_output=False \ 64 aligner_args.cpu_decoding=True \ 65 aligner_args.decode_batch_size=8 \ 66 aligner_args.ctc_cfg.prob_suppress_index=-1 \ 67 aligner_args.ctc_cfg.prob_suppress_value=0.5 \ 68 aligner_args.rnnt_cfg.predictor_window_size=10 \ 69 aligner_args.decoder_module_cfg.intersect_pruned=true \ 70 aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ 71 ... 72 73 """ 74 75 76 import os 77 from dataclasses import dataclass, is_dataclass 78 from typing import Optional 79 80 import pytorch_lightning as ptl 81 import torch 82 from omegaconf import MISSING, OmegaConf 83 84 from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter 85 from nemo.collections.asr.models import ASRModel 86 from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig 87 from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig 88 from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel 89 from nemo.core.config import TrainerConfig, hydra_runner 90 from nemo.utils import logging 91 from nemo.utils.get_rank import is_global_rank_zero 92 93 94 @dataclass 95 class ParallelAlignmentConfig: 96 model: Optional[str] = None # name 97 predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) 98 aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() 99 output_path: str = MISSING 100 model_stride: int = 8 101 102 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") 103 104 # there arguments will be ignored 105 return_predictions: bool = False 106 use_cer: bool = False 107 108 109 def match_train_config(predict_ds, train_ds): 110 # It copies the important configurations from the train dataset of the model 111 # into the predict_ds to be used for prediction. It is needed to match the training configurations. 112 if train_ds is None: 113 return 114 115 predict_ds.sample_rate = train_ds.get("sample_rate", 16000) 116 cfg_name_list = [ 117 "int_values", 118 "use_start_end_token", 119 "blank_index", 120 "unk_index", 121 "normalize", 122 "parser", 123 "eos_id", 124 "bos_id", 125 "pad_id", 126 ] 127 128 if is_dataclass(predict_ds): 129 predict_ds = OmegaConf.structured(predict_ds) 130 for cfg_name in cfg_name_list: 131 if hasattr(train_ds, cfg_name): 132 setattr(predict_ds, cfg_name, getattr(train_ds, cfg_name)) 133 134 return predict_ds 135 136 137 @hydra_runner(config_name="AlignmentConfig", schema=ParallelAlignmentConfig) 138 def main(cfg: ParallelAlignmentConfig): 139 if cfg.model.endswith(".nemo"): 140 logging.info("Attempting to initialize from .nemo file") 141 model = ASRModel.restore_from(restore_path=cfg.model, map_location="cpu") 142 elif cfg.model.endswith(".ckpt"): 143 logging.info("Attempting to initialize from .ckpt file") 144 model = ASRModel.load_from_checkpoint(checkpoint_path=cfg.model, map_location="cpu") 145 else: 146 logging.info( 147 "Attempting to initialize from a pretrained model as the model name does not have the extension of .nemo or .ckpt" 148 ) 149 model = ASRModel.from_pretrained(model_name=cfg.model, map_location="cpu") 150 151 trainer = ptl.Trainer(**cfg.trainer) 152 153 cfg.predict_ds.return_sample_id = True 154 cfg.return_predictions = False 155 cfg.use_cer = False 156 cfg.predict_ds = match_train_config(predict_ds=cfg.predict_ds, train_ds=model._cfg.train_ds) 157 data_loader = model._setup_dataloader_from_config(cfg.predict_ds) 158 159 os.makedirs(cfg.output_path, exist_ok=True) 160 # trainer.global_rank is not valid before predict() is called. Need this hack to find the correct global_rank. 161 global_rank = trainer.node_rank * trainer.num_devices + int(os.environ.get("LOCAL_RANK", 0)) 162 output_file = os.path.join(cfg.output_path, f"predictions_{global_rank}.json") 163 output_ctm_dir = os.path.join(cfg.output_path, "ctm") 164 predictor_writer = ASRCTMPredictionWriter( 165 dataset=data_loader.dataset, 166 output_file=output_file, 167 output_ctm_dir=output_ctm_dir, 168 time_per_frame=cfg.model_stride * model._cfg.preprocessor['window_stride'], 169 ) 170 trainer.callbacks.extend([predictor_writer]) 171 172 aligner_wrapper = AlignerWrapperModel(model=model, cfg=cfg.aligner_args) 173 trainer.predict(model=aligner_wrapper, dataloaders=data_loader, return_predictions=cfg.return_predictions) 174 samples_num = predictor_writer.close_output_file() 175 176 logging.info( 177 f"Prediction on rank {global_rank} is done for {samples_num} samples and results are stored in {output_file}." 178 ) 179 180 if torch.distributed.is_initialized(): 181 torch.distributed.barrier() 182 183 samples_num = 0 184 if is_global_rank_zero(): 185 output_file = os.path.join(cfg.output_path, f"predictions_all.json") 186 logging.info(f"Prediction files are being aggregated in {output_file}.") 187 with open(output_file, 'tw', encoding="utf-8") as outf: 188 for rank in range(trainer.world_size): 189 input_file = os.path.join(cfg.output_path, f"predictions_{rank}.json") 190 with open(input_file, 'r', encoding="utf-8") as inpf: 191 lines = inpf.readlines() 192 samples_num += len(lines) 193 outf.writelines(lines) 194 logging.info( 195 f"Prediction is done for {samples_num} samples in total on all workers and results are aggregated in {output_file}." 196 ) 197 198 199 if __name__ == '__main__': 200 main() 201 [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import re 17 from abc import abstractmethod 18 from dataclasses import dataclass, is_dataclass 19 from typing import Callable, Dict, List, Optional, Tuple, Union 20 21 import editdistance 22 import numpy as np 23 import torch 24 from omegaconf import OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.metrics.wer import move_dimension_to_the_front 28 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding as beam_decode 29 from nemo.collections.asr.parts.submodules import rnnt_greedy_decoding as greedy_decode 30 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 31 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 32 from nemo.utils import logging 33 34 __all__ = ['RNNTDecoding', 'RNNTWER'] 35 36 37 class AbstractRNNTDecoding(ConfidenceMixin): 38 """ 39 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 40 41 Args: 42 decoding_cfg: A dict-like object which contains the following key-value pairs. 43 strategy: str value which represents the type of decoding that can occur. 44 Possible values are : 45 - greedy, greedy_batch (for greedy decoding). 46 - beam, tsd, alsd (for beam search decoding). 47 48 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 49 tokens as well as the decoded string. Default is False in order to avoid double decoding 50 unless required. 51 52 preserve_alignments: Bool flag which preserves the history of logprobs generated during 53 decoding (sample / batched). When set to true, the Hypothesis will contain 54 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 55 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 56 57 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 58 with the `return_hypotheses` flag set to True. 59 60 The length of the list corresponds to the Acoustic Length (T). 61 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 62 U is the number of target tokens for the current timestep Ti. 63 64 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 65 word based timestamp mapping the output log-probabilities to discrete intervals of timestamps. 66 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 67 68 rnnt_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 69 Can take the following values - "char" for character/subword time stamps, "word" for word level 70 time stamps and "all" (default), for both character level and word level time stamps. 71 72 word_seperator: Str token representing the seperator between words. 73 74 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 75 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 76 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of ints. 77 78 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 79 scores. In order to obtain hypotheses with confidence scores, please utilize 80 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 81 82 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 83 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 84 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 85 86 The length of the list corresponds to the Acoustic Length (T). 87 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 88 U is the number of target tokens for the current timestep Ti. 89 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 90 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 91 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 92 93 The length of the list corresponds to the number of recognized tokens. 94 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 95 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 96 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 97 98 The length of the list corresponds to the number of recognized words. 99 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 100 from the `token_confidence`. 101 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 102 Valid options are `mean`, `min`, `max`, `prod`. 103 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 104 confidence scores. 105 106 name: The measure name (str). 107 Supported values: 108 - 'max_prob' for using the maximum token probability as a confidence. 109 - 'entropy' for using a normalized entropy of a log-likelihood vector. 110 111 entropy_type: Which type of entropy to use (str). 112 Used if confidence_measure_cfg.name is set to `entropy`. 113 Supported values: 114 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 115 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 116 Note that for this entropy, the alpha should comply the following inequality: 117 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 118 where V is the model vocabulary size. 119 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 120 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 121 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 122 More: https://en.wikipedia.org/wiki/Tsallis_entropy 123 - 'renyi' for the Rรฉnyi entropy. 124 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 125 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 126 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 127 128 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 129 When the alpha equals one, scaling is not applied to 'max_prob', 130 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 131 132 entropy_norm: A mapping of the entropy value to the interval [0,1]. 133 Supported values: 134 - 'lin' for using the linear mapping. 135 - 'exp' for using exponential mapping with linear shift. 136 137 The config may further contain the following sub-dictionaries: 138 "greedy": 139 max_symbols: int, describing the maximum number of target tokens to decode per 140 timestep during greedy decoding. Setting to larger values allows longer sentences 141 to be decoded, at the cost of increased execution time. 142 preserve_frame_confidence: Same as above, overrides above value. 143 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 144 145 "beam": 146 beam_size: int, defining the beam size for beam search. Must be >= 1. 147 If beam_size == 1, will perform cached greedy search. This might be slightly different 148 results compared to the greedy search above. 149 150 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 151 Set to True by default. 152 153 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 154 hypotheses after beam search has concluded. This flag is set by default. 155 156 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 157 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 158 at increased cost to execution time. 159 160 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 161 If an integer is provided, it can decode sequences of that particular maximum length. 162 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 163 where seq_len is the length of the acoustic model output (T). 164 165 NOTE: 166 If a float is provided, it can be greater than 1! 167 By default, a float of 2.0 is used so that a target sequence can be at most twice 168 as long as the acoustic model output length T. 169 170 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 171 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 172 173 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 174 in order to reduce expensive beam search cost later. int >= 0. 175 176 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 177 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 178 and affects the speed of inference since large values will perform large beam search in the next step. 179 180 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 181 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 182 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 183 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 184 expansion apart from the "most likely" candidate. 185 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 186 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 187 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 188 tuned on a validation set. 189 190 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 191 192 decoder: The Decoder/Prediction network module. 193 joint: The Joint network module. 194 blank_id: The id of the RNNT blank token. 195 """ 196 197 def __init__(self, decoding_cfg, decoder, joint, blank_id: int): 198 super(AbstractRNNTDecoding, self).__init__() 199 200 # Convert dataclass to config object 201 if is_dataclass(decoding_cfg): 202 decoding_cfg = OmegaConf.structured(decoding_cfg) 203 204 self.cfg = decoding_cfg 205 self.blank_id = blank_id 206 self.num_extra_outputs = joint.num_extra_outputs 207 self.big_blank_durations = self.cfg.get("big_blank_durations", None) 208 self.durations = self.cfg.get("durations", None) 209 self.compute_hypothesis_token_set = self.cfg.get("compute_hypothesis_token_set", False) 210 self.compute_langs = decoding_cfg.get('compute_langs', False) 211 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 212 self.joint_fused_batch_size = self.cfg.get('fused_batch_size', None) 213 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 214 self.word_seperator = self.cfg.get('word_seperator', ' ') 215 216 if self.durations is not None: # this means it's a TDT model. 217 if blank_id == 0: 218 raise ValueError("blank_id must equal len(non_blank_vocabs) for TDT models") 219 if self.big_blank_durations is not None: 220 raise ValueError("duration and big_blank_durations can't both be not None") 221 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 222 raise ValueError("currently only greedy and greedy_batch inference is supported for TDT models") 223 224 if self.big_blank_durations is not None: # this means it's a multi-blank model. 225 if blank_id == 0: 226 raise ValueError("blank_id must equal len(vocabs) for multi-blank RNN-T models") 227 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 228 raise ValueError( 229 "currently only greedy and greedy_batch inference is supported for multi-blank models" 230 ) 231 232 possible_strategies = ['greedy', 'greedy_batch', 'beam', 'tsd', 'alsd', 'maes'] 233 if self.cfg.strategy not in possible_strategies: 234 raise ValueError(f"Decoding strategy must be one of {possible_strategies}") 235 236 # Update preserve alignments 237 if self.preserve_alignments is None: 238 if self.cfg.strategy in ['greedy', 'greedy_batch']: 239 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 240 241 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 242 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 243 244 # Update compute timestamps 245 if self.compute_timestamps is None: 246 if self.cfg.strategy in ['greedy', 'greedy_batch']: 247 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 248 249 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 250 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 251 252 # Test if alignments are being preserved for RNNT 253 if self.compute_timestamps is True and self.preserve_alignments is False: 254 raise ValueError("If `compute_timesteps` flag is set, then `preserve_alignments` flag must also be set.") 255 256 # initialize confidence-related fields 257 self._init_confidence(self.cfg.get('confidence_cfg', None)) 258 259 # Confidence estimation is not implemented for these strategies 260 if ( 261 not self.preserve_frame_confidence 262 and self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes'] 263 and self.cfg.beam.get('preserve_frame_confidence', False) 264 ): 265 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 266 267 if self.cfg.strategy == 'greedy': 268 if self.big_blank_durations is None: 269 if self.durations is None: 270 self.decoding = greedy_decode.GreedyRNNTInfer( 271 decoder_model=decoder, 272 joint_model=joint, 273 blank_index=self.blank_id, 274 max_symbols_per_step=( 275 self.cfg.greedy.get('max_symbols', None) 276 or self.cfg.greedy.get('max_symbols_per_step', None) 277 ), 278 preserve_alignments=self.preserve_alignments, 279 preserve_frame_confidence=self.preserve_frame_confidence, 280 confidence_measure_cfg=self.confidence_measure_cfg, 281 ) 282 else: 283 self.decoding = greedy_decode.GreedyTDTInfer( 284 decoder_model=decoder, 285 joint_model=joint, 286 blank_index=self.blank_id, 287 durations=self.durations, 288 max_symbols_per_step=( 289 self.cfg.greedy.get('max_symbols', None) 290 or self.cfg.greedy.get('max_symbols_per_step', None) 291 ), 292 preserve_alignments=self.preserve_alignments, 293 preserve_frame_confidence=self.preserve_frame_confidence, 294 confidence_measure_cfg=self.confidence_measure_cfg, 295 ) 296 else: 297 self.decoding = greedy_decode.GreedyMultiblankRNNTInfer( 298 decoder_model=decoder, 299 joint_model=joint, 300 blank_index=self.blank_id, 301 big_blank_durations=self.big_blank_durations, 302 max_symbols_per_step=( 303 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 304 ), 305 preserve_alignments=self.preserve_alignments, 306 preserve_frame_confidence=self.preserve_frame_confidence, 307 confidence_measure_cfg=self.confidence_measure_cfg, 308 ) 309 310 elif self.cfg.strategy == 'greedy_batch': 311 if self.big_blank_durations is None: 312 if self.durations is None: 313 self.decoding = greedy_decode.GreedyBatchedRNNTInfer( 314 decoder_model=decoder, 315 joint_model=joint, 316 blank_index=self.blank_id, 317 max_symbols_per_step=( 318 self.cfg.greedy.get('max_symbols', None) 319 or self.cfg.greedy.get('max_symbols_per_step', None) 320 ), 321 preserve_alignments=self.preserve_alignments, 322 preserve_frame_confidence=self.preserve_frame_confidence, 323 confidence_measure_cfg=self.confidence_measure_cfg, 324 ) 325 else: 326 self.decoding = greedy_decode.GreedyBatchedTDTInfer( 327 decoder_model=decoder, 328 joint_model=joint, 329 blank_index=self.blank_id, 330 durations=self.durations, 331 max_symbols_per_step=( 332 self.cfg.greedy.get('max_symbols', None) 333 or self.cfg.greedy.get('max_symbols_per_step', None) 334 ), 335 preserve_alignments=self.preserve_alignments, 336 preserve_frame_confidence=self.preserve_frame_confidence, 337 confidence_measure_cfg=self.confidence_measure_cfg, 338 ) 339 340 else: 341 self.decoding = greedy_decode.GreedyBatchedMultiblankRNNTInfer( 342 decoder_model=decoder, 343 joint_model=joint, 344 blank_index=self.blank_id, 345 big_blank_durations=self.big_blank_durations, 346 max_symbols_per_step=( 347 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 348 ), 349 preserve_alignments=self.preserve_alignments, 350 preserve_frame_confidence=self.preserve_frame_confidence, 351 confidence_measure_cfg=self.confidence_measure_cfg, 352 ) 353 354 elif self.cfg.strategy == 'beam': 355 356 self.decoding = beam_decode.BeamRNNTInfer( 357 decoder_model=decoder, 358 joint_model=joint, 359 beam_size=self.cfg.beam.beam_size, 360 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 361 search_type='default', 362 score_norm=self.cfg.beam.get('score_norm', True), 363 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 364 preserve_alignments=self.preserve_alignments, 365 ) 366 367 elif self.cfg.strategy == 'tsd': 368 369 self.decoding = beam_decode.BeamRNNTInfer( 370 decoder_model=decoder, 371 joint_model=joint, 372 beam_size=self.cfg.beam.beam_size, 373 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 374 search_type='tsd', 375 score_norm=self.cfg.beam.get('score_norm', True), 376 tsd_max_sym_exp_per_step=self.cfg.beam.get('tsd_max_sym_exp', 10), 377 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 378 preserve_alignments=self.preserve_alignments, 379 ) 380 381 elif self.cfg.strategy == 'alsd': 382 383 self.decoding = beam_decode.BeamRNNTInfer( 384 decoder_model=decoder, 385 joint_model=joint, 386 beam_size=self.cfg.beam.beam_size, 387 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 388 search_type='alsd', 389 score_norm=self.cfg.beam.get('score_norm', True), 390 alsd_max_target_len=self.cfg.beam.get('alsd_max_target_len', 2), 391 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 392 preserve_alignments=self.preserve_alignments, 393 ) 394 395 elif self.cfg.strategy == 'maes': 396 397 self.decoding = beam_decode.BeamRNNTInfer( 398 decoder_model=decoder, 399 joint_model=joint, 400 beam_size=self.cfg.beam.beam_size, 401 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 402 search_type='maes', 403 score_norm=self.cfg.beam.get('score_norm', True), 404 maes_num_steps=self.cfg.beam.get('maes_num_steps', 2), 405 maes_prefix_alpha=self.cfg.beam.get('maes_prefix_alpha', 1), 406 maes_expansion_gamma=self.cfg.beam.get('maes_expansion_gamma', 2.3), 407 maes_expansion_beta=self.cfg.beam.get('maes_expansion_beta', 2.0), 408 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 409 preserve_alignments=self.preserve_alignments, 410 ngram_lm_model=self.cfg.beam.get('ngram_lm_model', None), 411 ngram_lm_alpha=self.cfg.beam.get('ngram_lm_alpha', 0.0), 412 hat_subtract_ilm=self.cfg.beam.get('hat_subtract_ilm', False), 413 hat_ilm_weight=self.cfg.beam.get('hat_ilm_weight', 0.0), 414 ) 415 416 else: 417 418 raise ValueError( 419 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 420 f"but was provided {self.cfg.strategy}" 421 ) 422 423 # Update the joint fused batch size or disable it entirely if needed. 424 self.update_joint_fused_batch_size() 425 426 def rnnt_decoder_predictions_tensor( 427 self, 428 encoder_output: torch.Tensor, 429 encoded_lengths: torch.Tensor, 430 return_hypotheses: bool = False, 431 partial_hypotheses: Optional[List[Hypothesis]] = None, 432 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 433 """ 434 Decode an encoder output by autoregressive decoding of the Decoder+Joint networks. 435 436 Args: 437 encoder_output: torch.Tensor of shape [B, D, T]. 438 encoded_lengths: torch.Tensor containing lengths of the padded encoder outputs. Shape [B]. 439 return_hypotheses: bool. If set to True it will return list of Hypothesis or NBestHypotheses 440 441 Returns: 442 If `return_best_hypothesis` is set: 443 A tuple (hypotheses, None): 444 hypotheses - list of Hypothesis (best hypothesis per sample). 445 Look at rnnt_utils.Hypothesis for more information. 446 447 If `return_best_hypothesis` is not set: 448 A tuple(hypotheses, all_hypotheses) 449 hypotheses - list of Hypothesis (best hypothesis per sample). 450 Look at rnnt_utils.Hypothesis for more information. 451 all_hypotheses - list of NBestHypotheses. Each NBestHypotheses further contains a sorted 452 list of all the hypotheses of the model per sample. 453 Look at rnnt_utils.NBestHypotheses for more information. 454 """ 455 # Compute hypotheses 456 with torch.inference_mode(): 457 hypotheses_list = self.decoding( 458 encoder_output=encoder_output, encoded_lengths=encoded_lengths, partial_hypotheses=partial_hypotheses 459 ) # type: [List[Hypothesis]] 460 461 # extract the hypotheses 462 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 463 464 prediction_list = hypotheses_list 465 466 if isinstance(prediction_list[0], NBestHypotheses): 467 hypotheses = [] 468 all_hypotheses = [] 469 470 for nbest_hyp in prediction_list: # type: NBestHypotheses 471 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 472 decoded_hyps = self.decode_hypothesis(n_hyps) # type: List[str] 473 474 # If computing timestamps 475 if self.compute_timestamps is True: 476 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 477 for hyp_idx in range(len(decoded_hyps)): 478 decoded_hyps[hyp_idx] = self.compute_rnnt_timestamps(decoded_hyps[hyp_idx], timestamp_type) 479 480 hypotheses.append(decoded_hyps[0]) # best hypothesis 481 all_hypotheses.append(decoded_hyps) 482 483 if return_hypotheses: 484 return hypotheses, all_hypotheses 485 486 best_hyp_text = [h.text for h in hypotheses] 487 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 488 return best_hyp_text, all_hyp_text 489 490 else: 491 hypotheses = self.decode_hypothesis(prediction_list) # type: List[str] 492 493 # If computing timestamps 494 if self.compute_timestamps is True: 495 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 496 for hyp_idx in range(len(hypotheses)): 497 hypotheses[hyp_idx] = self.compute_rnnt_timestamps(hypotheses[hyp_idx], timestamp_type) 498 499 if return_hypotheses: 500 # greedy decoding, can get high-level confidence scores 501 if self.preserve_frame_confidence and ( 502 self.preserve_word_confidence or self.preserve_token_confidence 503 ): 504 hypotheses = self.compute_confidence(hypotheses) 505 return hypotheses, None 506 507 best_hyp_text = [h.text for h in hypotheses] 508 return best_hyp_text, None 509 510 def decode_hypothesis(self, hypotheses_list: List[Hypothesis]) -> List[Union[Hypothesis, NBestHypotheses]]: 511 """ 512 Decode a list of hypotheses into a list of strings. 513 514 Args: 515 hypotheses_list: List of Hypothesis. 516 517 Returns: 518 A list of strings. 519 """ 520 for ind in range(len(hypotheses_list)): 521 # Extract the integer encoded hypothesis 522 prediction = hypotheses_list[ind].y_sequence 523 524 if type(prediction) != list: 525 prediction = prediction.tolist() 526 527 # RNN-T sample level is already preprocessed by implicit RNNT decoding 528 # Simply remove any blank and possibly big blank tokens 529 if self.big_blank_durations is not None: # multi-blank RNNT 530 num_extra_outputs = len(self.big_blank_durations) 531 prediction = [p for p in prediction if p < self.blank_id - num_extra_outputs] 532 elif self.durations is not None: # TDT model. 533 prediction = [p for p in prediction if p < self.blank_id] 534 else: # standard RNN-T 535 prediction = [p for p in prediction if p != self.blank_id] 536 537 # De-tokenize the integer tokens; if not computing timestamps 538 if self.compute_timestamps is True: 539 # keep the original predictions, wrap with the number of repetitions per token and alignments 540 # this is done so that `rnnt_decoder_predictions_tensor()` can process this hypothesis 541 # in order to compute exact time stamps. 542 alignments = copy.deepcopy(hypotheses_list[ind].alignments) 543 token_repetitions = [1] * len(alignments) # preserve number of repetitions per token 544 hypothesis = (prediction, alignments, token_repetitions) 545 else: 546 hypothesis = self.decode_tokens_to_str(prediction) 547 548 # TODO: remove 549 # collapse leading spaces before . , ? for PC models 550 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 551 552 if self.compute_hypothesis_token_set: 553 hypotheses_list[ind].tokens = self.decode_ids_to_tokens(prediction) 554 555 # De-tokenize the integer tokens 556 hypotheses_list[ind].text = hypothesis 557 558 return hypotheses_list 559 560 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 561 """ 562 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 563 Assumes that `frame_confidence` is present in the hypotheses. 564 565 Args: 566 hypotheses_list: List of Hypothesis. 567 568 Returns: 569 A list of hypotheses with high-level confidence scores. 570 """ 571 if self.exclude_blank_from_confidence: 572 for hyp in hypotheses_list: 573 hyp.token_confidence = hyp.non_blank_frame_confidence 574 else: 575 for hyp in hypotheses_list: 576 offset = 0 577 token_confidence = [] 578 if len(hyp.timestep) > 0: 579 for ts, te in zip(hyp.timestep, hyp.timestep[1:] + [len(hyp.frame_confidence)]): 580 if ts != te: 581 # <blank> tokens are considered to belong to the last non-blank token, if any. 582 token_confidence.append( 583 self._aggregate_confidence( 584 [hyp.frame_confidence[ts][offset]] 585 + [fc[0] for fc in hyp.frame_confidence[ts + 1 : te]] 586 ) 587 ) 588 offset = 0 589 else: 590 token_confidence.append(hyp.frame_confidence[ts][offset]) 591 offset += 1 592 hyp.token_confidence = token_confidence 593 if self.preserve_word_confidence: 594 for hyp in hypotheses_list: 595 hyp.word_confidence = self._aggregate_token_confidence(hyp) 596 return hypotheses_list 597 598 @abstractmethod 599 def decode_tokens_to_str(self, tokens: List[int]) -> str: 600 """ 601 Implemented by subclass in order to decoder a token id list into a string. 602 603 Args: 604 tokens: List of int representing the token ids. 605 606 Returns: 607 A decoded string. 608 """ 609 raise NotImplementedError() 610 611 @abstractmethod 612 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 613 """ 614 Implemented by subclass in order to decode a token id list into a token list. 615 A token list is the string representation of each token id. 616 617 Args: 618 tokens: List of int representing the token ids. 619 620 Returns: 621 A list of decoded tokens. 622 """ 623 raise NotImplementedError() 624 625 @abstractmethod 626 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 627 """ 628 Implemented by subclass in order to 629 compute the most likely language ID (LID) string given the tokens. 630 631 Args: 632 tokens: List of int representing the token ids. 633 634 Returns: 635 A decoded LID string. 636 """ 637 raise NotImplementedError() 638 639 @abstractmethod 640 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 641 """ 642 Implemented by subclass in order to 643 decode a token id list into language ID (LID) list. 644 645 Args: 646 tokens: List of int representing the token ids. 647 648 Returns: 649 A list of decoded LIDS. 650 """ 651 raise NotImplementedError() 652 653 def update_joint_fused_batch_size(self): 654 if self.joint_fused_batch_size is None: 655 # do nothing and let the Joint itself handle setting up of the fused batch 656 return 657 658 if not hasattr(self.decoding.joint, 'set_fused_batch_size'): 659 logging.warning( 660 "The joint module does not have `set_fused_batch_size(int)` as a setter function.\n" 661 "Ignoring update of joint fused batch size." 662 ) 663 return 664 665 if not hasattr(self.decoding.joint, 'set_fuse_loss_wer'): 666 logging.warning( 667 "The joint module does not have `set_fuse_loss_wer(bool, RNNTLoss, RNNTWER)` " 668 "as a setter function.\n" 669 "Ignoring update of joint fused batch size." 670 ) 671 return 672 673 if self.joint_fused_batch_size > 0: 674 self.decoding.joint.set_fused_batch_size(self.joint_fused_batch_size) 675 else: 676 logging.info("Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.") 677 self.decoding.joint.set_fuse_loss_wer(False) 678 679 def compute_rnnt_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 680 assert timestamp_type in ['char', 'word', 'all'] 681 682 # Unpack the temporary storage 683 decoded_prediction, alignments, token_repetitions = hypothesis.text 684 685 # Retrieve offsets 686 char_offsets = word_offsets = None 687 char_offsets = self._compute_offsets(hypothesis, token_repetitions, self.blank_id) 688 689 # finally, set the flattened decoded predictions to text field for later text decoding 690 hypothesis.text = decoded_prediction 691 692 # Assert number of offsets and hypothesis tokens are 1:1 match. 693 num_flattened_tokens = 0 694 for t in range(len(char_offsets)): 695 # Subtract one here for the extra RNNT BLANK token emitted to designate "End of timestep" 696 num_flattened_tokens += len(char_offsets[t]['char']) - 1 697 698 if num_flattened_tokens != len(hypothesis.text): 699 raise ValueError( 700 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 701 " have to be of the same length, but are: " 702 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 703 f" {len(hypothesis.text)}" 704 ) 705 706 encoded_char_offsets = copy.deepcopy(char_offsets) 707 708 # Correctly process the token ids to chars/subwords. 709 for i, offsets in enumerate(char_offsets): 710 decoded_chars = [] 711 for char in offsets['char'][:-1]: # ignore the RNNT Blank token at end of every timestep with -1 subset 712 decoded_chars.append(self.decode_tokens_to_str([int(char)])) 713 char_offsets[i]["char"] = decoded_chars 714 715 # detect char vs subword models 716 lens = [] 717 for v in char_offsets: 718 tokens = v["char"] 719 # each token may be either 1 unicode token or multiple unicode token 720 # for character based models, only 1 token is used 721 # for subword, more than one token can be used. 722 # Computing max, then summing up total lens is a test to check for char vs subword 723 # For char models, len(lens) == sum(lens) 724 # but this is violated for subword models. 725 max_len = max(len(c) for c in tokens) 726 lens.append(max_len) 727 728 # array of one or more chars implies subword based model with multiple char emitted per TxU step (via subword) 729 if sum(lens) > len(lens): 730 text_type = 'subword' 731 else: 732 # full array of ones implies character based model with 1 char emitted per TxU step 733 text_type = 'char' 734 735 # retrieve word offsets from character offsets 736 word_offsets = None 737 if timestamp_type in ['word', 'all']: 738 if text_type == 'char': 739 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 740 else: 741 # utilize the copy of char offsets with the correct integer ids for tokens 742 # so as to avoid tokenize -> detokenize -> compare -> merge steps. 743 word_offsets = self._get_word_offsets_subwords_sentencepiece( 744 encoded_char_offsets, 745 hypothesis, 746 decode_ids_to_tokens=self.decode_ids_to_tokens, 747 decode_tokens_to_str=self.decode_tokens_to_str, 748 ) 749 750 # attach results 751 if len(hypothesis.timestep) > 0: 752 timestep_info = hypothesis.timestep 753 else: 754 timestep_info = [] 755 756 # Setup defaults 757 hypothesis.timestep = {"timestep": timestep_info} 758 759 # Add char / subword time stamps 760 if char_offsets is not None and timestamp_type in ['char', 'all']: 761 hypothesis.timestep['char'] = char_offsets 762 763 # Add word time stamps 764 if word_offsets is not None and timestamp_type in ['word', 'all']: 765 hypothesis.timestep['word'] = word_offsets 766 767 # Convert the flattened token indices to text 768 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 769 770 return hypothesis 771 772 @staticmethod 773 def _compute_offsets( 774 hypothesis: Hypothesis, token_repetitions: List[int], rnnt_token: int 775 ) -> List[Dict[str, Union[str, int]]]: 776 """ 777 Utility method that calculates the indidual time indices where a token starts and ends. 778 779 Args: 780 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 781 emitted at every time step after rnnt collapse. 782 token_repetitions: A list of ints representing the number of repetitions of each emitted token. 783 rnnt_token: The integer of the rnnt blank token used during rnnt collapse. 784 785 Returns: 786 787 """ 788 start_index = 0 789 790 # If the exact timestep information is available, utilize the 1st non-rnnt blank token timestep 791 # as the start index. 792 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 793 start_index = max(0, hypothesis.timestep[0] - 1) 794 795 # Construct the start and end indices brackets 796 end_indices = np.asarray(token_repetitions).cumsum() 797 start_indices = np.concatenate(([start_index], end_indices[:-1])) 798 799 # Process the TxU dangling alignment tensor, containing pairs of (logits, label) 800 alignment_labels = [al_logits_labels for al_logits_labels in hypothesis.text[1]] 801 for t in range(len(alignment_labels)): 802 for u in range(len(alignment_labels[t])): 803 alignment_labels[t][u] = alignment_labels[t][u][1] # pick label from (logit, label) tuple 804 805 # Merge the results per token into a list of dictionaries 806 offsets = [ 807 {"char": a, "start_offset": s, "end_offset": e} 808 for a, s, e in zip(alignment_labels, start_indices, end_indices) 809 ] 810 811 # Filter out RNNT token (blank at [t][0] position). This is because blank can only occur at end of a 812 # time step for RNNT, so if 0th token is blank, then that timestep is skipped. 813 offsets = list(filter(lambda offsets: offsets["char"][0] != rnnt_token, offsets)) 814 return offsets 815 816 @staticmethod 817 def _get_word_offsets_chars( 818 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 819 ) -> Dict[str, Union[str, float]]: 820 """ 821 Utility method which constructs word time stamps out of character time stamps. 822 823 References: 824 This code is a port of the Hugging Face code for word time stamp construction. 825 826 Args: 827 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 828 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 829 830 Returns: 831 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 832 "end_offset". 833 """ 834 word_offsets = [] 835 836 last_state = "SPACE" 837 word = "" 838 start_offset = 0 839 end_offset = 0 840 for i, offset in enumerate(offsets): 841 chars = offset["char"] 842 for char in chars: 843 state = "SPACE" if char == word_delimiter_char else "WORD" 844 845 if state == last_state: 846 # If we are in the same state as before, we simply repeat what we've done before 847 end_offset = offset["end_offset"] 848 word += char 849 else: 850 # Switching state 851 if state == "SPACE": 852 # Finishing a word 853 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 854 else: 855 # Starting a new word 856 start_offset = offset["start_offset"] 857 end_offset = offset["end_offset"] 858 word = char 859 860 last_state = state 861 862 if last_state == "WORD": 863 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 864 865 return word_offsets 866 867 @staticmethod 868 def _get_word_offsets_subwords_sentencepiece( 869 offsets: Dict[str, Union[str, float]], 870 hypothesis: Hypothesis, 871 decode_ids_to_tokens: Callable[[List[int]], str], 872 decode_tokens_to_str: Callable[[List[int]], str], 873 ) -> Dict[str, Union[str, float]]: 874 """ 875 Utility method which constructs word time stamps out of sub-word time stamps. 876 877 **Note**: Only supports Sentencepiece based tokenizers ! 878 879 Args: 880 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 881 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 882 after rnnt collapse. 883 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 884 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 885 886 Returns: 887 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 888 "end_offset". 889 """ 890 word_offsets = [] 891 built_token = [] 892 previous_token_index = 0 893 # For every offset token 894 for i, offset in enumerate(offsets): 895 # For every subword token in offset token list (ignoring the RNNT Blank token at the end) 896 for char in offset['char'][:-1]: 897 char = int(char) 898 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if built_token: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 # This should only be done when these arrays contain more than one element. 931 if offsets and word_offsets: 932 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 933 934 # If there are any remaining tokens left, inject them all into the final word offset. 935 # The start offset of this token is the start time of the next token to process. 936 # The end offset of this token is the end time of the last token from offsets. 937 # Note that built_token is a flat list; but offsets contains a nested list which 938 # may have different dimensionality. 939 # As such, we can't rely on the length of the list of built_token to index offsets. 940 if built_token: 941 # start from the previous token index as this hasn't been committed to word_offsets yet 942 # if we still have content in built_token 943 start_offset = offsets[previous_token_index]["start_offset"] 944 word_offsets.append( 945 { 946 "word": decode_tokens_to_str(built_token), 947 "start_offset": start_offset, 948 "end_offset": offsets[-1]["end_offset"], 949 } 950 ) 951 built_token.clear() 952 953 return word_offsets 954 955 956 class RNNTDecoding(AbstractRNNTDecoding): 957 """ 958 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 959 960 Args: 961 decoding_cfg: A dict-like object which contains the following key-value pairs. 962 strategy: str value which represents the type of decoding that can occur. 963 Possible values are : 964 - greedy, greedy_batch (for greedy decoding). 965 - beam, tsd, alsd (for beam search decoding). 966 967 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 968 tokens as well as the decoded string. Default is False in order to avoid double decoding 969 unless required. 970 971 preserve_alignments: Bool flag which preserves the history of logprobs generated during 972 decoding (sample / batched). When set to true, the Hypothesis will contain 973 the non-null value for `logprobs` in it. Here, `alignments` is a List of List of 974 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 975 976 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 977 with the `return_hypotheses` flag set to True. 978 979 The length of the list corresponds to the Acoustic Length (T). 980 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 981 U is the number of target tokens for the current timestep Ti. 982 983 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 984 scores. In order to obtain hypotheses with confidence scores, please utilize 985 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 986 987 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 988 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 989 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 990 991 The length of the list corresponds to the Acoustic Length (T). 992 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 993 U is the number of target tokens for the current timestep Ti. 994 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 995 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 996 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 997 998 The length of the list corresponds to the number of recognized tokens. 999 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1000 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1001 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1002 1003 The length of the list corresponds to the number of recognized words. 1004 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1005 from the `token_confidence`. 1006 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1007 Valid options are `mean`, `min`, `max`, `prod`. 1008 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1009 confidence scores. 1010 1011 name: The measure name (str). 1012 Supported values: 1013 - 'max_prob' for using the maximum token probability as a confidence. 1014 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1015 1016 entropy_type: Which type of entropy to use (str). 1017 Used if confidence_measure_cfg.name is set to `entropy`. 1018 Supported values: 1019 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1020 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1021 Note that for this entropy, the alpha should comply the following inequality: 1022 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1023 where V is the model vocabulary size. 1024 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1025 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1026 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1027 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1028 - 'renyi' for the Rรฉnyi entropy. 1029 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1030 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1031 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1032 1033 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1034 When the alpha equals one, scaling is not applied to 'max_prob', 1035 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1036 1037 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1038 Supported values: 1039 - 'lin' for using the linear mapping. 1040 - 'exp' for using exponential mapping with linear shift. 1041 1042 The config may further contain the following sub-dictionaries: 1043 "greedy": 1044 max_symbols: int, describing the maximum number of target tokens to decode per 1045 timestep during greedy decoding. Setting to larger values allows longer sentences 1046 to be decoded, at the cost of increased execution time. 1047 1048 preserve_frame_confidence: Same as above, overrides above value. 1049 1050 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 1051 1052 "beam": 1053 beam_size: int, defining the beam size for beam search. Must be >= 1. 1054 If beam_size == 1, will perform cached greedy search. This might be slightly different 1055 results compared to the greedy search above. 1056 1057 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 1058 Set to True by default. 1059 1060 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1061 hypotheses after beam search has concluded. This flag is set by default. 1062 1063 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 1064 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 1065 at increased cost to execution time. 1066 1067 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 1068 If an integer is provided, it can decode sequences of that particular maximum length. 1069 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 1070 where seq_len is the length of the acoustic model output (T). 1071 1072 NOTE: 1073 If a float is provided, it can be greater than 1! 1074 By default, a float of 2.0 is used so that a target sequence can be at most twice 1075 as long as the acoustic model output length T. 1076 1077 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 1078 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 1079 1080 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 1081 in order to reduce expensive beam search cost later. int >= 0. 1082 1083 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 1084 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 1085 and affects the speed of inference since large values will perform large beam search in the next step. 1086 1087 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 1088 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 1089 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 1090 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 1091 expansion apart from the "most likely" candidate. 1092 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 1093 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 1094 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 1095 tuned on a validation set. 1096 1097 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 1098 1099 decoder: The Decoder/Prediction network module. 1100 joint: The Joint network module. 1101 vocabulary: The vocabulary (excluding the RNNT blank token) which will be used for decoding. 1102 """ 1103 1104 def __init__( 1105 self, decoding_cfg, decoder, joint, vocabulary, 1106 ): 1107 # we need to ensure blank is the last token in the vocab for the case of RNNT and Multi-blank RNNT. 1108 blank_id = len(vocabulary) + joint.num_extra_outputs 1109 1110 if hasattr(decoding_cfg, 'model_type') and decoding_cfg.model_type == 'tdt': 1111 blank_id = len(vocabulary) 1112 1113 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1114 1115 super(RNNTDecoding, self).__init__( 1116 decoding_cfg=decoding_cfg, decoder=decoder, joint=joint, blank_id=blank_id, 1117 ) 1118 1119 if isinstance(self.decoding, beam_decode.BeamRNNTInfer): 1120 self.decoding.set_decoding_type('char') 1121 1122 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1123 """ 1124 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1125 1126 Args: 1127 hypothesis: Hypothesis 1128 1129 Returns: 1130 A list of word-level confidence scores. 1131 """ 1132 return self._aggregate_token_confidence_chars(hypothesis.words, hypothesis.token_confidence) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c < self.blank_id - self.num_extra_outputs] 1159 return token_list 1160 1161 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 1162 """ 1163 Compute the most likely language ID (LID) string given the tokens. 1164 1165 Args: 1166 tokens: List of int representing the token ids. 1167 1168 Returns: 1169 A decoded LID string. 1170 """ 1171 lang = self.tokenizer.ids_to_lang(tokens) 1172 return lang 1173 1174 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 1175 """ 1176 Decode a token id list into language ID (LID) list. 1177 1178 Args: 1179 tokens: List of int representing the token ids. 1180 1181 Returns: 1182 A list of decoded LIDS. 1183 """ 1184 lang_list = self.tokenizer.ids_to_text_and_langs(tokens) 1185 return lang_list 1186 1187 1188 class RNNTWER(Metric): 1189 """ 1190 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. 1191 When doing distributed training/evaluation the result of res=WER(predictions, targets, target_lengths) calls 1192 will be all-reduced between all workers using SUM operations. 1193 Here contains two numbers res=[wer_numerator, wer_denominator]. WER=wer_numerator/wer_denominator. 1194 1195 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. 1196 Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1197 1198 Example: 1199 def validation_step(self, batch, batch_idx): 1200 ... 1201 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1202 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1203 return self.val_outputs 1204 1205 def on_validation_epoch_end(self): 1206 ... 1207 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1208 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1209 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1210 self.val_outputs.clear() # free memory 1211 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1212 1213 Args: 1214 decoding: RNNTDecoding object that will perform autoregressive decoding of the RNNT model. 1215 batch_dim_index: Index of the batch dimension. 1216 use_cer: Whether to use Character Error Rate isntead of Word Error Rate. 1217 log_prediction: Whether to log a single decoded sample per call. 1218 1219 Returns: 1220 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenshtein's 1221 distances for all prediction - reference pairs, total number of words in all references. 1222 """ 1223 1224 full_state_update = True 1225 1226 def __init__( 1227 self, decoding: RNNTDecoding, batch_dim_index=0, use_cer=False, log_prediction=True, dist_sync_on_step=False 1228 ): 1229 super(RNNTWER, self).__init__(dist_sync_on_step=dist_sync_on_step) 1230 self.decoding = decoding 1231 self.batch_dim_index = batch_dim_index 1232 self.use_cer = use_cer 1233 self.log_prediction = log_prediction 1234 self.blank_id = self.decoding.blank_id 1235 self.labels_map = self.decoding.labels_map 1236 1237 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1238 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1239 1240 def update( 1241 self, 1242 encoder_output: torch.Tensor, 1243 encoded_lengths: torch.Tensor, 1244 targets: torch.Tensor, 1245 target_lengths: torch.Tensor, 1246 ) -> torch.Tensor: 1247 words = 0 1248 scores = 0 1249 references = [] 1250 with torch.no_grad(): 1251 # prediction_cpu_tensor = tensors[0].long().cpu() 1252 targets_cpu_tensor = targets.long().cpu() 1253 targets_cpu_tensor = move_dimension_to_the_front(targets_cpu_tensor, self.batch_dim_index) 1254 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1255 1256 # iterate over batch 1257 for ind in range(targets_cpu_tensor.shape[0]): 1258 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1259 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1260 1261 reference = self.decoding.decode_tokens_to_str(target) 1262 references.append(reference) 1263 1264 hypotheses, _ = self.decoding.rnnt_decoder_predictions_tensor(encoder_output, encoded_lengths) 1265 1266 if self.log_prediction: 1267 logging.info(f"\n") 1268 logging.info(f"reference :{references[0]}") 1269 logging.info(f"predicted :{hypotheses[0]}") 1270 1271 for h, r in zip(hypotheses, references): 1272 if self.use_cer: 1273 h_list = list(h) 1274 r_list = list(r) 1275 else: 1276 h_list = h.split() 1277 r_list = r.split() 1278 words += len(r_list) 1279 # Compute Levenshtein's distance 1280 scores += editdistance.eval(h_list, r_list) 1281 1282 self.scores += torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1283 self.words += torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1284 # return torch.tensor([scores, words]).to(predictions.device) 1285 1286 def compute(self): 1287 wer = self.scores.float() / self.words 1288 return wer, self.scores.detach(), self.words.detach() 1289 1290 1291 @dataclass 1292 class RNNTDecodingConfig: 1293 model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" 1294 strategy: str = "greedy_batch" 1295 1296 compute_hypothesis_token_set: bool = False 1297 1298 # preserve decoding alignments 1299 preserve_alignments: Optional[bool] = None 1300 1301 # confidence config 1302 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 1307 # compute RNNT time stamps 1308 compute_timestamps: Optional[bool] = None 1309 1310 # compute language IDs 1311 compute_langs: bool = False 1312 1313 # token representing word seperator 1314 word_seperator: str = " " 1315 1316 # type of timestamps to calculate 1317 rnnt_timestamp_type: str = "all" # can be char, word or all for both 1318 1319 # greedy decoding config 1320 greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() 1321 1322 # beam decoding config 1323 beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) 1324 1325 # can be used to change temperature for decoding 1326 temperature: float = 1.0 1327 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/metrics/wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import re 16 from abc import abstractmethod 17 from dataclasses import dataclass, is_dataclass 18 from typing import Callable, Dict, List, Optional, Tuple, Union 19 20 import editdistance 21 import jiwer 22 import numpy as np 23 import torch 24 from omegaconf import DictConfig, OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding 28 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 29 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 30 from nemo.utils import logging, logging_mode 31 32 __all__ = ['word_error_rate', 'word_error_rate_detail', 'WER', 'move_dimension_to_the_front'] 33 34 35 def word_error_rate(hypotheses: List[str], references: List[str], use_cer=False) -> float: 36 """ 37 Computes Average Word Error rate between two texts represented as 38 corresponding lists of string. 39 40 Hypotheses and references must have same length. 41 42 Args: 43 hypotheses (list): list of hypotheses 44 references(list) : list of references 45 use_cer (bool): set True to enable cer 46 47 Returns: 48 wer (float): average word error rate 49 """ 50 scores = 0 51 words = 0 52 if len(hypotheses) != len(references): 53 raise ValueError( 54 "In word error rate calculation, hypotheses and reference" 55 " lists must have the same number of elements. But I got:" 56 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 57 ) 58 for h, r in zip(hypotheses, references): 59 if use_cer: 60 h_list = list(h) 61 r_list = list(r) 62 else: 63 h_list = h.split() 64 r_list = r.split() 65 words += len(r_list) 66 # May deprecate using editdistance in future release for here and rest of codebase 67 # once we confirm jiwer is reliable. 68 scores += editdistance.eval(h_list, r_list) 69 if words != 0: 70 wer = 1.0 * scores / words 71 else: 72 wer = float('inf') 73 return wer 74 75 76 def word_error_rate_detail( 77 hypotheses: List[str], references: List[str], use_cer=False 78 ) -> Tuple[float, int, float, float, float]: 79 """ 80 Computes Average Word Error Rate with details (insertion rate, deletion rate, substitution rate) 81 between two texts represented as corresponding lists of string. 82 83 Hypotheses and references must have same length. 84 85 Args: 86 hypotheses (list): list of hypotheses 87 references(list) : list of references 88 use_cer (bool): set True to enable cer 89 90 Returns: 91 wer (float): average word error rate 92 words (int): Total number of words/charactors of given reference texts 93 ins_rate (float): average insertion error rate 94 del_rate (float): average deletion error rate 95 sub_rate (float): average substitution error rate 96 """ 97 scores = 0 98 words = 0 99 ops_count = {'substitutions': 0, 'insertions': 0, 'deletions': 0} 100 101 if len(hypotheses) != len(references): 102 raise ValueError( 103 "In word error rate calculation, hypotheses and reference" 104 " lists must have the same number of elements. But I got:" 105 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 106 ) 107 108 for h, r in zip(hypotheses, references): 109 if use_cer: 110 h_list = list(h) 111 r_list = list(r) 112 else: 113 h_list = h.split() 114 r_list = r.split() 115 116 # To get rid of the issue that jiwer does not allow empty string 117 if len(r_list) == 0: 118 if len(h_list) != 0: 119 errors = len(h_list) 120 ops_count['insertions'] += errors 121 else: 122 errors = 0 123 else: 124 if use_cer: 125 measures = jiwer.cer(r, h, return_dict=True) 126 else: 127 measures = jiwer.compute_measures(r, h) 128 129 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 130 ops_count['insertions'] += measures['insertions'] 131 ops_count['deletions'] += measures['deletions'] 132 ops_count['substitutions'] += measures['substitutions'] 133 134 scores += errors 135 words += len(r_list) 136 137 if words != 0: 138 wer = 1.0 * scores / words 139 ins_rate = 1.0 * ops_count['insertions'] / words 140 del_rate = 1.0 * ops_count['deletions'] / words 141 sub_rate = 1.0 * ops_count['substitutions'] / words 142 else: 143 wer, ins_rate, del_rate, sub_rate = float('inf'), float('inf'), float('inf'), float('inf') 144 145 return wer, words, ins_rate, del_rate, sub_rate 146 147 148 def word_error_rate_per_utt(hypotheses: List[str], references: List[str], use_cer=False) -> Tuple[List[float], float]: 149 """ 150 Computes Word Error Rate per utterance and the average WER 151 between two texts represented as corresponding lists of string. 152 153 Hypotheses and references must have same length. 154 155 Args: 156 hypotheses (list): list of hypotheses 157 references(list) : list of references 158 use_cer (bool): set True to enable cer 159 160 Returns: 161 wer_per_utt (List[float]): word error rate per utterance 162 avg_wer (float): average word error rate 163 """ 164 scores = 0 165 words = 0 166 wer_per_utt = [] 167 168 if len(hypotheses) != len(references): 169 raise ValueError( 170 "In word error rate calculation, hypotheses and reference" 171 " lists must have the same number of elements. But I got:" 172 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 173 ) 174 175 for h, r in zip(hypotheses, references): 176 if use_cer: 177 h_list = list(h) 178 r_list = list(r) 179 else: 180 h_list = h.split() 181 r_list = r.split() 182 183 # To get rid of the issue that jiwer does not allow empty string 184 if len(r_list) == 0: 185 if len(h_list) != 0: 186 errors = len(h_list) 187 wer_per_utt.append(float('inf')) 188 else: 189 if use_cer: 190 measures = jiwer.cer(r, h, return_dict=True) 191 er = measures['cer'] 192 else: 193 measures = jiwer.compute_measures(r, h) 194 er = measures['wer'] 195 196 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 197 wer_per_utt.append(er) 198 199 scores += errors 200 words += len(r_list) 201 202 if words != 0: 203 avg_wer = 1.0 * scores / words 204 else: 205 avg_wer = float('inf') 206 207 return wer_per_utt, avg_wer 208 209 210 def move_dimension_to_the_front(tensor, dim_index): 211 all_dims = list(range(tensor.ndim)) 212 return tensor.permute(*([dim_index] + all_dims[:dim_index] + all_dims[dim_index + 1 :])) 213 214 215 class AbstractCTCDecoding(ConfidenceMixin): 216 """ 217 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs. 218 219 Args: 220 decoding_cfg: A dict-like object which contains the following key-value pairs. 221 strategy: str value which represents the type of decoding that can occur. 222 Possible values are : 223 - greedy (for greedy decoding). 224 - beam (for DeepSpeed KenLM based decoding). 225 226 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 227 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 228 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 229 230 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 231 Can take the following values - "char" for character/subword time stamps, "word" for word level 232 time stamps and "all" (default), for both character level and word level time stamps. 233 234 word_seperator: Str token representing the seperator between words. 235 236 preserve_alignments: Bool flag which preserves the history of logprobs generated during 237 decoding (sample / batched). When set to true, the Hypothesis will contain 238 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 239 240 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 241 scores. In order to obtain hypotheses with confidence scores, please utilize 242 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 243 244 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 245 generated during decoding. When set to true, the Hypothesis will contain 246 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 247 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 248 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 249 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 250 251 The length of the list corresponds to the number of recognized tokens. 252 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 253 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 254 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 255 256 The length of the list corresponds to the number of recognized words. 257 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 258 from the `token_confidence`. 259 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 260 Valid options are `mean`, `min`, `max`, `prod`. 261 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 262 confidence scores. 263 264 name: The measure name (str). 265 Supported values: 266 - 'max_prob' for using the maximum token probability as a confidence. 267 - 'entropy' for using a normalized entropy of a log-likelihood vector. 268 269 entropy_type: Which type of entropy to use (str). 270 Used if confidence_measure_cfg.name is set to `entropy`. 271 Supported values: 272 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 273 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 274 Note that for this entropy, the alpha should comply the following inequality: 275 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 276 where V is the model vocabulary size. 277 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 278 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 279 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 280 More: https://en.wikipedia.org/wiki/Tsallis_entropy 281 - 'renyi' for the Rรฉnyi entropy. 282 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 285 286 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 287 When the alpha equals one, scaling is not applied to 'max_prob', 288 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 289 290 entropy_norm: A mapping of the entropy value to the interval [0,1]. 291 Supported values: 292 - 'lin' for using the linear mapping. 293 - 'exp' for using exponential mapping with linear shift. 294 295 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 296 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 297 298 The config may further contain the following sub-dictionaries: 299 "greedy": 300 preserve_alignments: Same as above, overrides above value. 301 compute_timestamps: Same as above, overrides above value. 302 preserve_frame_confidence: Same as above, overrides above value. 303 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 304 305 "beam": 306 beam_size: int, defining the beam size for beam search. Must be >= 1. 307 If beam_size == 1, will perform cached greedy search. This might be slightly different 308 results compared to the greedy search above. 309 310 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 311 hypotheses after beam search has concluded. This flag is set by default. 312 313 beam_alpha: float, the strength of the Language model on the final score of a token. 314 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 315 316 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 317 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 318 319 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 320 If the path is invalid (file is not found at path), will raise a deferred error at the moment 321 of calculation of beam search, so that users may update / change the decoding strategy 322 to point to the correct file. 323 324 blank_id: The id of the RNNT blank token. 325 """ 326 327 def __init__(self, decoding_cfg, blank_id: int): 328 super().__init__() 329 330 # Convert dataclas to config 331 if is_dataclass(decoding_cfg): 332 decoding_cfg = OmegaConf.structured(decoding_cfg) 333 334 if not isinstance(decoding_cfg, DictConfig): 335 decoding_cfg = OmegaConf.create(decoding_cfg) 336 337 OmegaConf.set_struct(decoding_cfg, False) 338 339 # update minimal config 340 minimal_cfg = ['greedy'] 341 for item in minimal_cfg: 342 if item not in decoding_cfg: 343 decoding_cfg[item] = OmegaConf.create({}) 344 345 self.cfg = decoding_cfg 346 self.blank_id = blank_id 347 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 348 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 349 self.batch_dim_index = self.cfg.get('batch_dim_index', 0) 350 self.word_seperator = self.cfg.get('word_seperator', ' ') 351 352 possible_strategies = ['greedy', 'beam', 'pyctcdecode', 'flashlight'] 353 if self.cfg.strategy not in possible_strategies: 354 raise ValueError(f"Decoding strategy must be one of {possible_strategies}. Given {self.cfg.strategy}") 355 356 # Update preserve alignments 357 if self.preserve_alignments is None: 358 if self.cfg.strategy in ['greedy']: 359 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 360 else: 361 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 362 363 # Update compute timestamps 364 if self.compute_timestamps is None: 365 if self.cfg.strategy in ['greedy']: 366 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 367 elif self.cfg.strategy in ['beam']: 368 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 369 370 # initialize confidence-related fields 371 self._init_confidence(self.cfg.get('confidence_cfg', None)) 372 373 # Confidence estimation is not implemented for strategies other than `greedy` 374 if ( 375 not self.preserve_frame_confidence 376 and self.cfg.strategy != 'greedy' 377 and self.cfg.beam.get('preserve_frame_confidence', False) 378 ): 379 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 380 381 # we need timestamps to extract non-blank per-frame confidence 382 if self.compute_timestamps is not None: 383 self.compute_timestamps |= self.preserve_frame_confidence 384 385 if self.cfg.strategy == 'greedy': 386 387 self.decoding = ctc_greedy_decoding.GreedyCTCInfer( 388 blank_id=self.blank_id, 389 preserve_alignments=self.preserve_alignments, 390 compute_timestamps=self.compute_timestamps, 391 preserve_frame_confidence=self.preserve_frame_confidence, 392 confidence_measure_cfg=self.confidence_measure_cfg, 393 ) 394 395 elif self.cfg.strategy == 'beam': 396 397 self.decoding = ctc_beam_decoding.BeamCTCInfer( 398 blank_id=blank_id, 399 beam_size=self.cfg.beam.get('beam_size', 1), 400 search_type='default', 401 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 402 preserve_alignments=self.preserve_alignments, 403 compute_timestamps=self.compute_timestamps, 404 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 405 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 406 kenlm_path=self.cfg.beam.get('kenlm_path', None), 407 ) 408 409 self.decoding.override_fold_consecutive_value = False 410 411 elif self.cfg.strategy == 'pyctcdecode': 412 413 self.decoding = ctc_beam_decoding.BeamCTCInfer( 414 blank_id=blank_id, 415 beam_size=self.cfg.beam.get('beam_size', 1), 416 search_type='pyctcdecode', 417 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 418 preserve_alignments=self.preserve_alignments, 419 compute_timestamps=self.compute_timestamps, 420 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 421 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 422 kenlm_path=self.cfg.beam.get('kenlm_path', None), 423 pyctcdecode_cfg=self.cfg.beam.get('pyctcdecode_cfg', None), 424 ) 425 426 self.decoding.override_fold_consecutive_value = False 427 428 elif self.cfg.strategy == 'flashlight': 429 430 self.decoding = ctc_beam_decoding.BeamCTCInfer( 431 blank_id=blank_id, 432 beam_size=self.cfg.beam.get('beam_size', 1), 433 search_type='flashlight', 434 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 435 preserve_alignments=self.preserve_alignments, 436 compute_timestamps=self.compute_timestamps, 437 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 438 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 439 kenlm_path=self.cfg.beam.get('kenlm_path', None), 440 flashlight_cfg=self.cfg.beam.get('flashlight_cfg', None), 441 ) 442 443 self.decoding.override_fold_consecutive_value = False 444 445 else: 446 raise ValueError( 447 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 448 f"but was provided {self.cfg.strategy}" 449 ) 450 451 def ctc_decoder_predictions_tensor( 452 self, 453 decoder_outputs: torch.Tensor, 454 decoder_lengths: torch.Tensor = None, 455 fold_consecutive: bool = True, 456 return_hypotheses: bool = False, 457 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 458 """ 459 Decodes a sequence of labels to words 460 461 Args: 462 decoder_outputs: An integer torch.Tensor of shape [Batch, Time, {Vocabulary}] (if ``batch_index_dim == 0``) or [Time, Batch] 463 (if ``batch_index_dim == 1``) of integer indices that correspond to the index of some character in the 464 label set. 465 decoder_lengths: Optional tensor of length `Batch` which contains the integer lengths 466 of the sequence in the padded `predictions` tensor. 467 fold_consecutive: Bool, determine whether to perform "ctc collapse", folding consecutive tokens 468 into a single token. 469 return_hypotheses: Bool flag whether to return just the decoding predictions of the model 470 or a Hypothesis object that holds information such as the decoded `text`, 471 the `alignment` of emited by the CTC Model, and the `length` of the sequence (if available). 472 May also contain the log-probabilities of the decoder (if this method is called via 473 transcribe()) 474 475 Returns: 476 Either a list of str which represent the CTC decoded strings per sample, 477 or a list of Hypothesis objects containing additional information. 478 """ 479 480 if isinstance(decoder_outputs, torch.Tensor): 481 decoder_outputs = move_dimension_to_the_front(decoder_outputs, self.batch_dim_index) 482 483 if ( 484 hasattr(self.decoding, 'override_fold_consecutive_value') 485 and self.decoding.override_fold_consecutive_value is not None 486 ): 487 logging.info( 488 f"Beam search requires that consecutive ctc tokens are not folded. \n" 489 f"Overriding provided value of `fold_consecutive` = {fold_consecutive} to " 490 f"{self.decoding.override_fold_consecutive_value}", 491 mode=logging_mode.ONCE, 492 ) 493 fold_consecutive = self.decoding.override_fold_consecutive_value 494 495 with torch.inference_mode(): 496 # Resolve the forward step of the decoding strategy 497 hypotheses_list = self.decoding( 498 decoder_output=decoder_outputs, decoder_lengths=decoder_lengths 499 ) # type: List[List[Hypothesis]] 500 501 # extract the hypotheses 502 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 503 504 if isinstance(hypotheses_list[0], NBestHypotheses): 505 hypotheses = [] 506 all_hypotheses = [] 507 508 for nbest_hyp in hypotheses_list: # type: NBestHypotheses 509 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 510 decoded_hyps = self.decode_hypothesis( 511 n_hyps, fold_consecutive 512 ) # type: List[Union[Hypothesis, NBestHypotheses]] 513 514 # If computing timestamps 515 if self.compute_timestamps is True: 516 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 517 for hyp_idx in range(len(decoded_hyps)): 518 decoded_hyps[hyp_idx] = self.compute_ctc_timestamps(decoded_hyps[hyp_idx], timestamp_type) 519 520 hypotheses.append(decoded_hyps[0]) # best hypothesis 521 all_hypotheses.append(decoded_hyps) 522 523 if return_hypotheses: 524 return hypotheses, all_hypotheses 525 526 best_hyp_text = [h.text for h in hypotheses] 527 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 528 return best_hyp_text, all_hyp_text 529 530 else: 531 hypotheses = self.decode_hypothesis( 532 hypotheses_list, fold_consecutive 533 ) # type: List[Union[Hypothesis, NBestHypotheses]] 534 535 # If computing timestamps 536 if self.compute_timestamps is True: 537 # greedy decoding, can get high-level confidence scores 538 if return_hypotheses and (self.preserve_word_confidence or self.preserve_token_confidence): 539 hypotheses = self.compute_confidence(hypotheses) 540 else: 541 # remove unused token_repetitions from Hypothesis.text 542 for hyp in hypotheses: 543 hyp.text = hyp.text[:2] 544 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 545 for hyp_idx in range(len(hypotheses)): 546 hypotheses[hyp_idx] = self.compute_ctc_timestamps(hypotheses[hyp_idx], timestamp_type) 547 548 if return_hypotheses: 549 return hypotheses, None 550 551 best_hyp_text = [h.text for h in hypotheses] 552 return best_hyp_text, None 553 554 def decode_hypothesis( 555 self, hypotheses_list: List[Hypothesis], fold_consecutive: bool 556 ) -> List[Union[Hypothesis, NBestHypotheses]]: 557 """ 558 Decode a list of hypotheses into a list of strings. 559 560 Args: 561 hypotheses_list: List of Hypothesis. 562 fold_consecutive: Whether to collapse the ctc blank tokens or not. 563 564 Returns: 565 A list of strings. 566 """ 567 for ind in range(len(hypotheses_list)): 568 # Extract the integer encoded hypothesis 569 hyp = hypotheses_list[ind] 570 prediction = hyp.y_sequence 571 predictions_len = hyp.length if hyp.length > 0 else None 572 573 if fold_consecutive: 574 if type(prediction) != list: 575 prediction = prediction.numpy().tolist() 576 577 if predictions_len is not None: 578 prediction = prediction[:predictions_len] 579 580 # CTC decoding procedure 581 decoded_prediction = [] 582 token_lengths = [] # preserve token lengths 583 token_repetitions = [] # preserve number of repetitions per token 584 585 previous = self.blank_id 586 last_length = 0 587 last_repetition = 1 588 589 for pidx, p in enumerate(prediction): 590 if (p != previous or previous == self.blank_id) and p != self.blank_id: 591 decoded_prediction.append(p) 592 593 token_lengths.append(pidx - last_length) 594 last_length = pidx 595 token_repetitions.append(last_repetition) 596 last_repetition = 1 597 598 if p == previous and previous != self.blank_id: 599 last_repetition += 1 600 601 previous = p 602 603 if len(token_repetitions) > 0: 604 token_repetitions = token_repetitions[1:] + [last_repetition] 605 606 else: 607 if predictions_len is not None: 608 prediction = prediction[:predictions_len] 609 decoded_prediction = prediction[prediction != self.blank_id].tolist() 610 token_lengths = [1] * len(decoded_prediction) # preserve number of repetitions per token 611 token_repetitions = [1] * len(decoded_prediction) # preserve number of repetitions per token 612 613 # De-tokenize the integer tokens; if not computing timestamps 614 if self.compute_timestamps is True: 615 # keep the original predictions, wrap with the number of repetitions per token 616 # this is done so that `ctc_decoder_predictions_tensor()` can process this hypothesis 617 # in order to compute exact time stamps. 618 hypothesis = (decoded_prediction, token_lengths, token_repetitions) 619 else: 620 hypothesis = self.decode_tokens_to_str(decoded_prediction) 621 622 # TODO: remove 623 # collapse leading spaces before . , ? for PC models 624 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 625 626 # Preserve this wrapped hypothesis or decoded text tokens. 627 hypotheses_list[ind].text = hypothesis 628 629 return hypotheses_list 630 631 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 632 """ 633 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 634 Assumes that `frame_confidence` is present in the hypotheses. 635 636 Args: 637 hypotheses_list: List of Hypothesis. 638 639 Returns: 640 A list of hypotheses with high-level confidence scores. 641 """ 642 for hyp in hypotheses_list: 643 if not isinstance(hyp.text, tuple) or len(hyp.text) != 3: 644 # the method must have been called in the wrong place 645 raise ValueError( 646 """Wrong format of the `text` attribute of a hypothesis.\n 647 Expected: (decoded_prediction, token_repetitions)\n 648 The method invocation is expected between .decode_hypothesis() and .compute_ctc_timestamps()""" 649 ) 650 token_repetitions = hyp.text[2] 651 hyp.text = hyp.text[:2] 652 token_confidence = [] 653 if self.exclude_blank_from_confidence: 654 non_blank_frame_confidence = hyp.non_blank_frame_confidence 655 i = 0 656 for tr in token_repetitions: 657 # token repetition can be zero 658 j = i + tr 659 token_confidence.append(self._aggregate_confidence(non_blank_frame_confidence[i:j])) 660 i = j 661 else: 662 # <blank> tokens are considered to belong to the last non-blank token, if any. 663 token_lengths = hyp.text[1] 664 if len(token_lengths) > 0: 665 ts = token_lengths[0] 666 for tl in token_lengths[1:] + [len(hyp.frame_confidence)]: 667 token_confidence.append(self._aggregate_confidence(hyp.frame_confidence[ts : ts + tl])) 668 ts += tl 669 hyp.token_confidence = token_confidence 670 if self.preserve_word_confidence: 671 for hyp in hypotheses_list: 672 hyp.word_confidence = self._aggregate_token_confidence(hyp) 673 return hypotheses_list 674 675 @abstractmethod 676 def decode_tokens_to_str(self, tokens: List[int]) -> str: 677 """ 678 Implemented by subclass in order to decoder a token id list into a string. 679 680 Args: 681 tokens: List of int representing the token ids. 682 683 Returns: 684 A decoded string. 685 """ 686 raise NotImplementedError() 687 688 @abstractmethod 689 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 690 """ 691 Implemented by subclass in order to decode a token id list into a token list. 692 A token list is the string representation of each token id. 693 694 Args: 695 tokens: List of int representing the token ids. 696 697 Returns: 698 A list of decoded tokens. 699 """ 700 raise NotImplementedError() 701 702 def compute_ctc_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 703 """ 704 Method to compute time stamps at char/subword, and word level given some hypothesis. 705 Requires the input hypothesis to contain a `text` field that is the tuple. The tuple contains - 706 the ctc collapsed integer ids, and the number of repetitions of each token. 707 708 Args: 709 hypothesis: A Hypothesis object, with a wrapped `text` field. 710 The `text` field must contain a tuple with two values - 711 The ctc collapsed integer ids 712 A list of integers that represents the number of repetitions per token. 713 timestamp_type: A str value that represents the type of time stamp calculated. 714 Can be one of "char", "word" or "all" 715 716 Returns: 717 A Hypothesis object with a modified `timestep` value, which is now a dictionary containing 718 the time stamp information. 719 """ 720 assert timestamp_type in ['char', 'word', 'all'] 721 722 # Unpack the temporary storage, and set the decoded predictions 723 decoded_prediction, token_lengths = hypothesis.text 724 hypothesis.text = decoded_prediction 725 726 # Retrieve offsets 727 char_offsets = word_offsets = None 728 char_offsets = self._compute_offsets(hypothesis, token_lengths, self.blank_id) 729 730 # Assert number of offsets and hypothesis tokens are 1:1 match. 731 if len(char_offsets) != len(hypothesis.text): 732 raise ValueError( 733 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 734 " have to be of the same length, but are: " 735 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 736 f" {len(hypothesis.text)}" 737 ) 738 739 # Correctly process the token ids to chars/subwords. 740 for i, char in enumerate(hypothesis.text): 741 char_offsets[i]["char"] = self.decode_tokens_to_str([char]) 742 743 # detect char vs subword models 744 lens = [len(list(v["char"])) > 1 for v in char_offsets] 745 if any(lens): 746 text_type = 'subword' 747 else: 748 text_type = 'char' 749 750 # retrieve word offsets from character offsets 751 word_offsets = None 752 if timestamp_type in ['word', 'all']: 753 if text_type == 'char': 754 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 755 else: 756 word_offsets = self._get_word_offsets_subwords_sentencepiece( 757 char_offsets, 758 hypothesis, 759 decode_ids_to_tokens=self.decode_ids_to_tokens, 760 decode_tokens_to_str=self.decode_tokens_to_str, 761 ) 762 763 # attach results 764 if len(hypothesis.timestep) > 0: 765 timestep_info = hypothesis.timestep 766 else: 767 timestep_info = [] 768 769 # Setup defaults 770 hypothesis.timestep = {"timestep": timestep_info} 771 772 # Add char / subword time stamps 773 if char_offsets is not None and timestamp_type in ['char', 'all']: 774 hypothesis.timestep['char'] = char_offsets 775 776 # Add word time stamps 777 if word_offsets is not None and timestamp_type in ['word', 'all']: 778 hypothesis.timestep['word'] = word_offsets 779 780 # Convert the token indices to text 781 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 782 783 return hypothesis 784 785 @staticmethod 786 def _compute_offsets( 787 hypothesis: Hypothesis, token_lengths: List[int], ctc_token: int 788 ) -> List[Dict[str, Union[str, int]]]: 789 """ 790 Utility method that calculates the indidual time indices where a token starts and ends. 791 792 Args: 793 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 794 emitted at every time step after ctc collapse. 795 token_lengths: A list of ints representing the lengths of each emitted token. 796 ctc_token: The integer of the ctc blank token used during ctc collapse. 797 798 Returns: 799 800 """ 801 start_index = 0 802 803 # If the exact timestep information is available, utilize the 1st non-ctc blank token timestep 804 # as the start index. 805 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 806 start_index = max(0, hypothesis.timestep[0] - 1) 807 808 # Construct the start and end indices brackets 809 end_indices = np.asarray(token_lengths).cumsum() 810 start_indices = np.concatenate(([start_index], end_indices[:-1])) 811 812 # Merge the results per token into a list of dictionaries 813 offsets = [ 814 {"char": t, "start_offset": s, "end_offset": e} 815 for t, s, e in zip(hypothesis.text, start_indices, end_indices) 816 ] 817 818 # Filter out CTC token 819 offsets = list(filter(lambda offsets: offsets["char"] != ctc_token, offsets)) 820 return offsets 821 822 @staticmethod 823 def _get_word_offsets_chars( 824 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 825 ) -> Dict[str, Union[str, float]]: 826 """ 827 Utility method which constructs word time stamps out of character time stamps. 828 829 References: 830 This code is a port of the Hugging Face code for word time stamp construction. 831 832 Args: 833 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 834 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 835 836 Returns: 837 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 838 "end_offset". 839 """ 840 word_offsets = [] 841 842 last_state = "SPACE" 843 word = "" 844 start_offset = 0 845 end_offset = 0 846 for i, offset in enumerate(offsets): 847 char = offset["char"] 848 state = "SPACE" if char == word_delimiter_char else "WORD" 849 850 if state == last_state: 851 # If we are in the same state as before, we simply repeat what we've done before 852 end_offset = offset["end_offset"] 853 word += char 854 else: 855 # Switching state 856 if state == "SPACE": 857 # Finishing a word 858 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 859 else: 860 # Starting a new word 861 start_offset = offset["start_offset"] 862 end_offset = offset["end_offset"] 863 word = char 864 865 last_state = state 866 if last_state == "WORD": 867 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 868 869 return word_offsets 870 871 @staticmethod 872 def _get_word_offsets_subwords_sentencepiece( 873 offsets: Dict[str, Union[str, float]], 874 hypothesis: Hypothesis, 875 decode_ids_to_tokens: Callable[[List[int]], str], 876 decode_tokens_to_str: Callable[[List[int]], str], 877 ) -> Dict[str, Union[str, float]]: 878 """ 879 Utility method which constructs word time stamps out of sub-word time stamps. 880 881 **Note**: Only supports Sentencepiece based tokenizers ! 882 883 Args: 884 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 885 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 886 after ctc collapse. 887 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 888 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 889 890 Returns: 891 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 892 "end_offset". 893 """ 894 word_offsets = [] 895 built_token = [] 896 previous_token_index = 0 897 # For every collapsed sub-word token 898 for i, char in enumerate(hypothesis.text): 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if len(built_token) > 0: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 if len(word_offsets) == 0: 931 # alaptev: sometimes word_offsets can be empty 932 if len(built_token) > 0: 933 word_offsets.append( 934 { 935 "word": decode_tokens_to_str(built_token), 936 "start_offset": offsets[0]["start_offset"], 937 "end_offset": offsets[-1]["end_offset"], 938 } 939 ) 940 built_token.clear() 941 else: 942 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 943 944 # If there are any remaining tokens left, inject them all into the final word offset. 945 # Note: The start offset of this token is the start time of the first token inside build_token. 946 # Note: The end offset of this token is the end time of the last token inside build_token 947 if len(built_token) > 0: 948 word_offsets.append( 949 { 950 "word": decode_tokens_to_str(built_token), 951 "start_offset": offsets[-(len(built_token))]["start_offset"], 952 "end_offset": offsets[-1]["end_offset"], 953 } 954 ) 955 built_token.clear() 956 957 return word_offsets 958 959 @property 960 def preserve_alignments(self): 961 return self._preserve_alignments 962 963 @preserve_alignments.setter 964 def preserve_alignments(self, value): 965 self._preserve_alignments = value 966 967 if hasattr(self, 'decoding'): 968 self.decoding.preserve_alignments = value 969 970 @property 971 def compute_timestamps(self): 972 return self._compute_timestamps 973 974 @compute_timestamps.setter 975 def compute_timestamps(self, value): 976 self._compute_timestamps = value 977 978 if hasattr(self, 'decoding'): 979 self.decoding.compute_timestamps = value 980 981 @property 982 def preserve_frame_confidence(self): 983 return self._preserve_frame_confidence 984 985 @preserve_frame_confidence.setter 986 def preserve_frame_confidence(self, value): 987 self._preserve_frame_confidence = value 988 989 if hasattr(self, 'decoding'): 990 self.decoding.preserve_frame_confidence = value 991 992 993 class CTCDecoding(AbstractCTCDecoding): 994 """ 995 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs for character 996 based models. 997 998 Args: 999 decoding_cfg: A dict-like object which contains the following key-value pairs. 1000 strategy: str value which represents the type of decoding that can occur. 1001 Possible values are : 1002 - greedy (for greedy decoding). 1003 - beam (for DeepSpeed KenLM based decoding). 1004 1005 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 1006 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 1007 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 1008 1009 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 1010 Can take the following values - "char" for character/subword time stamps, "word" for word level 1011 time stamps and "all" (default), for both character level and word level time stamps. 1012 1013 word_seperator: Str token representing the seperator between words. 1014 1015 preserve_alignments: Bool flag which preserves the history of logprobs generated during 1016 decoding (sample / batched). When set to true, the Hypothesis will contain 1017 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 1018 1019 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 1020 scores. In order to obtain hypotheses with confidence scores, please utilize 1021 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 1022 1023 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 1024 generated during decoding. When set to true, the Hypothesis will contain 1025 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 1026 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 1027 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1028 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 1029 1030 The length of the list corresponds to the number of recognized tokens. 1031 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1032 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1033 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1034 1035 The length of the list corresponds to the number of recognized words. 1036 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1037 from the `token_confidence`. 1038 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1039 Valid options are `mean`, `min`, `max`, `prod`. 1040 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1041 confidence scores. 1042 1043 name: The measure name (str). 1044 Supported values: 1045 - 'max_prob' for using the maximum token probability as a confidence. 1046 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1047 1048 entropy_type: Which type of entropy to use (str). 1049 Used if confidence_measure_cfg.name is set to `entropy`. 1050 Supported values: 1051 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1052 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1053 Note that for this entropy, the alpha should comply the following inequality: 1054 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1055 where V is the model vocabulary size. 1056 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1057 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1058 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1059 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1060 - 'renyi' for the Rรฉnyi entropy. 1061 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1062 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1063 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1064 1065 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1066 When the alpha equals one, scaling is not applied to 'max_prob', 1067 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1068 1069 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1070 Supported values: 1071 - 'lin' for using the linear mapping. 1072 - 'exp' for using exponential mapping with linear shift. 1073 1074 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 1075 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 1076 1077 The config may further contain the following sub-dictionaries: 1078 "greedy": 1079 preserve_alignments: Same as above, overrides above value. 1080 compute_timestamps: Same as above, overrides above value. 1081 preserve_frame_confidence: Same as above, overrides above value. 1082 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 1083 1084 "beam": 1085 beam_size: int, defining the beam size for beam search. Must be >= 1. 1086 If beam_size == 1, will perform cached greedy search. This might be slightly different 1087 results compared to the greedy search above. 1088 1089 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1090 hypotheses after beam search has concluded. This flag is set by default. 1091 1092 beam_alpha: float, the strength of the Language model on the final score of a token. 1093 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1094 1095 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 1096 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1097 1098 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 1099 If the path is invalid (file is not found at path), will raise a deferred error at the moment 1100 of calculation of beam search, so that users may update / change the decoding strategy 1101 to point to the correct file. 1102 1103 blank_id: The id of the RNNT blank token. 1104 """ 1105 1106 def __init__( 1107 self, decoding_cfg, vocabulary, 1108 ): 1109 blank_id = len(vocabulary) 1110 self.vocabulary = vocabulary 1111 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1112 1113 super().__init__(decoding_cfg=decoding_cfg, blank_id=blank_id) 1114 1115 # Finalize Beam Search Decoding framework 1116 if isinstance(self.decoding, ctc_beam_decoding.AbstractBeamCTCInfer): 1117 self.decoding.set_vocabulary(self.vocabulary) 1118 self.decoding.set_decoding_type('char') 1119 1120 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1121 """ 1122 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1123 1124 Args: 1125 hypothesis: Hypothesis 1126 1127 Returns: 1128 A list of word-level confidence scores. 1129 """ 1130 return self._aggregate_token_confidence_chars( 1131 self.decode_tokens_to_str(hypothesis.text[0]).split(), hypothesis.token_confidence 1132 ) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c != self.blank_id] 1159 return token_list 1160 1161 1162 class WER(Metric): 1163 """ 1164 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference 1165 texts. When doing distributed training/evaluation the result of ``res=WER(predictions, targets, target_lengths)`` 1166 calls will be all-reduced between all workers using SUM operations. Here ``res`` contains three numbers 1167 ``res=[wer, total_levenstein_distance, total_number_of_words]``. 1168 1169 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step 1170 results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1171 1172 Example: 1173 def validation_step(self, batch, batch_idx): 1174 ... 1175 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1176 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1177 return self.val_outputs 1178 1179 def on_validation_epoch_end(self): 1180 ... 1181 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1182 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1183 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1184 self.val_outputs.clear() # free memory 1185 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1186 1187 Args: 1188 decoding: An instance of CTCDecoding. 1189 use_cer: Whether to use Character Error Rate instead of Word Error Rate. 1190 log_prediction: Whether to log a single decoded sample per call. 1191 fold_consecutive: Whether repeated consecutive characters should be folded into one when decoding. 1192 1193 Returns: 1194 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenstein's 1195 distances for all prediction - reference pairs, total number of words in all references. 1196 """ 1197 1198 full_state_update: bool = True 1199 1200 def __init__( 1201 self, 1202 decoding: CTCDecoding, 1203 use_cer=False, 1204 log_prediction=True, 1205 fold_consecutive=True, 1206 dist_sync_on_step=False, 1207 ): 1208 super().__init__(dist_sync_on_step=dist_sync_on_step) 1209 1210 self.decoding = decoding 1211 self.use_cer = use_cer 1212 self.log_prediction = log_prediction 1213 self.fold_consecutive = fold_consecutive 1214 1215 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1216 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1217 1218 def update( 1219 self, 1220 predictions: torch.Tensor, 1221 targets: torch.Tensor, 1222 target_lengths: torch.Tensor, 1223 predictions_lengths: torch.Tensor = None, 1224 ): 1225 """ 1226 Updates metric state. 1227 Args: 1228 predictions: an integer torch.Tensor of shape ``[Batch, Time, {Vocabulary}]`` (if ``batch_dim_index == 0``) or 1229 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1230 targets: an integer torch.Tensor of shape ``[Batch, Time]`` (if ``batch_dim_index == 0``) or 1231 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1232 target_lengths: an integer torch.Tensor of shape ``[Batch]`` 1233 predictions_lengths: an integer torch.Tensor of shape ``[Batch]`` 1234 """ 1235 words = 0 1236 scores = 0 1237 references = [] 1238 with torch.no_grad(): 1239 # prediction_cpu_tensor = tensors[0].long().cpu() 1240 targets_cpu_tensor = targets.long().cpu() 1241 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1242 1243 # iterate over batch 1244 for ind in range(targets_cpu_tensor.shape[0]): 1245 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1246 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1247 reference = self.decoding.decode_tokens_to_str(target) 1248 references.append(reference) 1249 1250 hypotheses, _ = self.decoding.ctc_decoder_predictions_tensor( 1251 predictions, predictions_lengths, fold_consecutive=self.fold_consecutive 1252 ) 1253 1254 if self.log_prediction: 1255 logging.info(f"\n") 1256 logging.info(f"reference:{references[0]}") 1257 logging.info(f"predicted:{hypotheses[0]}") 1258 1259 for h, r in zip(hypotheses, references): 1260 if self.use_cer: 1261 h_list = list(h) 1262 r_list = list(r) 1263 else: 1264 h_list = h.split() 1265 r_list = r.split() 1266 words += len(r_list) 1267 # Compute Levenstein's distance 1268 scores += editdistance.eval(h_list, r_list) 1269 1270 self.scores = torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1271 self.words = torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1272 # return torch.tensor([scores, words]).to(predictions.device) 1273 1274 def compute(self): 1275 scores = self.scores.detach().float() 1276 words = self.words.detach().float() 1277 return scores / words, scores, words 1278 1279 1280 @dataclass 1281 class CTCDecodingConfig: 1282 strategy: str = "greedy" 1283 1284 # preserve decoding alignments 1285 preserve_alignments: Optional[bool] = None 1286 1287 # compute ctc time stamps 1288 compute_timestamps: Optional[bool] = None 1289 1290 # token representing word seperator 1291 word_seperator: str = " " 1292 1293 # type of timestamps to calculate 1294 ctc_timestamp_type: str = "all" # can be char, word or all for both 1295 1296 # batch dimension 1297 batch_dim_index: int = 0 1298 1299 # greedy decoding config 1300 greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() 1301 1302 # beam decoding config 1303 beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) 1304 1305 # confidence config 1306 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1307 1308 # can be used to change temperature for decoding 1309 temperature: float = 1.0 1310 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/asr/models/configs/aligner_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig 18 19 20 @dataclass 21 class AlignerCTCConfig: 22 prob_suppress_index: int = -1 23 prob_suppress_value: float = 1.0 24 25 26 @dataclass 27 class AlignerRNNTConfig: 28 predictor_window_size: int = 0 29 predictor_step_size: int = 1 30 31 32 @dataclass 33 class AlignerWrapperModelConfig: 34 alignment_type: str = "forced" 35 word_output: bool = True 36 cpu_decoding: bool = False 37 decode_batch_size: int = 0 38 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() 39 rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() 40 41 42 @dataclass 43 class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): 44 decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() 45 [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 22 from nemo.collections.asr.modules.audio_preprocessing import ( 23 AudioToMelSpectrogramPreprocessorConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class ASRDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[Any] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[Any] = None 40 tarred_shard_strategy: str = "scatter" 41 shard_manifests: bool = False 42 shuffle_n: int = 0 43 44 # Optional 45 int_values: Optional[int] = None 46 augmentor: Optional[Dict[str, Any]] = None 47 max_duration: Optional[float] = None 48 min_duration: Optional[float] = None 49 max_utts: int = 0 50 blank_index: int = -1 51 unk_index: int = -1 52 normalize: bool = False 53 trim: bool = True 54 parser: Optional[str] = 'en' 55 eos_id: Optional[int] = None 56 bos_id: Optional[int] = None 57 pad_id: int = 0 58 use_start_end_token: bool = False 59 return_sample_id: Optional[bool] = False 60 61 # bucketing params 62 bucketing_strategy: str = "synced_randomized" 63 bucketing_batch_size: Optional[Any] = None 64 bucketing_weights: Optional[List[int]] = None 65 66 67 @dataclass 68 class EncDecCTCConfig(model_cfg.ModelConfig): 69 # Model global arguments 70 sample_rate: int = 16000 71 repeat: int = 1 72 dropout: float = 0.0 73 separable: bool = False 74 labels: List[str] = MISSING 75 76 # Dataset configs 77 train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) 78 validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 79 test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 80 81 # Optimizer / Scheduler config 82 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 83 84 # Model component configs 85 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 86 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 87 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 88 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 89 decoding: CTCDecodingConfig = CTCDecodingConfig() 90 91 92 @dataclass 93 class EncDecCTCModelConfig(model_cfg.NemoConfig): 94 model: EncDecCTCConfig = EncDecCTCConfig() 95 96 97 @dataclass 98 class CacheAwareStreamingConfig: 99 chunk_size: int = 0 # the size of each chunk at each step, it can be a list of two integers to specify different chunk sizes for the first step and others 100 shift_size: int = 0 # the size of the shift in each step, it can be a list of two integers to specify different shift sizes for the first step and others 101 102 cache_drop_size: int = 0 # the number of steps to drop from the cache 103 last_channel_cache_size: int = 0 # the size of the needed cache for last channel layers 104 105 valid_out_len: int = 0 # the number of the steps in the final output which are valid (have the same value as in the offline mode) 106 107 pre_encode_cache_size: int = 0 # the size of the needed cache for the pre-encoding part of the model to avoid caching inside the pre-encoding layers 108 drop_extra_pre_encoded: int = 0 # the number of steps to get dropped after the pre-encoding layer 109 110 last_channel_num: int = 0 # number of the last channel layers (like MHA layers) which need caching in the model 111 last_time_num: int = 0 # number of the last time layers (like convolutions) which need caching in the model 112 [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderClassificationConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class EncDecClassificationDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[str] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[str] = None 40 tarred_shard_strategy: str = "scatter" 41 shuffle_n: int = 0 42 43 # Optional 44 int_values: Optional[int] = None 45 augmentor: Optional[Dict[str, Any]] = None 46 max_duration: Optional[float] = None 47 min_duration: Optional[float] = None 48 cal_labels_occurrence: Optional[bool] = False 49 50 # VAD Optional 51 vad_stream: Optional[bool] = None 52 window_length_in_sec: float = 0.31 53 shift_length_in_sec: float = 0.01 54 normalize_audio: bool = False 55 is_regression_task: bool = False 56 57 # bucketing params 58 bucketing_strategy: str = "synced_randomized" 59 bucketing_batch_size: Optional[Any] = None 60 bucketing_weights: Optional[List[int]] = None 61 62 63 @dataclass 64 class EncDecClassificationConfig(model_cfg.ModelConfig): 65 # Model global arguments 66 sample_rate: int = 16000 67 repeat: int = 1 68 dropout: float = 0.0 69 separable: bool = True 70 kernel_size_factor: float = 1.0 71 labels: List[str] = MISSING 72 timesteps: int = MISSING 73 74 # Dataset configs 75 train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 76 manifest_filepath=None, shuffle=True, trim_silence=False 77 ) 78 validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 79 manifest_filepath=None, shuffle=False 80 ) 81 test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 82 manifest_filepath=None, shuffle=False 83 ) 84 85 # Optimizer / Scheduler config 86 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 87 88 # Model component configs 89 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() 90 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 91 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 92 audio_length=timesteps 93 ) 94 95 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 96 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 97 98 99 @dataclass 100 class EncDecClassificationModelConfig(model_cfg.NemoConfig): 101 model: EncDecClassificationConfig = EncDecClassificationConfig() 102 [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import asdict, dataclass 16 from typing import Any, Dict, Optional, Tuple, Union 17 18 19 @dataclass 20 class DiarizerComponentConfig: 21 """Dataclass to imitate HydraConfig dict when accessing parameters.""" 22 23 def get(self, name: str, default: Optional[Any] = None): 24 return getattr(self, name, default) 25 26 def __iter__(self): 27 for key in asdict(self): 28 yield key 29 30 def dict(self) -> Dict: 31 return asdict(self) 32 33 34 @dataclass 35 class ASRDiarizerCTCDecoderParams: 36 pretrained_language_model: Optional[str] = None # KenLM model file: .arpa model file or .bin binary file. 37 beam_width: int = 32 38 alpha: float = 0.5 39 beta: float = 2.5 40 41 42 @dataclass 43 class ASRRealigningLMParams: 44 # Provide a KenLM language model in .arpa format. 45 arpa_language_model: Optional[str] = None 46 # Min number of words for the left context. 47 min_number_of_words: int = 3 48 # Max number of words for the right context. 49 max_number_of_words: int = 10 50 # The threshold for the difference between two log probability values from two hypotheses. 51 logprob_diff_threshold: float = 1.2 52 53 54 @dataclass 55 class ASRDiarizerParams(DiarizerComponentConfig): 56 # if True, speech segmentation for diarization is based on word-timestamps from ASR inference. 57 asr_based_vad: bool = False 58 # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD. 59 asr_based_vad_threshold: float = 1.0 60 # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null. 61 asr_batch_size: Optional[int] = None 62 # Native decoder delay. null is recommended to use the default values for each ASR model. 63 decoder_delay_in_sec: Optional[float] = None 64 # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. 65 word_ts_anchor_offset: Optional[float] = None 66 # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. 67 word_ts_anchor_pos: str = "start" 68 # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. 69 fix_word_ts_with_VAD: bool = False 70 # If True, use colored text to distinguish speakers in the output transcript. 71 colored_text: bool = False 72 # If True, the start and end time of each speaker turn is printed in the output transcript. 73 print_time: bool = True 74 # If True, the output transcript breaks the line to fix the line width (default is 90 chars) 75 break_lines: bool = False 76 77 78 @dataclass 79 class ASRDiarizerConfig(DiarizerComponentConfig): 80 model_path: Optional[str] = "stt_en_conformer_ctc_large" 81 parameters: ASRDiarizerParams = ASRDiarizerParams() 82 ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() 83 realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() 84 85 86 @dataclass 87 class VADParams(DiarizerComponentConfig): 88 window_length_in_sec: float = 0.15 # Window length in sec for VAD context input 89 shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction 90 smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) 91 overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter 92 onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech 93 offset: float = 0.1 # Offset threshold for detecting the end of a speech 94 pad_onset: float = 0.1 # Adding durations before each speech segment 95 pad_offset: float = 0 # Adding durations after each speech segment 96 min_duration_on: float = 0 # Threshold for small non_speech deletion 97 min_duration_off: float = 0.2 # Threshold for short speech segment deletion 98 filter_speech_first: bool = True 99 100 101 @dataclass 102 class VADConfig(DiarizerComponentConfig): 103 model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name 104 external_vad_manifest: Optional[str] = None 105 parameters: VADParams = VADParams() 106 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): 110 # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] 111 window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) 112 # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] 113 shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) 114 # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] 115 multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) 116 # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. 117 save_embeddings: bool = True 118 119 120 @dataclass 121 class SpeakerEmbeddingsConfig(DiarizerComponentConfig): 122 # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) 123 model_path: Optional[str] = None 124 parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() 125 126 127 @dataclass 128 class ClusteringParams(DiarizerComponentConfig): 129 # If True, use num of speakers value provided in manifest file. 130 oracle_num_speakers: bool = False 131 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. 132 max_num_speakers: int = 8 133 # If the number of segments is lower than this number, enhanced speaker counting is activated. 134 enhanced_count_thres: int = 80 135 # Determines the range of p-value search: 0 < p <= max_rp_threshold. 136 max_rp_threshold: float = 0.25 137 # The higher the number, the more values will be examined with more time. 138 sparse_search_volume: int = 30 139 # If True, take a majority vote on multiple p-values to estimate the number of speakers. 140 maj_vote_spk_count: bool = False 141 142 143 @dataclass 144 class ClusteringConfig(DiarizerComponentConfig): 145 parameters: ClusteringParams = ClusteringParams() 146 147 148 @dataclass 149 class MSDDParams(DiarizerComponentConfig): 150 # If True, use speaker embedding model in checkpoint, else provided speaker embedding model in config will be used. 151 use_speaker_model_from_ckpt: bool = True 152 # Batch size for MSDD inference. 153 infer_batch_size: int = 25 154 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. 155 sigmoid_threshold: Tuple[float] = (0.7,) 156 # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. 157 seq_eval_mode: bool = False 158 # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. 159 split_infer: bool = True 160 # The length of split short sequence when split_infer is True. 161 diar_window_length: int = 50 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): 174 manifest_filepath: Optional[str] = None 175 out_dir: Optional[str] = None 176 oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps 177 collar: float = 0.25 # Collar value for scoring 178 ignore_overlap: bool = True # Consider or ignore overlap segments while scoring 179 vad: VADConfig = VADConfig() 180 speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() 181 clustering: ClusteringConfig = ClusteringConfig() 182 msdd_model: MSDDConfig = MSDDConfig() 183 asr: ASRDiarizerConfig = ASRDiarizerConfig() 184 185 186 @dataclass 187 class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): 188 diarizer: DiarizerConfig = DiarizerConfig() 189 device: str = "cpu" 190 verbose: bool = False 191 batch_size: int = 64 192 num_workers: int = 1 193 sample_rate: int = 16000 194 name: str = "" 195 196 @classmethod 197 def init_config(cls, diar_model_path: str, vad_model_path: str, map_location: str, verbose: bool): 198 return NeuralDiarizerInferenceConfig( 199 DiarizerConfig( 200 vad=VADConfig(model_path=vad_model_path), msdd_model=MSDDConfig(model_path=diar_model_path), 201 ), 202 device=map_location, 203 verbose=verbose, 204 ) 205 [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig 18 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig 19 from nemo.core.config.modelPT import NemoConfig 20 21 22 @dataclass 23 class GraphModuleConfig: 24 criterion_type: str = "ml" 25 loss_type: str = "ctc" 26 split_batch_size: int = 0 27 dec_type: str = "topo" 28 transcribe_training: bool = True 29 backend_cfg: BackendConfig = BackendConfig() 30 31 32 @dataclass 33 class EncDecK2SeqConfig(EncDecCTCConfig): 34 graph_module_cfg: GraphModuleConfig = GraphModuleConfig() 35 36 37 @dataclass 38 class EncDecK2SeqModelConfig(NemoConfig): 39 model: EncDecK2SeqConfig = EncDecK2SeqConfig() 40 [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import classification_models_config as clf_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ( 27 ConvASRDecoderClassificationConfig, 28 ConvASREncoderConfig, 29 JasperEncoderConfig, 30 ) 31 from nemo.core.config import modelPT as model_cfg 32 33 34 # fmt: off 35 def matchboxnet_3x1x64(): 36 config = [ 37 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 38 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 39 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 40 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 41 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 42 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 43 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 44 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 45 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 46 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 47 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 48 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 49 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 50 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 51 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 52 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 53 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 54 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 55 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 56 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 57 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 58 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 59 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 60 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 61 ] 62 return config 63 64 65 def matchboxnet_3x1x64_vad(): 66 config = [ 67 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 68 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 69 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 70 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 71 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 72 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 73 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 74 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 75 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 76 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 77 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 78 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 79 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 80 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 81 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 82 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 83 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 84 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 85 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 86 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 87 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 88 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 89 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 90 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 91 ] 92 return config 93 94 95 # fmt: on 96 97 98 @dataclass 99 class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): 100 # Model global arguments 101 sample_rate: int = 16000 102 repeat: int = 1 103 dropout: float = 0.0 104 separable: bool = True 105 kernel_size_factor: float = 1.0 106 timesteps: int = 128 107 labels: List[str] = MISSING 108 109 # Dataset configs 110 train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 111 manifest_filepath=None, shuffle=True, trim_silence=False 112 ) 113 validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 114 manifest_filepath=None, shuffle=False 115 ) 116 test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 117 manifest_filepath=None, shuffle=False 118 ) 119 120 # Optimizer / Scheduler config 121 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 122 123 # Model general component configs 124 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) 125 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( 126 freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 127 ) 128 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 129 audio_length=128 130 ) 131 132 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 133 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 134 135 136 @dataclass 137 class MatchboxNetVADModelConfig(MatchboxNetModelConfig): 138 timesteps: int = 64 139 labels: List[str] = field(default_factory=lambda: ['background', 'speech']) 140 141 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = None 142 143 144 class EncDecClassificationModelConfigBuilder(model_cfg.ModelConfigBuilder): 145 VALID_CONFIGS = ['matchboxnet_3x1x64', 'matchboxnet_3x1x64_vad'] 146 147 def __init__(self, name: str = 'matchboxnet_3x1x64', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 148 if name not in EncDecClassificationModelConfigBuilder.VALID_CONFIGS: 149 raise ValueError("`name` must be one of : \n" f"{EncDecClassificationModelConfigBuilder.VALID_CONFIGS}") 150 151 self.name = name 152 153 if 'matchboxnet_3x1x64_vad' in name: 154 if encoder_cfg_func is None: 155 encoder_cfg_func = matchboxnet_3x1x64_vad 156 157 model_cfg = MatchboxNetVADModelConfig( 158 repeat=1, 159 separable=True, 160 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 161 decoder=ConvASRDecoderClassificationConfig(), 162 ) 163 164 elif 'matchboxnet_3x1x64' in name: 165 if encoder_cfg_func is None: 166 encoder_cfg_func = matchboxnet_3x1x64 167 168 model_cfg = MatchboxNetModelConfig( 169 repeat=1, 170 separable=False, 171 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 172 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 173 decoder=ConvASRDecoderClassificationConfig(), 174 ) 175 176 else: 177 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 178 179 super(EncDecClassificationModelConfigBuilder, self).__init__(model_cfg) 180 self.model_cfg: clf_cfg.EncDecClassificationConfig = model_cfg # enable type hinting 181 182 def set_labels(self, labels: List[str]): 183 self.model_cfg.labels = labels 184 185 def set_separable(self, separable: bool): 186 self.model_cfg.separable = separable 187 188 def set_repeat(self, repeat: int): 189 self.model_cfg.repeat = repeat 190 191 def set_sample_rate(self, sample_rate: int): 192 self.model_cfg.sample_rate = sample_rate 193 194 def set_dropout(self, dropout: float = 0.0): 195 self.model_cfg.dropout = dropout 196 197 def set_timesteps(self, timesteps: int): 198 self.model_cfg.timesteps = timesteps 199 200 def set_is_regression_task(self, is_regression_task: bool): 201 self.model_cfg.is_regression_task = is_regression_task 202 203 # Note: Autocomplete for users wont work without these overrides 204 # But practically it is not needed since python will infer at runtime 205 206 # def set_train_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 207 # super().set_train_ds(cfg) 208 # 209 # def set_validation_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 210 # super().set_validation_ds(cfg) 211 # 212 # def set_test_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 213 # super().set_test_ds(cfg) 214 215 def _finalize_cfg(self): 216 # propagate labels 217 self.model_cfg.train_ds.labels = self.model_cfg.labels 218 self.model_cfg.validation_ds.labels = self.model_cfg.labels 219 self.model_cfg.test_ds.labels = self.model_cfg.labels 220 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 221 222 # propagate num classes 223 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 224 225 # propagate sample rate 226 self.model_cfg.sample_rate = self.model_cfg.sample_rate 227 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 228 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 229 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 230 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 231 232 # propagate filters 233 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 234 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 235 236 # propagate timeteps 237 if self.model_cfg.crop_or_pad_augment is not None: 238 self.model_cfg.crop_or_pad_augment.audio_length = self.model_cfg.timesteps 239 240 # propagate separable 241 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 242 layer.separable = self.model_cfg.separable 243 244 # propagate repeat 245 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 246 layer.repeat = self.model_cfg.repeat 247 248 # propagate dropout 249 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 250 layer.dropout = self.model_cfg.dropout 251 252 def build(self) -> clf_cfg.EncDecClassificationConfig: 253 return super().build() 254 [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import asr_models_config as ctc_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMelSpectrogramPreprocessorConfig, 23 SpectrogramAugmentationConfig, 24 ) 25 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig, JasperEncoderConfig 26 from nemo.core.config import modelPT as model_cfg 27 28 29 # fmt: off 30 def qn_15x5(): 31 config = [ 32 JasperEncoderConfig(filters=256, repeat=1, kernel=[33], stride=[2], dilation=[1], dropout=0.0, 33 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 34 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 35 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 36 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 37 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 38 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 39 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 40 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 41 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 42 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 43 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 44 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 45 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 46 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 47 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 48 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 49 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 50 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 51 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 52 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 53 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 54 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 55 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 56 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 57 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 58 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 59 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 60 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 61 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 62 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 63 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 64 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 65 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 66 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 67 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 68 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 69 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 70 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 71 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 72 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 73 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 74 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 75 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 76 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 77 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 78 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 79 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 80 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 81 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 82 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 83 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 84 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 85 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 86 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 87 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 88 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 89 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 90 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 91 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 92 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 93 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 94 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 95 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 96 JasperEncoderConfig(filters=512, repeat=1, kernel=[87], stride=[1], dilation=[2], dropout=0.0, 97 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 98 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 99 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 100 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 101 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 102 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 103 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 104 ] 105 return config 106 107 108 def jasper_10x5_dr(): 109 config = [ 110 JasperEncoderConfig(filters=256, repeat=1, kernel=[11], stride=[2], dilation=[1], dropout=0.2, 111 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 112 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 113 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 114 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 115 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 116 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 117 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 118 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 119 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 120 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 121 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 122 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 123 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 124 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 125 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 126 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 127 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 128 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 129 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 130 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 131 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 132 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 133 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 134 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 135 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 136 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 137 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 138 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 139 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 140 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 141 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 142 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 143 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 144 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 145 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 146 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 147 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 148 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 149 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 150 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 151 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 152 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 153 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 154 JasperEncoderConfig(filters=896, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.4, 155 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 156 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 157 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 158 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.4, 159 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 160 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 161 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 162 ] 163 return config 164 # fmt: on 165 166 167 @dataclass 168 class JasperModelConfig(ctc_cfg.EncDecCTCConfig): 169 # Model global arguments 170 sample_rate: int = 16000 171 repeat: int = 1 172 dropout: float = 0.0 173 separable: bool = False 174 labels: List[str] = MISSING 175 176 # Dataset configs 177 train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( 178 manifest_filepath=None, shuffle=True, trim_silence=True 179 ) 180 validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 181 test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 182 183 # Optimizer / Scheduler config 184 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 185 186 # Model general component configs 187 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 188 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 189 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 190 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 191 192 193 @dataclass 194 class QuartzNetModelConfig(JasperModelConfig): 195 separable: bool = True 196 197 198 class EncDecCTCModelConfigBuilder(model_cfg.ModelConfigBuilder): 199 VALID_CONFIGS = ['quartznet_15x5', 'quartznet_15x5_zh', 'jasper_10x5dr'] 200 201 def __init__(self, name: str = 'quartznet_15x5', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 202 if name not in EncDecCTCModelConfigBuilder.VALID_CONFIGS: 203 raise ValueError("`name` must be one of : \n" f"{EncDecCTCModelConfigBuilder.VALID_CONFIGS}") 204 205 self.name = name 206 207 if 'quartznet_15x5' in name: 208 if encoder_cfg_func is None: 209 encoder_cfg_func = qn_15x5 210 211 model_cfg = QuartzNetModelConfig( 212 repeat=5, 213 separable=True, 214 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 215 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 216 decoder=ConvASRDecoderConfig(), 217 ) 218 219 elif 'jasper_10x5' in name: 220 if encoder_cfg_func is None: 221 encoder_cfg_func = jasper_10x5_dr 222 223 model_cfg = JasperModelConfig( 224 repeat=5, 225 separable=False, 226 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 227 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 228 decoder=ConvASRDecoderConfig(), 229 ) 230 231 else: 232 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 233 234 super(EncDecCTCModelConfigBuilder, self).__init__(model_cfg) 235 self.model_cfg: ctc_cfg.EncDecCTCConfig = model_cfg # enable type hinting 236 237 if 'zh' in name: 238 self.set_dataset_normalize(normalize=False) 239 240 def set_labels(self, labels: List[str]): 241 self.model_cfg.labels = labels 242 243 def set_separable(self, separable: bool): 244 self.model_cfg.separable = separable 245 246 def set_repeat(self, repeat: int): 247 self.model_cfg.repeat = repeat 248 249 def set_sample_rate(self, sample_rate: int): 250 self.model_cfg.sample_rate = sample_rate 251 252 def set_dropout(self, dropout: float = 0.0): 253 self.model_cfg.dropout = dropout 254 255 def set_dataset_normalize(self, normalize: bool): 256 self.model_cfg.train_ds.normalize = normalize 257 self.model_cfg.validation_ds.normalize = normalize 258 self.model_cfg.test_ds.normalize = normalize 259 260 # Note: Autocomplete for users wont work without these overrides 261 # But practically it is not needed since python will infer at runtime 262 263 # def set_train_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 264 # super().set_train_ds(cfg) 265 # 266 # def set_validation_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 267 # super().set_validation_ds(cfg) 268 # 269 # def set_test_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 270 # super().set_test_ds(cfg) 271 272 def _finalize_cfg(self): 273 # propagate labels 274 self.model_cfg.train_ds.labels = self.model_cfg.labels 275 self.model_cfg.validation_ds.labels = self.model_cfg.labels 276 self.model_cfg.test_ds.labels = self.model_cfg.labels 277 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 278 279 # propagate num classes 280 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 281 282 # propagate sample rate 283 self.model_cfg.sample_rate = self.model_cfg.sample_rate 284 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 285 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 286 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 287 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 288 289 # propagate filters 290 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 291 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 292 293 # propagate separable 294 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 295 layer.separable = self.model_cfg.separable 296 297 # propagate repeat 298 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 299 layer.repeat = self.model_cfg.repeat 300 301 # propagate dropout 302 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 303 layer.dropout = self.model_cfg.dropout 304 305 def build(self) -> ctc_cfg.EncDecCTCConfig: 306 return super().build() 307 [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import random 17 from abc import ABC, abstractmethod 18 from dataclasses import dataclass 19 from typing import Any, Dict, Optional, Tuple 20 21 import torch 22 from packaging import version 23 24 from nemo.collections.asr.parts.numba.spec_augment import SpecAugmentNumba, spec_augment_launch_heuristics 25 from nemo.collections.asr.parts.preprocessing.features import ( 26 FilterbankFeatures, 27 FilterbankFeaturesTA, 28 make_seq_mask_like, 29 ) 30 from nemo.collections.asr.parts.submodules.spectr_augment import SpecAugment, SpecCutout 31 from nemo.core.classes import Exportable, NeuralModule, typecheck 32 from nemo.core.neural_types import ( 33 AudioSignal, 34 LengthsType, 35 MelSpectrogramType, 36 MFCCSpectrogramType, 37 NeuralType, 38 SpectrogramType, 39 ) 40 from nemo.core.utils import numba_utils 41 from nemo.core.utils.numba_utils import __NUMBA_MINIMUM_VERSION__ 42 from nemo.utils import logging 43 44 try: 45 import torchaudio 46 import torchaudio.functional 47 import torchaudio.transforms 48 49 TORCHAUDIO_VERSION = version.parse(torchaudio.__version__) 50 TORCHAUDIO_VERSION_MIN = version.parse('0.5') 51 52 HAVE_TORCHAUDIO = True 53 except ModuleNotFoundError: 54 HAVE_TORCHAUDIO = False 55 56 __all__ = [ 57 'AudioToMelSpectrogramPreprocessor', 58 'AudioToSpectrogram', 59 'SpectrogramToAudio', 60 'AudioToMFCCPreprocessor', 61 'SpectrogramAugmentation', 62 'MaskedPatchAugmentation', 63 'CropOrPadSpectrogramAugmentation', 64 ] 65 66 67 class AudioPreprocessor(NeuralModule, ABC): 68 """ 69 An interface for Neural Modules that performs audio pre-processing, 70 transforming the wav files to features. 71 """ 72 73 def __init__(self, win_length, hop_length): 74 super().__init__() 75 76 self.win_length = win_length 77 self.hop_length = hop_length 78 79 self.torch_windows = { 80 'hann': torch.hann_window, 81 'hamming': torch.hamming_window, 82 'blackman': torch.blackman_window, 83 'bartlett': torch.bartlett_window, 84 'ones': torch.ones, 85 None: torch.ones, 86 } 87 88 @typecheck() 89 @torch.no_grad() 90 def forward(self, input_signal, length): 91 processed_signal, processed_length = self.get_features(input_signal, length) 92 93 return processed_signal, processed_length 94 95 @abstractmethod 96 def get_features(self, input_signal, length): 97 # Called by forward(). Subclasses should implement this. 98 pass 99 100 101 class AudioToMelSpectrogramPreprocessor(AudioPreprocessor, Exportable): 102 """Featurizer module that converts wavs to mel spectrograms. 103 104 Args: 105 sample_rate (int): Sample rate of the input audio data. 106 Defaults to 16000 107 window_size (float): Size of window for fft in seconds 108 Defaults to 0.02 109 window_stride (float): Stride of window for fft in seconds 110 Defaults to 0.01 111 n_window_size (int): Size of window for fft in samples 112 Defaults to None. Use one of window_size or n_window_size. 113 n_window_stride (int): Stride of window for fft in samples 114 Defaults to None. Use one of window_stride or n_window_stride. 115 window (str): Windowing function for fft. can be one of ['hann', 116 'hamming', 'blackman', 'bartlett'] 117 Defaults to "hann" 118 normalize (str): Can be one of ['per_feature', 'all_features']; all 119 other options disable feature normalization. 'all_features' 120 normalizes the entire spectrogram to be mean 0 with std 1. 121 'pre_features' normalizes per channel / freq instead. 122 Defaults to "per_feature" 123 n_fft (int): Length of FT window. If None, it uses the smallest power 124 of 2 that is larger than n_window_size. 125 Defaults to None 126 preemph (float): Amount of pre emphasis to add to audio. Can be 127 disabled by passing None. 128 Defaults to 0.97 129 features (int): Number of mel spectrogram freq bins to output. 130 Defaults to 64 131 lowfreq (int): Lower bound on mel basis in Hz. 132 Defaults to 0 133 highfreq (int): Lower bound on mel basis in Hz. 134 Defaults to None 135 log (bool): Log features. 136 Defaults to True 137 log_zero_guard_type(str): Need to avoid taking the log of zero. There 138 are two options: "add" or "clamp". 139 Defaults to "add". 140 log_zero_guard_value(float, or str): Add or clamp requires the number 141 to add with or clamp to. log_zero_guard_value can either be a float 142 or "tiny" or "eps". torch.finfo is used if "tiny" or "eps" is 143 passed. 144 Defaults to 2**-24. 145 dither (float): Amount of white-noise dithering. 146 Defaults to 1e-5 147 pad_to (int): Ensures that the output size of the time dimension is 148 a multiple of pad_to. 149 Defaults to 16 150 frame_splicing (int): Defaults to 1 151 exact_pad (bool): If True, sets stft center to False and adds padding, such that num_frames = audio_length 152 // hop_length. Defaults to False. 153 pad_value (float): The value that shorter mels are padded with. 154 Defaults to 0 155 mag_power (float): The power that the linear spectrogram is raised to 156 prior to multiplication with mel basis. 157 Defaults to 2 for a power spec 158 rng : Random number generator 159 nb_augmentation_prob (float) : Probability with which narrowband augmentation would be applied to 160 samples in the batch. 161 Defaults to 0.0 162 nb_max_freq (int) : Frequency above which all frequencies will be masked for narrowband augmentation. 163 Defaults to 4000 164 use_torchaudio: Whether to use the `torchaudio` implementation. 165 mel_norm: Normalization used for mel filterbank weights. 166 Defaults to 'slaney' (area normalization) 167 stft_exact_pad: Deprecated argument, kept for compatibility with older checkpoints. 168 stft_conv: Deprecated argument, kept for compatibility with older checkpoints. 169 """ 170 171 def save_to(self, save_path: str): 172 pass 173 174 @classmethod 175 def restore_from(cls, restore_path: str): 176 pass 177 178 @property 179 def input_types(self): 180 """Returns definitions of module input ports. 181 """ 182 return { 183 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 184 "length": NeuralType( 185 tuple('B'), LengthsType() 186 ), # Please note that length should be in samples not seconds. 187 } 188 189 @property 190 def output_types(self): 191 """Returns definitions of module output ports. 192 193 processed_signal: 194 0: AxisType(BatchTag) 195 1: AxisType(MelSpectrogramSignalTag) 196 2: AxisType(ProcessedTimeTag) 197 processed_length: 198 0: AxisType(BatchTag) 199 """ 200 return { 201 "processed_signal": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 202 "processed_length": NeuralType(tuple('B'), LengthsType()), 203 } 204 205 def __init__( 206 self, 207 sample_rate=16000, 208 window_size=0.02, 209 window_stride=0.01, 210 n_window_size=None, 211 n_window_stride=None, 212 window="hann", 213 normalize="per_feature", 214 n_fft=None, 215 preemph=0.97, 216 features=64, 217 lowfreq=0, 218 highfreq=None, 219 log=True, 220 log_zero_guard_type="add", 221 log_zero_guard_value=2 ** -24, 222 dither=1e-5, 223 pad_to=16, 224 frame_splicing=1, 225 exact_pad=False, 226 pad_value=0, 227 mag_power=2.0, 228 rng=None, 229 nb_augmentation_prob=0.0, 230 nb_max_freq=4000, 231 use_torchaudio: bool = False, 232 mel_norm="slaney", 233 stft_exact_pad=False, # Deprecated arguments; kept for config compatibility 234 stft_conv=False, # Deprecated arguments; kept for config compatibility 235 ): 236 super().__init__(n_window_size, n_window_stride) 237 238 self._sample_rate = sample_rate 239 if window_size and n_window_size: 240 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 241 if window_stride and n_window_stride: 242 raise ValueError( 243 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 244 ) 245 if window_size: 246 n_window_size = int(window_size * self._sample_rate) 247 if window_stride: 248 n_window_stride = int(window_stride * self._sample_rate) 249 250 # Given the long and similar argument list, point to the class and instantiate it by reference 251 if not use_torchaudio: 252 featurizer_class = FilterbankFeatures 253 else: 254 featurizer_class = FilterbankFeaturesTA 255 self.featurizer = featurizer_class( 256 sample_rate=self._sample_rate, 257 n_window_size=n_window_size, 258 n_window_stride=n_window_stride, 259 window=window, 260 normalize=normalize, 261 n_fft=n_fft, 262 preemph=preemph, 263 nfilt=features, 264 lowfreq=lowfreq, 265 highfreq=highfreq, 266 log=log, 267 log_zero_guard_type=log_zero_guard_type, 268 log_zero_guard_value=log_zero_guard_value, 269 dither=dither, 270 pad_to=pad_to, 271 frame_splicing=frame_splicing, 272 exact_pad=exact_pad, 273 pad_value=pad_value, 274 mag_power=mag_power, 275 rng=rng, 276 nb_augmentation_prob=nb_augmentation_prob, 277 nb_max_freq=nb_max_freq, 278 mel_norm=mel_norm, 279 stft_exact_pad=stft_exact_pad, # Deprecated arguments; kept for config compatibility 280 stft_conv=stft_conv, # Deprecated arguments; kept for config compatibility 281 ) 282 283 def input_example(self, max_batch: int = 8, max_dim: int = 32000, min_length: int = 200): 284 batch_size = torch.randint(low=1, high=max_batch, size=[1]).item() 285 max_length = torch.randint(low=min_length, high=max_dim, size=[1]).item() 286 signals = torch.rand(size=[batch_size, max_length]) * 2 - 1 287 lengths = torch.randint(low=min_length, high=max_dim, size=[batch_size]) 288 lengths[0] = max_length 289 return signals, lengths 290 291 def get_features(self, input_signal, length): 292 return self.featurizer(input_signal, length) 293 294 @property 295 def filter_banks(self): 296 return self.featurizer.filter_banks 297 298 299 class AudioToMFCCPreprocessor(AudioPreprocessor): 300 """Preprocessor that converts wavs to MFCCs. 301 Uses torchaudio.transforms.MFCC. 302 303 Args: 304 sample_rate: The sample rate of the audio. 305 Defaults to 16000. 306 window_size: Size of window for fft in seconds. Used to calculate the 307 win_length arg for mel spectrogram. 308 Defaults to 0.02 309 window_stride: Stride of window for fft in seconds. Used to caculate 310 the hop_length arg for mel spect. 311 Defaults to 0.01 312 n_window_size: Size of window for fft in samples 313 Defaults to None. Use one of window_size or n_window_size. 314 n_window_stride: Stride of window for fft in samples 315 Defaults to None. Use one of window_stride or n_window_stride. 316 window: Windowing function for fft. can be one of ['hann', 317 'hamming', 'blackman', 'bartlett', 'none', 'null']. 318 Defaults to 'hann' 319 n_fft: Length of FT window. If None, it uses the smallest power of 2 320 that is larger than n_window_size. 321 Defaults to None 322 lowfreq (int): Lower bound on mel basis in Hz. 323 Defaults to 0 324 highfreq (int): Lower bound on mel basis in Hz. 325 Defaults to None 326 n_mels: Number of mel filterbanks. 327 Defaults to 64 328 n_mfcc: Number of coefficients to retain 329 Defaults to 64 330 dct_type: Type of discrete cosine transform to use 331 norm: Type of norm to use 332 log: Whether to use log-mel spectrograms instead of db-scaled. 333 Defaults to True. 334 """ 335 336 @property 337 def input_types(self): 338 """Returns definitions of module input ports. 339 """ 340 return { 341 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 342 "length": NeuralType(tuple('B'), LengthsType()), 343 } 344 345 @property 346 def output_types(self): 347 """Returns definitions of module output ports. 348 """ 349 return { 350 "processed_signal": NeuralType(('B', 'D', 'T'), MFCCSpectrogramType()), 351 "processed_length": NeuralType(tuple('B'), LengthsType()), 352 } 353 354 def save_to(self, save_path: str): 355 pass 356 357 @classmethod 358 def restore_from(cls, restore_path: str): 359 pass 360 361 def __init__( 362 self, 363 sample_rate=16000, 364 window_size=0.02, 365 window_stride=0.01, 366 n_window_size=None, 367 n_window_stride=None, 368 window='hann', 369 n_fft=None, 370 lowfreq=0.0, 371 highfreq=None, 372 n_mels=64, 373 n_mfcc=64, 374 dct_type=2, 375 norm='ortho', 376 log=True, 377 ): 378 self._sample_rate = sample_rate 379 if not HAVE_TORCHAUDIO: 380 logging.error('Could not import torchaudio. Some features might not work.') 381 382 raise ModuleNotFoundError( 383 "torchaudio is not installed but is necessary for " 384 "AudioToMFCCPreprocessor. We recommend you try " 385 "building it from source for the PyTorch version you have." 386 ) 387 if window_size and n_window_size: 388 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 389 if window_stride and n_window_stride: 390 raise ValueError( 391 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 392 ) 393 # Get win_length (n_window_size) and hop_length (n_window_stride) 394 if window_size: 395 n_window_size = int(window_size * self._sample_rate) 396 if window_stride: 397 n_window_stride = int(window_stride * self._sample_rate) 398 399 super().__init__(n_window_size, n_window_stride) 400 401 mel_kwargs = {} 402 403 mel_kwargs['f_min'] = lowfreq 404 mel_kwargs['f_max'] = highfreq 405 mel_kwargs['n_mels'] = n_mels 406 407 mel_kwargs['n_fft'] = n_fft or 2 ** math.ceil(math.log2(n_window_size)) 408 409 mel_kwargs['win_length'] = n_window_size 410 mel_kwargs['hop_length'] = n_window_stride 411 412 # Set window_fn. None defaults to torch.ones. 413 window_fn = self.torch_windows.get(window, None) 414 if window_fn is None: 415 raise ValueError( 416 f"Window argument for AudioProcessor is invalid: {window}." 417 f"For no window function, use 'ones' or None." 418 ) 419 mel_kwargs['window_fn'] = window_fn 420 421 # Use torchaudio's implementation of MFCCs as featurizer 422 self.featurizer = torchaudio.transforms.MFCC( 423 sample_rate=self._sample_rate, 424 n_mfcc=n_mfcc, 425 dct_type=dct_type, 426 norm=norm, 427 log_mels=log, 428 melkwargs=mel_kwargs, 429 ) 430 431 def get_features(self, input_signal, length): 432 features = self.featurizer(input_signal) 433 seq_len = torch.ceil(length.to(torch.float32) / self.hop_length).to(dtype=torch.long) 434 return features, seq_len 435 436 437 class SpectrogramAugmentation(NeuralModule): 438 """ 439 Performs time and freq cuts in one of two ways. 440 SpecAugment zeroes out vertical and horizontal sections as described in 441 SpecAugment (https://arxiv.org/abs/1904.08779). Arguments for use with 442 SpecAugment are `freq_masks`, `time_masks`, `freq_width`, and `time_width`. 443 SpecCutout zeroes out rectangulars as described in Cutout 444 (https://arxiv.org/abs/1708.04552). Arguments for use with Cutout are 445 `rect_masks`, `rect_freq`, and `rect_time`. 446 447 Args: 448 freq_masks (int): how many frequency segments should be cut. 449 Defaults to 0. 450 time_masks (int): how many time segments should be cut 451 Defaults to 0. 452 freq_width (int): maximum number of frequencies to be cut in one 453 segment. 454 Defaults to 10. 455 time_width (int): maximum number of time steps to be cut in one 456 segment 457 Defaults to 10. 458 rect_masks (int): how many rectangular masks should be cut 459 Defaults to 0. 460 rect_freq (int): maximum size of cut rectangles along the frequency 461 dimension 462 Defaults to 5. 463 rect_time (int): maximum size of cut rectangles along the time 464 dimension 465 Defaults to 25. 466 """ 467 468 @property 469 def input_types(self): 470 """Returns definitions of module input types 471 """ 472 return { 473 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 474 "length": NeuralType(tuple('B'), LengthsType()), 475 } 476 477 @property 478 def output_types(self): 479 """Returns definitions of module output types 480 """ 481 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 482 483 def __init__( 484 self, 485 freq_masks=0, 486 time_masks=0, 487 freq_width=10, 488 time_width=10, 489 rect_masks=0, 490 rect_time=5, 491 rect_freq=20, 492 rng=None, 493 mask_value=0.0, 494 use_numba_spec_augment: bool = True, 495 ): 496 super().__init__() 497 498 if rect_masks > 0: 499 self.spec_cutout = SpecCutout(rect_masks=rect_masks, rect_time=rect_time, rect_freq=rect_freq, rng=rng,) 500 # self.spec_cutout.to(self._device) 501 else: 502 self.spec_cutout = lambda input_spec: input_spec 503 if freq_masks + time_masks > 0: 504 self.spec_augment = SpecAugment( 505 freq_masks=freq_masks, 506 time_masks=time_masks, 507 freq_width=freq_width, 508 time_width=time_width, 509 rng=rng, 510 mask_value=mask_value, 511 ) 512 else: 513 self.spec_augment = lambda input_spec, length: input_spec 514 515 # Check if numba is supported, and use a Numba kernel if it is 516 if use_numba_spec_augment and numba_utils.numba_cuda_is_supported(__NUMBA_MINIMUM_VERSION__): 517 logging.info('Numba CUDA SpecAugment kernel is being used') 518 self.spec_augment_numba = SpecAugmentNumba( 519 freq_masks=freq_masks, 520 time_masks=time_masks, 521 freq_width=freq_width, 522 time_width=time_width, 523 rng=rng, 524 mask_value=mask_value, 525 ) 526 else: 527 self.spec_augment_numba = None 528 529 @typecheck() 530 def forward(self, input_spec, length): 531 augmented_spec = self.spec_cutout(input_spec=input_spec) 532 533 # To run the Numba kernel, correct numba version is required as well as 534 # tensor must be on GPU and length must be provided 535 if self.spec_augment_numba is not None and spec_augment_launch_heuristics(augmented_spec, length): 536 augmented_spec = self.spec_augment_numba(input_spec=augmented_spec, length=length) 537 else: 538 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 539 return augmented_spec 540 541 542 class MaskedPatchAugmentation(NeuralModule): 543 """ 544 Zeroes out fixed size time patches of the spectrogram. 545 All samples in batch are guaranteed to have the same amount of masked time steps. 546 Optionally also performs frequency masking in the same way as SpecAugment. 547 Args: 548 patch_size (int): up to how many time steps does one patch consist of. 549 Defaults to 48. 550 mask_patches (float): how many patches should be masked in each sample. 551 if >= 1., interpreted as number of patches (after converting to int) 552 if <1., interpreted as fraction of total tokens to be masked (number of patches is rounded up) 553 Defaults to 10. 554 freq_masks (int): how many frequency segments should be cut. 555 Defaults to 0. 556 freq_width (int): maximum number of frequencies to be cut in a segment. 557 Defaults to 0. 558 """ 559 560 @property 561 def input_types(self): 562 """Returns definitions of module input types 563 """ 564 return { 565 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 566 "length": NeuralType(tuple('B'), LengthsType()), 567 } 568 569 @property 570 def output_types(self): 571 """Returns definitions of module output types 572 """ 573 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 574 575 def __init__( 576 self, patch_size: int = 48, mask_patches: float = 10.0, freq_masks: int = 0, freq_width: int = 0, 577 ): 578 super().__init__() 579 self.patch_size = patch_size 580 if mask_patches >= 1: 581 self.mask_patches = int(mask_patches) 582 elif mask_patches >= 0: 583 self._mask_fraction = mask_patches 584 self.mask_patches = None 585 else: 586 raise ValueError('mask_patches cannot be negative') 587 588 if freq_masks > 0: 589 self.spec_augment = SpecAugment(freq_masks=freq_masks, time_masks=0, freq_width=freq_width, time_width=0,) 590 else: 591 self.spec_augment = None 592 593 @typecheck() 594 def forward(self, input_spec, length): 595 augmented_spec = input_spec 596 597 min_len = torch.min(length) 598 599 if self.mask_patches is None: 600 # masking specified as fraction 601 len_fraction = int(min_len * self._mask_fraction) 602 mask_patches = len_fraction // self.patch_size + int(len_fraction % self.patch_size != 0) 603 else: 604 mask_patches = self.mask_patches 605 606 if min_len < self.patch_size * mask_patches: 607 mask_patches = min_len // self.patch_size 608 609 for idx in range(input_spec.shape[0]): 610 cur_len = length[idx] 611 patches = range(cur_len // self.patch_size) 612 masked_patches = random.sample(patches, mask_patches) 613 614 for mp in masked_patches: 615 augmented_spec[idx, :, mp * self.patch_size : (mp + 1) * self.patch_size] = 0.0 616 617 if self.spec_augment is not None: 618 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 619 620 return augmented_spec 621 622 623 class CropOrPadSpectrogramAugmentation(NeuralModule): 624 """ 625 Pad or Crop the incoming Spectrogram to a certain shape. 626 627 Args: 628 audio_length (int): the final number of timesteps that is required. 629 The signal will be either padded or cropped temporally to this 630 size. 631 """ 632 633 def __init__(self, audio_length): 634 super(CropOrPadSpectrogramAugmentation, self).__init__() 635 self.audio_length = audio_length 636 637 @typecheck() 638 @torch.no_grad() 639 def forward(self, input_signal, length): 640 image = input_signal 641 num_images = image.shape[0] 642 643 audio_length = self.audio_length 644 image_len = image.shape[-1] 645 646 # Crop long signal 647 if image_len > audio_length: # randomly slice 648 cutout_images = [] 649 offset = torch.randint(low=0, high=image_len - audio_length + 1, size=[num_images]) 650 651 for idx, offset in enumerate(offset): 652 cutout_images.append(image[idx : idx + 1, :, offset : offset + audio_length]) 653 654 image = torch.cat(cutout_images, dim=0) 655 del cutout_images 656 657 else: # symmetrically pad short signal with zeros 658 pad_left = (audio_length - image_len) // 2 659 pad_right = (audio_length - image_len) // 2 660 661 if (audio_length - image_len) % 2 == 1: 662 pad_right += 1 663 664 image = torch.nn.functional.pad(image, [pad_left, pad_right], mode="constant", value=0) 665 666 # Replace dynamic length sequences with static number of timesteps 667 length = (length * 0) + audio_length 668 669 return image, length 670 671 @property 672 def input_types(self): 673 """Returns definitions of module output ports. 674 """ 675 return { 676 "input_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 677 "length": NeuralType(tuple('B'), LengthsType()), 678 } 679 680 @property 681 def output_types(self): 682 """Returns definitions of module output ports. 683 """ 684 return { 685 "processed_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 686 "processed_length": NeuralType(tuple('B'), LengthsType()), 687 } 688 689 def save_to(self, save_path: str): 690 pass 691 692 @classmethod 693 def restore_from(cls, restore_path: str): 694 pass 695 696 697 class AudioToSpectrogram(NeuralModule): 698 """Transform a batch of input multi-channel signals into a batch of 699 STFT-based spectrograms. 700 701 Args: 702 fft_length: length of FFT 703 hop_length: length of hops/shifts of the sliding window 704 power: exponent for magnitude spectrogram. Default `None` will 705 return a complex-valued spectrogram 706 """ 707 708 def __init__(self, fft_length: int, hop_length: int, power: Optional[float] = None): 709 if not HAVE_TORCHAUDIO: 710 logging.error('Could not import torchaudio. Some features might not work.') 711 712 raise ModuleNotFoundError( 713 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 714 ) 715 716 super().__init__() 717 718 # For now, assume FFT length is divisible by two 719 if fft_length % 2 != 0: 720 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 721 722 self.stft = torchaudio.transforms.Spectrogram( 723 n_fft=fft_length, hop_length=hop_length, power=power, pad_mode='constant' 724 ) 725 726 # number of subbands 727 self.F = fft_length // 2 + 1 728 729 @property 730 def num_subbands(self) -> int: 731 return self.F 732 733 @property 734 def input_types(self) -> Dict[str, NeuralType]: 735 """Returns definitions of module output ports. 736 """ 737 return { 738 "input": NeuralType(('B', 'C', 'T'), AudioSignal()), 739 "input_length": NeuralType(('B',), LengthsType(), optional=True), 740 } 741 742 @property 743 def output_types(self) -> Dict[str, NeuralType]: 744 """Returns definitions of module output ports. 745 """ 746 return { 747 "output": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 748 "output_length": NeuralType(('B',), LengthsType()), 749 } 750 751 @typecheck() 752 def forward( 753 self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None 754 ) -> Tuple[torch.Tensor, torch.Tensor]: 755 """Convert a batch of C-channel input signals 756 into a batch of complex-valued spectrograms. 757 758 Args: 759 input: Time-domain input signal with C channels, shape (B, C, T) 760 input_length: Length of valid entries along the time dimension, shape (B,) 761 762 Returns: 763 Output spectrogram with F subbands and N time frames, shape (B, C, F, N) 764 and output length with shape (B,). 765 """ 766 B, T = input.size(0), input.size(-1) 767 input = input.view(B, -1, T) 768 769 # STFT output (B, C, F, N) 770 with torch.cuda.amp.autocast(enabled=False): 771 output = self.stft(input.float()) 772 773 if input_length is not None: 774 # Mask padded frames 775 output_length = self.get_output_length(input_length=input_length) 776 777 length_mask: torch.Tensor = make_seq_mask_like( 778 lengths=output_length, like=output, time_dim=-1, valid_ones=False 779 ) 780 output = output.masked_fill(length_mask, 0.0) 781 else: 782 # Assume all frames are valid for all examples in the batch 783 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 784 785 return output, output_length 786 787 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 788 """Get length of valid frames for the output. 789 790 Args: 791 input_length: number of valid samples, shape (B,) 792 793 Returns: 794 Number of valid frames, shape (B,) 795 """ 796 output_length = input_length.div(self.stft.hop_length, rounding_mode='floor').add(1).long() 797 return output_length 798 799 800 class SpectrogramToAudio(NeuralModule): 801 """Transform a batch of input multi-channel spectrograms into a batch of 802 time-domain multi-channel signals. 803 804 Args: 805 fft_length: length of FFT 806 hop_length: length of hops/shifts of the sliding window 807 power: exponent for magnitude spectrogram. Default `None` will 808 return a complex-valued spectrogram 809 """ 810 811 def __init__(self, fft_length: int, hop_length: int): 812 if not HAVE_TORCHAUDIO: 813 logging.error('Could not import torchaudio. Some features might not work.') 814 815 raise ModuleNotFoundError( 816 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 817 ) 818 819 super().__init__() 820 821 # For now, assume FFT length is divisible by two 822 if fft_length % 2 != 0: 823 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 824 825 self.istft = torchaudio.transforms.InverseSpectrogram( 826 n_fft=fft_length, hop_length=hop_length, pad_mode='constant' 827 ) 828 829 self.F = fft_length // 2 + 1 830 831 @property 832 def num_subbands(self) -> int: 833 return self.F 834 835 @property 836 def input_types(self) -> Dict[str, NeuralType]: 837 """Returns definitions of module output ports. 838 """ 839 return { 840 "input": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 841 "input_length": NeuralType(('B',), LengthsType(), optional=True), 842 } 843 844 @property 845 def output_types(self) -> Dict[str, NeuralType]: 846 """Returns definitions of module output ports. 847 """ 848 return { 849 "output": NeuralType(('B', 'C', 'T'), AudioSignal()), 850 "output_length": NeuralType(('B',), LengthsType()), 851 } 852 853 @typecheck() 854 def forward(self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None) -> torch.Tensor: 855 """Convert input complex-valued spectrogram to a time-domain 856 signal. Multi-channel IO is supported. 857 858 Args: 859 input: Input spectrogram for C channels, shape (B, C, F, N) 860 input_length: Length of valid entries along the time dimension, shape (B,) 861 862 Returns: 863 Time-domain signal with T time-domain samples and C channels, (B, C, T) 864 and output length with shape (B,). 865 """ 866 B, F, N = input.size(0), input.size(-2), input.size(-1) 867 assert F == self.F, f'Number of subbands F={F} not matching self.F={self.F}' 868 input = input.view(B, -1, F, N) 869 870 # iSTFT output (B, C, T) 871 with torch.cuda.amp.autocast(enabled=False): 872 output = self.istft(input.cfloat()) 873 874 if input_length is not None: 875 # Mask padded samples 876 output_length = self.get_output_length(input_length=input_length) 877 878 length_mask: torch.Tensor = make_seq_mask_like( 879 lengths=output_length, like=output, time_dim=-1, valid_ones=False 880 ) 881 output = output.masked_fill(length_mask, 0.0) 882 else: 883 # Assume all frames are valid for all examples in the batch 884 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 885 886 return output, output_length 887 888 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 889 """Get length of valid samples for the output. 890 891 Args: 892 input_length: number of valid frames, shape (B,) 893 894 Returns: 895 Number of valid samples, shape (B,) 896 """ 897 output_length = input_length.sub(1).mul(self.istft.hop_length).long() 898 return output_length 899 900 901 @dataclass 902 class AudioToMelSpectrogramPreprocessorConfig: 903 _target_: str = "nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor" 904 sample_rate: int = 16000 905 window_size: float = 0.02 906 window_stride: float = 0.01 907 n_window_size: Optional[int] = None 908 n_window_stride: Optional[int] = None 909 window: str = "hann" 910 normalize: str = "per_feature" 911 n_fft: Optional[int] = None 912 preemph: float = 0.97 913 features: int = 64 914 lowfreq: int = 0 915 highfreq: Optional[int] = None 916 log: bool = True 917 log_zero_guard_type: str = "add" 918 log_zero_guard_value: float = 2 ** -24 919 dither: float = 1e-5 920 pad_to: int = 16 921 frame_splicing: int = 1 922 exact_pad: bool = False 923 pad_value: int = 0 924 mag_power: float = 2.0 925 rng: Optional[str] = None 926 nb_augmentation_prob: float = 0.0 927 nb_max_freq: int = 4000 928 use_torchaudio: bool = False 929 mel_norm: str = "slaney" 930 stft_exact_pad: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 931 stft_conv: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 932 933 934 @dataclass 935 class AudioToMFCCPreprocessorConfig: 936 _target_: str = 'nemo.collections.asr.modules.AudioToMFCCPreprocessor' 937 sample_rate: int = 16000 938 window_size: float = 0.02 939 window_stride: float = 0.01 940 n_window_size: Optional[int] = None 941 n_window_stride: Optional[int] = None 942 window: str = 'hann' 943 n_fft: Optional[int] = None 944 lowfreq: Optional[float] = 0.0 945 highfreq: Optional[float] = None 946 n_mels: int = 64 947 n_mfcc: int = 64 948 dct_type: int = 2 949 norm: str = 'ortho' 950 log: bool = True 951 952 953 @dataclass 954 class SpectrogramAugmentationConfig: 955 _target_: str = "nemo.collections.asr.modules.SpectrogramAugmentation" 956 freq_masks: int = 0 957 time_masks: int = 0 958 freq_width: int = 0 959 time_width: Optional[Any] = 0 960 rect_masks: int = 0 961 rect_time: int = 0 962 rect_freq: int = 0 963 mask_value: float = 0 964 rng: Optional[Any] = None # random.Random() type 965 use_numba_spec_augment: bool = True 966 967 968 @dataclass 969 class CropOrPadSpectrogramAugmentationConfig: 970 audio_length: int 971 _target_: str = "nemo.collections.asr.modules.CropOrPadSpectrogramAugmentation" 972 973 974 @dataclass 975 class MaskedPatchAugmentationConfig: 976 patch_size: int = 48 977 mask_patches: float = 10.0 978 freq_masks: int = 0 979 freq_width: int = 0 980 _target_: str = "nemo.collections.asr.modules.MaskedPatchAugmentation" 981 [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/parts/k2/classes.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from abc import ABC 16 from dataclasses import dataclass 17 from typing import Any, Optional, Tuple 18 19 import torch 20 from omegaconf import DictConfig 21 22 from nemo.utils import logging 23 24 25 @dataclass 26 class GraphIntersectDenseConfig: 27 """Graph dense intersection config. 28 """ 29 30 search_beam: float = 20.0 31 output_beam: float = 10.0 32 min_active_states: int = 30 33 max_active_states: int = 10000 34 35 36 @dataclass 37 class GraphModuleConfig: 38 """Config for graph modules. 39 Typically used with graph losses and decoders. 40 """ 41 42 topo_type: str = "default" 43 topo_with_self_loops: bool = True 44 token_lm: Optional[Any] = None 45 intersect_pruned: bool = False 46 intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() 47 boost_coeff: float = 0.0 48 predictor_window_size: int = 0 49 predictor_step_size: int = 1 50 51 52 class ASRK2Mixin(ABC): 53 """k2 Mixin class that simplifies the construction of various models with k2-based losses. 54 55 It does the following: 56 - Sets up the graph loss and decoder (methods _init_k2 and update_k2_modules). 57 - Registers external graphs, if needed. 58 - Augments forward(...) with optional graph decoding to get accurate predictions. 59 """ 60 61 def _init_k2(self): 62 """ 63 k2-related initialization implementation. 64 65 This method is expected to run after the __init__ which sets self._cfg 66 self._cfg is expected to have the attribute graph_module_cfg 67 """ 68 if not hasattr(self, "_cfg"): 69 raise ValueError("self._cfg must be set before calling _init_k2().") 70 if not hasattr(self._cfg, "graph_module_cfg") or self._cfg.graph_module_cfg is None: 71 raise ValueError("self._cfg.graph_module_cfg must be set and cannot be None.") 72 self.graph_module_cfg = self._cfg.graph_module_cfg 73 74 # register token_lm for MAPLoss 75 criterion_type = self.graph_module_cfg.get("criterion_type", "ml") 76 self.use_graph_lm = criterion_type == "map" 77 if self.use_graph_lm: 78 token_lm_path = self.graph_module_cfg.backend_cfg.get("token_lm", None) 79 if token_lm_path is None: 80 raise ValueError( 81 f"graph_module_cfg.backend_cfg.token_lm is empty. It must be set for criterion_type == `{criterion_type}`" 82 ) 83 token_lm_path = self.register_artifact('graph_module_cfg.backend_cfg.token_lm', token_lm_path) 84 self.graph_module_cfg.backend_cfg["token_lm"] = token_lm_path 85 86 self.update_k2_modules(self.graph_module_cfg) 87 88 def update_k2_modules(self, input_cfg: DictConfig): 89 """ 90 Helper function to initialize or update k2 loss and transcribe_decoder. 91 92 Args: 93 input_cfg: DictConfig to take new parameters from. Schema is expected as in 94 nemo.collections.asr.models.configs.k2_sequence_models_config.GraphModuleConfig 95 """ 96 del self.loss 97 if hasattr(self, "transcribe_decoder"): 98 del self.transcribe_decoder 99 100 if hasattr(self, "joint"): 101 # RNNT 102 num_classes = self.joint.num_classes_with_blank - 1 103 else: 104 # CTC, MMI, ... 105 num_classes = self.decoder.num_classes_with_blank - 1 106 remove_consecutive = input_cfg.backend_cfg.get("topo_with_self_loops", True) and input_cfg.backend_cfg.get( 107 "topo_type", "default" 108 ) not in ["forced_blank", "identity",] 109 self._wer.remove_consecutive = remove_consecutive 110 111 from nemo.collections.asr.losses.lattice_losses import LatticeLoss 112 113 self.loss = LatticeLoss( 114 num_classes=num_classes, 115 reduction=self._cfg.get("ctc_reduction", "mean_batch"), 116 backend="k2", 117 criterion_type=input_cfg.get("criterion_type", "ml"), 118 loss_type=input_cfg.get("loss_type", "ctc"), 119 split_batch_size=input_cfg.get("split_batch_size", 0), 120 graph_module_cfg=input_cfg.backend_cfg, 121 ) 122 123 criterion_type = self.loss.criterion_type 124 self.use_graph_lm = criterion_type == "map" 125 transcribe_training = input_cfg.get("transcribe_training", False) 126 if transcribe_training and criterion_type == "ml": 127 logging.warning( 128 f"""You do not need to use transcribe_training=`{transcribe_training}` 129 with criterion_type=`{criterion_type}`. transcribe_training will be set to False.""" 130 ) 131 transcribe_training = False 132 self.transcribe_training = transcribe_training 133 if self.use_graph_lm: 134 from nemo.collections.asr.modules.graph_decoder import ViterbiDecoderWithGraph 135 136 self.transcribe_decoder = ViterbiDecoderWithGraph( 137 num_classes=num_classes, 138 backend="k2", 139 dec_type="token_lm", 140 return_type="1best", 141 return_ilabels=True, 142 output_aligned=True, 143 split_batch_size=input_cfg.get("split_batch_size", 0), 144 graph_module_cfg=input_cfg.backend_cfg, 145 ) 146 147 def _forward_k2_post_processing( 148 self, log_probs: torch.Tensor, encoded_length: torch.Tensor, greedy_predictions: torch.Tensor 149 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: 150 """ 151 k2-related post-processing parf of .forward() 152 153 Args: 154 log_probs: The log probabilities tensor of shape [B, T, D]. 155 encoded_length: The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 156 greedy_predictions: The greedy token predictions of the model of shape [B, T] 157 158 Returns: 159 A tuple of 3 elements - 160 1) The log probabilities tensor of shape [B, T, D]. 161 2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 162 3) The greedy token predictions of the model of shape [B, T] (via argmax) 163 """ 164 # greedy_predictions from .forward() are incorrect for criterion_type=`map` 165 # getting correct greedy_predictions, if needed 166 if self.use_graph_lm and (not self.training or self.transcribe_training): 167 greedy_predictions, encoded_length, _ = self.transcribe_decoder.forward( 168 log_probs=log_probs, log_probs_length=encoded_length 169 ) 170 return log_probs, encoded_length, greedy_predictions 171 [end of nemo/collections/asr/parts/k2/classes.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from dataclasses import dataclass 17 from typing import Any, Optional 18 19 import torch 20 from torch import nn as nn 21 22 from nemo.collections.asr.parts.submodules import multi_head_attention as mha 23 from nemo.collections.common.parts import adapter_modules 24 from nemo.core.classes.mixins import adapter_mixin_strategies 25 26 27 class MHAResidualAddAdapterStrategy(adapter_mixin_strategies.ResidualAddAdapterStrategy): 28 """ 29 An implementation of residual addition of an adapter module with its input for the MHA Adapters. 30 """ 31 32 def forward(self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin'): 33 """ 34 A basic strategy, comprising of a residual connection over the input, after forward pass by 35 the underlying adapter. Additional work is done to pack and unpack the dictionary of inputs and outputs. 36 37 Note: The `value` tensor is added to the output of the attention adapter as the residual connection. 38 39 Args: 40 input: A dictionary of multiple input arguments for the adapter module. 41 `query`, `key`, `value`: Original output tensor of the module, or the output of the 42 previous adapter (if more than one adapters are enabled). 43 `mask`: Attention mask. 44 `pos_emb`: Optional positional embedding for relative encoding. 45 adapter: The adapter module that is currently required to perform the forward pass. 46 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 47 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 48 49 Returns: 50 The result tensor, after one of the active adapters has finished its forward passes. 51 """ 52 out = self.compute_output(input, adapter, module=module) 53 54 # If not in training mode, or probability of stochastic depth is 0, skip step. 55 p = self.stochastic_depth 56 if not module.training or p == 0.0: 57 pass 58 else: 59 out = self.apply_stochastic_depth(out, input['value'], adapter, module=module) 60 61 # Return the residual connection output = input + adapter(input) 62 result = input['value'] + out 63 64 # If l2_lambda is activated, register the loss value 65 self.compute_auxiliary_losses(result, input['value'], adapter, module=module) 66 67 return result 68 69 def compute_output( 70 self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin' 71 ) -> torch.Tensor: 72 """ 73 Compute the output of a single adapter to some input. 74 75 Args: 76 input: Original output tensor of the module, or the output of the previous adapter (if more than 77 one adapters are enabled). 78 adapter: The adapter module that is currently required to perform the forward pass. 79 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 80 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 81 82 Returns: 83 The result tensor, after one of the active adapters has finished its forward passes. 84 """ 85 if isinstance(input, (list, tuple)): 86 out = adapter(*input) 87 elif isinstance(input, dict): 88 out = adapter(**input) 89 else: 90 out = adapter(input) 91 return out 92 93 94 @dataclass 95 class MHAResidualAddAdapterStrategyConfig(adapter_mixin_strategies.ResidualAddAdapterStrategyConfig): 96 _target_: str = "{0}.{1}".format( 97 MHAResidualAddAdapterStrategy.__module__, MHAResidualAddAdapterStrategy.__name__ 98 ) # mandatory field 99 100 101 class MultiHeadAttentionAdapter(mha.MultiHeadAttention, adapter_modules.AdapterModuleUtil): 102 """Multi-Head Attention layer of Transformer. 103 Args: 104 n_head (int): number of heads 105 n_feat (int): size of the features 106 dropout_rate (float): dropout rate 107 proj_dim (int, optional): Optional integer value for projection before computing attention. 108 If None, then there is no projection (equivalent to proj_dim = n_feat). 109 If > 0, then will project the n_feat to proj_dim before calculating attention. 110 If <0, then will equal n_head, so that each head has a projected dimension of 1. 111 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 112 """ 113 114 def __init__( 115 self, 116 n_head: int, 117 n_feat: int, 118 dropout_rate: float, 119 proj_dim: Optional[int] = None, 120 adapter_strategy: MHAResidualAddAdapterStrategy = None, 121 ): 122 super().__init__(n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, max_cache_len=0) 123 124 self.pre_norm = nn.LayerNorm(n_feat) 125 126 # Set the projection dim to number of heads automatically 127 if proj_dim is not None and proj_dim < 1: 128 proj_dim = n_head 129 130 self.proj_dim = proj_dim 131 132 # Recompute weights for projection dim 133 if self.proj_dim is not None: 134 if self.proj_dim % n_head != 0: 135 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 136 137 self.d_k = self.proj_dim // n_head 138 self.s_d_k = math.sqrt(self.d_k) 139 self.linear_q = nn.Linear(n_feat, self.proj_dim) 140 self.linear_k = nn.Linear(n_feat, self.proj_dim) 141 self.linear_v = nn.Linear(n_feat, self.proj_dim) 142 self.linear_out = nn.Linear(self.proj_dim, n_feat) 143 144 # Setup adapter strategy 145 self.setup_adapter_strategy(adapter_strategy) 146 147 # reset parameters for Q to be identity operation 148 self.reset_parameters() 149 150 def forward(self, query, key, value, mask, pos_emb=None, cache=None): 151 """Compute 'Scaled Dot Product Attention'. 152 Args: 153 query (torch.Tensor): (batch, time1, size) 154 key (torch.Tensor): (batch, time2, size) 155 value(torch.Tensor): (batch, time2, size) 156 mask (torch.Tensor): (batch, time1, time2) 157 cache (torch.Tensor) : (batch, time_cache, size) 158 159 returns: 160 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 161 cache (torch.Tensor) : (batch, time_cache_next, size) 162 """ 163 # Need to perform duplicate computations as at this point the tensors have been 164 # separated by the adapter forward 165 query = self.pre_norm(query) 166 key = self.pre_norm(key) 167 value = self.pre_norm(value) 168 169 return super().forward(query, key, value, mask, pos_emb, cache=cache) 170 171 def reset_parameters(self): 172 with torch.no_grad(): 173 nn.init.zeros_(self.linear_out.weight) 174 nn.init.zeros_(self.linear_out.bias) 175 176 def get_default_strategy_config(self) -> 'dataclass': 177 return MHAResidualAddAdapterStrategyConfig() 178 179 180 @dataclass 181 class MultiHeadAttentionAdapterConfig: 182 n_head: int 183 n_feat: int 184 dropout_rate: float = 0.0 185 proj_dim: Optional[int] = None 186 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 187 _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) 188 189 190 class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): 191 """Multi-Head Attention layer of Transformer-XL with support of relative positional encoding. 192 Paper: https://arxiv.org/abs/1901.02860 193 Args: 194 n_head (int): number of heads 195 n_feat (int): size of the features 196 dropout_rate (float): dropout rate 197 proj_dim (int, optional): Optional integer value for projection before computing attention. 198 If None, then there is no projection (equivalent to proj_dim = n_feat). 199 If > 0, then will project the n_feat to proj_dim before calculating attention. 200 If <0, then will equal n_head, so that each head has a projected dimension of 1. 201 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 202 """ 203 204 def __init__( 205 self, 206 n_head: int, 207 n_feat: int, 208 dropout_rate: float, 209 proj_dim: Optional[int] = None, 210 adapter_strategy: MHAResidualAddAdapterStrategyConfig = None, 211 ): 212 super().__init__( 213 n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, pos_bias_u=None, pos_bias_v=None, max_cache_len=0 214 ) 215 216 self.pre_norm = nn.LayerNorm(n_feat) 217 218 # Set the projection dim to number of heads automatically 219 if proj_dim is not None and proj_dim < 1: 220 proj_dim = n_head 221 222 self.proj_dim = proj_dim 223 224 # Recompute weights for projection dim 225 if self.proj_dim is not None: 226 if self.proj_dim % n_head != 0: 227 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 228 229 self.d_k = self.proj_dim // n_head 230 self.s_d_k = math.sqrt(self.d_k) 231 self.linear_q = nn.Linear(n_feat, self.proj_dim) 232 self.linear_k = nn.Linear(n_feat, self.proj_dim) 233 self.linear_v = nn.Linear(n_feat, self.proj_dim) 234 self.linear_out = nn.Linear(self.proj_dim, n_feat) 235 self.linear_pos = nn.Linear(n_feat, self.proj_dim, bias=False) 236 self.pos_bias_u = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 237 self.pos_bias_v = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 238 239 # Setup adapter strategy 240 self.setup_adapter_strategy(adapter_strategy) 241 242 # reset parameters for Q to be identity operation 243 self.reset_parameters() 244 245 def forward(self, query, key, value, mask, pos_emb, cache=None): 246 """Compute 'Scaled Dot Product Attention' with rel. positional encoding. 247 Args: 248 query (torch.Tensor): (batch, time1, size) 249 key (torch.Tensor): (batch, time2, size) 250 value(torch.Tensor): (batch, time2, size) 251 mask (torch.Tensor): (batch, time1, time2) 252 pos_emb (torch.Tensor) : (batch, time1, size) 253 cache (torch.Tensor) : (batch, time_cache, size) 254 Returns: 255 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 256 cache_next (torch.Tensor) : (batch, time_cache_next, size) 257 """ 258 # Need to perform duplicate computations as at this point the tensors have been 259 # separated by the adapter forward 260 query = self.pre_norm(query) 261 key = self.pre_norm(key) 262 value = self.pre_norm(value) 263 264 return super().forward(query, key, value, mask, pos_emb, cache=cache) 265 266 def reset_parameters(self): 267 with torch.no_grad(): 268 nn.init.zeros_(self.linear_out.weight) 269 nn.init.zeros_(self.linear_out.bias) 270 271 # NOTE: This exact procedure apparently highly important. 272 # Above operation is safe to do as self.linear_out.weight *= 0.0 (similar for bias) 273 # However: 274 # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 275 # For some reason at init sometimes it will cause the value of the tensor to become NaN 276 # All operations to compute matrix_ac and matrix_bd will then fail. 277 nn.init.zeros_(self.pos_bias_u) 278 nn.init.zeros_(self.pos_bias_v) 279 280 def get_default_strategy_config(self) -> 'dataclass': 281 return MHAResidualAddAdapterStrategyConfig() 282 283 284 @dataclass 285 class RelPositionMultiHeadAttentionAdapterConfig: 286 n_head: int 287 n_feat: int 288 dropout_rate: float = 0.0 289 proj_dim: Optional[int] = None 290 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 291 _target_: str = "{0}.{1}".format( 292 RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ 293 ) 294 295 296 class PositionalEncodingAdapter(mha.PositionalEncoding, adapter_modules.AdapterModuleUtil): 297 298 """ 299 Absolute positional embedding adapter. 300 301 .. note:: 302 303 Absolute positional embedding value is added to the input tensor *without residual connection* ! 304 Therefore, the input is changed, if you only require the positional embedding, drop the returned `x` ! 305 306 Args: 307 d_model (int): The input dimension of x. 308 max_len (int): The max sequence length. 309 xscale (float): The input scaling factor. Defaults to 1.0. 310 adapter_strategy (AbstractAdapterStrategy): By default, ReturnResultAdapterStrategyConfig. 311 An adapter composition function object. 312 NOTE: Since this is a positional encoding, it will not add a residual ! 313 """ 314 315 def __init__( 316 self, 317 d_model: int, 318 max_len: int = 5000, 319 xscale=1.0, 320 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 321 ): 322 323 super().__init__( 324 d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, 325 ) 326 327 # Setup adapter strategy 328 self.setup_adapter_strategy(adapter_strategy) 329 330 def get_default_strategy_config(self) -> 'dataclass': 331 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 332 333 334 @dataclass 335 class PositionalEncodingAdapterConfig: 336 d_model: int 337 max_len: int = 5000 338 xscale: float = 1.0 339 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 340 _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) 341 342 343 class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): 344 """ 345 Relative positional encoding for TransformerXL's layers 346 See : Appendix B in https://arxiv.org/abs/1901.02860 347 348 .. note:: 349 350 Relative positional embedding value is **not** added to the input tensor ! 351 Therefore, the input should be updated changed, if you only require the positional embedding, drop the returned `x` ! 352 353 Args: 354 d_model (int): embedding dim 355 max_len (int): maximum input length 356 xscale (bool): whether to scale the input by sqrt(d_model) 357 adapter_strategy: By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. 358 """ 359 360 def __init__( 361 self, 362 d_model: int, 363 max_len: int = 5000, 364 xscale=1.0, 365 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 366 ): 367 super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) 368 369 # Setup adapter strategy 370 self.setup_adapter_strategy(adapter_strategy) 371 372 def get_default_strategy_config(self) -> 'dataclass': 373 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 374 375 376 @dataclass 377 class RelPositionalEncodingAdapterConfig: 378 d_model: int 379 max_len: int = 5000 380 xscale: float = 1.0 381 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 382 _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) 383 [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import os 17 from dataclasses import dataclass 18 from typing import List, Optional, Tuple, Union 19 20 import torch 21 22 from nemo.collections.asr.parts.utils import rnnt_utils 23 from nemo.collections.common.tokenizers.tokenizer_spec import TokenizerSpec 24 from nemo.core.classes import Typing, typecheck 25 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 26 from nemo.utils import logging 27 28 DEFAULT_TOKEN_OFFSET = 100 29 30 31 def pack_hypotheses( 32 hypotheses: List[rnnt_utils.NBestHypotheses], logitlen: torch.Tensor, 33 ) -> List[rnnt_utils.NBestHypotheses]: 34 35 if logitlen is not None: 36 if hasattr(logitlen, 'cpu'): 37 logitlen_cpu = logitlen.to('cpu') 38 else: 39 logitlen_cpu = logitlen 40 41 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.NBestHypotheses 42 for candidate_idx, cand in enumerate(hyp.n_best_hypotheses): 43 cand.y_sequence = torch.tensor(cand.y_sequence, dtype=torch.long) 44 45 if logitlen is not None: 46 cand.length = logitlen_cpu[idx] 47 48 if cand.dec_state is not None: 49 cand.dec_state = _states_to_device(cand.dec_state) 50 51 return hypotheses 52 53 54 def _states_to_device(dec_state, device='cpu'): 55 if torch.is_tensor(dec_state): 56 dec_state = dec_state.to(device) 57 58 elif isinstance(dec_state, (list, tuple)): 59 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 60 61 return dec_state 62 63 64 class AbstractBeamCTCInfer(Typing): 65 """A beam CTC decoder. 66 67 Provides a common abstraction for sample level beam decoding. 68 69 Args: 70 blank_id: int, index of the blank token. Can be 0 or len(vocabulary). 71 beam_size: int, size of the beam used in the underlying beam search engine. 72 73 """ 74 75 @property 76 def input_types(self): 77 """Returns definitions of module input ports. 78 """ 79 return { 80 "decoder_output": NeuralType(('B', 'T', 'D'), LogprobsType()), 81 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 82 } 83 84 @property 85 def output_types(self): 86 """Returns definitions of module output ports. 87 """ 88 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 89 90 def __init__(self, blank_id: int, beam_size: int): 91 self.blank_id = blank_id 92 93 if beam_size < 1: 94 raise ValueError("Beam search size cannot be less than 1!") 95 96 self.beam_size = beam_size 97 98 # Variables set by corresponding setter methods 99 self.vocab = None 100 self.decoding_type = None 101 self.tokenizer = None 102 103 # Utility maps for vocabulary 104 self.vocab_index_map = None 105 self.index_vocab_map = None 106 107 # Internal variable, used to prevent double reduction of consecutive tokens (ctc collapse) 108 self.override_fold_consecutive_value = None 109 110 def set_vocabulary(self, vocab: List[str]): 111 """ 112 Set the vocabulary of the decoding framework. 113 114 Args: 115 vocab: List of str. Each token corresponds to its location in the vocabulary emitted by the model. 116 Note that this vocabulary must NOT contain the "BLANK" token. 117 """ 118 self.vocab = vocab 119 self.vocab_index_map = {v: i for i, v in enumerate(vocab)} 120 self.index_vocab_map = {i: v for i, v in enumerate(vocab)} 121 122 def set_decoding_type(self, decoding_type: str): 123 """ 124 Sets the decoding type of the framework. Can support either char or subword models. 125 126 Args: 127 decoding_type: Str corresponding to decoding type. Only supports "char" and "subword". 128 """ 129 decoding_type = decoding_type.lower() 130 supported_types = ['char', 'subword'] 131 132 if decoding_type not in supported_types: 133 raise ValueError( 134 f"Unsupported decoding type. Supported types = {supported_types}.\n" f"Given = {decoding_type}" 135 ) 136 137 self.decoding_type = decoding_type 138 139 def set_tokenizer(self, tokenizer: TokenizerSpec): 140 """ 141 Set the tokenizer of the decoding framework. 142 143 Args: 144 tokenizer: NeMo tokenizer object, which inherits from TokenizerSpec. 145 """ 146 self.tokenizer = tokenizer 147 148 @typecheck() 149 def forward( 150 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 151 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 152 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 153 Output token is generated auto-repressively. 154 155 Args: 156 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 157 decoder_lengths: list of int representing the length of each sequence 158 output sequence. 159 160 Returns: 161 packed list containing batch number of sentences (Hypotheses). 162 """ 163 raise NotImplementedError() 164 165 def __call__(self, *args, **kwargs): 166 return self.forward(*args, **kwargs) 167 168 169 class BeamCTCInfer(AbstractBeamCTCInfer): 170 """A greedy CTC decoder. 171 172 Provides a common abstraction for sample level and batch level greedy decoding. 173 174 Args: 175 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 176 preserve_alignments: Bool flag which preserves the history of logprobs generated during 177 decoding (sample / batched). When set to true, the Hypothesis will contain 178 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 179 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 180 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 181 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 182 183 """ 184 185 def __init__( 186 self, 187 blank_id: int, 188 beam_size: int, 189 search_type: str = "default", 190 return_best_hypothesis: bool = True, 191 preserve_alignments: bool = False, 192 compute_timestamps: bool = False, 193 beam_alpha: float = 1.0, 194 beam_beta: float = 0.0, 195 kenlm_path: str = None, 196 flashlight_cfg: Optional['FlashlightConfig'] = None, 197 pyctcdecode_cfg: Optional['PyCTCDecodeConfig'] = None, 198 ): 199 super().__init__(blank_id=blank_id, beam_size=beam_size) 200 201 self.search_type = search_type 202 self.return_best_hypothesis = return_best_hypothesis 203 self.preserve_alignments = preserve_alignments 204 self.compute_timestamps = compute_timestamps 205 206 if self.compute_timestamps: 207 raise ValueError(f"Currently this flag is not supported for beam search algorithms.") 208 209 self.vocab = None # This must be set by specific method by user before calling forward() ! 210 211 if search_type == "default" or search_type == "nemo": 212 self.search_algorithm = self.default_beam_search 213 elif search_type == "pyctcdecode": 214 self.search_algorithm = self._pyctcdecode_beam_search 215 elif search_type == "flashlight": 216 self.search_algorithm = self.flashlight_beam_search 217 else: 218 raise NotImplementedError( 219 f"The search type ({search_type}) supplied is not supported!\n" 220 f"Please use one of : (default, nemo, pyctcdecode)" 221 ) 222 223 # Log the beam search algorithm 224 logging.info(f"Beam search algorithm: {search_type}") 225 226 self.beam_alpha = beam_alpha 227 self.beam_beta = beam_beta 228 229 # Default beam search args 230 self.kenlm_path = kenlm_path 231 232 # PyCTCDecode params 233 if pyctcdecode_cfg is None: 234 pyctcdecode_cfg = PyCTCDecodeConfig() 235 self.pyctcdecode_cfg = pyctcdecode_cfg # type: PyCTCDecodeConfig 236 237 if flashlight_cfg is None: 238 flashlight_cfg = FlashlightConfig() 239 self.flashlight_cfg = flashlight_cfg 240 241 # Default beam search scorer functions 242 self.default_beam_scorer = None 243 self.pyctcdecode_beam_scorer = None 244 self.flashlight_beam_scorer = None 245 self.token_offset = 0 246 247 @typecheck() 248 def forward( 249 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 250 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 251 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 252 Output token is generated auto-repressively. 253 254 Args: 255 decoder_output: A tensor of size (batch, timesteps, features). 256 decoder_lengths: list of int representing the length of each sequence 257 output sequence. 258 259 Returns: 260 packed list containing batch number of sentences (Hypotheses). 261 """ 262 if self.vocab is None: 263 raise RuntimeError("Please set the vocabulary with `set_vocabulary()` before calling this function.") 264 265 if self.decoding_type is None: 266 raise ValueError("Please set the decoding type with `set_decoding_type()` before calling this function.") 267 268 with torch.no_grad(), torch.inference_mode(): 269 # Process each sequence independently 270 prediction_tensor = decoder_output 271 272 if prediction_tensor.ndim != 3: 273 raise ValueError( 274 f"`decoder_output` must be a tensor of shape [B, T, V] (log probs, float). " 275 f"Provided shape = {prediction_tensor.shape}" 276 ) 277 278 # determine type of input - logprobs or labels 279 out_len = decoder_lengths if decoder_lengths is not None else None 280 hypotheses = self.search_algorithm(prediction_tensor, out_len) 281 282 # Pack results into Hypotheses 283 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 284 285 # Pack the result 286 if self.return_best_hypothesis and isinstance(packed_result[0], rnnt_utils.NBestHypotheses): 287 packed_result = [res.n_best_hypotheses[0] for res in packed_result] # type: Hypothesis 288 289 return (packed_result,) 290 291 @torch.no_grad() 292 def default_beam_search( 293 self, x: torch.Tensor, out_len: torch.Tensor 294 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 295 """ 296 Open Seq2Seq Beam Search Algorithm (DeepSpeed) 297 298 Args: 299 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 300 and V is the vocabulary size. The tensor contains log-probabilities. 301 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 302 303 Returns: 304 A list of NBestHypotheses objects, one for each sequence in the batch. 305 """ 306 if self.compute_timestamps: 307 raise ValueError( 308 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 309 ) 310 311 if self.default_beam_scorer is None: 312 # Check for filepath 313 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 314 raise FileNotFoundError( 315 f"KenLM binary file not found at : {self.kenlm_path}. " 316 f"Please set a valid path in the decoding config." 317 ) 318 319 # perform token offset for subword models 320 if self.decoding_type == 'subword': 321 vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 322 else: 323 # char models 324 vocab = self.vocab 325 326 # Must import at runtime to avoid circular dependency due to module level import. 327 from nemo.collections.asr.modules.beam_search_decoder import BeamSearchDecoderWithLM 328 329 self.default_beam_scorer = BeamSearchDecoderWithLM( 330 vocab=vocab, 331 lm_path=self.kenlm_path, 332 beam_width=self.beam_size, 333 alpha=self.beam_alpha, 334 beta=self.beam_beta, 335 num_cpus=max(1, os.cpu_count()), 336 input_tensor=False, 337 ) 338 339 x = x.to('cpu') 340 341 with typecheck.disable_checks(): 342 data = [x[sample_id, : out_len[sample_id], :].softmax(dim=-1) for sample_id in range(len(x))] 343 beams_batch = self.default_beam_scorer.forward(log_probs=data, log_probs_length=None) 344 345 # For each sample in the batch 346 nbest_hypotheses = [] 347 for beams_idx, beams in enumerate(beams_batch): 348 # For each beam candidate / hypothesis in each sample 349 hypotheses = [] 350 for candidate_idx, candidate in enumerate(beams): 351 hypothesis = rnnt_utils.Hypothesis( 352 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 353 ) 354 355 # For subword encoding, NeMo will double encode the subword (multiple tokens) into a 356 # singular unicode id. In doing so, we preserve the semantic of the unicode token, and 357 # compress the size of the final KenLM ARPA / Binary file. 358 # In order to do double encoding, we shift the subword by some token offset. 359 # This step is ignored for character based models. 360 if self.decoding_type == 'subword': 361 pred_token_ids = [ord(c) - self.token_offset for c in candidate[1]] 362 else: 363 # Char models 364 pred_token_ids = [self.vocab_index_map[c] for c in candidate[1]] 365 366 # We preserve the token ids and the score for this hypothesis 367 hypothesis.y_sequence = pred_token_ids 368 hypothesis.score = candidate[0] 369 370 # If alignment must be preserved, we preserve a view of the output logprobs. 371 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 372 # require specific processing for each sample in the beam. 373 # This is done to preserve memory. 374 if self.preserve_alignments: 375 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 376 377 hypotheses.append(hypothesis) 378 379 # Wrap the result in NBestHypothesis. 380 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 381 nbest_hypotheses.append(hypotheses) 382 383 return nbest_hypotheses 384 385 @torch.no_grad() 386 def _pyctcdecode_beam_search( 387 self, x: torch.Tensor, out_len: torch.Tensor 388 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 389 """ 390 PyCTCDecode Beam Search Algorithm. Should support Char and Subword models. 391 392 Args: 393 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 394 and V is the vocabulary size. The tensor contains log-probabilities. 395 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 396 397 Returns: 398 A list of NBestHypotheses objects, one for each sequence in the batch. 399 """ 400 if self.compute_timestamps: 401 raise ValueError( 402 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 403 ) 404 405 try: 406 import pyctcdecode 407 except (ImportError, ModuleNotFoundError): 408 raise ImportError( 409 f"Could not load `pyctcdecode` library. Please install it from pip using :\n" 410 f"pip install --upgrade pyctcdecode" 411 ) 412 413 if self.pyctcdecode_beam_scorer is None: 414 self.pyctcdecode_beam_scorer = pyctcdecode.build_ctcdecoder( 415 labels=self.vocab, kenlm_model_path=self.kenlm_path, alpha=self.beam_alpha, beta=self.beam_beta 416 ) # type: pyctcdecode.BeamSearchDecoderCTC 417 418 x = x.to('cpu').numpy() 419 420 with typecheck.disable_checks(): 421 beams_batch = [] 422 for sample_id in range(len(x)): 423 logprobs = x[sample_id, : out_len[sample_id], :] 424 result = self.pyctcdecode_beam_scorer.decode_beams( 425 logprobs, 426 beam_width=self.beam_size, 427 beam_prune_logp=self.pyctcdecode_cfg.beam_prune_logp, 428 token_min_logp=self.pyctcdecode_cfg.token_min_logp, 429 prune_history=self.pyctcdecode_cfg.prune_history, 430 hotwords=self.pyctcdecode_cfg.hotwords, 431 hotword_weight=self.pyctcdecode_cfg.hotword_weight, 432 lm_start_state=None, 433 ) # Output format: text, last_lm_state, text_frames, logit_score, lm_score 434 beams_batch.append(result) 435 436 nbest_hypotheses = [] 437 for beams_idx, beams in enumerate(beams_batch): 438 hypotheses = [] 439 for candidate_idx, candidate in enumerate(beams): 440 # Candidate = (text, last_lm_state, text_frames, logit_score, lm_score) 441 hypothesis = rnnt_utils.Hypothesis( 442 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 443 ) 444 445 # TODO: Requires token ids to be returned rather than text. 446 if self.decoding_type == 'subword': 447 if self.tokenizer is None: 448 raise ValueError("Tokenizer must be provided for subword decoding. Use set_tokenizer().") 449 450 pred_token_ids = self.tokenizer.text_to_ids(candidate[0]) 451 else: 452 if self.vocab is None: 453 raise ValueError("Vocab must be provided for character decoding. Use set_vocab().") 454 455 chars = list(candidate[0]) 456 pred_token_ids = [self.vocab_index_map[c] for c in chars] 457 458 hypothesis.y_sequence = pred_token_ids 459 hypothesis.text = candidate[0] # text 460 hypothesis.score = candidate[4] # score 461 462 # Inject word level timestamps 463 hypothesis.timestep = candidate[2] # text_frames 464 465 if self.preserve_alignments: 466 hypothesis.alignments = torch.from_numpy(x[beams_idx][: out_len[beams_idx]]) 467 468 hypotheses.append(hypothesis) 469 470 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 471 nbest_hypotheses.append(hypotheses) 472 473 return nbest_hypotheses 474 475 @torch.no_grad() 476 def flashlight_beam_search( 477 self, x: torch.Tensor, out_len: torch.Tensor 478 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 479 """ 480 Flashlight Beam Search Algorithm. Should support Char and Subword models. 481 482 Args: 483 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 484 and V is the vocabulary size. The tensor contains log-probabilities. 485 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 486 487 Returns: 488 A list of NBestHypotheses objects, one for each sequence in the batch. 489 """ 490 if self.compute_timestamps: 491 raise ValueError( 492 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 493 ) 494 495 if self.flashlight_beam_scorer is None: 496 # Check for filepath 497 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 498 raise FileNotFoundError( 499 f"KenLM binary file not found at : {self.kenlm_path}. " 500 f"Please set a valid path in the decoding config." 501 ) 502 503 # perform token offset for subword models 504 # if self.decoding_type == 'subword': 505 # vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 506 # else: 507 # # char models 508 # vocab = self.vocab 509 510 # Must import at runtime to avoid circular dependency due to module level import. 511 from nemo.collections.asr.modules.flashlight_decoder import FlashLightKenLMBeamSearchDecoder 512 513 self.flashlight_beam_scorer = FlashLightKenLMBeamSearchDecoder( 514 lm_path=self.kenlm_path, 515 vocabulary=self.vocab, 516 tokenizer=self.tokenizer, 517 lexicon_path=self.flashlight_cfg.lexicon_path, 518 boost_path=self.flashlight_cfg.boost_path, 519 beam_size=self.beam_size, 520 beam_size_token=self.flashlight_cfg.beam_size_token, 521 beam_threshold=self.flashlight_cfg.beam_threshold, 522 lm_weight=self.beam_alpha, 523 word_score=self.beam_beta, 524 unk_weight=self.flashlight_cfg.unk_weight, 525 sil_weight=self.flashlight_cfg.sil_weight, 526 ) 527 528 x = x.to('cpu') 529 530 with typecheck.disable_checks(): 531 beams_batch = self.flashlight_beam_scorer.forward(log_probs=x) 532 533 # For each sample in the batch 534 nbest_hypotheses = [] 535 for beams_idx, beams in enumerate(beams_batch): 536 # For each beam candidate / hypothesis in each sample 537 hypotheses = [] 538 for candidate_idx, candidate in enumerate(beams): 539 hypothesis = rnnt_utils.Hypothesis( 540 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 541 ) 542 543 # We preserve the token ids and the score for this hypothesis 544 hypothesis.y_sequence = candidate['tokens'].tolist() 545 hypothesis.score = candidate['score'] 546 547 # If alignment must be preserved, we preserve a view of the output logprobs. 548 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 549 # require specific processing for each sample in the beam. 550 # This is done to preserve memory. 551 if self.preserve_alignments: 552 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 553 554 hypotheses.append(hypothesis) 555 556 # Wrap the result in NBestHypothesis. 557 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 558 nbest_hypotheses.append(hypotheses) 559 560 return nbest_hypotheses 561 562 def set_decoding_type(self, decoding_type: str): 563 super().set_decoding_type(decoding_type) 564 565 # Please check train_kenlm.py in scripts/asr_language_modeling/ to find out why we need 566 # TOKEN_OFFSET for BPE-based models 567 if self.decoding_type == 'subword': 568 self.token_offset = DEFAULT_TOKEN_OFFSET 569 570 571 @dataclass 572 class PyCTCDecodeConfig: 573 # These arguments cannot be imported from pyctcdecode (optional dependency) 574 # Therefore we copy the values explicitly 575 # Taken from pyctcdecode.constant 576 beam_prune_logp: float = -10.0 577 token_min_logp: float = -5.0 578 prune_history: bool = False 579 hotwords: Optional[List[str]] = None 580 hotword_weight: float = 10.0 581 582 583 @dataclass 584 class FlashlightConfig: 585 lexicon_path: Optional[str] = None 586 boost_path: Optional[str] = None 587 beam_size_token: int = 16 588 beam_threshold: float = 20.0 589 unk_weight: float = -math.inf 590 sil_weight: float = 0.0 591 592 593 @dataclass 594 class BeamCTCInferConfig: 595 beam_size: int 596 search_type: str = 'default' 597 preserve_alignments: bool = False 598 compute_timestamps: bool = False 599 return_best_hypothesis: bool = True 600 601 beam_alpha: float = 1.0 602 beam_beta: float = 0.0 603 kenlm_path: Optional[str] = None 604 605 flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() 606 pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() 607 [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import List, Optional 17 18 import torch 19 from omegaconf import DictConfig, OmegaConf 20 21 from nemo.collections.asr.parts.utils import rnnt_utils 22 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMeasureConfig, ConfidenceMeasureMixin 23 from nemo.core.classes import Typing, typecheck 24 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 25 from nemo.utils import logging 26 27 28 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 29 30 if logitlen is not None: 31 if hasattr(logitlen, 'cpu'): 32 logitlen_cpu = logitlen.to('cpu') 33 else: 34 logitlen_cpu = logitlen 35 36 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 37 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 38 39 if logitlen is not None: 40 hyp.length = logitlen_cpu[idx] 41 42 if hyp.dec_state is not None: 43 hyp.dec_state = _states_to_device(hyp.dec_state) 44 45 return hypotheses 46 47 48 def _states_to_device(dec_state, device='cpu'): 49 if torch.is_tensor(dec_state): 50 dec_state = dec_state.to(device) 51 52 elif isinstance(dec_state, (list, tuple)): 53 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 54 55 return dec_state 56 57 58 class GreedyCTCInfer(Typing, ConfidenceMeasureMixin): 59 """A greedy CTC decoder. 60 61 Provides a common abstraction for sample level and batch level greedy decoding. 62 63 Args: 64 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 65 preserve_alignments: Bool flag which preserves the history of logprobs generated during 66 decoding (sample / batched). When set to true, the Hypothesis will contain 67 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 68 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 69 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 70 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 71 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 72 generated during decoding. When set to true, the Hypothesis will contain 73 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 74 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 75 confidence scores. 76 77 name: The measure name (str). 78 Supported values: 79 - 'max_prob' for using the maximum token probability as a confidence. 80 - 'entropy' for using a normalized entropy of a log-likelihood vector. 81 82 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 83 Supported values: 84 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 85 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 86 Note that for this entropy, the alpha should comply the following inequality: 87 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 88 where V is the model vocabulary size. 89 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 90 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 91 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 92 More: https://en.wikipedia.org/wiki/Tsallis_entropy 93 - 'renyi' for the Rรฉnyi entropy. 94 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 95 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 96 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 97 98 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 99 When the alpha equals one, scaling is not applied to 'max_prob', 100 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 101 102 entropy_norm: A mapping of the entropy value to the interval [0,1]. 103 Supported values: 104 - 'lin' for using the linear mapping. 105 - 'exp' for using exponential mapping with linear shift. 106 107 """ 108 109 @property 110 def input_types(self): 111 """Returns definitions of module input ports. 112 """ 113 # Input can be of dimention - 114 # ('B', 'T', 'D') [Log probs] or ('B', 'T') [Labels] 115 116 return { 117 "decoder_output": NeuralType(None, LogprobsType()), 118 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 119 } 120 121 @property 122 def output_types(self): 123 """Returns definitions of module output ports. 124 """ 125 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 126 127 def __init__( 128 self, 129 blank_id: int, 130 preserve_alignments: bool = False, 131 compute_timestamps: bool = False, 132 preserve_frame_confidence: bool = False, 133 confidence_measure_cfg: Optional[DictConfig] = None, 134 ): 135 super().__init__() 136 137 self.blank_id = blank_id 138 self.preserve_alignments = preserve_alignments 139 # we need timestamps to extract non-blank per-frame confidence 140 self.compute_timestamps = compute_timestamps | preserve_frame_confidence 141 self.preserve_frame_confidence = preserve_frame_confidence 142 143 # set confidence calculation measure 144 self._init_confidence_measure(confidence_measure_cfg) 145 146 @typecheck() 147 def forward( 148 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 149 ): 150 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 151 Output token is generated auto-repressively. 152 153 Args: 154 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 155 decoder_lengths: list of int representing the length of each sequence 156 output sequence. 157 158 Returns: 159 packed list containing batch number of sentences (Hypotheses). 160 """ 161 with torch.inference_mode(): 162 hypotheses = [] 163 # Process each sequence independently 164 prediction_cpu_tensor = decoder_output.cpu() 165 166 if prediction_cpu_tensor.ndim < 2 or prediction_cpu_tensor.ndim > 3: 167 raise ValueError( 168 f"`decoder_output` must be a tensor of shape [B, T] (labels, int) or " 169 f"[B, T, V] (log probs, float). Provided shape = {prediction_cpu_tensor.shape}" 170 ) 171 172 # determine type of input - logprobs or labels 173 if prediction_cpu_tensor.ndim == 2: # labels 174 greedy_decode = self._greedy_decode_labels 175 else: 176 greedy_decode = self._greedy_decode_logprobs 177 178 for ind in range(prediction_cpu_tensor.shape[0]): 179 out_len = decoder_lengths[ind] if decoder_lengths is not None else None 180 hypothesis = greedy_decode(prediction_cpu_tensor[ind], out_len) 181 hypotheses.append(hypothesis) 182 183 # Pack results into Hypotheses 184 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 185 186 return (packed_result,) 187 188 @torch.no_grad() 189 def _greedy_decode_logprobs(self, x: torch.Tensor, out_len: torch.Tensor): 190 # x: [T, D] 191 # out_len: [seq_len] 192 193 # Initialize blank state and empty label set in Hypothesis 194 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 195 prediction = x.detach().cpu() 196 197 if out_len is not None: 198 prediction = prediction[:out_len] 199 200 prediction_logprobs, prediction_labels = prediction.max(dim=-1) 201 202 non_blank_ids = prediction_labels != self.blank_id 203 hypothesis.y_sequence = prediction_labels.numpy().tolist() 204 hypothesis.score = (prediction_logprobs[non_blank_ids]).sum() 205 206 if self.preserve_alignments: 207 # Preserve the logprobs, as well as labels after argmax 208 hypothesis.alignments = (prediction.clone(), prediction_labels.clone()) 209 210 if self.compute_timestamps: 211 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 212 213 if self.preserve_frame_confidence: 214 hypothesis.frame_confidence = self._get_confidence(prediction) 215 216 return hypothesis 217 218 @torch.no_grad() 219 def _greedy_decode_labels(self, x: torch.Tensor, out_len: torch.Tensor): 220 # x: [T] 221 # out_len: [seq_len] 222 223 # Initialize blank state and empty label set in Hypothesis 224 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 225 prediction_labels = x.detach().cpu() 226 227 if out_len is not None: 228 prediction_labels = prediction_labels[:out_len] 229 230 non_blank_ids = prediction_labels != self.blank_id 231 hypothesis.y_sequence = prediction_labels.numpy().tolist() 232 hypothesis.score = -1.0 233 234 if self.preserve_alignments: 235 raise ValueError("Requested for alignments, but predictions provided were labels, not log probabilities.") 236 237 if self.compute_timestamps: 238 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 239 240 if self.preserve_frame_confidence: 241 raise ValueError( 242 "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." 243 ) 244 245 return hypothesis 246 247 def __call__(self, *args, **kwargs): 248 return self.forward(*args, **kwargs) 249 250 251 @dataclass 252 class GreedyCTCInferConfig: 253 preserve_alignments: bool = False 254 compute_timestamps: bool = False 255 preserve_frame_confidence: bool = False 256 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 257 confidence_method_cfg: str = "DEPRECATED" 258 259 def __post_init__(self): 260 # OmegaConf.structured ensures that post_init check is always executed 261 self.confidence_measure_cfg = OmegaConf.structured( 262 self.confidence_measure_cfg 263 if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) 264 else ConfidenceMeasureConfig(**self.confidence_measure_cfg) 265 ) 266 if self.confidence_method_cfg != "DEPRECATED": 267 logging.warning( 268 "`confidence_method_cfg` is deprecated and will be removed in the future. " 269 "Please use `confidence_measure_cfg` instead." 270 ) 271 272 # TODO (alaptev): delete the following two lines sometime in the future 273 logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") 274 # OmegaConf.structured ensures that post_init check is always executed 275 self.confidence_measure_cfg = OmegaConf.structured( 276 self.confidence_method_cfg 277 if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) 278 else ConfidenceMeasureConfig(**self.confidence_method_cfg) 279 ) 280 [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 # Copyright 2017 Johns Hopkins University (Shinji Watanabe) 16 # 17 # Licensed under the Apache License, Version 2.0 (the "License"); 18 # you may not use this file except in compliance with the License. 19 # You may obtain a copy of the License at 20 # 21 # http://www.apache.org/licenses/LICENSE-2.0 22 # 23 # Unless required by applicable law or agreed to in writing, software 24 # distributed under the License is distributed on an "AS IS" BASIS, 25 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 26 # See the License for the specific language governing permissions and 27 # limitations under the License. 28 29 from dataclasses import dataclass 30 from typing import List, Optional, Tuple, Union 31 32 import numpy as np 33 import torch 34 from omegaconf import DictConfig, OmegaConf 35 36 from nemo.collections.asr.modules import rnnt_abstract 37 from nemo.collections.asr.parts.utils import rnnt_utils 38 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMeasureConfig, ConfidenceMeasureMixin 39 from nemo.collections.common.parts.rnn import label_collate 40 from nemo.core.classes import Typing, typecheck 41 from nemo.core.neural_types import AcousticEncodedRepresentation, ElementType, HypothesisType, LengthsType, NeuralType 42 from nemo.utils import logging 43 44 45 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 46 47 if hasattr(logitlen, 'cpu'): 48 logitlen_cpu = logitlen.to('cpu') 49 else: 50 logitlen_cpu = logitlen 51 52 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 53 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 54 hyp.length = logitlen_cpu[idx] 55 56 if hyp.dec_state is not None: 57 hyp.dec_state = _states_to_device(hyp.dec_state) 58 59 return hypotheses 60 61 62 def _states_to_device(dec_state, device='cpu'): 63 if torch.is_tensor(dec_state): 64 dec_state = dec_state.to(device) 65 66 elif isinstance(dec_state, (list, tuple)): 67 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 68 69 return dec_state 70 71 72 class _GreedyRNNTInfer(Typing, ConfidenceMeasureMixin): 73 """A greedy transducer decoder. 74 75 Provides a common abstraction for sample level and batch level greedy decoding. 76 77 Args: 78 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 79 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 80 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 81 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 82 to a sequence in a single time step; if set to None then there is 83 no limit. 84 preserve_alignments: Bool flag which preserves the history of alignments generated during 85 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 86 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 87 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 88 89 The length of the list corresponds to the Acoustic Length (T). 90 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 91 U is the number of target tokens for the current timestep Ti. 92 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 93 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 94 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 95 96 The length of the list corresponds to the Acoustic Length (T). 97 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 98 U is the number of target tokens for the current timestep Ti. 99 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 100 confidence scores. 101 102 name: The measure name (str). 103 Supported values: 104 - 'max_prob' for using the maximum token probability as a confidence. 105 - 'entropy' for using a normalized entropy of a log-likelihood vector. 106 107 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 108 Supported values: 109 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 110 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 111 Note that for this entropy, the alpha should comply the following inequality: 112 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 113 where V is the model vocabulary size. 114 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 115 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 116 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 117 More: https://en.wikipedia.org/wiki/Tsallis_entropy 118 - 'renyi' for the Rรฉnyi entropy. 119 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 120 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 121 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 122 123 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 124 When the alpha equals one, scaling is not applied to 'max_prob', 125 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 126 127 entropy_norm: A mapping of the entropy value to the interval [0,1]. 128 Supported values: 129 - 'lin' for using the linear mapping. 130 - 'exp' for using exponential mapping with linear shift. 131 """ 132 133 @property 134 def input_types(self): 135 """Returns definitions of module input ports. 136 """ 137 return { 138 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), 139 "encoded_lengths": NeuralType(tuple('B'), LengthsType()), 140 "partial_hypotheses": [NeuralType(elements_type=HypothesisType(), optional=True)], # must always be last 141 } 142 143 @property 144 def output_types(self): 145 """Returns definitions of module output ports. 146 """ 147 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 148 149 def __init__( 150 self, 151 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 152 joint_model: rnnt_abstract.AbstractRNNTJoint, 153 blank_index: int, 154 max_symbols_per_step: Optional[int] = None, 155 preserve_alignments: bool = False, 156 preserve_frame_confidence: bool = False, 157 confidence_measure_cfg: Optional[DictConfig] = None, 158 ): 159 super().__init__() 160 self.decoder = decoder_model 161 self.joint = joint_model 162 163 self._blank_index = blank_index 164 self._SOS = blank_index # Start of single index 165 self.max_symbols = max_symbols_per_step 166 self.preserve_alignments = preserve_alignments 167 self.preserve_frame_confidence = preserve_frame_confidence 168 169 # set confidence calculation measure 170 self._init_confidence_measure(confidence_measure_cfg) 171 172 def __call__(self, *args, **kwargs): 173 return self.forward(*args, **kwargs) 174 175 @torch.no_grad() 176 def _pred_step( 177 self, 178 label: Union[torch.Tensor, int], 179 hidden: Optional[torch.Tensor], 180 add_sos: bool = False, 181 batch_size: Optional[int] = None, 182 ) -> Tuple[torch.Tensor, torch.Tensor]: 183 """ 184 Common prediction step based on the AbstractRNNTDecoder implementation. 185 186 Args: 187 label: (int/torch.Tensor): Label or "Start-of-Signal" token. 188 hidden: (Optional torch.Tensor): RNN State vector 189 add_sos (bool): Whether to add a zero vector at the begging as "start of sentence" token. 190 batch_size: Batch size of the output tensor. 191 192 Returns: 193 g: (B, U, H) if add_sos is false, else (B, U + 1, H) 194 hid: (h, c) where h is the final sequence hidden state and c is 195 the final cell state: 196 h (tensor), shape (L, B, H) 197 c (tensor), shape (L, B, H) 198 """ 199 if isinstance(label, torch.Tensor): 200 # label: [batch, 1] 201 if label.dtype != torch.long: 202 label = label.long() 203 204 else: 205 # Label is an integer 206 if label == self._SOS: 207 return self.decoder.predict(None, hidden, add_sos=add_sos, batch_size=batch_size) 208 209 label = label_collate([[label]]) 210 211 # output: [B, 1, K] 212 return self.decoder.predict(label, hidden, add_sos=add_sos, batch_size=batch_size) 213 214 def _joint_step(self, enc, pred, log_normalize: Optional[bool] = None): 215 """ 216 Common joint step based on AbstractRNNTJoint implementation. 217 218 Args: 219 enc: Output of the Encoder model. A torch.Tensor of shape [B, 1, H1] 220 pred: Output of the Decoder model. A torch.Tensor of shape [B, 1, H2] 221 log_normalize: Whether to log normalize or not. None will log normalize only for CPU. 222 223 Returns: 224 logits of shape (B, T=1, U=1, V + 1) 225 """ 226 with torch.no_grad(): 227 logits = self.joint.joint(enc, pred) 228 229 if log_normalize is None: 230 if not logits.is_cuda: # Use log softmax only if on CPU 231 logits = logits.log_softmax(dim=len(logits.shape) - 1) 232 else: 233 if log_normalize: 234 logits = logits.log_softmax(dim=len(logits.shape) - 1) 235 236 return logits 237 238 239 class GreedyRNNTInfer(_GreedyRNNTInfer): 240 """A greedy transducer decoder. 241 242 Sequence level greedy decoding, performed auto-regressively. 243 244 Args: 245 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 246 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 247 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 248 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 249 to a sequence in a single time step; if set to None then there is 250 no limit. 251 preserve_alignments: Bool flag which preserves the history of alignments generated during 252 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 253 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 254 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 255 256 The length of the list corresponds to the Acoustic Length (T). 257 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 258 U is the number of target tokens for the current timestep Ti. 259 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 260 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 261 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 262 263 The length of the list corresponds to the Acoustic Length (T). 264 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 265 U is the number of target tokens for the current timestep Ti. 266 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 267 confidence scores. 268 269 name: The measure name (str). 270 Supported values: 271 - 'max_prob' for using the maximum token probability as a confidence. 272 - 'entropy' for using a normalized entropy of a log-likelihood vector. 273 274 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 275 Supported values: 276 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 277 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 278 Note that for this entropy, the alpha should comply the following inequality: 279 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 280 where V is the model vocabulary size. 281 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 282 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/Tsallis_entropy 285 - 'renyi' for the Rรฉnyi entropy. 286 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 287 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 288 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 289 290 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 291 When the alpha equals one, scaling is not applied to 'max_prob', 292 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 293 294 entropy_norm: A mapping of the entropy value to the interval [0,1]. 295 Supported values: 296 - 'lin' for using the linear mapping. 297 - 'exp' for using exponential mapping with linear shift. 298 """ 299 300 def __init__( 301 self, 302 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 303 joint_model: rnnt_abstract.AbstractRNNTJoint, 304 blank_index: int, 305 max_symbols_per_step: Optional[int] = None, 306 preserve_alignments: bool = False, 307 preserve_frame_confidence: bool = False, 308 confidence_measure_cfg: Optional[DictConfig] = None, 309 ): 310 super().__init__( 311 decoder_model=decoder_model, 312 joint_model=joint_model, 313 blank_index=blank_index, 314 max_symbols_per_step=max_symbols_per_step, 315 preserve_alignments=preserve_alignments, 316 preserve_frame_confidence=preserve_frame_confidence, 317 confidence_measure_cfg=confidence_measure_cfg, 318 ) 319 320 @typecheck() 321 def forward( 322 self, 323 encoder_output: torch.Tensor, 324 encoded_lengths: torch.Tensor, 325 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 326 ): 327 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 328 Output token is generated auto-regressively. 329 330 Args: 331 encoder_output: A tensor of size (batch, features, timesteps). 332 encoded_lengths: list of int representing the length of each sequence 333 output sequence. 334 335 Returns: 336 packed list containing batch number of sentences (Hypotheses). 337 """ 338 # Preserve decoder and joint training state 339 decoder_training_state = self.decoder.training 340 joint_training_state = self.joint.training 341 342 with torch.inference_mode(): 343 # Apply optional preprocessing 344 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 345 346 self.decoder.eval() 347 self.joint.eval() 348 349 hypotheses = [] 350 # Process each sequence independently 351 with self.decoder.as_frozen(), self.joint.as_frozen(): 352 for batch_idx in range(encoder_output.size(0)): 353 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 354 logitlen = encoded_lengths[batch_idx] 355 356 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 357 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 358 hypotheses.append(hypothesis) 359 360 # Pack results into Hypotheses 361 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 362 363 self.decoder.train(decoder_training_state) 364 self.joint.train(joint_training_state) 365 366 return (packed_result,) 367 368 @torch.no_grad() 369 def _greedy_decode( 370 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 371 ): 372 # x: [T, 1, D] 373 # out_len: [seq_len] 374 375 # Initialize blank state and empty label set in Hypothesis 376 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 377 378 if partial_hypotheses is not None: 379 hypothesis.last_token = partial_hypotheses.last_token 380 hypothesis.y_sequence = ( 381 partial_hypotheses.y_sequence.cpu().tolist() 382 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 383 else partial_hypotheses.y_sequence 384 ) 385 if partial_hypotheses.dec_state is not None: 386 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 387 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 388 389 if self.preserve_alignments: 390 # Alignments is a 2-dimensional dangling list representing T x U 391 hypothesis.alignments = [[]] 392 393 if self.preserve_frame_confidence: 394 hypothesis.frame_confidence = [[]] 395 396 # For timestep t in X_t 397 for time_idx in range(out_len): 398 # Extract encoder embedding at timestep t 399 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 400 f = x.narrow(dim=0, start=time_idx, length=1) 401 402 # Setup exit flags and counter 403 not_blank = True 404 symbols_added = 0 405 # While blank is not predicted, or we dont run out of max symbols per timestep 406 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 407 # In the first timestep, we initialize the network with RNNT Blank 408 # In later timesteps, we provide previous predicted label as input. 409 if hypothesis.last_token is None and hypothesis.dec_state is None: 410 last_label = self._SOS 411 else: 412 last_label = label_collate([[hypothesis.last_token]]) 413 414 # Perform prediction network and joint network steps. 415 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 416 # If preserving per-frame confidence, log_normalize must be true 417 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 418 0, 0, 0, : 419 ] 420 421 del g 422 423 # torch.max(0) op doesnt exist for FP 16. 424 if logp.dtype != torch.float32: 425 logp = logp.float() 426 427 # get index k, of max prob 428 v, k = logp.max(0) 429 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 430 431 if self.preserve_alignments: 432 # insert logprobs into last timestep 433 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 434 435 if self.preserve_frame_confidence: 436 # insert confidence into last timestep 437 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 438 439 del logp 440 441 # If blank token is predicted, exit inner loop, move onto next timestep t 442 if k == self._blank_index: 443 not_blank = False 444 else: 445 # Append token to label set, update RNN state. 446 hypothesis.y_sequence.append(k) 447 hypothesis.score += float(v) 448 hypothesis.timestep.append(time_idx) 449 hypothesis.dec_state = hidden_prime 450 hypothesis.last_token = k 451 452 # Increment token counter. 453 symbols_added += 1 454 455 if self.preserve_alignments: 456 # convert Ti-th logits into a torch array 457 hypothesis.alignments.append([]) # blank buffer for next timestep 458 459 if self.preserve_frame_confidence: 460 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 461 462 # Remove trailing empty list of Alignments 463 if self.preserve_alignments: 464 if len(hypothesis.alignments[-1]) == 0: 465 del hypothesis.alignments[-1] 466 467 # Remove trailing empty list of per-frame confidence 468 if self.preserve_frame_confidence: 469 if len(hypothesis.frame_confidence[-1]) == 0: 470 del hypothesis.frame_confidence[-1] 471 472 # Unpack the hidden states 473 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 474 475 return hypothesis 476 477 478 class GreedyBatchedRNNTInfer(_GreedyRNNTInfer): 479 """A batch level greedy transducer decoder. 480 481 Batch level greedy decoding, performed auto-regressively. 482 483 Args: 484 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 485 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 486 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 487 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 488 to a sequence in a single time step; if set to None then there is 489 no limit. 490 preserve_alignments: Bool flag which preserves the history of alignments generated during 491 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 492 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 493 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 494 495 The length of the list corresponds to the Acoustic Length (T). 496 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 497 U is the number of target tokens for the current timestep Ti. 498 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 499 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 500 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 501 502 The length of the list corresponds to the Acoustic Length (T). 503 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 504 U is the number of target tokens for the current timestep Ti. 505 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 506 confidence scores. 507 508 name: The measure name (str). 509 Supported values: 510 - 'max_prob' for using the maximum token probability as a confidence. 511 - 'entropy' for using a normalized entropy of a log-likelihood vector. 512 513 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 514 Supported values: 515 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 516 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 517 Note that for this entropy, the alpha should comply the following inequality: 518 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 519 where V is the model vocabulary size. 520 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 521 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 522 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 523 More: https://en.wikipedia.org/wiki/Tsallis_entropy 524 - 'renyi' for the Rรฉnyi entropy. 525 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 526 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 527 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 528 529 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 530 When the alpha equals one, scaling is not applied to 'max_prob', 531 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 532 533 entropy_norm: A mapping of the entropy value to the interval [0,1]. 534 Supported values: 535 - 'lin' for using the linear mapping. 536 - 'exp' for using exponential mapping with linear shift. 537 """ 538 539 def __init__( 540 self, 541 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 542 joint_model: rnnt_abstract.AbstractRNNTJoint, 543 blank_index: int, 544 max_symbols_per_step: Optional[int] = None, 545 preserve_alignments: bool = False, 546 preserve_frame_confidence: bool = False, 547 confidence_measure_cfg: Optional[DictConfig] = None, 548 ): 549 super().__init__( 550 decoder_model=decoder_model, 551 joint_model=joint_model, 552 blank_index=blank_index, 553 max_symbols_per_step=max_symbols_per_step, 554 preserve_alignments=preserve_alignments, 555 preserve_frame_confidence=preserve_frame_confidence, 556 confidence_measure_cfg=confidence_measure_cfg, 557 ) 558 559 # Depending on availability of `blank_as_pad` support 560 # switch between more efficient batch decoding technique 561 if self.decoder.blank_as_pad: 562 self._greedy_decode = self._greedy_decode_blank_as_pad 563 else: 564 self._greedy_decode = self._greedy_decode_masked 565 566 @typecheck() 567 def forward( 568 self, 569 encoder_output: torch.Tensor, 570 encoded_lengths: torch.Tensor, 571 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 572 ): 573 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 574 Output token is generated auto-regressively. 575 576 Args: 577 encoder_output: A tensor of size (batch, features, timesteps). 578 encoded_lengths: list of int representing the length of each sequence 579 output sequence. 580 581 Returns: 582 packed list containing batch number of sentences (Hypotheses). 583 """ 584 # Preserve decoder and joint training state 585 decoder_training_state = self.decoder.training 586 joint_training_state = self.joint.training 587 588 with torch.inference_mode(): 589 # Apply optional preprocessing 590 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 591 logitlen = encoded_lengths 592 593 self.decoder.eval() 594 self.joint.eval() 595 596 with self.decoder.as_frozen(), self.joint.as_frozen(): 597 inseq = encoder_output # [B, T, D] 598 hypotheses = self._greedy_decode( 599 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 600 ) 601 602 # Pack the hypotheses results 603 packed_result = pack_hypotheses(hypotheses, logitlen) 604 605 self.decoder.train(decoder_training_state) 606 self.joint.train(joint_training_state) 607 608 return (packed_result,) 609 610 def _greedy_decode_blank_as_pad( 611 self, 612 x: torch.Tensor, 613 out_len: torch.Tensor, 614 device: torch.device, 615 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 616 ): 617 if partial_hypotheses is not None: 618 raise NotImplementedError("`partial_hypotheses` support is not supported") 619 620 with torch.inference_mode(): 621 # x: [B, T, D] 622 # out_len: [B] 623 # device: torch.device 624 625 # Initialize list of Hypothesis 626 batchsize = x.shape[0] 627 hypotheses = [ 628 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 629 ] 630 631 # Initialize Hidden state matrix (shared by entire batch) 632 hidden = None 633 634 # If alignments need to be preserved, register a dangling list to hold the values 635 if self.preserve_alignments: 636 # alignments is a 3-dimensional dangling list representing B x T x U 637 for hyp in hypotheses: 638 hyp.alignments = [[]] 639 640 # If confidence scores need to be preserved, register a dangling list to hold the values 641 if self.preserve_frame_confidence: 642 # frame_confidence is a 3-dimensional dangling list representing B x T x U 643 for hyp in hypotheses: 644 hyp.frame_confidence = [[]] 645 646 # Last Label buffer + Last Label without blank buffer 647 # batch level equivalent of the last_label 648 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 649 650 # Mask buffers 651 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 652 653 # Get max sequence length 654 max_out_len = out_len.max() 655 for time_idx in range(max_out_len): 656 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 657 658 # Prepare t timestamp batch variables 659 not_blank = True 660 symbols_added = 0 661 662 # Reset blank mask 663 blank_mask.mul_(False) 664 665 # Update blank mask with time mask 666 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 667 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 668 blank_mask = time_idx >= out_len 669 # Start inner loop 670 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 671 # Batch prediction and joint network steps 672 # If very first prediction step, submit SOS tag (blank) to pred_step. 673 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 674 if time_idx == 0 and symbols_added == 0 and hidden is None: 675 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 676 else: 677 # Perform batch step prediction of decoder, getting new states and scores ("g") 678 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 679 680 # Batched joint step - Output = [B, V + 1] 681 # If preserving per-frame confidence, log_normalize must be true 682 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 683 :, 0, 0, : 684 ] 685 686 if logp.dtype != torch.float32: 687 logp = logp.float() 688 689 # Get index k, of max prob for batch 690 v, k = logp.max(1) 691 del g 692 693 # Update blank mask with current predicted blanks 694 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 695 k_is_blank = k == self._blank_index 696 blank_mask.bitwise_or_(k_is_blank) 697 all_blanks = torch.all(blank_mask) 698 699 del k_is_blank 700 701 # If preserving alignments, check if sequence length of sample has been reached 702 # before adding alignment 703 if self.preserve_alignments: 704 # Insert logprobs into last timestep per sample 705 logp_vals = logp.to('cpu') 706 logp_ids = logp_vals.max(1)[1] 707 for batch_idx, is_blank in enumerate(blank_mask): 708 # we only want to update non-blanks, unless we are at the last step in the loop where 709 # all elements produced blanks, otherwise there will be duplicate predictions 710 # saved in alignments 711 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 712 hypotheses[batch_idx].alignments[-1].append( 713 (logp_vals[batch_idx], logp_ids[batch_idx]) 714 ) 715 del logp_vals 716 717 # If preserving per-frame confidence, check if sequence length of sample has been reached 718 # before adding confidence scores 719 if self.preserve_frame_confidence: 720 # Insert probabilities into last timestep per sample 721 confidence = self._get_confidence(logp) 722 for batch_idx, is_blank in enumerate(blank_mask): 723 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 724 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 725 del logp 726 727 # If all samples predict / have predicted prior blanks, exit loop early 728 # This is equivalent to if single sample predicted k 729 if all_blanks: 730 not_blank = False 731 else: 732 # Collect batch indices where blanks occurred now/past 733 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 734 735 # Recover prior state for all samples which predicted blank now/past 736 if hidden is not None: 737 # LSTM has 2 states 738 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 739 740 elif len(blank_indices) > 0 and hidden is None: 741 # Reset state if there were some blank and other non-blank predictions in batch 742 # Original state is filled with zeros so we just multiply 743 # LSTM has 2 states 744 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 745 746 # Recover prior predicted label for all samples which predicted blank now/past 747 k[blank_indices] = last_label[blank_indices, 0] 748 749 # Update new label and hidden state for next iteration 750 last_label = k.clone().view(-1, 1) 751 hidden = hidden_prime 752 753 # Update predicted labels, accounting for time mask 754 # If blank was predicted even once, now or in the past, 755 # Force the current predicted label to also be blank 756 # This ensures that blanks propogate across all timesteps 757 # once they have occured (normally stopping condition of sample level loop). 758 for kidx, ki in enumerate(k): 759 if blank_mask[kidx] == 0: 760 hypotheses[kidx].y_sequence.append(ki) 761 hypotheses[kidx].timestep.append(time_idx) 762 hypotheses[kidx].score += float(v[kidx]) 763 symbols_added += 1 764 765 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 766 # Then preserve U at current timestep Ti 767 # Finally, forward the timestep history to Ti+1 for that sample 768 # All of this should only be done iff the current time index <= sample-level AM length. 769 # Otherwise ignore and move to next sample / next timestep. 770 if self.preserve_alignments: 771 772 # convert Ti-th logits into a torch array 773 for batch_idx in range(batchsize): 774 775 # this checks if current timestep <= sample-level AM length 776 # If current timestep > sample-level AM length, no alignments will be added 777 # Therefore the list of Uj alignments is empty here. 778 if len(hypotheses[batch_idx].alignments[-1]) > 0: 779 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 780 781 # Do the same if preserving per-frame confidence 782 if self.preserve_frame_confidence: 783 784 for batch_idx in range(batchsize): 785 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 786 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 787 788 # Remove trailing empty list of alignments at T_{am-len} x Uj 789 if self.preserve_alignments: 790 for batch_idx in range(batchsize): 791 if len(hypotheses[batch_idx].alignments[-1]) == 0: 792 del hypotheses[batch_idx].alignments[-1] 793 794 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 795 if self.preserve_frame_confidence: 796 for batch_idx in range(batchsize): 797 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 798 del hypotheses[batch_idx].frame_confidence[-1] 799 800 # Preserve states 801 for batch_idx in range(batchsize): 802 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 803 804 return hypotheses 805 806 def _greedy_decode_masked( 807 self, 808 x: torch.Tensor, 809 out_len: torch.Tensor, 810 device: torch.device, 811 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 812 ): 813 if partial_hypotheses is not None: 814 raise NotImplementedError("`partial_hypotheses` support is not supported") 815 816 # x: [B, T, D] 817 # out_len: [B] 818 # device: torch.device 819 820 # Initialize state 821 batchsize = x.shape[0] 822 hypotheses = [ 823 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 824 ] 825 826 # Initialize Hidden state matrix (shared by entire batch) 827 hidden = None 828 829 # If alignments need to be preserved, register a danling list to hold the values 830 if self.preserve_alignments: 831 # alignments is a 3-dimensional dangling list representing B x T x U 832 for hyp in hypotheses: 833 hyp.alignments = [[]] 834 else: 835 alignments = None 836 837 # If confidence scores need to be preserved, register a danling list to hold the values 838 if self.preserve_frame_confidence: 839 # frame_confidence is a 3-dimensional dangling list representing B x T x U 840 for hyp in hypotheses: 841 hyp.frame_confidence = [[]] 842 843 # Last Label buffer + Last Label without blank buffer 844 # batch level equivalent of the last_label 845 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 846 last_label_without_blank = last_label.clone() 847 848 # Mask buffers 849 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 850 851 # Get max sequence length 852 max_out_len = out_len.max() 853 854 with torch.inference_mode(): 855 for time_idx in range(max_out_len): 856 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 857 858 # Prepare t timestamp batch variables 859 not_blank = True 860 symbols_added = 0 861 862 # Reset blank mask 863 blank_mask.mul_(False) 864 865 # Update blank mask with time mask 866 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 867 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 868 blank_mask = time_idx >= out_len 869 870 # Start inner loop 871 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 872 # Batch prediction and joint network steps 873 # If very first prediction step, submit SOS tag (blank) to pred_step. 874 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 875 if time_idx == 0 and symbols_added == 0 and hidden is None: 876 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 877 else: 878 # Set a dummy label for the blank value 879 # This value will be overwritten by "blank" again the last label update below 880 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 881 last_label_without_blank_mask = last_label == self._blank_index 882 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 883 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 884 ~last_label_without_blank_mask 885 ] 886 887 # Perform batch step prediction of decoder, getting new states and scores ("g") 888 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 889 890 # Batched joint step - Output = [B, V + 1] 891 # If preserving per-frame confidence, log_normalize must be true 892 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 893 :, 0, 0, : 894 ] 895 896 if logp.dtype != torch.float32: 897 logp = logp.float() 898 899 # Get index k, of max prob for batch 900 v, k = logp.max(1) 901 del g 902 903 # Update blank mask with current predicted blanks 904 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 905 k_is_blank = k == self._blank_index 906 blank_mask.bitwise_or_(k_is_blank) 907 all_blanks = torch.all(blank_mask) 908 909 # If preserving alignments, check if sequence length of sample has been reached 910 # before adding alignment 911 if self.preserve_alignments: 912 # Insert logprobs into last timestep per sample 913 logp_vals = logp.to('cpu') 914 logp_ids = logp_vals.max(1)[1] 915 for batch_idx, is_blank in enumerate(blank_mask): 916 # we only want to update non-blanks, unless we are at the last step in the loop where 917 # all elements produced blanks, otherwise there will be duplicate predictions 918 # saved in alignments 919 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 920 hypotheses[batch_idx].alignments[-1].append( 921 (logp_vals[batch_idx], logp_ids[batch_idx]) 922 ) 923 924 del logp_vals 925 926 # If preserving per-frame confidence, check if sequence length of sample has been reached 927 # before adding confidence scores 928 if self.preserve_frame_confidence: 929 # Insert probabilities into last timestep per sample 930 confidence = self._get_confidence(logp) 931 for batch_idx, is_blank in enumerate(blank_mask): 932 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 933 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 934 del logp 935 936 # If all samples predict / have predicted prior blanks, exit loop early 937 # This is equivalent to if single sample predicted k 938 if blank_mask.all(): 939 not_blank = False 940 else: 941 # Collect batch indices where blanks occurred now/past 942 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 943 944 # Recover prior state for all samples which predicted blank now/past 945 if hidden is not None: 946 # LSTM has 2 states 947 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 948 949 elif len(blank_indices) > 0 and hidden is None: 950 # Reset state if there were some blank and other non-blank predictions in batch 951 # Original state is filled with zeros so we just multiply 952 # LSTM has 2 states 953 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 954 955 # Recover prior predicted label for all samples which predicted blank now/past 956 k[blank_indices] = last_label[blank_indices, 0] 957 958 # Update new label and hidden state for next iteration 959 last_label = k.view(-1, 1) 960 hidden = hidden_prime 961 962 # Update predicted labels, accounting for time mask 963 # If blank was predicted even once, now or in the past, 964 # Force the current predicted label to also be blank 965 # This ensures that blanks propogate across all timesteps 966 # once they have occured (normally stopping condition of sample level loop). 967 for kidx, ki in enumerate(k): 968 if blank_mask[kidx] == 0: 969 hypotheses[kidx].y_sequence.append(ki) 970 hypotheses[kidx].timestep.append(time_idx) 971 hypotheses[kidx].score += float(v[kidx]) 972 973 symbols_added += 1 974 975 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 976 # Then preserve U at current timestep Ti 977 # Finally, forward the timestep history to Ti+1 for that sample 978 # All of this should only be done iff the current time index <= sample-level AM length. 979 # Otherwise ignore and move to next sample / next timestep. 980 if self.preserve_alignments: 981 982 # convert Ti-th logits into a torch array 983 for batch_idx in range(batchsize): 984 985 # this checks if current timestep <= sample-level AM length 986 # If current timestep > sample-level AM length, no alignments will be added 987 # Therefore the list of Uj alignments is empty here. 988 if len(hypotheses[batch_idx].alignments[-1]) > 0: 989 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 990 991 # Do the same if preserving per-frame confidence 992 if self.preserve_frame_confidence: 993 994 for batch_idx in range(batchsize): 995 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 996 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 997 998 # Remove trailing empty list of alignments at T_{am-len} x Uj 999 if self.preserve_alignments: 1000 for batch_idx in range(batchsize): 1001 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1002 del hypotheses[batch_idx].alignments[-1] 1003 1004 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1005 if self.preserve_frame_confidence: 1006 for batch_idx in range(batchsize): 1007 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1008 del hypotheses[batch_idx].frame_confidence[-1] 1009 1010 # Preserve states 1011 for batch_idx in range(batchsize): 1012 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1013 1014 return hypotheses 1015 1016 1017 class ExportedModelGreedyBatchedRNNTInfer: 1018 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = None): 1019 self.encoder_model_path = encoder_model 1020 self.decoder_joint_model_path = decoder_joint_model 1021 self.max_symbols_per_step = max_symbols_per_step 1022 1023 # Will be populated at runtime 1024 self._blank_index = None 1025 1026 def __call__(self, audio_signal: torch.Tensor, length: torch.Tensor): 1027 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 1028 Output token is generated auto-regressively. 1029 1030 Args: 1031 encoder_output: A tensor of size (batch, features, timesteps). 1032 encoded_lengths: list of int representing the length of each sequence 1033 output sequence. 1034 1035 Returns: 1036 packed list containing batch number of sentences (Hypotheses). 1037 """ 1038 with torch.no_grad(): 1039 # Apply optional preprocessing 1040 encoder_output, encoded_lengths = self.run_encoder(audio_signal=audio_signal, length=length) 1041 1042 if torch.is_tensor(encoder_output): 1043 encoder_output = encoder_output.transpose(1, 2) 1044 else: 1045 encoder_output = encoder_output.transpose([0, 2, 1]) # (B, T, D) 1046 logitlen = encoded_lengths 1047 1048 inseq = encoder_output # [B, T, D] 1049 hypotheses, timestamps = self._greedy_decode(inseq, logitlen) 1050 1051 # Pack the hypotheses results 1052 packed_result = [rnnt_utils.Hypothesis(score=-1.0, y_sequence=[]) for _ in range(len(hypotheses))] 1053 for i in range(len(packed_result)): 1054 packed_result[i].y_sequence = torch.tensor(hypotheses[i], dtype=torch.long) 1055 packed_result[i].length = timestamps[i] 1056 1057 del hypotheses 1058 1059 return packed_result 1060 1061 def _greedy_decode(self, x, out_len): 1062 # x: [B, T, D] 1063 # out_len: [B] 1064 1065 # Initialize state 1066 batchsize = x.shape[0] 1067 hidden = self._get_initial_states(batchsize) 1068 target_lengths = torch.ones(batchsize, dtype=torch.int32) 1069 1070 # Output string buffer 1071 label = [[] for _ in range(batchsize)] 1072 timesteps = [[] for _ in range(batchsize)] 1073 1074 # Last Label buffer + Last Label without blank buffer 1075 # batch level equivalent of the last_label 1076 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long).numpy() 1077 if torch.is_tensor(x): 1078 last_label = torch.from_numpy(last_label).to(self.device) 1079 1080 # Mask buffers 1081 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool).numpy() 1082 1083 # Get max sequence length 1084 max_out_len = out_len.max() 1085 for time_idx in range(max_out_len): 1086 f = x[:, time_idx : time_idx + 1, :] # [B, 1, D] 1087 1088 if torch.is_tensor(f): 1089 f = f.transpose(1, 2) 1090 else: 1091 f = f.transpose([0, 2, 1]) 1092 1093 # Prepare t timestamp batch variables 1094 not_blank = True 1095 symbols_added = 0 1096 1097 # Reset blank mask 1098 blank_mask *= False 1099 1100 # Update blank mask with time mask 1101 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1102 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1103 blank_mask = time_idx >= out_len 1104 # Start inner loop 1105 while not_blank and (self.max_symbols_per_step is None or symbols_added < self.max_symbols_per_step): 1106 1107 # Batch prediction and joint network steps 1108 # If very first prediction step, submit SOS tag (blank) to pred_step. 1109 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1110 if time_idx == 0 and symbols_added == 0: 1111 g = torch.tensor([self._blank_index] * batchsize, dtype=torch.int32).view(-1, 1) 1112 else: 1113 if torch.is_tensor(last_label): 1114 g = last_label.type(torch.int32) 1115 else: 1116 g = last_label.astype(np.int32) 1117 1118 # Batched joint step - Output = [B, V + 1] 1119 joint_out, hidden_prime = self.run_decoder_joint(f, g, target_lengths, *hidden) 1120 logp, pred_lengths = joint_out 1121 logp = logp[:, 0, 0, :] 1122 1123 # Get index k, of max prob for batch 1124 if torch.is_tensor(logp): 1125 v, k = logp.max(1) 1126 else: 1127 k = np.argmax(logp, axis=1).astype(np.int32) 1128 1129 # Update blank mask with current predicted blanks 1130 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1131 k_is_blank = k == self._blank_index 1132 blank_mask |= k_is_blank 1133 1134 del k_is_blank 1135 del logp 1136 1137 # If all samples predict / have predicted prior blanks, exit loop early 1138 # This is equivalent to if single sample predicted k 1139 if blank_mask.all(): 1140 not_blank = False 1141 1142 else: 1143 # Collect batch indices where blanks occurred now/past 1144 if torch.is_tensor(blank_mask): 1145 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1146 else: 1147 blank_indices = blank_mask.astype(np.int32).nonzero() 1148 1149 if type(blank_indices) in (list, tuple): 1150 blank_indices = blank_indices[0] 1151 1152 # Recover prior state for all samples which predicted blank now/past 1153 if hidden is not None: 1154 # LSTM has 2 states 1155 for state_id in range(len(hidden)): 1156 hidden_prime[state_id][:, blank_indices, :] = hidden[state_id][:, blank_indices, :] 1157 1158 elif len(blank_indices) > 0 and hidden is None: 1159 # Reset state if there were some blank and other non-blank predictions in batch 1160 # Original state is filled with zeros so we just multiply 1161 # LSTM has 2 states 1162 for state_id in range(len(hidden_prime)): 1163 hidden_prime[state_id][:, blank_indices, :] *= 0.0 1164 1165 # Recover prior predicted label for all samples which predicted blank now/past 1166 k[blank_indices] = last_label[blank_indices, 0] 1167 1168 # Update new label and hidden state for next iteration 1169 if torch.is_tensor(k): 1170 last_label = k.clone().reshape(-1, 1) 1171 else: 1172 last_label = k.copy().reshape(-1, 1) 1173 hidden = hidden_prime 1174 1175 # Update predicted labels, accounting for time mask 1176 # If blank was predicted even once, now or in the past, 1177 # Force the current predicted label to also be blank 1178 # This ensures that blanks propogate across all timesteps 1179 # once they have occured (normally stopping condition of sample level loop). 1180 for kidx, ki in enumerate(k): 1181 if blank_mask[kidx] == 0: 1182 label[kidx].append(ki) 1183 timesteps[kidx].append(time_idx) 1184 1185 symbols_added += 1 1186 1187 return label, timesteps 1188 1189 def _setup_blank_index(self): 1190 raise NotImplementedError() 1191 1192 def run_encoder(self, audio_signal, length): 1193 raise NotImplementedError() 1194 1195 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1196 raise NotImplementedError() 1197 1198 def _get_initial_states(self, batchsize): 1199 raise NotImplementedError() 1200 1201 1202 class ONNXGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1203 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = 10): 1204 super().__init__( 1205 encoder_model=encoder_model, 1206 decoder_joint_model=decoder_joint_model, 1207 max_symbols_per_step=max_symbols_per_step, 1208 ) 1209 1210 try: 1211 import onnx 1212 import onnxruntime 1213 except (ModuleNotFoundError, ImportError): 1214 raise ImportError(f"`onnx` or `onnxruntime` could not be imported, please install the libraries.\n") 1215 1216 if torch.cuda.is_available(): 1217 # Try to use onnxruntime-gpu 1218 providers = ['TensorrtExecutionProvider', 'CUDAExecutionProvider'] 1219 else: 1220 # Fall back to CPU and onnxruntime-cpu 1221 providers = ['CPUExecutionProvider'] 1222 1223 onnx_session_opt = onnxruntime.SessionOptions() 1224 onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL 1225 1226 onnx_model = onnx.load(self.encoder_model_path) 1227 onnx.checker.check_model(onnx_model, full_check=True) 1228 self.encoder_model = onnx_model 1229 self.encoder = onnxruntime.InferenceSession( 1230 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1231 ) 1232 1233 onnx_model = onnx.load(self.decoder_joint_model_path) 1234 onnx.checker.check_model(onnx_model, full_check=True) 1235 self.decoder_joint_model = onnx_model 1236 self.decoder_joint = onnxruntime.InferenceSession( 1237 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1238 ) 1239 1240 logging.info("Successfully loaded encoder, decoder and joint onnx models !") 1241 1242 # Will be populated at runtime 1243 self._blank_index = None 1244 self.max_symbols_per_step = max_symbols_per_step 1245 1246 self._setup_encoder_input_output_keys() 1247 self._setup_decoder_joint_input_output_keys() 1248 self._setup_blank_index() 1249 1250 def _setup_encoder_input_output_keys(self): 1251 self.encoder_inputs = list(self.encoder_model.graph.input) 1252 self.encoder_outputs = list(self.encoder_model.graph.output) 1253 1254 def _setup_decoder_joint_input_output_keys(self): 1255 self.decoder_joint_inputs = list(self.decoder_joint_model.graph.input) 1256 self.decoder_joint_outputs = list(self.decoder_joint_model.graph.output) 1257 1258 def _setup_blank_index(self): 1259 # ASSUME: Single input with no time length information 1260 dynamic_dim = 257 1261 shapes = self.encoder_inputs[0].type.tensor_type.shape.dim 1262 ip_shape = [] 1263 for shape in shapes: 1264 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1265 ip_shape.append(dynamic_dim) # replace dynamic axes with constant 1266 else: 1267 ip_shape.append(int(shape.dim_value)) 1268 1269 enc_logits, encoded_length = self.run_encoder( 1270 audio_signal=torch.randn(*ip_shape), length=torch.randint(0, 1, size=(dynamic_dim,)) 1271 ) 1272 1273 # prepare states 1274 states = self._get_initial_states(batchsize=dynamic_dim) 1275 1276 # run decoder 1 step 1277 joint_out, states = self.run_decoder_joint(enc_logits, None, None, *states) 1278 log_probs, lengths = joint_out 1279 1280 self._blank_index = log_probs.shape[-1] - 1 # last token of vocab size is blank token 1281 logging.info( 1282 f"Enc-Dec-Joint step was evaluated, blank token id = {self._blank_index}; vocab size = {log_probs.shape[-1]}" 1283 ) 1284 1285 def run_encoder(self, audio_signal, length): 1286 if hasattr(audio_signal, 'cpu'): 1287 audio_signal = audio_signal.cpu().numpy() 1288 1289 if hasattr(length, 'cpu'): 1290 length = length.cpu().numpy() 1291 1292 ip = { 1293 self.encoder_inputs[0].name: audio_signal, 1294 self.encoder_inputs[1].name: length, 1295 } 1296 enc_out = self.encoder.run(None, ip) 1297 enc_out, encoded_length = enc_out # ASSUME: single output 1298 return enc_out, encoded_length 1299 1300 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1301 # ASSUME: Decoder is RNN Transducer 1302 if targets is None: 1303 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32) 1304 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32) 1305 1306 if hasattr(targets, 'cpu'): 1307 targets = targets.cpu().numpy() 1308 1309 if hasattr(target_length, 'cpu'): 1310 target_length = target_length.cpu().numpy() 1311 1312 ip = { 1313 self.decoder_joint_inputs[0].name: enc_logits, 1314 self.decoder_joint_inputs[1].name: targets, 1315 self.decoder_joint_inputs[2].name: target_length, 1316 } 1317 1318 num_states = 0 1319 if states is not None and len(states) > 0: 1320 num_states = len(states) 1321 for idx, state in enumerate(states): 1322 if hasattr(state, 'cpu'): 1323 state = state.cpu().numpy() 1324 1325 ip[self.decoder_joint_inputs[len(ip)].name] = state 1326 1327 dec_out = self.decoder_joint.run(None, ip) 1328 1329 # unpack dec output 1330 if num_states > 0: 1331 new_states = dec_out[-num_states:] 1332 dec_out = dec_out[:-num_states] 1333 else: 1334 new_states = None 1335 1336 return dec_out, new_states 1337 1338 def _get_initial_states(self, batchsize): 1339 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1340 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1341 num_states = len(input_state_nodes) 1342 if num_states == 0: 1343 return 1344 1345 input_states = [] 1346 for state_id in range(num_states): 1347 node = input_state_nodes[state_id] 1348 ip_shape = [] 1349 for shape_idx, shape in enumerate(node.type.tensor_type.shape.dim): 1350 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1351 ip_shape.append(batchsize) # replace dynamic axes with constant 1352 else: 1353 ip_shape.append(int(shape.dim_value)) 1354 1355 input_states.append(torch.zeros(*ip_shape)) 1356 1357 return input_states 1358 1359 1360 class TorchscriptGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1361 def __init__( 1362 self, 1363 encoder_model: str, 1364 decoder_joint_model: str, 1365 cfg: DictConfig, 1366 device: str, 1367 max_symbols_per_step: Optional[int] = 10, 1368 ): 1369 super().__init__( 1370 encoder_model=encoder_model, 1371 decoder_joint_model=decoder_joint_model, 1372 max_symbols_per_step=max_symbols_per_step, 1373 ) 1374 1375 self.cfg = cfg 1376 self.device = device 1377 1378 self.encoder = torch.jit.load(self.encoder_model_path, map_location=self.device) 1379 self.decoder_joint = torch.jit.load(self.decoder_joint_model_path, map_location=self.device) 1380 1381 logging.info("Successfully loaded encoder, decoder and joint torchscript models !") 1382 1383 # Will be populated at runtime 1384 self._blank_index = None 1385 self.max_symbols_per_step = max_symbols_per_step 1386 1387 self._setup_encoder_input_keys() 1388 self._setup_decoder_joint_input_keys() 1389 self._setup_blank_index() 1390 1391 def _setup_encoder_input_keys(self): 1392 arguments = self.encoder.forward.schema.arguments[1:] 1393 self.encoder_inputs = [arg for arg in arguments] 1394 1395 def _setup_decoder_joint_input_keys(self): 1396 arguments = self.decoder_joint.forward.schema.arguments[1:] 1397 self.decoder_joint_inputs = [arg for arg in arguments] 1398 1399 def _setup_blank_index(self): 1400 self._blank_index = len(self.cfg.joint.vocabulary) 1401 1402 logging.info(f"Blank token id = {self._blank_index}; vocab size = {len(self.cfg.joint.vocabulary) + 1}") 1403 1404 def run_encoder(self, audio_signal, length): 1405 enc_out = self.encoder(audio_signal, length) 1406 enc_out, encoded_length = enc_out # ASSUME: single output 1407 return enc_out, encoded_length 1408 1409 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1410 # ASSUME: Decoder is RNN Transducer 1411 if targets is None: 1412 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32, device=enc_logits.device) 1413 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32, device=enc_logits.device) 1414 1415 num_states = 0 1416 if states is not None and len(states) > 0: 1417 num_states = len(states) 1418 1419 dec_out = self.decoder_joint(enc_logits, targets, target_length, *states) 1420 1421 # unpack dec output 1422 if num_states > 0: 1423 new_states = dec_out[-num_states:] 1424 dec_out = dec_out[:-num_states] 1425 else: 1426 new_states = None 1427 1428 return dec_out, new_states 1429 1430 def _get_initial_states(self, batchsize): 1431 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1432 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1433 num_states = len(input_state_nodes) 1434 if num_states == 0: 1435 return 1436 1437 input_states = [] 1438 for state_id in range(num_states): 1439 # Hardcode shape size for LSTM (1 is for num layers in LSTM, which is flattened for export) 1440 ip_shape = [1, batchsize, self.cfg.model_defaults.pred_hidden] 1441 input_states.append(torch.zeros(*ip_shape, device=self.device)) 1442 1443 return input_states 1444 1445 1446 class GreedyMultiblankRNNTInfer(GreedyRNNTInfer): 1447 """A greedy transducer decoder for multi-blank RNN-T. 1448 1449 Sequence level greedy decoding, performed auto-regressively. 1450 1451 Args: 1452 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1453 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1454 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1455 big_blank_durations: a list containing durations for big blanks the model supports. 1456 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1457 to a sequence in a single time step; if set to None then there is 1458 no limit. 1459 preserve_alignments: Bool flag which preserves the history of alignments generated during 1460 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1461 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1462 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1463 The length of the list corresponds to the Acoustic Length (T). 1464 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1465 U is the number of target tokens for the current timestep Ti. 1466 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1467 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1468 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1469 The length of the list corresponds to the Acoustic Length (T). 1470 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1471 U is the number of target tokens for the current timestep Ti. 1472 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1473 confidence scores. 1474 1475 name: The measure name (str). 1476 Supported values: 1477 - 'max_prob' for using the maximum token probability as a confidence. 1478 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1479 1480 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 1481 Supported values: 1482 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1483 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1484 Note that for this entropy, the alpha should comply the following inequality: 1485 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1486 where V is the model vocabulary size. 1487 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1488 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1489 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1490 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1491 - 'renyi' for the Rรฉnyi entropy. 1492 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1493 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1494 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1495 1496 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1497 When the alpha equals one, scaling is not applied to 'max_prob', 1498 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1499 1500 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1501 Supported values: 1502 - 'lin' for using the linear mapping. 1503 - 'exp' for using exponential mapping with linear shift. 1504 """ 1505 1506 def __init__( 1507 self, 1508 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1509 joint_model: rnnt_abstract.AbstractRNNTJoint, 1510 blank_index: int, 1511 big_blank_durations: list, 1512 max_symbols_per_step: Optional[int] = None, 1513 preserve_alignments: bool = False, 1514 preserve_frame_confidence: bool = False, 1515 confidence_measure_cfg: Optional[DictConfig] = None, 1516 ): 1517 super().__init__( 1518 decoder_model=decoder_model, 1519 joint_model=joint_model, 1520 blank_index=blank_index, 1521 max_symbols_per_step=max_symbols_per_step, 1522 preserve_alignments=preserve_alignments, 1523 preserve_frame_confidence=preserve_frame_confidence, 1524 confidence_measure_cfg=confidence_measure_cfg, 1525 ) 1526 self.big_blank_durations = big_blank_durations 1527 self._SOS = blank_index - len(big_blank_durations) 1528 1529 @torch.no_grad() 1530 def _greedy_decode( 1531 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 1532 ): 1533 # x: [T, 1, D] 1534 # out_len: [seq_len] 1535 1536 # Initialize blank state and empty label set in Hypothesis 1537 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 1538 1539 if partial_hypotheses is not None: 1540 hypothesis.last_token = partial_hypotheses.last_token 1541 hypothesis.y_sequence = ( 1542 partial_hypotheses.y_sequence.cpu().tolist() 1543 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 1544 else partial_hypotheses.y_sequence 1545 ) 1546 if partial_hypotheses.dec_state is not None: 1547 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 1548 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 1549 1550 if self.preserve_alignments: 1551 # Alignments is a 2-dimensional dangling list representing T x U 1552 hypothesis.alignments = [[]] 1553 1554 if self.preserve_frame_confidence: 1555 hypothesis.frame_confidence = [[]] 1556 1557 # if this variable is > 1, it means the last emission was a big-blank and we need to skip frames. 1558 big_blank_duration = 1 1559 1560 # For timestep t in X_t 1561 for time_idx in range(out_len): 1562 if big_blank_duration > 1: 1563 # skip frames until big_blank_duration == 1. 1564 big_blank_duration -= 1 1565 continue 1566 # Extract encoder embedding at timestep t 1567 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 1568 f = x.narrow(dim=0, start=time_idx, length=1) 1569 1570 # Setup exit flags and counter 1571 not_blank = True 1572 symbols_added = 0 1573 1574 # While blank is not predicted, or we dont run out of max symbols per timestep 1575 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1576 # In the first timestep, we initialize the network with RNNT Blank 1577 # In later timesteps, we provide previous predicted label as input. 1578 if hypothesis.last_token is None and hypothesis.dec_state is None: 1579 last_label = self._SOS 1580 else: 1581 last_label = label_collate([[hypothesis.last_token]]) 1582 1583 # Perform prediction network and joint network steps. 1584 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 1585 # If preserving per-frame confidence, log_normalize must be true 1586 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1587 0, 0, 0, : 1588 ] 1589 1590 del g 1591 1592 # torch.max(0) op doesnt exist for FP 16. 1593 if logp.dtype != torch.float32: 1594 logp = logp.float() 1595 1596 # get index k, of max prob 1597 v, k = logp.max(0) 1598 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 1599 1600 # Note, we have non-blanks in the vocab first, followed by big blanks, and standard blank at last. 1601 # here we check if it's a big blank and if yes, set the duration variable. 1602 if k >= self._blank_index - len(self.big_blank_durations) and k < self._blank_index: 1603 big_blank_duration = self.big_blank_durations[self._blank_index - k - 1] 1604 1605 if self.preserve_alignments: 1606 # insert logprobs into last timestep 1607 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 1608 1609 if self.preserve_frame_confidence: 1610 # insert confidence into last timestep 1611 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 1612 1613 del logp 1614 1615 # If any type of blank token is predicted, exit inner loop, move onto next timestep t 1616 if k >= self._blank_index - len(self.big_blank_durations): 1617 not_blank = False 1618 else: 1619 # Append token to label set, update RNN state. 1620 hypothesis.y_sequence.append(k) 1621 hypothesis.score += float(v) 1622 hypothesis.timestep.append(time_idx) 1623 hypothesis.dec_state = hidden_prime 1624 hypothesis.last_token = k 1625 1626 # Increment token counter. 1627 symbols_added += 1 1628 1629 if self.preserve_alignments: 1630 # convert Ti-th logits into a torch array 1631 hypothesis.alignments.append([]) # blank buffer for next timestep 1632 1633 if self.preserve_frame_confidence: 1634 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 1635 1636 # Remove trailing empty list of Alignments 1637 if self.preserve_alignments: 1638 if len(hypothesis.alignments[-1]) == 0: 1639 del hypothesis.alignments[-1] 1640 1641 # Remove trailing empty list of per-frame confidence 1642 if self.preserve_frame_confidence: 1643 if len(hypothesis.frame_confidence[-1]) == 0: 1644 del hypothesis.frame_confidence[-1] 1645 1646 # Unpack the hidden states 1647 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 1648 1649 return hypothesis 1650 1651 1652 class GreedyBatchedMultiblankRNNTInfer(GreedyBatchedRNNTInfer): 1653 """A batch level greedy transducer decoder. 1654 Batch level greedy decoding, performed auto-regressively. 1655 Args: 1656 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1657 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1658 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1659 big_blank_durations: a list containing durations for big blanks the model supports. 1660 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1661 to a sequence in a single time step; if set to None then there is 1662 no limit. 1663 preserve_alignments: Bool flag which preserves the history of alignments generated during 1664 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1665 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1666 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1667 The length of the list corresponds to the Acoustic Length (T). 1668 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1669 U is the number of target tokens for the current timestep Ti. 1670 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1671 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1672 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1673 The length of the list corresponds to the Acoustic Length (T). 1674 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1675 U is the number of target tokens for the current timestep Ti. 1676 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1677 confidence scores. 1678 1679 name: The measure name (str). 1680 Supported values: 1681 - 'max_prob' for using the maximum token probability as a confidence. 1682 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1683 1684 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 1685 Supported values: 1686 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1687 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1688 Note that for this entropy, the alpha should comply the following inequality: 1689 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1690 where V is the model vocabulary size. 1691 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1692 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1693 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1694 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1695 - 'renyi' for the Rรฉnyi entropy. 1696 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1697 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1698 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1699 1700 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1701 When the alpha equals one, scaling is not applied to 'max_prob', 1702 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1703 1704 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1705 Supported values: 1706 - 'lin' for using the linear mapping. 1707 - 'exp' for using exponential mapping with linear shift. 1708 """ 1709 1710 def __init__( 1711 self, 1712 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1713 joint_model: rnnt_abstract.AbstractRNNTJoint, 1714 blank_index: int, 1715 big_blank_durations: List[int], 1716 max_symbols_per_step: Optional[int] = None, 1717 preserve_alignments: bool = False, 1718 preserve_frame_confidence: bool = False, 1719 confidence_measure_cfg: Optional[DictConfig] = None, 1720 ): 1721 super().__init__( 1722 decoder_model=decoder_model, 1723 joint_model=joint_model, 1724 blank_index=blank_index, 1725 max_symbols_per_step=max_symbols_per_step, 1726 preserve_alignments=preserve_alignments, 1727 preserve_frame_confidence=preserve_frame_confidence, 1728 confidence_measure_cfg=confidence_measure_cfg, 1729 ) 1730 self.big_blank_durations = big_blank_durations 1731 1732 # Depending on availability of `blank_as_pad` support 1733 # switch between more efficient batch decoding technique 1734 if self.decoder.blank_as_pad: 1735 self._greedy_decode = self._greedy_decode_blank_as_pad 1736 else: 1737 self._greedy_decode = self._greedy_decode_masked 1738 self._SOS = blank_index - len(big_blank_durations) 1739 1740 def _greedy_decode_blank_as_pad( 1741 self, 1742 x: torch.Tensor, 1743 out_len: torch.Tensor, 1744 device: torch.device, 1745 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1746 ): 1747 if partial_hypotheses is not None: 1748 raise NotImplementedError("`partial_hypotheses` support is not supported") 1749 1750 with torch.inference_mode(): 1751 # x: [B, T, D] 1752 # out_len: [B] 1753 # device: torch.device 1754 1755 # Initialize list of Hypothesis 1756 batchsize = x.shape[0] 1757 hypotheses = [ 1758 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1759 ] 1760 1761 # Initialize Hidden state matrix (shared by entire batch) 1762 hidden = None 1763 1764 # If alignments need to be preserved, register a danling list to hold the values 1765 if self.preserve_alignments: 1766 # alignments is a 3-dimensional dangling list representing B x T x U 1767 for hyp in hypotheses: 1768 hyp.alignments = [[]] 1769 1770 # If confidence scores need to be preserved, register a danling list to hold the values 1771 if self.preserve_frame_confidence: 1772 # frame_confidence is a 3-dimensional dangling list representing B x T x U 1773 for hyp in hypotheses: 1774 hyp.frame_confidence = [[]] 1775 1776 # Last Label buffer + Last Label without blank buffer 1777 # batch level equivalent of the last_label 1778 last_label = torch.full([batchsize, 1], fill_value=self._SOS, dtype=torch.long, device=device) 1779 1780 # this mask is true for if the emission is *any type* of blank. 1781 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 1782 1783 # Get max sequence length 1784 max_out_len = out_len.max() 1785 1786 # We have a mask for each big blank. A mask is "true" means: the previous emission is exactly the big-blank 1787 # with the corresponding duration, or has larger duration. E.g., for big_blank_mask for duration 2, it will 1788 # be set true if the previous emission was a big blank with duration 4, or 3 or 2; but false if prevoius 1789 # emission was a standard blank (with duration = 1). 1790 big_blank_masks = [torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device)] * len( 1791 self.big_blank_durations 1792 ) 1793 1794 # if this variable > 1, it means the previous emission was big-blank and we need to skip frames. 1795 big_blank_duration = 1 1796 1797 for time_idx in range(max_out_len): 1798 if big_blank_duration > 1: 1799 # skip frames until big_blank_duration == 1 1800 big_blank_duration -= 1 1801 continue 1802 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 1803 1804 # Prepare t timestamp batch variables 1805 not_blank = True 1806 symbols_added = 0 1807 1808 # Reset all blank masks 1809 blank_mask.mul_(False) 1810 for i in range(len(big_blank_masks)): 1811 big_blank_masks[i].mul_(False) 1812 1813 # Update blank mask with time mask 1814 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1815 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1816 blank_mask = time_idx >= out_len 1817 for i in range(len(big_blank_masks)): 1818 big_blank_masks[i] = time_idx >= out_len 1819 1820 # Start inner loop 1821 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1822 # Batch prediction and joint network steps 1823 # If very first prediction step, submit SOS tag (blank) to pred_step. 1824 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1825 if time_idx == 0 and symbols_added == 0 and hidden is None: 1826 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 1827 else: 1828 # Perform batch step prediction of decoder, getting new states and scores ("g") 1829 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 1830 1831 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 1832 # If preserving per-frame confidence, log_normalize must be true 1833 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1834 :, 0, 0, : 1835 ] 1836 1837 if logp.dtype != torch.float32: 1838 logp = logp.float() 1839 1840 # Get index k, of max prob for batch 1841 v, k = logp.max(1) 1842 del g 1843 1844 # Update blank mask with current predicted blanks 1845 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1846 k_is_blank = k >= self._blank_index - len(self.big_blank_durations) 1847 blank_mask.bitwise_or_(k_is_blank) 1848 1849 for i in range(len(big_blank_masks)): 1850 # using <= since as we mentioned before, the mask doesn't store exact matches. 1851 # instead, it is True when the predicted blank's duration is >= the duration that the 1852 # mask corresponds to. 1853 k_is_big_blank = k <= self._blank_index - 1 - i 1854 1855 # need to do a bitwise_and since it could also be a non-blank. 1856 k_is_big_blank.bitwise_and_(k_is_blank) 1857 big_blank_masks[i].bitwise_or_(k_is_big_blank) 1858 1859 del k_is_blank 1860 1861 # If preserving alignments, check if sequence length of sample has been reached 1862 # before adding alignment 1863 if self.preserve_alignments: 1864 # Insert logprobs into last timestep per sample 1865 logp_vals = logp.to('cpu') 1866 logp_ids = logp_vals.max(1)[1] 1867 for batch_idx in range(batchsize): 1868 if time_idx < out_len[batch_idx]: 1869 hypotheses[batch_idx].alignments[-1].append( 1870 (logp_vals[batch_idx], logp_ids[batch_idx]) 1871 ) 1872 del logp_vals 1873 1874 # If preserving per-frame confidence, check if sequence length of sample has been reached 1875 # before adding confidence scores 1876 if self.preserve_frame_confidence: 1877 # Insert probabilities into last timestep per sample 1878 confidence = self._get_confidence(logp) 1879 for batch_idx in range(batchsize): 1880 if time_idx < out_len[batch_idx]: 1881 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 1882 del logp 1883 1884 # If all samples predict / have predicted prior blanks, exit loop early 1885 # This is equivalent to if single sample predicted k 1886 if blank_mask.all(): 1887 not_blank = False 1888 else: 1889 # Collect batch indices where blanks occurred now/past 1890 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1891 1892 # Recover prior state for all samples which predicted blank now/past 1893 if hidden is not None: 1894 # LSTM has 2 states 1895 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 1896 1897 elif len(blank_indices) > 0 and hidden is None: 1898 # Reset state if there were some blank and other non-blank predictions in batch 1899 # Original state is filled with zeros so we just multiply 1900 # LSTM has 2 states 1901 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 1902 1903 # Recover prior predicted label for all samples which predicted blank now/past 1904 k[blank_indices] = last_label[blank_indices, 0] 1905 1906 # Update new label and hidden state for next iteration 1907 last_label = k.clone().view(-1, 1) 1908 hidden = hidden_prime 1909 1910 # Update predicted labels, accounting for time mask 1911 # If blank was predicted even once, now or in the past, 1912 # Force the current predicted label to also be blank 1913 # This ensures that blanks propogate across all timesteps 1914 # once they have occured (normally stopping condition of sample level loop). 1915 for kidx, ki in enumerate(k): 1916 if blank_mask[kidx] == 0: 1917 hypotheses[kidx].y_sequence.append(ki) 1918 hypotheses[kidx].timestep.append(time_idx) 1919 hypotheses[kidx].score += float(v[kidx]) 1920 1921 symbols_added += 1 1922 1923 for i in range(len(big_blank_masks) + 1): 1924 # The task here is find the shortest blank duration of all batches. 1925 # so we start from the shortest blank duration and go up, 1926 # and stop once we found the duration whose corresponding mask isn't all True. 1927 if i == len(big_blank_masks) or not big_blank_masks[i].all(): 1928 big_blank_duration = self.big_blank_durations[i - 1] if i > 0 else 1 1929 break 1930 1931 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 1932 # Then preserve U at current timestep Ti 1933 # Finally, forward the timestep history to Ti+1 for that sample 1934 # All of this should only be done iff the current time index <= sample-level AM length. 1935 # Otherwise ignore and move to next sample / next timestep. 1936 if self.preserve_alignments: 1937 1938 # convert Ti-th logits into a torch array 1939 for batch_idx in range(batchsize): 1940 1941 # this checks if current timestep <= sample-level AM length 1942 # If current timestep > sample-level AM length, no alignments will be added 1943 # Therefore the list of Uj alignments is empty here. 1944 if len(hypotheses[batch_idx].alignments[-1]) > 0: 1945 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 1946 1947 # Do the same if preserving per-frame confidence 1948 if self.preserve_frame_confidence: 1949 1950 for batch_idx in range(batchsize): 1951 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 1952 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 1953 1954 # Remove trailing empty list of alignments at T_{am-len} x Uj 1955 if self.preserve_alignments: 1956 for batch_idx in range(batchsize): 1957 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1958 del hypotheses[batch_idx].alignments[-1] 1959 1960 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1961 if self.preserve_frame_confidence: 1962 for batch_idx in range(batchsize): 1963 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1964 del hypotheses[batch_idx].frame_confidence[-1] 1965 1966 # Preserve states 1967 for batch_idx in range(batchsize): 1968 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1969 1970 return hypotheses 1971 1972 def _greedy_decode_masked( 1973 self, 1974 x: torch.Tensor, 1975 out_len: torch.Tensor, 1976 device: torch.device, 1977 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1978 ): 1979 if partial_hypotheses is not None: 1980 raise NotImplementedError("`partial_hypotheses` support is not supported") 1981 1982 if self.big_blank_durations != [1] * len(self.big_blank_durations): 1983 raise NotImplementedError( 1984 "Efficient frame-skipping version for multi-blank masked decoding is not supported." 1985 ) 1986 1987 # x: [B, T, D] 1988 # out_len: [B] 1989 # device: torch.device 1990 1991 # Initialize state 1992 batchsize = x.shape[0] 1993 hypotheses = [ 1994 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1995 ] 1996 1997 # Initialize Hidden state matrix (shared by entire batch) 1998 hidden = None 1999 2000 # If alignments need to be preserved, register a danling list to hold the values 2001 if self.preserve_alignments: 2002 # alignments is a 3-dimensional dangling list representing B x T x U 2003 for hyp in hypotheses: 2004 hyp.alignments = [[]] 2005 else: 2006 hyp.alignments = None 2007 2008 # If confidence scores need to be preserved, register a danling list to hold the values 2009 if self.preserve_frame_confidence: 2010 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2011 for hyp in hypotheses: 2012 hyp.frame_confidence = [[]] 2013 2014 # Last Label buffer + Last Label without blank buffer 2015 # batch level equivalent of the last_label 2016 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2017 last_label_without_blank = last_label.clone() 2018 2019 # Mask buffers 2020 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2021 2022 # Get max sequence length 2023 max_out_len = out_len.max() 2024 2025 with torch.inference_mode(): 2026 for time_idx in range(max_out_len): 2027 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2028 2029 # Prepare t timestamp batch variables 2030 not_blank = True 2031 symbols_added = 0 2032 2033 # Reset blank mask 2034 blank_mask.mul_(False) 2035 2036 # Update blank mask with time mask 2037 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2038 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2039 blank_mask = time_idx >= out_len 2040 2041 # Start inner loop 2042 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 2043 # Batch prediction and joint network steps 2044 # If very first prediction step, submit SOS tag (blank) to pred_step. 2045 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2046 if time_idx == 0 and symbols_added == 0 and hidden is None: 2047 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2048 else: 2049 # Set a dummy label for the blank value 2050 # This value will be overwritten by "blank" again the last label update below 2051 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 2052 last_label_without_blank_mask = last_label >= self._blank_index 2053 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 2054 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 2055 ~last_label_without_blank_mask 2056 ] 2057 2058 # Perform batch step prediction of decoder, getting new states and scores ("g") 2059 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 2060 2061 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2062 # If preserving per-frame confidence, log_normalize must be true 2063 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 2064 :, 0, 0, : 2065 ] 2066 2067 if logp.dtype != torch.float32: 2068 logp = logp.float() 2069 2070 # Get index k, of max prob for batch 2071 v, k = logp.max(1) 2072 del g 2073 2074 # Update blank mask with current predicted blanks 2075 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2076 k_is_blank = k == self._blank_index 2077 blank_mask.bitwise_or_(k_is_blank) 2078 2079 # If preserving alignments, check if sequence length of sample has been reached 2080 # before adding alignment 2081 if self.preserve_alignments: 2082 # Insert logprobs into last timestep per sample 2083 logp_vals = logp.to('cpu') 2084 logp_ids = logp_vals.max(1)[1] 2085 for batch_idx in range(batchsize): 2086 if time_idx < out_len[batch_idx]: 2087 hypotheses[batch_idx].alignments[-1].append( 2088 (logp_vals[batch_idx], logp_ids[batch_idx]) 2089 ) 2090 del logp_vals 2091 2092 # If preserving per-frame confidence, check if sequence length of sample has been reached 2093 # before adding confidence scores 2094 if self.preserve_frame_confidence: 2095 # Insert probabilities into last timestep per sample 2096 confidence = self._get_confidence(logp) 2097 for batch_idx in range(batchsize): 2098 if time_idx < out_len[batch_idx]: 2099 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 2100 del logp 2101 2102 # If all samples predict / have predicted prior blanks, exit loop early 2103 # This is equivalent to if single sample predicted k 2104 if blank_mask.all(): 2105 not_blank = False 2106 else: 2107 # Collect batch indices where blanks occurred now/past 2108 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2109 2110 # Recover prior state for all samples which predicted blank now/past 2111 if hidden is not None: 2112 # LSTM has 2 states 2113 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2114 2115 elif len(blank_indices) > 0 and hidden is None: 2116 # Reset state if there were some blank and other non-blank predictions in batch 2117 # Original state is filled with zeros so we just multiply 2118 # LSTM has 2 states 2119 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2120 2121 # Recover prior predicted label for all samples which predicted blank now/past 2122 k[blank_indices] = last_label[blank_indices, 0] 2123 2124 # Update new label and hidden state for next iteration 2125 last_label = k.view(-1, 1) 2126 hidden = hidden_prime 2127 2128 # Update predicted labels, accounting for time mask 2129 # If blank was predicted even once, now or in the past, 2130 # Force the current predicted label to also be blank 2131 # This ensures that blanks propogate across all timesteps 2132 # once they have occured (normally stopping condition of sample level loop). 2133 for kidx, ki in enumerate(k): 2134 if blank_mask[kidx] == 0: 2135 hypotheses[kidx].y_sequence.append(ki) 2136 hypotheses[kidx].timestep.append(time_idx) 2137 hypotheses[kidx].score += float(v[kidx]) 2138 2139 symbols_added += 1 2140 2141 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 2142 # Then preserve U at current timestep Ti 2143 # Finally, forward the timestep history to Ti+1 for that sample 2144 # All of this should only be done iff the current time index <= sample-level AM length. 2145 # Otherwise ignore and move to next sample / next timestep. 2146 if self.preserve_alignments: 2147 2148 # convert Ti-th logits into a torch array 2149 for batch_idx in range(batchsize): 2150 2151 # this checks if current timestep <= sample-level AM length 2152 # If current timestep > sample-level AM length, no alignments will be added 2153 # Therefore the list of Uj alignments is empty here. 2154 if len(hypotheses[batch_idx].alignments[-1]) > 0: 2155 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 2156 2157 # Do the same if preserving per-frame confidence 2158 if self.preserve_frame_confidence: 2159 2160 for batch_idx in range(batchsize): 2161 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 2162 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 2163 2164 # Remove trailing empty list of alignments at T_{am-len} x Uj 2165 if self.preserve_alignments: 2166 for batch_idx in range(batchsize): 2167 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2168 del hypotheses[batch_idx].alignments[-1] 2169 2170 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2171 if self.preserve_frame_confidence: 2172 for batch_idx in range(batchsize): 2173 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2174 del hypotheses[batch_idx].frame_confidence[-1] 2175 2176 # Preserve states 2177 for batch_idx in range(batchsize): 2178 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2179 2180 return hypotheses 2181 2182 2183 @dataclass 2184 class GreedyRNNTInferConfig: 2185 max_symbols_per_step: Optional[int] = 10 2186 preserve_alignments: bool = False 2187 preserve_frame_confidence: bool = False 2188 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 2189 confidence_method_cfg: str = "DEPRECATED" 2190 2191 def __post_init__(self): 2192 # OmegaConf.structured ensures that post_init check is always executed 2193 self.confidence_measure_cfg = OmegaConf.structured( 2194 self.confidence_measure_cfg 2195 if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) 2196 else ConfidenceMeasureConfig(**self.confidence_measure_cfg) 2197 ) 2198 if self.confidence_method_cfg != "DEPRECATED": 2199 logging.warning( 2200 "`confidence_method_cfg` is deprecated and will be removed in the future. " 2201 "Please use `confidence_measure_cfg` instead." 2202 ) 2203 2204 # TODO (alaptev): delete the following two lines sometime in the future 2205 logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") 2206 # OmegaConf.structured ensures that post_init check is always executed 2207 self.confidence_measure_cfg = OmegaConf.structured( 2208 self.confidence_method_cfg 2209 if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) 2210 else ConfidenceMeasureConfig(**self.confidence_method_cfg) 2211 ) 2212 self.confidence_method_cfg = "DEPRECATED" 2213 2214 2215 @dataclass 2216 class GreedyBatchedRNNTInferConfig: 2217 max_symbols_per_step: Optional[int] = 10 2218 preserve_alignments: bool = False 2219 preserve_frame_confidence: bool = False 2220 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 2221 confidence_method_cfg: str = "DEPRECATED" 2222 2223 def __post_init__(self): 2224 # OmegaConf.structured ensures that post_init check is always executed 2225 self.confidence_measure_cfg = OmegaConf.structured( 2226 self.confidence_measure_cfg 2227 if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) 2228 else ConfidenceMeasureConfig(**self.confidence_measure_cfg) 2229 ) 2230 if self.confidence_method_cfg != "DEPRECATED": 2231 logging.warning( 2232 "`confidence_method_cfg` is deprecated and will be removed in the future. " 2233 "Please use `confidence_measure_cfg` instead." 2234 ) 2235 2236 # TODO (alaptev): delete the following two lines sometime in the future 2237 logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") 2238 # OmegaConf.structured ensures that post_init check is always executed 2239 self.confidence_measure_cfg = OmegaConf.structured( 2240 self.confidence_method_cfg 2241 if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) 2242 else ConfidenceMeasureConfig(**self.confidence_method_cfg) 2243 ) 2244 self.confidence_method_cfg = "DEPRECATED" 2245 2246 2247 class GreedyTDTInfer(_GreedyRNNTInfer): 2248 """A greedy TDT decoder. 2249 2250 Sequence level greedy decoding, performed auto-regressively. 2251 2252 Args: 2253 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2254 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2255 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2256 durations: a list containing durations for TDT. 2257 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2258 to a sequence in a single time step; if set to None then there is 2259 no limit. 2260 preserve_alignments: Bool flag which preserves the history of alignments generated during 2261 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2262 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2263 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2264 The length of the list corresponds to the Acoustic Length (T). 2265 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2266 U is the number of target tokens for the current timestep Ti. 2267 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2268 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2269 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2270 The length of the list corresponds to the Acoustic Length (T). 2271 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2272 U is the number of target tokens for the current timestep Ti. 2273 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 2274 confidence scores. 2275 2276 name: The measure name (str). 2277 Supported values: 2278 - 'max_prob' for using the maximum token probability as a confidence. 2279 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2280 2281 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 2282 Supported values: 2283 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2284 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2285 Note that for this entropy, the alpha should comply the following inequality: 2286 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2287 where V is the model vocabulary size. 2288 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2289 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2290 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2291 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2292 - 'renyi' for the Rรฉnyi entropy. 2293 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2294 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2295 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2296 2297 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2298 When the alpha equals one, scaling is not applied to 'max_prob', 2299 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2300 2301 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2302 Supported values: 2303 - 'lin' for using the linear mapping. 2304 - 'exp' for using exponential mapping with linear shift. 2305 """ 2306 2307 def __init__( 2308 self, 2309 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2310 joint_model: rnnt_abstract.AbstractRNNTJoint, 2311 blank_index: int, 2312 durations: list, 2313 max_symbols_per_step: Optional[int] = None, 2314 preserve_alignments: bool = False, 2315 preserve_frame_confidence: bool = False, 2316 confidence_measure_cfg: Optional[DictConfig] = None, 2317 ): 2318 super().__init__( 2319 decoder_model=decoder_model, 2320 joint_model=joint_model, 2321 blank_index=blank_index, 2322 max_symbols_per_step=max_symbols_per_step, 2323 preserve_alignments=preserve_alignments, 2324 preserve_frame_confidence=preserve_frame_confidence, 2325 confidence_measure_cfg=confidence_measure_cfg, 2326 ) 2327 self.durations = durations 2328 2329 @typecheck() 2330 def forward( 2331 self, 2332 encoder_output: torch.Tensor, 2333 encoded_lengths: torch.Tensor, 2334 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2335 ): 2336 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2337 Output token is generated auto-regressively. 2338 Args: 2339 encoder_output: A tensor of size (batch, features, timesteps). 2340 encoded_lengths: list of int representing the length of each sequence 2341 output sequence. 2342 Returns: 2343 packed list containing batch number of sentences (Hypotheses). 2344 """ 2345 # Preserve decoder and joint training state 2346 decoder_training_state = self.decoder.training 2347 joint_training_state = self.joint.training 2348 2349 with torch.inference_mode(): 2350 # Apply optional preprocessing 2351 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2352 2353 self.decoder.eval() 2354 self.joint.eval() 2355 2356 hypotheses = [] 2357 # Process each sequence independently 2358 with self.decoder.as_frozen(), self.joint.as_frozen(): 2359 for batch_idx in range(encoder_output.size(0)): 2360 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 2361 logitlen = encoded_lengths[batch_idx] 2362 2363 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 2364 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 2365 hypotheses.append(hypothesis) 2366 2367 # Pack results into Hypotheses 2368 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 2369 2370 self.decoder.train(decoder_training_state) 2371 self.joint.train(joint_training_state) 2372 2373 return (packed_result,) 2374 2375 @torch.no_grad() 2376 def _greedy_decode( 2377 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 2378 ): 2379 # x: [T, 1, D] 2380 # out_len: [seq_len] 2381 2382 # Initialize blank state and empty label set in Hypothesis 2383 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 2384 2385 if partial_hypotheses is not None: 2386 hypothesis.last_token = partial_hypotheses.last_token 2387 hypothesis.y_sequence = ( 2388 partial_hypotheses.y_sequence.cpu().tolist() 2389 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 2390 else partial_hypotheses.y_sequence 2391 ) 2392 if partial_hypotheses.dec_state is not None: 2393 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 2394 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 2395 2396 if self.preserve_alignments: 2397 # Alignments is a 2-dimensional dangling list representing T x U 2398 hypothesis.alignments = [[]] 2399 2400 if self.preserve_frame_confidence: 2401 hypothesis.frame_confidence = [[]] 2402 2403 time_idx = 0 2404 while time_idx < out_len: 2405 # Extract encoder embedding at timestep t 2406 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 2407 f = x.narrow(dim=0, start=time_idx, length=1) 2408 2409 # Setup exit flags and counter 2410 not_blank = True 2411 symbols_added = 0 2412 2413 need_loop = True 2414 # While blank is not predicted, or we dont run out of max symbols per timestep 2415 while need_loop and (self.max_symbols is None or symbols_added < self.max_symbols): 2416 # In the first timestep, we initialize the network with RNNT Blank 2417 # In later timesteps, we provide previous predicted label as input. 2418 if hypothesis.last_token is None and hypothesis.dec_state is None: 2419 last_label = self._SOS 2420 else: 2421 last_label = label_collate([[hypothesis.last_token]]) 2422 2423 # Perform prediction network and joint network steps. 2424 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 2425 # If preserving per-frame confidence, log_normalize must be true 2426 logits = self._joint_step(f, g, log_normalize=False) 2427 logp = logits[0, 0, 0, : -len(self.durations)] 2428 if self.preserve_frame_confidence: 2429 logp = torch.log_softmax(logp, -1) 2430 2431 duration_logp = torch.log_softmax(logits[0, 0, 0, -len(self.durations) :], dim=-1) 2432 del g 2433 2434 # torch.max(0) op doesnt exist for FP 16. 2435 if logp.dtype != torch.float32: 2436 logp = logp.float() 2437 2438 # get index k, of max prob 2439 v, k = logp.max(0) 2440 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 2441 2442 d_v, d_k = duration_logp.max(0) 2443 d_k = d_k.item() 2444 2445 skip = self.durations[d_k] 2446 2447 if self.preserve_alignments: 2448 # insert logprobs into last timestep 2449 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 2450 2451 if self.preserve_frame_confidence: 2452 # insert confidence into last timestep 2453 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 2454 2455 del logp 2456 2457 # If blank token is predicted, exit inner loop, move onto next timestep t 2458 if k == self._blank_index: 2459 not_blank = False 2460 else: 2461 # Append token to label set, update RNN state. 2462 hypothesis.y_sequence.append(k) 2463 hypothesis.score += float(v) 2464 hypothesis.timestep.append(time_idx) 2465 hypothesis.dec_state = hidden_prime 2466 hypothesis.last_token = k 2467 2468 # Increment token counter. 2469 symbols_added += 1 2470 time_idx += skip 2471 need_loop = skip == 0 2472 2473 # this rarely happens, but we manually increment the `skip` number 2474 # if blank is emitted and duration=0 is predicted. This prevents possible 2475 # infinite loops. 2476 if skip == 0: 2477 skip = 1 2478 2479 if self.preserve_alignments: 2480 # convert Ti-th logits into a torch array 2481 hypothesis.alignments.append([]) # blank buffer for next timestep 2482 2483 if self.preserve_frame_confidence: 2484 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 2485 2486 if symbols_added == self.max_symbols: 2487 time_idx += 1 2488 2489 # Remove trailing empty list of Alignments 2490 if self.preserve_alignments: 2491 if len(hypothesis.alignments[-1]) == 0: 2492 del hypothesis.alignments[-1] 2493 2494 # Remove trailing empty list of per-frame confidence 2495 if self.preserve_frame_confidence: 2496 if len(hypothesis.frame_confidence[-1]) == 0: 2497 del hypothesis.frame_confidence[-1] 2498 2499 # Unpack the hidden states 2500 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 2501 2502 return hypothesis 2503 2504 2505 class GreedyBatchedTDTInfer(_GreedyRNNTInfer): 2506 """A batch level greedy TDT decoder. 2507 Batch level greedy decoding, performed auto-regressively. 2508 Args: 2509 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2510 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2511 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2512 durations: a list containing durations. 2513 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2514 to a sequence in a single time step; if set to None then there is 2515 no limit. 2516 preserve_alignments: Bool flag which preserves the history of alignments generated during 2517 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2518 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2519 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2520 The length of the list corresponds to the Acoustic Length (T). 2521 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2522 U is the number of target tokens for the current timestep Ti. 2523 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2524 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2525 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2526 The length of the list corresponds to the Acoustic Length (T). 2527 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2528 U is the number of target tokens for the current timestep Ti. 2529 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 2530 confidence scores. 2531 2532 name: The measure name (str). 2533 Supported values: 2534 - 'max_prob' for using the maximum token probability as a confidence. 2535 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2536 2537 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 2538 Supported values: 2539 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2540 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2541 Note that for this entropy, the alpha should comply the following inequality: 2542 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2543 where V is the model vocabulary size. 2544 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2545 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2546 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2547 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2548 - 'renyi' for the Rรฉnyi entropy. 2549 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2550 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2551 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2552 2553 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2554 When the alpha equals one, scaling is not applied to 'max_prob', 2555 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2556 2557 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2558 Supported values: 2559 - 'lin' for using the linear mapping. 2560 - 'exp' for using exponential mapping with linear shift. 2561 """ 2562 2563 def __init__( 2564 self, 2565 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2566 joint_model: rnnt_abstract.AbstractRNNTJoint, 2567 blank_index: int, 2568 durations: List[int], 2569 max_symbols_per_step: Optional[int] = None, 2570 preserve_alignments: bool = False, 2571 preserve_frame_confidence: bool = False, 2572 confidence_measure_cfg: Optional[DictConfig] = None, 2573 ): 2574 super().__init__( 2575 decoder_model=decoder_model, 2576 joint_model=joint_model, 2577 blank_index=blank_index, 2578 max_symbols_per_step=max_symbols_per_step, 2579 preserve_alignments=preserve_alignments, 2580 preserve_frame_confidence=preserve_frame_confidence, 2581 confidence_measure_cfg=confidence_measure_cfg, 2582 ) 2583 self.durations = durations 2584 2585 # Depending on availability of `blank_as_pad` support 2586 # switch between more efficient batch decoding technique 2587 if self.decoder.blank_as_pad: 2588 self._greedy_decode = self._greedy_decode_blank_as_pad 2589 else: 2590 self._greedy_decode = self._greedy_decode_masked 2591 2592 @typecheck() 2593 def forward( 2594 self, 2595 encoder_output: torch.Tensor, 2596 encoded_lengths: torch.Tensor, 2597 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2598 ): 2599 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2600 Output token is generated auto-regressively. 2601 Args: 2602 encoder_output: A tensor of size (batch, features, timesteps). 2603 encoded_lengths: list of int representing the length of each sequence 2604 output sequence. 2605 Returns: 2606 packed list containing batch number of sentences (Hypotheses). 2607 """ 2608 # Preserve decoder and joint training state 2609 decoder_training_state = self.decoder.training 2610 joint_training_state = self.joint.training 2611 2612 with torch.inference_mode(): 2613 # Apply optional preprocessing 2614 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2615 logitlen = encoded_lengths 2616 2617 self.decoder.eval() 2618 self.joint.eval() 2619 2620 with self.decoder.as_frozen(), self.joint.as_frozen(): 2621 inseq = encoder_output # [B, T, D] 2622 hypotheses = self._greedy_decode( 2623 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 2624 ) 2625 2626 # Pack the hypotheses results 2627 packed_result = pack_hypotheses(hypotheses, logitlen) 2628 2629 self.decoder.train(decoder_training_state) 2630 self.joint.train(joint_training_state) 2631 2632 return (packed_result,) 2633 2634 def _greedy_decode_blank_as_pad( 2635 self, 2636 x: torch.Tensor, 2637 out_len: torch.Tensor, 2638 device: torch.device, 2639 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2640 ): 2641 if partial_hypotheses is not None: 2642 raise NotImplementedError("`partial_hypotheses` support is not supported") 2643 2644 with torch.inference_mode(): 2645 # x: [B, T, D] 2646 # out_len: [B] 2647 # device: torch.device 2648 2649 # Initialize list of Hypothesis 2650 batchsize = x.shape[0] 2651 hypotheses = [ 2652 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 2653 ] 2654 2655 # Initialize Hidden state matrix (shared by entire batch) 2656 hidden = None 2657 2658 # If alignments need to be preserved, register a danling list to hold the values 2659 if self.preserve_alignments: 2660 # alignments is a 3-dimensional dangling list representing B x T x U 2661 for hyp in hypotheses: 2662 hyp.alignments = [[]] 2663 2664 # If confidence scores need to be preserved, register a danling list to hold the values 2665 if self.preserve_frame_confidence: 2666 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2667 for hyp in hypotheses: 2668 hyp.frame_confidence = [[]] 2669 2670 # Last Label buffer + Last Label without blank buffer 2671 # batch level equivalent of the last_label 2672 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2673 2674 # Mask buffers 2675 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2676 2677 # Get max sequence length 2678 max_out_len = out_len.max() 2679 2680 # skip means the number of frames the next decoding step should "jump" to. When skip == 1 2681 # it means the next decoding step will just use the next input frame. 2682 skip = 1 2683 for time_idx in range(max_out_len): 2684 if skip > 1: # if skip > 1 at the current step, we decrement it and skip the current frame. 2685 skip -= 1 2686 continue 2687 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2688 2689 # need_to_stay is a boolean indicates whether the next decoding step should remain in the same frame. 2690 need_to_stay = True 2691 symbols_added = 0 2692 2693 # Reset blank mask 2694 blank_mask.mul_(False) 2695 2696 # Update blank mask with time mask 2697 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2698 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2699 blank_mask = time_idx >= out_len 2700 2701 # Start inner loop 2702 while need_to_stay and (self.max_symbols is None or symbols_added < self.max_symbols): 2703 # Batch prediction and joint network steps 2704 # If very first prediction step, submit SOS tag (blank) to pred_step. 2705 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2706 if time_idx == 0 and symbols_added == 0 and hidden is None: 2707 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2708 else: 2709 # Perform batch step prediction of decoder, getting new states and scores ("g") 2710 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 2711 2712 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2713 # Note: log_normalize must not be True here since the joiner output is contanetation of both token logits and duration logits, 2714 # and they need to be normalized independently. 2715 joined = self._joint_step(f, g, log_normalize=None) 2716 logp = joined[:, 0, 0, : -len(self.durations)] 2717 duration_logp = joined[:, 0, 0, -len(self.durations) :] 2718 2719 if logp.dtype != torch.float32: 2720 logp = logp.float() 2721 duration_logp = duration_logp.float() 2722 2723 # get the max for both token and duration predictions. 2724 v, k = logp.max(1) 2725 dv, dk = duration_logp.max(1) 2726 2727 # here we set the skip value to be the minimum of all predicted durations, hense the "torch.min(dk)" call there. 2728 # Please refer to Section 5.2 of our paper https://arxiv.org/pdf/2304.06795.pdf for explanation of this. 2729 skip = self.durations[int(torch.min(dk))] 2730 2731 # this is a special case: if all batches emit blanks, we require that skip be at least 1 2732 # so we don't loop forever at the current frame. 2733 if blank_mask.all(): 2734 if skip == 0: 2735 skip = 1 2736 2737 need_to_stay = skip == 0 2738 del g 2739 2740 # Update blank mask with current predicted blanks 2741 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2742 k_is_blank = k == self._blank_index 2743 blank_mask.bitwise_or_(k_is_blank) 2744 2745 del k_is_blank 2746 del logp, duration_logp 2747 2748 # If all samples predict / have predicted prior blanks, exit loop early 2749 # This is equivalent to if single sample predicted k 2750 if not blank_mask.all(): 2751 # Collect batch indices where blanks occurred now/past 2752 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2753 2754 # Recover prior state for all samples which predicted blank now/past 2755 if hidden is not None: 2756 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2757 2758 elif len(blank_indices) > 0 and hidden is None: 2759 # Reset state if there were some blank and other non-blank predictions in batch 2760 # Original state is filled with zeros so we just multiply 2761 # LSTM has 2 states 2762 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2763 2764 # Recover prior predicted label for all samples which predicted blank now/past 2765 k[blank_indices] = last_label[blank_indices, 0] 2766 2767 # Update new label and hidden state for next iteration 2768 last_label = k.clone().view(-1, 1) 2769 hidden = hidden_prime 2770 2771 # Update predicted labels, accounting for time mask 2772 # If blank was predicted even once, now or in the past, 2773 # Force the current predicted label to also be blank 2774 # This ensures that blanks propogate across all timesteps 2775 # once they have occured (normally stopping condition of sample level loop). 2776 for kidx, ki in enumerate(k): 2777 if blank_mask[kidx] == 0: 2778 hypotheses[kidx].y_sequence.append(ki) 2779 hypotheses[kidx].timestep.append(time_idx) 2780 hypotheses[kidx].score += float(v[kidx]) 2781 2782 symbols_added += 1 2783 2784 # Remove trailing empty list of alignments at T_{am-len} x Uj 2785 if self.preserve_alignments: 2786 for batch_idx in range(batchsize): 2787 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2788 del hypotheses[batch_idx].alignments[-1] 2789 2790 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2791 if self.preserve_frame_confidence: 2792 for batch_idx in range(batchsize): 2793 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2794 del hypotheses[batch_idx].frame_confidence[-1] 2795 2796 # Preserve states 2797 for batch_idx in range(batchsize): 2798 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2799 2800 return hypotheses 2801 2802 def _greedy_decode_masked( 2803 self, 2804 x: torch.Tensor, 2805 out_len: torch.Tensor, 2806 device: torch.device, 2807 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2808 ): 2809 raise NotImplementedError("masked greedy-batched decode is not supported for TDT models.") 2810 [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from abc import ABC, abstractmethod 17 from dataclasses import dataclass 18 from functools import partial 19 from typing import List, Optional 20 21 import torch 22 from omegaconf import DictConfig, OmegaConf 23 24 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 25 from nemo.utils import logging 26 27 28 class ConfidenceMeasureConstants: 29 NAMES = ("max_prob", "entropy") 30 ENTROPY_TYPES = ("gibbs", "tsallis", "renyi") 31 ENTROPY_NORMS = ("lin", "exp") 32 33 @classmethod 34 def print(cls): 35 return ( 36 cls.__name__ 37 + ": " 38 + str({"NAMES": cls.NAMES, "ENTROPY_TYPES": cls.ENTROPY_TYPES, "ENTROPY_NORMS": cls.ENTROPY_NORMS}) 39 ) 40 41 42 class ConfidenceConstants: 43 AGGREGATIONS = ("mean", "min", "max", "prod") 44 45 @classmethod 46 def print(cls): 47 return cls.__name__ + ": " + str({"AGGREGATIONS": cls.AGGREGATIONS}) 48 49 50 @dataclass 51 class ConfidenceMeasureConfig: 52 """A Config which contains the measure name and settings to compute per-frame confidence scores. 53 54 Args: 55 name: The measure name (str). 56 Supported values: 57 - 'max_prob' for using the maximum token probability as a confidence. 58 - 'entropy' for using a normalized entropy of a log-likelihood vector. 59 60 entropy_type: Which type of entropy to use (str). 61 Used if confidence_measure_cfg.name is set to `entropy`. 62 Supported values: 63 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 64 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 65 Note that for this entropy, the alpha should comply the following inequality: 66 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 67 where V is the model vocabulary size. 68 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 69 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 70 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 71 More: https://en.wikipedia.org/wiki/Tsallis_entropy 72 - 'renyi' for the Rรฉnyi entropy. 73 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 74 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 75 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 76 77 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 78 When the alpha equals one, scaling is not applied to 'max_prob', 79 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 80 81 entropy_norm: A mapping of the entropy value to the interval [0,1]. 82 Supported values: 83 - 'lin' for using the linear mapping. 84 - 'exp' for using exponential mapping with linear shift. 85 """ 86 87 name: str = "entropy" 88 entropy_type: str = "tsallis" 89 alpha: float = 0.33 90 entropy_norm: str = "exp" 91 temperature: str = "DEPRECATED" 92 93 def __post_init__(self): 94 if self.temperature != "DEPRECATED": 95 logging.warning( 96 "`temperature` is deprecated and will be removed in the future. Please use `alpha` instead." 97 ) 98 99 # TODO (alaptev): delete the following two lines sometime in the future 100 logging.warning("Re-writing `alpha` with the value of `temperature`.") 101 # self.temperature has type str 102 self.alpha = float(self.temperature) 103 self.temperature = "DEPRECATED" 104 if self.name not in ConfidenceMeasureConstants.NAMES: 105 raise ValueError( 106 f"`name` must be one of the following: " 107 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.NAMES) + '`'}. Provided: `{self.name}`" 108 ) 109 if self.entropy_type not in ConfidenceMeasureConstants.ENTROPY_TYPES: 110 raise ValueError( 111 f"`entropy_type` must be one of the following: " 112 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.ENTROPY_TYPES) + '`'}. Provided: `{self.entropy_type}`" 113 ) 114 if self.alpha <= 0.0: 115 raise ValueError(f"`alpha` must be > 0. Provided: {self.alpha}") 116 if self.entropy_norm not in ConfidenceMeasureConstants.ENTROPY_NORMS: 117 raise ValueError( 118 f"`entropy_norm` must be one of the following: " 119 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.ENTROPY_NORMS) + '`'}. Provided: `{self.entropy_norm}`" 120 ) 121 122 123 @dataclass 124 class ConfidenceConfig: 125 """A config which contains the following key-value pairs related to confidence scores. 126 127 Args: 128 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 129 generated during decoding. When set to true, the Hypothesis will contain 130 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 131 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 132 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 133 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 134 135 The length of the list corresponds to the number of recognized tokens. 136 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 137 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 138 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 139 140 The length of the list corresponds to the number of recognized words. 141 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 142 from the `token_confidence`. 143 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 144 Valid options are `mean`, `min`, `max`, `prod`. 145 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 146 confidence scores. 147 148 name: The measure name (str). 149 Supported values: 150 - 'max_prob' for using the maximum token probability as a confidence. 151 - 'entropy' for using a normalized entropy of a log-likelihood vector. 152 153 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 154 Supported values: 155 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 156 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 157 Note that for this entropy, the alpha should comply the following inequality: 158 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 159 where V is the model vocabulary size. 160 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 161 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 162 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 163 More: https://en.wikipedia.org/wiki/Tsallis_entropy 164 - 'renyi' for the Rรฉnyi entropy. 165 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 166 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 167 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 168 169 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 170 When the alpha equals one, scaling is not applied to 'max_prob', 171 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 172 173 entropy_norm: A mapping of the entropy value to the interval [0,1]. 174 Supported values: 175 - 'lin' for using the linear mapping. 176 - 'exp' for using exponential mapping with linear shift. 177 """ 178 179 preserve_frame_confidence: bool = False 180 preserve_token_confidence: bool = False 181 preserve_word_confidence: bool = False 182 exclude_blank: bool = True 183 aggregation: str = "min" 184 measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() 185 method_cfg: str = "DEPRECATED" 186 187 def __post_init__(self): 188 # OmegaConf.structured ensures that post_init check is always executed 189 self.measure_cfg = OmegaConf.structured( 190 self.measure_cfg 191 if isinstance(self.measure_cfg, ConfidenceMeasureConfig) 192 else ConfidenceMeasureConfig(**self.measure_cfg) 193 ) 194 if self.method_cfg != "DEPRECATED": 195 logging.warning( 196 "`method_cfg` is deprecated and will be removed in the future. Please use `measure_cfg` instead." 197 ) 198 199 # TODO (alaptev): delete the following two lines sometime in the future 200 logging.warning("Re-writing `measure_cfg` with the value of `method_cfg`.") 201 # OmegaConf.structured ensures that post_init check is always executed 202 self.measure_cfg = OmegaConf.structured( 203 self.method_cfg 204 if isinstance(self.method_cfg, ConfidenceMeasureConfig) 205 else ConfidenceMeasureConfig(**self.method_cfg) 206 ) 207 self.method_cfg = "DEPRECATED" 208 if self.aggregation not in ConfidenceConstants.AGGREGATIONS: 209 raise ValueError( 210 f"`aggregation` has to be one of the following: " 211 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.AGGREGATIONS) + '`'}. Provided: `{self.aggregation}`" 212 ) 213 214 215 def get_confidence_measure_bank(): 216 """Generate a dictionary with confidence measure functionals. 217 218 Supported confidence measures: 219 max_prob: normalized maximum probability 220 entropy_gibbs_lin: Gibbs entropy with linear normalization 221 entropy_gibbs_exp: Gibbs entropy with exponential normalization 222 entropy_tsallis_lin: Tsallis entropy with linear normalization 223 entropy_tsallis_exp: Tsallis entropy with exponential normalization 224 entropy_renyi_lin: Rรฉnyi entropy with linear normalization 225 entropy_renyi_exp: Rรฉnyi entropy with exponential normalization 226 227 Returns: 228 dictionary with lambda functions. 229 """ 230 # helper functions 231 # Gibbs entropy is implemented without alpha 232 neg_entropy_gibbs = lambda x: (x.exp() * x).sum(-1) 233 neg_entropy_alpha = lambda x, t: (x * t).exp().sum(-1) 234 neg_entropy_alpha_gibbs = lambda x, t: ((x * t).exp() * x).sum(-1) 235 # too big for a lambda 236 def entropy_tsallis_exp(x, v, t): 237 exp_neg_max_ent = math.exp((1 - math.pow(v, 1 - t)) / (1 - t)) 238 return (((1 - neg_entropy_alpha(x, t)) / (1 - t)).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 239 240 def entropy_gibbs_exp(x, v, t): 241 exp_neg_max_ent = math.pow(v, -t * math.pow(v, 1 - t)) 242 return ((neg_entropy_alpha_gibbs(x, t) * t).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 243 244 # use Gibbs entropies for Tsallis and Rรฉnyi with t == 1.0 245 entropy_gibbs_lin_baseline = lambda x, v: 1 + neg_entropy_gibbs(x) / math.log(v) 246 entropy_gibbs_exp_baseline = lambda x, v: (neg_entropy_gibbs(x).exp() * v - 1) / (v - 1) 247 # fill the measure bank 248 confidence_measure_bank = {} 249 # Maximum probability measure is implemented without alpha 250 confidence_measure_bank["max_prob"] = ( 251 lambda x, v, t: (x.max(dim=-1)[0].exp() * v - 1) / (v - 1) 252 if t == 1.0 253 else ((x.max(dim=-1)[0] * t).exp() * math.pow(v, t) - 1) / (math.pow(v, t) - 1) 254 ) 255 confidence_measure_bank["entropy_gibbs_lin"] = ( 256 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 257 if t == 1.0 258 else 1 + neg_entropy_alpha_gibbs(x, t) / math.log(v) / math.pow(v, 1 - t) 259 ) 260 confidence_measure_bank["entropy_gibbs_exp"] = ( 261 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_gibbs_exp(x, v, t) 262 ) 263 confidence_measure_bank["entropy_tsallis_lin"] = ( 264 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 265 if t == 1.0 266 else 1 + (1 - neg_entropy_alpha(x, t)) / (math.pow(v, 1 - t) - 1) 267 ) 268 confidence_measure_bank["entropy_tsallis_exp"] = ( 269 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_tsallis_exp(x, v, t) 270 ) 271 confidence_measure_bank["entropy_renyi_lin"] = ( 272 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 273 if t == 1.0 274 else 1 + neg_entropy_alpha(x, t).log2() / (t - 1) / math.log(v, 2) 275 ) 276 confidence_measure_bank["entropy_renyi_exp"] = ( 277 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) 278 if t == 1.0 279 else (neg_entropy_alpha(x, t).pow(1 / (t - 1)) * v - 1) / (v - 1) 280 ) 281 return confidence_measure_bank 282 283 284 def get_confidence_aggregation_bank(): 285 """Generate a dictionary with confidence aggregation functions. 286 287 Supported confidence measures: 288 min: minimum 289 max: maximum 290 mean: arithmetic mean 291 prod: product 292 293 Returns: 294 dictionary with functions. 295 """ 296 confidence_aggregation_bank = {"mean": lambda x: sum(x) / len(x), "min": min, "max": max} 297 # python 3.7 and earlier do not have math.prod 298 if hasattr(math, "prod"): 299 confidence_aggregation_bank["prod"] = math.prod 300 else: 301 import operator 302 from functools import reduce 303 304 confidence_aggregation_bank["prod"] = lambda x: reduce(operator.mul, x, 1) 305 return confidence_aggregation_bank 306 307 308 class ConfidenceMeasureMixin(ABC): 309 """Confidence Measure Mixin class. 310 311 It initializes per-frame confidence measure. 312 """ 313 314 def _init_confidence_measure(self, confidence_measure_cfg: Optional[DictConfig] = None): 315 """Initialize per-frame confidence measure from config. 316 """ 317 # OmegaConf.structured ensures that post_init check is always executed 318 confidence_measure_cfg = OmegaConf.structured( 319 ConfidenceMeasureConfig() 320 if confidence_measure_cfg is None 321 else ConfidenceMeasureConfig(**confidence_measure_cfg) 322 ) 323 324 # set confidence calculation measure 325 # we suppose that self.blank_id == len(vocabulary) 326 self.num_tokens = (self.blank_id if hasattr(self, "blank_id") else self._blank_index) + 1 327 self.alpha = confidence_measure_cfg.alpha 328 329 # init confidence measure bank 330 self.confidence_measure_bank = get_confidence_measure_bank() 331 332 measure = None 333 # construct measure_name 334 measure_name = "" 335 if confidence_measure_cfg.name == "max_prob": 336 measure_name = "max_prob" 337 elif confidence_measure_cfg.name == "entropy": 338 measure_name = '_'.join( 339 [confidence_measure_cfg.name, confidence_measure_cfg.entropy_type, confidence_measure_cfg.entropy_norm] 340 ) 341 else: 342 raise ValueError(f"Unsupported `confidence_measure_cfg.name`: `{confidence_measure_cfg.name}`") 343 if measure_name not in self.confidence_measure_bank: 344 raise ValueError(f"Unsupported measure setup: `{measure_name}`") 345 measure = partial(self.confidence_measure_bank[measure_name], v=self.num_tokens, t=self.alpha) 346 self._get_confidence = lambda x: measure(torch.nan_to_num(x)).tolist() 347 348 349 class ConfidenceMixin(ABC): 350 """Confidence Mixin class. 351 352 It is responsible for confidence estimation method initialization and high-level confidence score calculation. 353 """ 354 355 def _init_confidence(self, confidence_cfg: Optional[DictConfig] = None): 356 """Initialize confidence-related fields and confidence aggregation function from config. 357 """ 358 # OmegaConf.structured ensures that post_init check is always executed 359 confidence_cfg = OmegaConf.structured( 360 ConfidenceConfig() if confidence_cfg is None else ConfidenceConfig(**confidence_cfg) 361 ) 362 self.confidence_measure_cfg = confidence_cfg.measure_cfg 363 364 # extract the config 365 self.preserve_word_confidence = confidence_cfg.get('preserve_word_confidence', False) 366 # set preserve_frame_confidence and preserve_token_confidence to True 367 # if preserve_word_confidence is True 368 self.preserve_token_confidence = ( 369 confidence_cfg.get('preserve_token_confidence', False) | self.preserve_word_confidence 370 ) 371 # set preserve_frame_confidence to True if preserve_token_confidence is True 372 self.preserve_frame_confidence = ( 373 confidence_cfg.get('preserve_frame_confidence', False) | self.preserve_token_confidence 374 ) 375 self.exclude_blank_from_confidence = confidence_cfg.get('exclude_blank', True) 376 self.word_confidence_aggregation = confidence_cfg.get('aggregation', "min") 377 378 # define aggregation functions 379 self.confidence_aggregation_bank = get_confidence_aggregation_bank() 380 self._aggregate_confidence = self.confidence_aggregation_bank[self.word_confidence_aggregation] 381 382 # Update preserve frame confidence 383 if self.preserve_frame_confidence is False: 384 if self.cfg.strategy in ['greedy', 'greedy_batch']: 385 self.preserve_frame_confidence = self.cfg.greedy.get('preserve_frame_confidence', False) 386 # OmegaConf.structured ensures that post_init check is always executed 387 confidence_measure_cfg = OmegaConf.structured(self.cfg.greedy).get('confidence_measure_cfg', None) 388 self.confidence_measure_cfg = ( 389 OmegaConf.structured(ConfidenceMeasureConfig()) 390 if confidence_measure_cfg is None 391 else OmegaConf.structured(ConfidenceMeasureConfig(**confidence_measure_cfg)) 392 ) 393 394 @abstractmethod 395 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 396 """Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 397 Assumes that `frame_confidence` is present in the hypotheses. 398 399 Args: 400 hypotheses_list: List of Hypothesis. 401 402 Returns: 403 A list of hypotheses with high-level confidence scores. 404 """ 405 raise NotImplementedError() 406 407 @abstractmethod 408 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 409 """Implemented by subclass in order to aggregate token confidence to a word-level confidence. 410 411 Args: 412 hypothesis: Hypothesis 413 414 Returns: 415 A list of word-level confidence scores. 416 """ 417 raise NotImplementedError() 418 419 def _aggregate_token_confidence_chars(self, words: List[str], token_confidence: List[float]) -> List[float]: 420 """Implementation of token confidence aggregation for character-based models. 421 422 Args: 423 words: List of words of a hypothesis. 424 token_confidence: List of token-level confidence scores of a hypothesis. 425 426 Returns: 427 A list of word-level confidence scores. 428 """ 429 word_confidence = [] 430 i = 0 431 for word in words: 432 word_len = len(word) 433 word_confidence.append(self._aggregate_confidence(token_confidence[i : i + word_len])) 434 # we assume that there is exactly one space token between words and exclude it from word confidence 435 i += word_len + 1 436 return word_confidence 437 438 def _aggregate_token_confidence_subwords_sentencepiece( 439 self, words: List[str], token_confidence: List[float], token_ids: List[int] 440 ) -> List[float]: 441 """Implementation of token confidence aggregation for subword-based models. 442 443 **Note**: Only supports Sentencepiece based tokenizers ! 444 445 Args: 446 words: List of words of a hypothesis. 447 token_confidence: List of token-level confidence scores of a hypothesis. 448 token_ids: List of token ids of a hypothesis. 449 450 Returns: 451 A list of word-level confidence scores. 452 """ 453 word_confidence = [] 454 # run only if there are final words 455 if len(words) > 0: 456 j = 0 457 prev_unk = False 458 prev_underline = False 459 for i, token_id in enumerate(token_ids): 460 token = self.decode_ids_to_tokens([int(token_id)])[0] 461 token_text = self.decode_tokens_to_str([int(token_id)]) 462 # treat `<unk>` as a separate word regardless of the next token 463 # to match the result of `tokenizer.ids_to_text` 464 if (token != token_text or prev_unk) and i > j: 465 # do not add confidence for `โ–` if the current token starts with `โ–` 466 # to match the result of `tokenizer.ids_to_text` 467 if not prev_underline: 468 word_confidence.append(self._aggregate_confidence(token_confidence[j:i])) 469 j = i 470 prev_unk = token == '<unk>' 471 prev_underline = token == 'โ–' 472 if not prev_underline: 473 word_confidence.append(self._aggregate_confidence(token_confidence[j : len(token_ids)])) 474 if len(words) != len(word_confidence): 475 raise RuntimeError( 476 f"""Something went wrong with word-level confidence aggregation.\n 477 Please check these values for debugging:\n 478 len(words): {len(words)},\n 479 len(word_confidence): {len(word_confidence)},\n 480 recognized text: `{' '.join(words)}`""" 481 ) 482 return word_confidence 483 [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of nemo/collections/common/parts/adapter_modules.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, is_dataclass 16 from typing import Any, Optional 17 18 from hydra.utils import instantiate 19 from omegaconf import OmegaConf 20 from torch import nn as nn 21 22 from nemo.collections.common.parts.utils import activation_registry 23 from nemo.core.classes.mixins import access_mixins, adapter_mixin_strategies 24 25 26 class AdapterModuleUtil(access_mixins.AccessMixin): 27 """ 28 Base class of Adapter Modules, providing common functionality to all Adapter Modules. 29 """ 30 31 def setup_adapter_strategy(self, adapter_strategy: Optional[adapter_mixin_strategies.AbstractAdapterStrategy]): 32 """ 33 Setup adapter strategy of this class, enabling dynamic change in the way the adapter output is 34 merged with the input. 35 36 When called successfully, will assign the variable `adapter_strategy` to the module. 37 38 Args: 39 adapter_strategy: Can be a None or an implementation of AbstractAdapterStrategy. 40 """ 41 # set default adapter strategy 42 if adapter_strategy is None: 43 adapter_strategy = self.get_default_strategy_config() 44 45 if is_dataclass(adapter_strategy): 46 adapter_strategy = OmegaConf.structured(adapter_strategy) 47 OmegaConf.set_struct(adapter_strategy, False) 48 49 # The config must have the `_target_` field pointing to the actual adapter strategy class 50 # which will load that strategy dynamically to this module. 51 if isinstance(adapter_strategy, dict) or OmegaConf.is_config(adapter_strategy): 52 self.adapter_strategy = instantiate(adapter_strategy) 53 elif isinstance(adapter_strategy, adapter_mixin_strategies.AbstractAdapterStrategy): 54 self.adapter_strategy = adapter_strategy 55 else: 56 raise AttributeError(f'`adapter_strategy` provided is invalid : {adapter_strategy}') 57 58 def get_default_strategy_config(self) -> 'dataclass': 59 """ 60 Returns a default adapter module strategy. 61 """ 62 return adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 63 64 def adapter_unfreeze(self,): 65 """ 66 Sets the requires grad for all parameters in the adapter to True. 67 This method should be overridden for any custom unfreeze behavior that is required. 68 For example, if not all params of the adapter should be unfrozen. 69 """ 70 for param in self.parameters(): 71 param.requires_grad_(True) 72 73 74 class LinearAdapter(nn.Module, AdapterModuleUtil): 75 76 """ 77 Simple Linear Feedforward Adapter module with LayerNorm and singe hidden layer with activation function. 78 Note: The adapter explicitly initializes its final layer with all zeros in order to avoid affecting the 79 original model when all adapters are disabled. 80 81 Args: 82 in_features: Input dimension of the module. Note that for adapters, input_dim == output_dim. 83 dim: Hidden dimension of the feed forward network. 84 activation: Str name for an activation function. 85 norm_position: Str, can be `pre` or `post`. Defaults to `pre`. Determines whether the normalization 86 will occur in the first layer or the last layer. Certain architectures may prefer one over the other. 87 dropout: float value, whether to perform dropout on the output of the last layer of the adapter. 88 adapter_strategy: By default, ResidualAddAdapterStrategyConfig. An adapter composition function object. 89 """ 90 91 def __init__( 92 self, 93 in_features: int, 94 dim: int, 95 activation: str = 'swish', 96 norm_position: str = 'pre', 97 dropout: float = 0.0, 98 adapter_strategy: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig = None, 99 ): 100 super().__init__() 101 102 activation = activation_registry[activation]() 103 # If the activation can be executed in place, do so. 104 if hasattr(activation, 'inplace'): 105 activation.inplace = True 106 107 assert norm_position in ['pre', 'post'] 108 self.norm_position = norm_position 109 110 if norm_position == 'pre': 111 self.module = nn.Sequential( 112 nn.LayerNorm(in_features), 113 nn.Linear(in_features, dim, bias=False), 114 activation, 115 nn.Linear(dim, in_features, bias=False), 116 ) 117 118 elif norm_position == 'post': 119 self.module = nn.Sequential( 120 nn.Linear(in_features, dim, bias=False), 121 activation, 122 nn.Linear(dim, in_features, bias=False), 123 nn.LayerNorm(in_features), 124 ) 125 126 if dropout > 0.0: 127 self.dropout = nn.Dropout(dropout) 128 else: 129 self.dropout = None 130 131 # Setup adapter strategy 132 self.setup_adapter_strategy(adapter_strategy) 133 134 # reset parameters 135 self.reset_parameters() 136 137 def reset_parameters(self): 138 # Final layer initializations must be 0 139 if self.norm_position == 'pre': 140 self.module[-1].weight.data *= 0 141 142 elif self.norm_position == 'post': 143 self.module[-1].weight.data *= 0 144 self.module[-1].bias.data *= 0 145 146 def forward(self, x): 147 x = self.module(x) 148 149 # Add dropout if available 150 if self.dropout is not None: 151 x = self.dropout(x) 152 153 return x 154 155 156 @dataclass 157 class LinearAdapterConfig: 158 in_features: int 159 dim: int 160 activation: str = 'swish' 161 norm_position: str = 'pre' 162 dropout: float = 0.0 163 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 164 _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) 165 [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import re 15 from typing import List 16 17 import ipadic 18 import MeCab 19 from pangu import spacing 20 from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer 21 22 23 class EnJaProcessor: 24 """ 25 Tokenizer, Detokenizer and Normalizer utilities for Japanese & English 26 Args: 27 lang_id: One of ['en', 'ja']. 28 """ 29 30 def __init__(self, lang_id: str): 31 self.lang_id = lang_id 32 self.moses_tokenizer = MosesTokenizer(lang=lang_id) 33 self.moses_detokenizer = MosesDetokenizer(lang=lang_id) 34 self.normalizer = MosesPunctNormalizer( 35 lang=lang_id, pre_replace_unicode_punct=True, post_remove_control_chars=True 36 ) 37 38 def detokenize(self, tokens: List[str]) -> str: 39 """ 40 Detokenizes a list of tokens 41 Args: 42 tokens: list of strings as tokens 43 Returns: 44 detokenized Japanese or English string 45 """ 46 return self.moses_detokenizer.detokenize(tokens) 47 48 def tokenize(self, text) -> str: 49 """ 50 Tokenizes text using Moses. Returns a string of tokens. 51 """ 52 tokens = self.moses_tokenizer.tokenize(text) 53 return ' '.join(tokens) 54 55 def normalize(self, text) -> str: 56 # Normalization doesn't handle Japanese periods correctly; 57 # 'ใ€‚'becomes '.'. 58 if self.lang_id == 'en': 59 return self.normalizer.normalize(text) 60 else: 61 return text 62 63 64 class JaMecabProcessor: 65 """ 66 Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English 67 """ 68 69 def __init__(self): 70 self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") 71 72 def detokenize(self, text: List[str]) -> str: 73 RE_WS_IN_FW = re.compile( 74 r'([\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])\s+(?=[\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])' 75 ) 76 77 detokenize = lambda s: spacing(RE_WS_IN_FW.sub(r'\1', s)).strip() 78 return detokenize(' '.join(text)) 79 80 def tokenize(self, text) -> str: 81 """ 82 Tokenizes text using Moses. Returns a string of tokens. 83 """ 84 return self.mecab_tokenizer.parse(text).strip() 85 86 def normalize(self, text) -> str: 87 return text 88 [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Optional, Tuple 17 18 from omegaconf.omegaconf import MISSING 19 20 from nemo.collections.nlp.data.machine_translation.machine_translation_dataset import TranslationDataConfig 21 from nemo.collections.nlp.models.enc_dec_nlp_model import EncDecNLPModelConfig 22 from nemo.collections.nlp.modules.common.token_classifier import TokenClassifierConfig 23 from nemo.collections.nlp.modules.common.tokenizer_utils import TokenizerConfig 24 from nemo.collections.nlp.modules.common.transformer.transformer import ( 25 NeMoTransformerConfig, 26 NeMoTransformerEncoderConfig, 27 ) 28 from nemo.collections.nlp.modules.common.transformer.transformer_bottleneck import ( 29 NeMoTransformerBottleneckDecoderConfig, 30 NeMoTransformerBottleneckEncoderConfig, 31 ) 32 from nemo.core.config.modelPT import OptimConfig, SchedConfig 33 34 35 @dataclass 36 class MTSchedConfig(SchedConfig): 37 name: str = 'InverseSquareRootAnnealing' 38 warmup_ratio: Optional[float] = None 39 last_epoch: int = -1 40 41 42 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). 43 @dataclass 44 class MTOptimConfig(OptimConfig): 45 name: str = 'adam' 46 lr: float = 1e-3 47 betas: Tuple[float, float] = (0.9, 0.98) 48 weight_decay: float = 0.0 49 sched: Optional[MTSchedConfig] = MTSchedConfig() 50 51 52 @dataclass 53 class MTEncDecModelConfig(EncDecNLPModelConfig): 54 # machine translation configurations 55 num_val_examples: int = 3 56 num_test_examples: int = 3 57 max_generation_delta: int = 10 58 label_smoothing: Optional[float] = 0.0 59 beam_size: int = 4 60 len_pen: float = 0.0 61 src_language: Any = 'en' # Any = str or List[str] 62 tgt_language: Any = 'en' # Any = str or List[str] 63 find_unused_parameters: Optional[bool] = True 64 shared_tokenizer: Optional[bool] = True 65 multilingual: Optional[bool] = False 66 preproc_out_dir: Optional[str] = None 67 validate_input_ids: Optional[bool] = True 68 shared_embeddings: bool = False 69 70 # network architecture configuration 71 encoder_tokenizer: Any = MISSING 72 encoder: Any = MISSING 73 74 decoder_tokenizer: Any = MISSING 75 decoder: Any = MISSING 76 77 head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) 78 79 # dataset configurations 80 train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 81 src_file_name=MISSING, 82 tgt_file_name=MISSING, 83 tokens_in_batch=512, 84 clean=True, 85 shuffle=True, 86 cache_ids=False, 87 use_cache=False, 88 ) 89 validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 90 src_file_name=MISSING, 91 tgt_file_name=MISSING, 92 tokens_in_batch=512, 93 clean=False, 94 shuffle=False, 95 cache_ids=False, 96 use_cache=False, 97 ) 98 test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 99 src_file_name=MISSING, 100 tgt_file_name=MISSING, 101 tokens_in_batch=512, 102 clean=False, 103 shuffle=False, 104 cache_ids=False, 105 use_cache=False, 106 ) 107 optim: Optional[OptimConfig] = MTOptimConfig() 108 109 110 @dataclass 111 class AAYNBaseConfig(MTEncDecModelConfig): 112 113 # Attention is All You Need Base Configuration 114 encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 115 decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 116 117 encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( 118 library='nemo', 119 model_name=None, 120 pretrained=False, 121 hidden_size=512, 122 inner_size=2048, 123 num_layers=6, 124 num_attention_heads=8, 125 ffn_dropout=0.1, 126 attn_score_dropout=0.1, 127 attn_layer_dropout=0.1, 128 ) 129 130 decoder: NeMoTransformerConfig = NeMoTransformerConfig( 131 library='nemo', 132 model_name=None, 133 pretrained=False, 134 hidden_size=512, 135 inner_size=2048, 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass 145 class MTBottleneckModelConfig(AAYNBaseConfig): 146 model_type: str = 'nll' 147 min_logv: float = -6 148 latent_size: int = -1 # -1 will take value of encoder hidden 149 non_recon_warmup_batches: int = 200000 150 recon_per_token: bool = True 151 log_timing: bool = True 152 153 encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( 154 library='nemo', 155 model_name=None, 156 pretrained=False, 157 hidden_size=512, 158 inner_size=2048, 159 num_layers=6, 160 num_attention_heads=8, 161 ffn_dropout=0.1, 162 attn_score_dropout=0.1, 163 attn_layer_dropout=0.1, 164 arch='seq2seq', 165 hidden_steps=32, 166 hidden_blocks=1, 167 hidden_init_method='params', 168 ) 169 170 decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( 171 library='nemo', 172 model_name=None, 173 pretrained=False, 174 inner_size=2048, 175 num_layers=6, 176 num_attention_heads=8, 177 ffn_dropout=0.1, 178 attn_score_dropout=0.1, 179 attn_layer_dropout=0.1, 180 arch='seq2seq', 181 ) 182 [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, Optional 17 18 from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict 19 20 from nemo.collections.common.parts.adapter_modules import LinearAdapterConfig 21 from nemo.collections.nlp.data.token_classification.punctuation_capitalization_dataset import ( 22 PunctuationCapitalizationEvalDataConfig, 23 PunctuationCapitalizationTrainDataConfig, 24 legacy_data_config_to_new_data_config, 25 ) 26 from nemo.core.config import TrainerConfig 27 from nemo.core.config.modelPT import NemoConfig 28 from nemo.utils.exp_manager import ExpManagerConfig 29 30 31 @dataclass 32 class FreezeConfig: 33 is_enabled: bool = False 34 """Freeze audio encoder weight and add Conformer Layers on top of it""" 35 d_model: Optional[int] = 256 36 """`d_model` parameter of ``ConformerLayer``""" 37 d_ff: Optional[int] = 1024 38 """``d_ff`` parameter of ``ConformerLayer``""" 39 num_layers: Optional[int] = 8 40 """``num_layers`` number of ``ConformerLayer`` modules to add on top of audio encoder""" 41 42 43 @dataclass 44 class AdapterConfig: 45 config: Optional[LinearAdapterConfig] = None 46 """Linear adapter config see ``collections.common.parts.LinearAdapterConfig``""" 47 enable: bool = False 48 """Use adapters for audio encoder""" 49 50 51 @dataclass 52 class FusionConfig: 53 num_layers: Optional[int] = 4 54 """"Number of layers to use in fusion""" 55 num_attention_heads: Optional[int] = 4 56 """Number of attention heads to use in fusion""" 57 inner_size: Optional[int] = 2048 58 """Fusion inner size""" 59 60 61 @dataclass 62 class AudioEncoderConfig: 63 pretrained_model: str = MISSING 64 """A configuration for restoring pretrained audio encoder""" 65 freeze: Optional[FreezeConfig] = None 66 adapter: Optional[AdapterConfig] = None 67 fusion: Optional[FusionConfig] = None 68 69 70 @dataclass 71 class TokenizerConfig: 72 """A structure and default values of source text tokenizer.""" 73 74 vocab_file: Optional[str] = None 75 """A path to vocabulary file which is used in ``'word'``, ``'char'``, and HuggingFace tokenizers""" 76 77 tokenizer_name: str = MISSING 78 """A name of the tokenizer used for tokenization of source sequences. Possible options are ``'sentencepiece'``, 79 ``'word'``, ``'char'``, HuggingFace tokenizers (e.g. ``'bert-base-uncased'``). For more options see function 80 ``nemo.collections.nlp.modules.common.get_tokenizer``. The tokenizer must have properties ``cls_id``, ``pad_id``, 81 ``sep_id``, ``unk_id``.""" 82 83 special_tokens: Optional[Dict[str, str]] = None 84 """A dictionary with special tokens passed to constructors of ``'char'``, ``'word'``, ``'sentencepiece'``, and 85 various HuggingFace tokenizers.""" 86 87 tokenizer_model: Optional[str] = None 88 """A path to a tokenizer model required for ``'sentencepiece'`` tokenizer.""" 89 90 91 @dataclass 92 class LanguageModelConfig: 93 """ 94 A structure and default values of language model configuration of punctuation and capitalization model. BERT like 95 HuggingFace models are supported. Provide a valid ``pretrained_model_name`` and, optionally, you may 96 reinitialize model via ``config_file`` or ``config``. 97 98 Alternatively you can initialize the language model using ``lm_checkpoint``. 99 100 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 101 """ 102 103 pretrained_model_name: str = MISSING 104 """A mandatory parameter containing name of HuggingFace pretrained model. For example, ``'bert-base-uncased'``.""" 105 106 config_file: Optional[str] = None 107 """A path to a file with HuggingFace model config which is used to reinitialize language model.""" 108 109 config: Optional[Dict] = None 110 """A HuggingFace config which is used to reinitialize language model.""" 111 112 lm_checkpoint: Optional[str] = None 113 """A path to a ``torch`` checkpoint of a language model.""" 114 115 116 @dataclass 117 class HeadConfig: 118 """ 119 A structure and default values of configuration of capitalization or punctuation model head. This config defines a 120 multilayer perceptron which is applied to output of a language model. Number of units in the hidden layer is equal 121 to the dimension of the language model. 122 123 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 124 """ 125 126 num_fc_layers: int = 1 127 """A number of hidden layers in a multilayer perceptron.""" 128 129 fc_dropout: float = 0.1 130 """A dropout used in an MLP.""" 131 132 activation: str = 'relu' 133 """An activation used in hidden layers.""" 134 135 use_transformer_init: bool = True 136 """Whether to initialize the weights of the classifier head with the approach that was used for language model 137 initialization.""" 138 139 140 @dataclass 141 class ClassLabelsConfig: 142 """ 143 A structure and default values of a mandatory part of config which contains names of files which are saved in .nemo 144 checkpoint. These files can also be used for passing label vocabulary to the model. For using them as label 145 vocabularies you will need to provide path these files in parameter 146 ``model.common_dataset_parameters.label_vocab_dir``. Each line in labels files 147 contains 1 label. The values are sorted, ``<line number>==<label id>``, starting from ``0``. A label with ``0`` id 148 must contain neutral label which must be equal to ``model.common_dataset_parameters.pad_label``. 149 150 This config is a part of :class:`~CommonDatasetParametersConfig`. 151 """ 152 153 punct_labels_file: str = MISSING 154 """A name of punctuation labels file.""" 155 156 capit_labels_file: str = MISSING 157 """A name of capitalization labels file.""" 158 159 160 @dataclass 161 class CommonDatasetParametersConfig: 162 """ 163 A structure and default values of common dataset parameters config which includes label and loss mask information. 164 If you omit parameters ``punct_label_ids``, ``capit_label_ids``, ``label_vocab_dir``, then labels will be inferred 165 from a training dataset or loaded from a checkpoint. 166 167 Parameters ``ignore_extra_tokens`` and ``ignore_start_end`` are responsible for forming loss mask. A loss mask 168 defines on which tokens loss is computed. 169 170 This parameter is a part of config :class:`~PunctuationCapitalizationModelConfig`. 171 """ 172 173 pad_label: str = MISSING 174 """A mandatory parameter which should contain label used for punctuation and capitalization label padding. It 175 also serves as a neutral label for both punctuation and capitalization. If any of ``punct_label_ids``, 176 ``capit_label_ids`` parameters is provided, then ``pad_label`` must have ``0`` id in them. In addition, if ``label_vocab_dir`` 177 is provided, then ``pad_label`` must be on the first lines in files ``class_labels.punct_labels_file`` and 178 ``class_labels.capit_labels_file``.""" 179 180 ignore_extra_tokens: bool = False 181 """Whether to compute loss on not first tokens in words. If this parameter is ``True``, then loss mask is ``False`` 182 for all tokens in a word except the first.""" 183 184 ignore_start_end: bool = True 185 """If ``False``, then loss is computed on [CLS] and [SEP] tokens.""" 186 187 punct_label_ids: Optional[Dict[str, int]] = None 188 """A dictionary with punctuation label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this 189 parameter and pass label ids through ``class_labels.punct_labels_file`` or let the model to infer label ids from 190 dataset or load them from checkpoint.""" 191 192 capit_label_ids: Optional[Dict[str, int]] = None 193 """A dictionary with capitalization label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit 194 this parameter and pass label ids through ``class_labels.capit_labels_file`` or let model to infer label ids from 195 dataset or load them from checkpoint.""" 196 197 label_vocab_dir: Optional[str] = None 198 """A path to directory which contains class labels files. See :class:`ClassLabelsConfig`. If this parameter is 199 provided, then labels will be loaded from files which are located in ``label_vocab_dir`` and have names specified 200 in ``model.class_labels`` configuration section. A label specified in ``pad_label`` has to be on the first lines 201 of ``model.class_labels`` files.""" 202 203 204 @dataclass 205 class PunctuationCapitalizationModelConfig: 206 """ 207 A configuration of 208 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 209 model. 210 211 See an example of model config in 212 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 213 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 214 215 This config is a part of :class:`~PunctuationCapitalizationConfig`. 216 """ 217 218 class_labels: ClassLabelsConfig = ClassLabelsConfig() 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" 229 230 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 231 """A configuration for creating validation datasets and data loaders.""" 232 233 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 234 """A configuration for creating test datasets and data loaders.""" 235 236 punct_head: HeadConfig = HeadConfig() 237 """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" 238 239 capit_head: HeadConfig = HeadConfig() 240 """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" 241 242 tokenizer: Any = TokenizerConfig() 243 """A configuration for source text tokenizer.""" 244 245 language_model: LanguageModelConfig = LanguageModelConfig() 246 """A configuration of a BERT-like language model which serves as a model body.""" 247 248 optim: Optional[Any] = None 249 """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For 250 description see `Optimizers 251 <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/core/core.html#optimizers>`_ section in 252 documentation and `primer <https://github.com/NVIDIA/NeMo/blob/main/tutorials/00_NeMo_Primer.ipynb>_ tutorial.""" 253 254 255 @dataclass 256 class PunctuationCapitalizationLexicalAudioModelConfig(PunctuationCapitalizationModelConfig): 257 """ 258 A configuration of 259 :class:`~nemo.collections.nlp.models.token_classification.punctuation_lexical_audio_capitalization_model.PunctuationCapitalizationLexicalAudioModel` 260 model. 261 262 See an example of model config in 263 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 264 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_lexical_audio_config.yaml>`_ 265 266 Audio encoder can be frozen during training with ``freeze_audio_encoder`` parameter. 267 Adapter can be added to audio encoder with ``use_adapters`` and ``adapter_config`` parameters. 268 More conformer layers can be added on top of pretrained audio encoder with ``frozen_conf_d_model``, ``frozen_conf_d_ff`` and ``frozen_conf_num_layers`` parameters. 269 """ 270 271 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 272 """A configuration for creating training dataset and data loader.""" 273 274 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 275 """A configuration for creating validation datasets and data loaders.""" 276 277 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 278 """A configuration for creating test datasets and data loaders.""" 279 280 audio_encoder: Optional[AudioEncoderConfig] = None 281 282 restore_lexical_encoder_from: Optional[str] = None 283 """"Path to .nemo checkpoint to load weights from""" # add more comments 284 285 use_weighted_loss: Optional[bool] = False 286 """If set to ``True`` CrossEntropyLoss will be weighted""" 287 288 289 @dataclass 290 class PunctuationCapitalizationConfig(NemoConfig): 291 """ 292 A config for punctuation model training and testing. 293 294 See an example of full config in 295 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 296 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 297 """ 298 299 pretrained_model: Optional[str] = None 300 """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options 301 by calling method 302 :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. 303 """ 304 305 name: Optional[str] = 'Punctuation_and_Capitalization' 306 """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" 307 308 do_training: bool = True 309 """Whether to perform training of the model.""" 310 311 do_testing: bool = False 312 """Whether ot perform testing of the model.""" 313 314 model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() 315 """A configuration for the 316 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 317 model.""" 318 319 trainer: Optional[TrainerConfig] = TrainerConfig() 320 """Contains ``Trainer`` Lightning class constructor parameters.""" 321 322 exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) 323 """A configuration with various NeMo training options such as output directories, resuming from checkpoint, 324 tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" 325 326 327 @dataclass 328 class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): 329 model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() 330 331 332 def is_legacy_model_config(model_cfg: DictConfig) -> bool: 333 """ 334 Test if model config is old style config. Old style configs are configs which were used before 335 ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of 336 ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support 337 tarred datasets. 338 339 Args: 340 model_cfg: model configuration 341 342 Returns: 343 whether ``model_config`` is legacy 344 """ 345 return 'common_dataset_parameters' not in model_cfg 346 347 348 def legacy_model_config_to_new_model_config(model_cfg: DictConfig) -> DictConfig: 349 """ 350 Transform old style config into 351 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig`. 352 Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style 353 datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. 354 Old style configs do not support tarred datasets. 355 356 Args: 357 model_cfg: old style config 358 359 Returns: 360 model config which follows dataclass 361 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig` 362 """ 363 train_ds = model_cfg.get('train_ds') 364 validation_ds = model_cfg.get('validation_ds') 365 test_ds = model_cfg.get('test_ds') 366 dataset = model_cfg.dataset 367 punct_head_config = model_cfg.get('punct_head', {}) 368 capit_head_config = model_cfg.get('capit_head', {}) 369 omega_conf = OmegaConf.structured( 370 PunctuationCapitalizationModelConfig( 371 class_labels=model_cfg.class_labels, 372 common_dataset_parameters=CommonDatasetParametersConfig( 373 pad_label=dataset.pad_label, 374 ignore_extra_tokens=dataset.get( 375 'ignore_extra_tokens', CommonDatasetParametersConfig.ignore_extra_tokens 376 ), 377 ignore_start_end=dataset.get('ignore_start_end', CommonDatasetParametersConfig.ignore_start_end), 378 punct_label_ids=model_cfg.punct_label_ids, 379 capit_label_ids=model_cfg.capit_label_ids, 380 ), 381 train_ds=None 382 if train_ds is None 383 else legacy_data_config_to_new_data_config(train_ds, dataset, train=True), 384 validation_ds=None 385 if validation_ds is None 386 else legacy_data_config_to_new_data_config(validation_ds, dataset, train=False), 387 test_ds=None if test_ds is None else legacy_data_config_to_new_data_config(test_ds, dataset, train=False), 388 punct_head=HeadConfig( 389 num_fc_layers=punct_head_config.get('punct_num_fc_layers', HeadConfig.num_fc_layers), 390 fc_dropout=punct_head_config.get('fc_dropout', HeadConfig.fc_dropout), 391 activation=punct_head_config.get('activation', HeadConfig.activation), 392 use_transformer_init=punct_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 393 ), 394 capit_head=HeadConfig( 395 num_fc_layers=capit_head_config.get('capit_num_fc_layers', HeadConfig.num_fc_layers), 396 fc_dropout=capit_head_config.get('fc_dropout', HeadConfig.fc_dropout), 397 activation=capit_head_config.get('activation', HeadConfig.activation), 398 use_transformer_init=capit_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 399 ), 400 tokenizer=model_cfg.tokenizer, 401 language_model=model_cfg.language_model, 402 optim=model_cfg.optim, 403 ) 404 ) 405 with open_dict(omega_conf): 406 retain_during_legacy_conversion = model_cfg.get('retain_during_legacy_conversion', {}) 407 for key in retain_during_legacy_conversion.keys(): 408 omega_conf[key] = retain_during_legacy_conversion[key] 409 return omega_conf 410 [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """Transformer based language model.""" 16 from MeCab import Model 17 from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule 18 from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule 19 from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( 20 MegatronRetrievalTransformerEncoderModule, 21 ) 22 from nemo.collections.nlp.modules.common.megatron.utils import ( 23 ApexGuardDefaults, 24 init_method_normal, 25 scaled_init_method_normal, 26 ) 27 28 try: 29 from apex.transformer.enums import AttnMaskType, ModelType 30 31 HAVE_APEX = True 32 except (ImportError, ModuleNotFoundError): 33 HAVE_APEX = False 34 # fake missing classes with None attributes 35 AttnMaskType = ApexGuardDefaults() 36 ModelType = ApexGuardDefaults() 37 38 try: 39 from megatron.core import ModelParallelConfig 40 41 HAVE_MEGATRON_CORE = True 42 43 except (ImportError, ModuleNotFoundError): 44 45 ModelParallelConfig = ApexGuardDefaults 46 47 HAVE_MEGATRON_CORE = False 48 49 __all__ = [] 50 51 AVAILABLE_ENCODERS = ["transformer", "perceiver", "retro"] 52 53 54 def get_encoder_model( 55 config: ModelParallelConfig, 56 arch, 57 hidden_size, 58 ffn_hidden_size, 59 num_layers, 60 num_attention_heads, 61 apply_query_key_layer_scaling=False, 62 kv_channels=None, 63 init_method=None, 64 scaled_init_method=None, 65 encoder_attn_mask_type=AttnMaskType.padding, 66 pre_process=True, 67 post_process=True, 68 init_method_std=0.02, 69 megatron_amp_O2=False, 70 hidden_dropout=0.1, 71 attention_dropout=0.1, 72 ffn_dropout=0.0, 73 precision=16, 74 fp32_residual_connection=False, 75 activations_checkpoint_method=None, 76 activations_checkpoint_num_layers=1, 77 activations_checkpoint_granularity=None, 78 layernorm_epsilon=1e-5, 79 bias_activation_fusion=True, 80 bias_dropout_add_fusion=True, 81 masked_softmax_fusion=True, 82 persist_layer_norm=False, 83 openai_gelu=False, 84 activation="gelu", 85 onnx_safe=False, 86 bias=True, 87 normalization="layernorm", 88 headscale=False, 89 transformer_block_type="pre_ln", 90 hidden_steps=32, 91 parent_model_type=ModelType.encoder_or_decoder, 92 layer_type=None, 93 chunk_size=64, 94 num_self_attention_per_cross_attention=1, 95 layer_number_offset=0, # this is use only for attention norm_factor scaling 96 megatron_legacy=False, 97 normalize_attention_scores=True, 98 sequence_parallel=False, 99 num_moe_experts=1, 100 moe_frequency=1, 101 moe_dropout=0.0, 102 turn_off_rop=False, # turn off the RoP positional embedding 103 version=1, # model version 104 position_embedding_type='learned_absolute', 105 use_flash_attention=False, 106 ): 107 """Build language model and return along with the key to save.""" 108 109 if kv_channels is None: 110 assert ( 111 hidden_size % num_attention_heads == 0 112 ), 'hidden_size must be divisible by num_attention_heads if kv_channels is None' 113 kv_channels = hidden_size // num_attention_heads 114 115 if init_method is None: 116 init_method = init_method_normal(init_method_std) 117 118 if scaled_init_method is None: 119 scaled_init_method = scaled_init_method_normal(init_method_std, num_layers) 120 121 if arch == "transformer": 122 # Language encoder. 123 encoder = MegatronTransformerEncoderModule( 124 config=config, 125 init_method=init_method, 126 output_layer_init_method=scaled_init_method, 127 hidden_size=hidden_size, 128 num_layers=num_layers, 129 num_attention_heads=num_attention_heads, 130 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 131 kv_channels=kv_channels, 132 ffn_hidden_size=ffn_hidden_size, 133 encoder_attn_mask_type=encoder_attn_mask_type, 134 pre_process=pre_process, 135 post_process=post_process, 136 megatron_amp_O2=megatron_amp_O2, 137 hidden_dropout=hidden_dropout, 138 attention_dropout=attention_dropout, 139 ffn_dropout=ffn_dropout, 140 precision=precision, 141 fp32_residual_connection=fp32_residual_connection, 142 activations_checkpoint_method=activations_checkpoint_method, 143 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 144 activations_checkpoint_granularity=activations_checkpoint_granularity, 145 layernorm_epsilon=layernorm_epsilon, 146 bias_activation_fusion=bias_activation_fusion, 147 bias_dropout_add_fusion=bias_dropout_add_fusion, 148 masked_softmax_fusion=masked_softmax_fusion, 149 persist_layer_norm=persist_layer_norm, 150 openai_gelu=openai_gelu, 151 onnx_safe=onnx_safe, 152 activation=activation, 153 bias=bias, 154 normalization=normalization, 155 transformer_block_type=transformer_block_type, 156 headscale=headscale, 157 parent_model_type=parent_model_type, 158 megatron_legacy=megatron_legacy, 159 normalize_attention_scores=normalize_attention_scores, 160 num_moe_experts=num_moe_experts, 161 moe_frequency=moe_frequency, 162 moe_dropout=moe_dropout, 163 position_embedding_type=position_embedding_type, 164 use_flash_attention=use_flash_attention, 165 ) 166 elif arch == "retro": 167 encoder = MegatronRetrievalTransformerEncoderModule( 168 config=config, 169 init_method=init_method, 170 output_layer_init_method=scaled_init_method, 171 hidden_size=hidden_size, 172 num_layers=num_layers, 173 num_attention_heads=num_attention_heads, 174 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 175 kv_channels=kv_channels, 176 layer_type=layer_type, 177 ffn_hidden_size=ffn_hidden_size, 178 pre_process=pre_process, 179 post_process=post_process, 180 megatron_amp_O2=megatron_amp_O2, 181 hidden_dropout=hidden_dropout, 182 attention_dropout=attention_dropout, 183 precision=precision, 184 fp32_residual_connection=fp32_residual_connection, 185 activations_checkpoint_method=activations_checkpoint_method, 186 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 187 activations_checkpoint_granularity=activations_checkpoint_granularity, 188 layernorm_epsilon=layernorm_epsilon, 189 bias_activation_fusion=bias_activation_fusion, 190 bias_dropout_add_fusion=bias_dropout_add_fusion, 191 masked_softmax_fusion=masked_softmax_fusion, 192 persist_layer_norm=persist_layer_norm, 193 openai_gelu=openai_gelu, 194 onnx_safe=onnx_safe, 195 activation=activation, 196 bias=bias, 197 normalization=normalization, 198 transformer_block_type=transformer_block_type, 199 parent_model_type=parent_model_type, 200 chunk_size=chunk_size, 201 layer_number_offset=layer_number_offset, 202 megatron_legacy=megatron_legacy, 203 normalize_attention_scores=normalize_attention_scores, 204 turn_off_rop=turn_off_rop, 205 version=version, 206 ) 207 elif arch == "perceiver": 208 encoder = MegatronPerceiverEncoderModule( 209 config=config, 210 init_method=init_method, 211 output_layer_init_method=scaled_init_method, 212 hidden_size=hidden_size, 213 num_layers=num_layers, 214 num_attention_heads=num_attention_heads, 215 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 216 kv_channels=kv_channels, 217 ffn_hidden_size=ffn_hidden_size, 218 encoder_attn_mask_type=encoder_attn_mask_type, 219 pre_process=pre_process, 220 post_process=post_process, 221 megatron_amp_O2=megatron_amp_O2, 222 hidden_dropout=hidden_dropout, 223 attention_dropout=attention_dropout, 224 ffn_dropout=ffn_dropout, 225 precision=precision, 226 fp32_residual_connection=fp32_residual_connection, 227 activations_checkpoint_method=activations_checkpoint_method, 228 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 229 activations_checkpoint_granularity=activations_checkpoint_granularity, 230 layernorm_epsilon=layernorm_epsilon, 231 bias_activation_fusion=bias_activation_fusion, 232 bias_dropout_add_fusion=bias_dropout_add_fusion, 233 masked_softmax_fusion=masked_softmax_fusion, 234 persist_layer_norm=persist_layer_norm, 235 openai_gelu=openai_gelu, 236 onnx_safe=onnx_safe, 237 activation=activation, 238 bias=bias, 239 normalization=normalization, 240 transformer_block_type=transformer_block_type, 241 headscale=headscale, 242 parent_model_type=parent_model_type, 243 hidden_steps=hidden_steps, 244 num_self_attention_per_cross_attention=num_self_attention_per_cross_attention, 245 megatron_legacy=megatron_legacy, 246 normalize_attention_scores=normalize_attention_scores, 247 ) 248 else: 249 raise ValueError(f"Unknown encoder arch = {arch}. Available encoder arch = {AVAILABLE_ENCODERS}") 250 251 return encoder 252 [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/tts/models/fastpitch.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import contextlib 15 from dataclasses import dataclass 16 from pathlib import Path 17 from typing import List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import DictConfig, OmegaConf, open_dict 22 from pytorch_lightning import Trainer 23 from pytorch_lightning.loggers import TensorBoardLogger 24 25 from nemo.collections.common.parts.preprocessing import parsers 26 from nemo.collections.tts.losses.aligner_loss import BinLoss, ForwardSumLoss 27 from nemo.collections.tts.losses.fastpitchloss import DurationLoss, EnergyLoss, MelLoss, PitchLoss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.modules.fastpitch import FastPitchModule 30 from nemo.collections.tts.parts.mixins import FastPitchAdapterModelMixin 31 from nemo.collections.tts.parts.utils.callbacks import LoggingCallback 32 from nemo.collections.tts.parts.utils.helpers import ( 33 batch_from_ragged, 34 g2p_backward_compatible_support, 35 plot_alignment_to_numpy, 36 plot_spectrogram_to_numpy, 37 process_batch, 38 sample_tts_input, 39 ) 40 from nemo.core.classes import Exportable 41 from nemo.core.classes.common import PretrainedModelInfo, typecheck 42 from nemo.core.neural_types.elements import ( 43 Index, 44 LengthsType, 45 MelSpectrogramType, 46 ProbsType, 47 RegressionValuesType, 48 TokenDurationType, 49 TokenIndex, 50 TokenLogDurationType, 51 ) 52 from nemo.core.neural_types.neural_type import NeuralType 53 from nemo.utils import logging, model_utils 54 55 56 @dataclass 57 class G2PConfig: 58 _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" 59 phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" 60 heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" 61 phoneme_probability: float = 0.5 62 63 64 @dataclass 65 class TextTokenizer: 66 _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" 67 punct: bool = True 68 stresses: bool = True 69 chars: bool = True 70 apostrophe: bool = True 71 pad_with_space: bool = True 72 add_blank_at: bool = True 73 g2p: G2PConfig = G2PConfig() 74 75 76 @dataclass 77 class TextTokenizerConfig: 78 text_tokenizer: TextTokenizer = TextTokenizer() 79 80 81 class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): 82 """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" 83 84 def __init__(self, cfg: DictConfig, trainer: Trainer = None): 85 # Convert to Hydra 1.0 compatible DictConfig 86 cfg = model_utils.convert_model_config_to_dict_config(cfg) 87 cfg = model_utils.maybe_update_config_version(cfg) 88 89 # Setup normalizer 90 self.normalizer = None 91 self.text_normalizer_call = None 92 self.text_normalizer_call_kwargs = {} 93 self._setup_normalizer(cfg) 94 95 self.learn_alignment = cfg.get("learn_alignment", False) 96 97 # Setup vocabulary (=tokenizer) and input_fft_kwargs (supported only with self.learn_alignment=True) 98 input_fft_kwargs = {} 99 if self.learn_alignment: 100 self.vocab = None 101 102 self.ds_class = cfg.train_ds.dataset._target_ 103 self.ds_class_name = self.ds_class.split(".")[-1] 104 if not self.ds_class in [ 105 "nemo.collections.tts.data.dataset.TTSDataset", 106 "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset", 107 "nemo.collections.tts.torch.data.TTSDataset", 108 ]: 109 raise ValueError(f"Unknown dataset class: {self.ds_class}.") 110 111 self._setup_tokenizer(cfg) 112 assert self.vocab is not None 113 input_fft_kwargs["n_embed"] = len(self.vocab.tokens) 114 input_fft_kwargs["padding_idx"] = self.vocab.pad 115 116 self._parser = None 117 self._tb_logger = None 118 super().__init__(cfg=cfg, trainer=trainer) 119 120 self.bin_loss_warmup_epochs = cfg.get("bin_loss_warmup_epochs", 100) 121 self.log_images = cfg.get("log_images", False) 122 self.log_train_images = False 123 124 default_prosody_loss_scale = 0.1 if self.learn_alignment else 1.0 125 dur_loss_scale = cfg.get("dur_loss_scale", default_prosody_loss_scale) 126 pitch_loss_scale = cfg.get("pitch_loss_scale", default_prosody_loss_scale) 127 energy_loss_scale = cfg.get("energy_loss_scale", default_prosody_loss_scale) 128 129 self.mel_loss_fn = MelLoss() 130 self.pitch_loss_fn = PitchLoss(loss_scale=pitch_loss_scale) 131 self.duration_loss_fn = DurationLoss(loss_scale=dur_loss_scale) 132 self.energy_loss_fn = EnergyLoss(loss_scale=energy_loss_scale) 133 134 self.aligner = None 135 if self.learn_alignment: 136 aligner_loss_scale = cfg.get("aligner_loss_scale", 1.0) 137 self.aligner = instantiate(self._cfg.alignment_module) 138 self.forward_sum_loss_fn = ForwardSumLoss(loss_scale=aligner_loss_scale) 139 self.bin_loss_fn = BinLoss(loss_scale=aligner_loss_scale) 140 141 self.preprocessor = instantiate(self._cfg.preprocessor) 142 input_fft = instantiate(self._cfg.input_fft, **input_fft_kwargs) 143 output_fft = instantiate(self._cfg.output_fft) 144 duration_predictor = instantiate(self._cfg.duration_predictor) 145 pitch_predictor = instantiate(self._cfg.pitch_predictor) 146 speaker_encoder = instantiate(self._cfg.get("speaker_encoder", None)) 147 energy_embedding_kernel_size = cfg.get("energy_embedding_kernel_size", 0) 148 energy_predictor = instantiate(self._cfg.get("energy_predictor", None)) 149 150 # [TODO] may remove if we change the pre-trained config 151 # cfg: condition_types = [ "add" ] 152 n_speakers = cfg.get("n_speakers", 0) 153 speaker_emb_condition_prosody = cfg.get("speaker_emb_condition_prosody", False) 154 speaker_emb_condition_decoder = cfg.get("speaker_emb_condition_decoder", False) 155 speaker_emb_condition_aligner = cfg.get("speaker_emb_condition_aligner", False) 156 min_token_duration = cfg.get("min_token_duration", 0) 157 use_log_energy = cfg.get("use_log_energy", True) 158 if n_speakers > 1 and "add" not in input_fft.cond_input.condition_types: 159 input_fft.cond_input.condition_types.append("add") 160 if speaker_emb_condition_prosody: 161 duration_predictor.cond_input.condition_types.append("add") 162 pitch_predictor.cond_input.condition_types.append("add") 163 if speaker_emb_condition_decoder: 164 output_fft.cond_input.condition_types.append("add") 165 if speaker_emb_condition_aligner and self.aligner is not None: 166 self.aligner.cond_input.condition_types.append("add") 167 168 self.fastpitch = FastPitchModule( 169 input_fft, 170 output_fft, 171 duration_predictor, 172 pitch_predictor, 173 energy_predictor, 174 self.aligner, 175 speaker_encoder, 176 n_speakers, 177 cfg.symbols_embedding_dim, 178 cfg.pitch_embedding_kernel_size, 179 energy_embedding_kernel_size, 180 cfg.n_mel_channels, 181 min_token_duration, 182 cfg.max_token_duration, 183 use_log_energy, 184 ) 185 self._input_types = self._output_types = None 186 self.export_config = { 187 "emb_range": (0, self.fastpitch.encoder.word_emb.num_embeddings), 188 "enable_volume": False, 189 "enable_ragged_batches": False, 190 } 191 if self.fastpitch.speaker_emb is not None: 192 self.export_config["num_speakers"] = cfg.n_speakers 193 194 self.log_config = cfg.get("log_config", None) 195 196 # Adapter modules setup (from FastPitchAdapterModelMixin) 197 self.setup_adapters() 198 199 def _get_default_text_tokenizer_conf(self): 200 text_tokenizer: TextTokenizerConfig = TextTokenizerConfig() 201 return OmegaConf.create(OmegaConf.to_yaml(text_tokenizer)) 202 203 def _setup_normalizer(self, cfg): 204 if "text_normalizer" in cfg: 205 normalizer_kwargs = {} 206 207 if "whitelist" in cfg.text_normalizer: 208 normalizer_kwargs["whitelist"] = self.register_artifact( 209 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 210 ) 211 try: 212 import nemo_text_processing 213 214 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 215 except Exception as e: 216 logging.error(e) 217 raise ImportError( 218 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 219 ) 220 221 self.text_normalizer_call = self.normalizer.normalize 222 if "text_normalizer_call_kwargs" in cfg: 223 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 224 225 def _setup_tokenizer(self, cfg): 226 text_tokenizer_kwargs = {} 227 228 if "g2p" in cfg.text_tokenizer: 229 # for backward compatibility 230 if ( 231 self._is_model_being_restored() 232 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 233 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 234 ): 235 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 236 cfg.text_tokenizer.g2p["_target_"] 237 ) 238 239 g2p_kwargs = {} 240 241 if "phoneme_dict" in cfg.text_tokenizer.g2p: 242 g2p_kwargs["phoneme_dict"] = self.register_artifact( 243 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 244 ) 245 246 if "heteronyms" in cfg.text_tokenizer.g2p: 247 g2p_kwargs["heteronyms"] = self.register_artifact( 248 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 249 ) 250 251 # for backward compatability 252 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 253 254 # TODO @xueyang: rename the instance of tokenizer because vocab is misleading. 255 self.vocab = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 256 257 @property 258 def tb_logger(self): 259 if self._tb_logger is None: 260 if self.logger is None and self.logger.experiment is None: 261 return None 262 tb_logger = self.logger.experiment 263 for logger in self.trainer.loggers: 264 if isinstance(logger, TensorBoardLogger): 265 tb_logger = logger.experiment 266 break 267 self._tb_logger = tb_logger 268 return self._tb_logger 269 270 @property 271 def parser(self): 272 if self._parser is not None: 273 return self._parser 274 275 if self.learn_alignment: 276 self._parser = self.vocab.encode 277 else: 278 self._parser = parsers.make_parser( 279 labels=self._cfg.labels, 280 name='en', 281 unk_id=-1, 282 blank_id=-1, 283 do_normalize=True, 284 abbreviation_version="fastpitch", 285 make_table=False, 286 ) 287 return self._parser 288 289 def parse(self, str_input: str, normalize=True) -> torch.tensor: 290 if self.training: 291 logging.warning("parse() is meant to be called in eval mode.") 292 293 if normalize and self.text_normalizer_call is not None: 294 str_input = self.text_normalizer_call(str_input, **self.text_normalizer_call_kwargs) 295 296 if self.learn_alignment: 297 eval_phon_mode = contextlib.nullcontext() 298 if hasattr(self.vocab, "set_phone_prob"): 299 eval_phon_mode = self.vocab.set_phone_prob(prob=1.0) 300 301 # Disable mixed g2p representation if necessary 302 with eval_phon_mode: 303 tokens = self.parser(str_input) 304 else: 305 tokens = self.parser(str_input) 306 307 x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device) 308 return x 309 310 @typecheck( 311 input_types={ 312 "text": NeuralType(('B', 'T_text'), TokenIndex()), 313 "durs": NeuralType(('B', 'T_text'), TokenDurationType()), 314 "pitch": NeuralType(('B', 'T_audio'), RegressionValuesType()), 315 "energy": NeuralType(('B', 'T_audio'), RegressionValuesType(), optional=True), 316 "speaker": NeuralType(('B'), Index(), optional=True), 317 "pace": NeuralType(optional=True), 318 "spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 319 "attn_prior": NeuralType(('B', 'T_spec', 'T_text'), ProbsType(), optional=True), 320 "mel_lens": NeuralType(('B'), LengthsType(), optional=True), 321 "input_lens": NeuralType(('B'), LengthsType(), optional=True), 322 # reference_* data is used for multi-speaker FastPitch training 323 "reference_spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 324 "reference_spec_lens": NeuralType(('B'), LengthsType(), optional=True), 325 } 326 ) 327 def forward( 328 self, 329 *, 330 text, 331 durs=None, 332 pitch=None, 333 energy=None, 334 speaker=None, 335 pace=1.0, 336 spec=None, 337 attn_prior=None, 338 mel_lens=None, 339 input_lens=None, 340 reference_spec=None, 341 reference_spec_lens=None, 342 ): 343 return self.fastpitch( 344 text=text, 345 durs=durs, 346 pitch=pitch, 347 energy=energy, 348 speaker=speaker, 349 pace=pace, 350 spec=spec, 351 attn_prior=attn_prior, 352 mel_lens=mel_lens, 353 input_lens=input_lens, 354 reference_spec=reference_spec, 355 reference_spec_lens=reference_spec_lens, 356 ) 357 358 @typecheck(output_types={"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType())}) 359 def generate_spectrogram( 360 self, 361 tokens: 'torch.tensor', 362 speaker: Optional[int] = None, 363 pace: float = 1.0, 364 reference_spec: Optional['torch.tensor'] = None, 365 reference_spec_lens: Optional['torch.tensor'] = None, 366 ) -> torch.tensor: 367 if self.training: 368 logging.warning("generate_spectrogram() is meant to be called in eval mode.") 369 if isinstance(speaker, int): 370 speaker = torch.tensor([speaker]).to(self.device) 371 spect, *_ = self( 372 text=tokens, 373 durs=None, 374 pitch=None, 375 speaker=speaker, 376 pace=pace, 377 reference_spec=reference_spec, 378 reference_spec_lens=reference_spec_lens, 379 ) 380 return spect 381 382 def training_step(self, batch, batch_idx): 383 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 384 None, 385 None, 386 None, 387 None, 388 None, 389 None, 390 ) 391 if self.learn_alignment: 392 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 393 batch_dict = batch 394 else: 395 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 396 audio = batch_dict.get("audio") 397 audio_lens = batch_dict.get("audio_lens") 398 text = batch_dict.get("text") 399 text_lens = batch_dict.get("text_lens") 400 attn_prior = batch_dict.get("align_prior_matrix", None) 401 pitch = batch_dict.get("pitch", None) 402 energy = batch_dict.get("energy", None) 403 speaker = batch_dict.get("speaker_id", None) 404 reference_audio = batch_dict.get("reference_audio", None) 405 reference_audio_len = batch_dict.get("reference_audio_lens", None) 406 else: 407 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 408 409 mels, spec_len = self.preprocessor(input_signal=audio, length=audio_lens) 410 reference_spec, reference_spec_len = None, None 411 if reference_audio is not None: 412 reference_spec, reference_spec_len = self.preprocessor( 413 input_signal=reference_audio, length=reference_audio_len 414 ) 415 416 ( 417 mels_pred, 418 _, 419 _, 420 log_durs_pred, 421 pitch_pred, 422 attn_soft, 423 attn_logprob, 424 attn_hard, 425 attn_hard_dur, 426 pitch, 427 energy_pred, 428 energy_tgt, 429 ) = self( 430 text=text, 431 durs=durs, 432 pitch=pitch, 433 energy=energy, 434 speaker=speaker, 435 pace=1.0, 436 spec=mels if self.learn_alignment else None, 437 reference_spec=reference_spec, 438 reference_spec_lens=reference_spec_len, 439 attn_prior=attn_prior, 440 mel_lens=spec_len, 441 input_lens=text_lens, 442 ) 443 if durs is None: 444 durs = attn_hard_dur 445 446 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 447 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 448 loss = mel_loss + dur_loss 449 if self.learn_alignment: 450 ctc_loss = self.forward_sum_loss_fn(attn_logprob=attn_logprob, in_lens=text_lens, out_lens=spec_len) 451 bin_loss_weight = min(self.current_epoch / self.bin_loss_warmup_epochs, 1.0) * 1.0 452 bin_loss = self.bin_loss_fn(hard_attention=attn_hard, soft_attention=attn_soft) * bin_loss_weight 453 loss += ctc_loss + bin_loss 454 455 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 456 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 457 loss += pitch_loss + energy_loss 458 459 self.log("t_loss", loss) 460 self.log("t_mel_loss", mel_loss) 461 self.log("t_dur_loss", dur_loss) 462 self.log("t_pitch_loss", pitch_loss) 463 if energy_tgt is not None: 464 self.log("t_energy_loss", energy_loss) 465 if self.learn_alignment: 466 self.log("t_ctc_loss", ctc_loss) 467 self.log("t_bin_loss", bin_loss) 468 469 # Log images to tensorboard 470 if self.log_images and self.log_train_images and isinstance(self.logger, TensorBoardLogger): 471 self.log_train_images = False 472 473 self.tb_logger.add_image( 474 "train_mel_target", 475 plot_spectrogram_to_numpy(mels[0].data.cpu().float().numpy()), 476 self.global_step, 477 dataformats="HWC", 478 ) 479 spec_predict = mels_pred[0].data.cpu().float().numpy() 480 self.tb_logger.add_image( 481 "train_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 482 ) 483 if self.learn_alignment: 484 attn = attn_hard[0].data.cpu().float().numpy().squeeze() 485 self.tb_logger.add_image( 486 "train_attn", plot_alignment_to_numpy(attn.T), self.global_step, dataformats="HWC", 487 ) 488 soft_attn = attn_soft[0].data.cpu().float().numpy().squeeze() 489 self.tb_logger.add_image( 490 "train_soft_attn", plot_alignment_to_numpy(soft_attn.T), self.global_step, dataformats="HWC", 491 ) 492 493 return loss 494 495 def validation_step(self, batch, batch_idx): 496 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 497 None, 498 None, 499 None, 500 None, 501 None, 502 None, 503 ) 504 if self.learn_alignment: 505 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 506 batch_dict = batch 507 else: 508 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 509 audio = batch_dict.get("audio") 510 audio_lens = batch_dict.get("audio_lens") 511 text = batch_dict.get("text") 512 text_lens = batch_dict.get("text_lens") 513 attn_prior = batch_dict.get("align_prior_matrix", None) 514 pitch = batch_dict.get("pitch", None) 515 energy = batch_dict.get("energy", None) 516 speaker = batch_dict.get("speaker_id", None) 517 reference_audio = batch_dict.get("reference_audio", None) 518 reference_audio_len = batch_dict.get("reference_audio_lens", None) 519 else: 520 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 521 522 mels, mel_lens = self.preprocessor(input_signal=audio, length=audio_lens) 523 reference_spec, reference_spec_len = None, None 524 if reference_audio is not None: 525 reference_spec, reference_spec_len = self.preprocessor( 526 input_signal=reference_audio, length=reference_audio_len 527 ) 528 529 # Calculate val loss on ground truth durations to better align L2 loss in time 530 (mels_pred, _, _, log_durs_pred, pitch_pred, _, _, _, attn_hard_dur, pitch, energy_pred, energy_tgt,) = self( 531 text=text, 532 durs=durs, 533 pitch=pitch, 534 energy=energy, 535 speaker=speaker, 536 pace=1.0, 537 spec=mels if self.learn_alignment else None, 538 reference_spec=reference_spec, 539 reference_spec_lens=reference_spec_len, 540 attn_prior=attn_prior, 541 mel_lens=mel_lens, 542 input_lens=text_lens, 543 ) 544 if durs is None: 545 durs = attn_hard_dur 546 547 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 548 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 549 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 550 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 551 loss = mel_loss + dur_loss + pitch_loss + energy_loss 552 553 val_outputs = { 554 "val_loss": loss, 555 "mel_loss": mel_loss, 556 "dur_loss": dur_loss, 557 "pitch_loss": pitch_loss, 558 "energy_loss": energy_loss if energy_tgt is not None else None, 559 "mel_target": mels if batch_idx == 0 else None, 560 "mel_pred": mels_pred if batch_idx == 0 else None, 561 } 562 self.validation_step_outputs.append(val_outputs) 563 return val_outputs 564 565 def on_validation_epoch_end(self): 566 collect = lambda key: torch.stack([x[key] for x in self.validation_step_outputs]).mean() 567 val_loss = collect("val_loss") 568 mel_loss = collect("mel_loss") 569 dur_loss = collect("dur_loss") 570 pitch_loss = collect("pitch_loss") 571 self.log("val_loss", val_loss, sync_dist=True) 572 self.log("val_mel_loss", mel_loss, sync_dist=True) 573 self.log("val_dur_loss", dur_loss, sync_dist=True) 574 self.log("val_pitch_loss", pitch_loss, sync_dist=True) 575 if self.validation_step_outputs[0]["energy_loss"] is not None: 576 energy_loss = collect("energy_loss") 577 self.log("val_energy_loss", energy_loss, sync_dist=True) 578 579 _, _, _, _, _, spec_target, spec_predict = self.validation_step_outputs[0].values() 580 581 if self.log_images and isinstance(self.logger, TensorBoardLogger): 582 self.tb_logger.add_image( 583 "val_mel_target", 584 plot_spectrogram_to_numpy(spec_target[0].data.cpu().float().numpy()), 585 self.global_step, 586 dataformats="HWC", 587 ) 588 spec_predict = spec_predict[0].data.cpu().float().numpy() 589 self.tb_logger.add_image( 590 "val_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 591 ) 592 self.log_train_images = True 593 self.validation_step_outputs.clear() # free memory) 594 595 def _setup_train_dataloader(self, cfg): 596 phon_mode = contextlib.nullcontext() 597 if hasattr(self.vocab, "set_phone_prob"): 598 phon_mode = self.vocab.set_phone_prob(self.vocab.phoneme_probability) 599 600 with phon_mode: 601 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 602 603 sampler = dataset.get_sampler(cfg.dataloader_params.batch_size) 604 return torch.utils.data.DataLoader( 605 dataset, collate_fn=dataset.collate_fn, sampler=sampler, **cfg.dataloader_params 606 ) 607 608 def _setup_test_dataloader(self, cfg): 609 phon_mode = contextlib.nullcontext() 610 if hasattr(self.vocab, "set_phone_prob"): 611 phon_mode = self.vocab.set_phone_prob(0.0) 612 613 with phon_mode: 614 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 615 616 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 617 618 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 619 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 620 raise ValueError(f"No dataset for {name}") 621 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 622 raise ValueError(f"No dataloader_params for {name}") 623 if shuffle_should_be: 624 if 'shuffle' not in cfg.dataloader_params: 625 logging.warning( 626 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 627 "config. Manually setting to True" 628 ) 629 with open_dict(cfg.dataloader_params): 630 cfg.dataloader_params.shuffle = True 631 elif not cfg.dataloader_params.shuffle: 632 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 633 elif cfg.dataloader_params.shuffle: 634 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 635 636 if self.ds_class == "nemo.collections.tts.data.dataset.TTSDataset": 637 phon_mode = contextlib.nullcontext() 638 if hasattr(self.vocab, "set_phone_prob"): 639 phon_mode = self.vocab.set_phone_prob(prob=None if name == "val" else self.vocab.phoneme_probability) 640 641 with phon_mode: 642 dataset = instantiate( 643 cfg.dataset, 644 text_normalizer=self.normalizer, 645 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 646 text_tokenizer=self.vocab, 647 ) 648 else: 649 dataset = instantiate(cfg.dataset) 650 651 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 652 653 def setup_training_data(self, cfg): 654 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 655 self._train_dl = self._setup_train_dataloader(cfg) 656 else: 657 self._train_dl = self.__setup_dataloader_from_config(cfg) 658 659 def setup_validation_data(self, cfg): 660 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 661 self._validation_dl = self._setup_test_dataloader(cfg) 662 else: 663 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="val") 664 665 def setup_test_data(self, cfg): 666 """Omitted.""" 667 pass 668 669 def configure_callbacks(self): 670 if not self.log_config: 671 return [] 672 673 sample_ds_class = self.log_config.dataset._target_ 674 if sample_ds_class != "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 675 raise ValueError(f"Logging callback only supported for TextToSpeechDataset, got {sample_ds_class}") 676 677 data_loader = self._setup_test_dataloader(self.log_config) 678 679 generators = instantiate(self.log_config.generators) 680 log_dir = Path(self.log_config.log_dir) if self.log_config.log_dir else None 681 log_callback = LoggingCallback( 682 generators=generators, 683 data_loader=data_loader, 684 log_epochs=self.log_config.log_epochs, 685 epoch_frequency=self.log_config.epoch_frequency, 686 output_dir=log_dir, 687 loggers=self.trainer.loggers, 688 log_tensorboard=self.log_config.log_tensorboard, 689 log_wandb=self.log_config.log_wandb, 690 ) 691 692 return [log_callback] 693 694 @classmethod 695 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 696 """ 697 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 698 Returns: 699 List of available pre-trained models. 700 """ 701 list_of_models = [] 702 703 # en-US, single speaker, 22050Hz, LJSpeech (ARPABET). 704 model = PretrainedModelInfo( 705 pretrained_model_name="tts_en_fastpitch", 706 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/1.8.1/files/tts_en_fastpitch_align.nemo", 707 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is ARPABET-based.", 708 class_=cls, 709 ) 710 list_of_models.append(model) 711 712 # en-US, single speaker, 22050Hz, LJSpeech (IPA). 713 model = PretrainedModelInfo( 714 pretrained_model_name="tts_en_fastpitch_ipa", 715 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/IPA_1.13.0/files/tts_en_fastpitch_align_ipa.nemo", 716 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is IPA-based.", 717 class_=cls, 718 ) 719 list_of_models.append(model) 720 721 # en-US, multi-speaker, 44100Hz, HiFiTTS. 722 model = PretrainedModelInfo( 723 pretrained_model_name="tts_en_fastpitch_multispeaker", 724 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_multispeaker_fastpitchhifigan/versions/1.10.0/files/tts_en_fastpitch_multispeaker.nemo", 725 description="This model is trained on HiFITTS sampled at 44100Hz with and can be used to generate male and female English voices with an American accent.", 726 class_=cls, 727 ) 728 list_of_models.append(model) 729 730 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 21.02 731 model = PretrainedModelInfo( 732 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2102", 733 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2102.nemo", 734 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 21.02 Dataset sampled at 22050Hz and can be used to generate male German voices.", 735 class_=cls, 736 ) 737 list_of_models.append(model) 738 739 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 22.10 740 model = PretrainedModelInfo( 741 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2210", 742 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2210.nemo", 743 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 22.10 Dataset sampled at 22050Hz and can be used to generate male German voices.", 744 class_=cls, 745 ) 746 list_of_models.append(model) 747 748 # de-DE, multi-speaker, 5 speakers, 44100 Hz, HUI-Audio-Corpus-German Clean. 749 model = PretrainedModelInfo( 750 pretrained_model_name="tts_de_fastpitch_multispeaker_5", 751 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitch_multispeaker_5/versions/1.11.0/files/tts_de_fastpitch_multispeaker_5.nemo", 752 description="This model is trained on 5 speakers in HUI-Audio-Corpus-German clean subset sampled at 44100Hz with and can be used to generate male and female German voices.", 753 class_=cls, 754 ) 755 list_of_models.append(model) 756 757 # es, 174 speakers, 44100Hz, OpenSLR (IPA) 758 model = PretrainedModelInfo( 759 pretrained_model_name="tts_es_fastpitch_multispeaker", 760 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_es_multispeaker_fastpitchhifigan/versions/1.15.0/files/tts_es_fastpitch_multispeaker.nemo", 761 description="This model is trained on 174 speakers in 6 crowdsourced Latin American Spanish OpenSLR datasets sampled at 44100Hz and can be used to generate male and female Spanish voices with Latin American accents.", 762 class_=cls, 763 ) 764 list_of_models.append(model) 765 766 # zh, single female speaker, 22050Hz, SFSpeech Bilingual Chinese/English dataset, improved model using richer 767 # dict and jieba word segmenter for polyphone disambiguation. 768 model = PretrainedModelInfo( 769 pretrained_model_name="tts_zh_fastpitch_sfspeech", 770 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_zh_fastpitch_hifigan_sfspeech/versions/1.15.0/files/tts_zh_fastpitch_sfspeech.nemo", 771 description="This model is trained on a single female speaker in SFSpeech Bilingual Chinese/English dataset" 772 " sampled at 22050Hz and can be used to generate female Mandarin Chinese voices. It is improved" 773 " using richer dict and jieba word segmenter for polyphone disambiguation.", 774 class_=cls, 775 ) 776 list_of_models.append(model) 777 778 # en, multi speaker, LibriTTS, 16000 Hz 779 # stft 25ms 10ms matching ASR params 780 # for use during Enhlish ASR training/adaptation 781 model = PretrainedModelInfo( 782 pretrained_model_name="tts_en_fastpitch_for_asr_finetuning", 783 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch_spectrogram_enhancer_for_asr_finetuning/versions/1.20.0/files/tts_en_fastpitch_for_asr_finetuning.nemo", 784 description="This model is trained on LibriSpeech, train-960 subset." 785 " STFT parameters follow those commonly used in ASR: 25 ms window, 10 ms hop." 786 " This model is supposed to be used with its companion SpetrogramEnhancer for " 787 " ASR fine-tuning. Usage for regular TTS tasks is not advised.", 788 class_=cls, 789 ) 790 list_of_models.append(model) 791 792 return list_of_models 793 794 # Methods for model exportability 795 def _prepare_for_export(self, **kwargs): 796 super()._prepare_for_export(**kwargs) 797 798 tensor_shape = ('T') if self.export_config["enable_ragged_batches"] else ('B', 'T') 799 800 # Define input_types and output_types as required by export() 801 self._input_types = { 802 "text": NeuralType(tensor_shape, TokenIndex()), 803 "pitch": NeuralType(tensor_shape, RegressionValuesType()), 804 "pace": NeuralType(tensor_shape), 805 "volume": NeuralType(tensor_shape, optional=True), 806 "batch_lengths": NeuralType(('B'), optional=True), 807 "speaker": NeuralType(('B'), Index(), optional=True), 808 } 809 self._output_types = { 810 "spect": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 811 "num_frames": NeuralType(('B'), TokenDurationType()), 812 "durs_predicted": NeuralType(('B', 'T'), TokenDurationType()), 813 "log_durs_predicted": NeuralType(('B', 'T'), TokenLogDurationType()), 814 "pitch_predicted": NeuralType(('B', 'T'), RegressionValuesType()), 815 } 816 if self.export_config["enable_volume"]: 817 self._output_types["volume_aligned"] = NeuralType(('B', 'T'), RegressionValuesType()) 818 819 def _export_teardown(self): 820 self._input_types = self._output_types = None 821 822 @property 823 def disabled_deployment_input_names(self): 824 """Implement this method to return a set of input names disabled for export""" 825 disabled_inputs = set() 826 if self.fastpitch.speaker_emb is None: 827 disabled_inputs.add("speaker") 828 if not self.export_config["enable_ragged_batches"]: 829 disabled_inputs.add("batch_lengths") 830 if not self.export_config["enable_volume"]: 831 disabled_inputs.add("volume") 832 return disabled_inputs 833 834 @property 835 def input_types(self): 836 return self._input_types 837 838 @property 839 def output_types(self): 840 return self._output_types 841 842 def input_example(self, max_batch=1, max_dim=44): 843 """ 844 Generates input examples for tracing etc. 845 Returns: 846 A tuple of input examples. 847 """ 848 par = next(self.fastpitch.parameters()) 849 inputs = sample_tts_input(self.export_config, par.device, max_batch=max_batch, max_dim=max_dim) 850 if 'enable_ragged_batches' not in self.export_config: 851 inputs.pop('batch_lengths', None) 852 return (inputs,) 853 854 def forward_for_export(self, text, pitch, pace, volume=None, batch_lengths=None, speaker=None): 855 if self.export_config["enable_ragged_batches"]: 856 text, pitch, pace, volume_tensor, lens = batch_from_ragged( 857 text, pitch, pace, batch_lengths, padding_idx=self.fastpitch.encoder.padding_idx, volume=volume 858 ) 859 if volume is not None: 860 volume = volume_tensor 861 return self.fastpitch.infer(text=text, pitch=pitch, pace=pace, volume=volume, speaker=speaker) 862 863 def interpolate_speaker( 864 self, original_speaker_1, original_speaker_2, weight_speaker_1, weight_speaker_2, new_speaker_id 865 ): 866 """ 867 This method performs speaker interpolation between two original speakers the model is trained on. 868 869 Inputs: 870 original_speaker_1: Integer speaker ID of first existing speaker in the model 871 original_speaker_2: Integer speaker ID of second existing speaker in the model 872 weight_speaker_1: Floating point weight associated in to first speaker during weight combination 873 weight_speaker_2: Floating point weight associated in to second speaker during weight combination 874 new_speaker_id: Integer speaker ID of new interpolated speaker in the model 875 """ 876 if self.fastpitch.speaker_emb is None: 877 raise Exception( 878 "Current FastPitch model is not a multi-speaker FastPitch model. Speaker interpolation can only \ 879 be performed with a multi-speaker model" 880 ) 881 n_speakers = self.fastpitch.speaker_emb.weight.data.size()[0] 882 if original_speaker_1 >= n_speakers or original_speaker_2 >= n_speakers or new_speaker_id >= n_speakers: 883 raise Exception( 884 f"Parameters original_speaker_1, original_speaker_2, new_speaker_id should be less than the total \ 885 total number of speakers FastPitch was trained on (n_speakers = {n_speakers})." 886 ) 887 speaker_emb_1 = ( 888 self.fastpitch.speaker_emb(torch.tensor(original_speaker_1, dtype=torch.int32).cuda()).clone().detach() 889 ) 890 speaker_emb_2 = ( 891 self.fastpitch.speaker_emb(torch.tensor(original_speaker_2, dtype=torch.int32).cuda()).clone().detach() 892 ) 893 new_speaker_emb = weight_speaker_1 * speaker_emb_1 + weight_speaker_2 * speaker_emb_2 894 self.fastpitch.speaker_emb.weight.data[new_speaker_id] = new_speaker_emb 895 [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/tts/models/tacotron2.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import contextlib 16 from dataclasses import dataclass 17 from typing import Any, Dict, List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import MISSING, DictConfig, OmegaConf, open_dict 22 from omegaconf.errors import ConfigAttributeError 23 from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger 24 from torch import nn 25 26 from nemo.collections.common.parts.preprocessing import parsers 27 from nemo.collections.tts.losses.tacotron2loss import Tacotron2Loss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.parts.utils.helpers import ( 30 g2p_backward_compatible_support, 31 get_mask_from_lengths, 32 tacotron2_log_to_tb_func, 33 tacotron2_log_to_wandb_func, 34 ) 35 from nemo.core.classes.common import PretrainedModelInfo, typecheck 36 from nemo.core.neural_types.elements import ( 37 AudioSignal, 38 EmbeddedTextType, 39 LengthsType, 40 LogitsType, 41 MelSpectrogramType, 42 SequenceToSequenceAlignmentType, 43 ) 44 from nemo.core.neural_types.neural_type import NeuralType 45 from nemo.utils import logging, model_utils 46 47 48 @dataclass 49 class Preprocessor: 50 _target_: str = MISSING 51 pad_value: float = MISSING 52 53 54 @dataclass 55 class Tacotron2Config: 56 preprocessor: Preprocessor = Preprocessor() 57 encoder: Dict[Any, Any] = MISSING 58 decoder: Dict[Any, Any] = MISSING 59 postnet: Dict[Any, Any] = MISSING 60 labels: List = MISSING 61 train_ds: Optional[Dict[Any, Any]] = None 62 validation_ds: Optional[Dict[Any, Any]] = None 63 64 65 class Tacotron2Model(SpectrogramGenerator): 66 """Tacotron 2 Model that is used to generate mel spectrograms from text""" 67 68 def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None): 69 # Convert to Hydra 1.0 compatible DictConfig 70 cfg = model_utils.convert_model_config_to_dict_config(cfg) 71 cfg = model_utils.maybe_update_config_version(cfg) 72 73 # setup normalizer 74 self.normalizer = None 75 self.text_normalizer_call = None 76 self.text_normalizer_call_kwargs = {} 77 self._setup_normalizer(cfg) 78 79 # setup tokenizer 80 self.tokenizer = None 81 if hasattr(cfg, 'text_tokenizer'): 82 self._setup_tokenizer(cfg) 83 84 self.num_tokens = len(self.tokenizer.tokens) 85 self.tokenizer_pad = self.tokenizer.pad 86 self.tokenizer_unk = self.tokenizer.oov 87 # assert self.tokenizer is not None 88 else: 89 self.num_tokens = len(cfg.labels) + 3 90 91 super().__init__(cfg=cfg, trainer=trainer) 92 93 schema = OmegaConf.structured(Tacotron2Config) 94 # ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes 95 if isinstance(cfg, dict): 96 cfg = OmegaConf.create(cfg) 97 elif not isinstance(cfg, DictConfig): 98 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 99 # Ensure passed cfg is compliant with schema 100 try: 101 OmegaConf.merge(cfg, schema) 102 self.pad_value = cfg.preprocessor.pad_value 103 except ConfigAttributeError: 104 self.pad_value = cfg.preprocessor.params.pad_value 105 logging.warning( 106 "Your config is using an old NeMo yaml configuration. Please ensure that the yaml matches the " 107 "current version in the main branch for future compatibility." 108 ) 109 110 self._parser = None 111 self.audio_to_melspec_precessor = instantiate(cfg.preprocessor) 112 self.text_embedding = nn.Embedding(self.num_tokens, 512) 113 self.encoder = instantiate(self._cfg.encoder) 114 self.decoder = instantiate(self._cfg.decoder) 115 self.postnet = instantiate(self._cfg.postnet) 116 self.loss = Tacotron2Loss() 117 self.calculate_loss = True 118 119 @property 120 def parser(self): 121 if self._parser is not None: 122 return self._parser 123 124 ds_class_name = self._cfg.train_ds.dataset._target_.split(".")[-1] 125 if ds_class_name == "TTSDataset": 126 self._parser = None 127 elif hasattr(self._cfg, "labels"): 128 self._parser = parsers.make_parser( 129 labels=self._cfg.labels, 130 name='en', 131 unk_id=-1, 132 blank_id=-1, 133 do_normalize=True, 134 abbreviation_version="fastpitch", 135 make_table=False, 136 ) 137 else: 138 raise ValueError("Wanted to setup parser, but model does not have necessary paramaters") 139 140 return self._parser 141 142 def parse(self, text: str, normalize=True) -> torch.Tensor: 143 if self.training: 144 logging.warning("parse() is meant to be called in eval mode.") 145 if normalize and self.text_normalizer_call is not None: 146 text = self.text_normalizer_call(text, **self.text_normalizer_call_kwargs) 147 148 eval_phon_mode = contextlib.nullcontext() 149 if hasattr(self.tokenizer, "set_phone_prob"): 150 eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0) 151 152 with eval_phon_mode: 153 if self.tokenizer is not None: 154 tokens = self.tokenizer.encode(text) 155 else: 156 tokens = self.parser(text) 157 # Old parser doesn't add bos and eos ids, so maunally add it 158 tokens = [len(self._cfg.labels)] + tokens + [len(self._cfg.labels) + 1] 159 tokens_tensor = torch.tensor(tokens).unsqueeze_(0).to(self.device) 160 return tokens_tensor 161 162 @property 163 def input_types(self): 164 if self.training: 165 return { 166 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 167 "token_len": NeuralType(('B'), LengthsType()), 168 "audio": NeuralType(('B', 'T'), AudioSignal()), 169 "audio_len": NeuralType(('B'), LengthsType()), 170 } 171 else: 172 return { 173 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 174 "token_len": NeuralType(('B'), LengthsType()), 175 "audio": NeuralType(('B', 'T'), AudioSignal(), optional=True), 176 "audio_len": NeuralType(('B'), LengthsType(), optional=True), 177 } 178 179 @property 180 def output_types(self): 181 if not self.calculate_loss and not self.training: 182 return { 183 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 184 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 185 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 186 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 187 "pred_length": NeuralType(('B'), LengthsType()), 188 } 189 return { 190 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 191 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 192 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 193 "spec_target": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 194 "spec_target_len": NeuralType(('B'), LengthsType()), 195 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 196 } 197 198 @typecheck() 199 def forward(self, *, tokens, token_len, audio=None, audio_len=None): 200 if audio is not None and audio_len is not None: 201 spec_target, spec_target_len = self.audio_to_melspec_precessor(audio, audio_len) 202 else: 203 if self.training or self.calculate_loss: 204 raise ValueError( 205 f"'audio' and 'audio_len' can not be None when either 'self.training' or 'self.calculate_loss' is True." 206 ) 207 208 token_embedding = self.text_embedding(tokens).transpose(1, 2) 209 encoder_embedding = self.encoder(token_embedding=token_embedding, token_len=token_len) 210 211 if self.training: 212 spec_pred_dec, gate_pred, alignments = self.decoder( 213 memory=encoder_embedding, decoder_inputs=spec_target, memory_lengths=token_len 214 ) 215 else: 216 spec_pred_dec, gate_pred, alignments, pred_length = self.decoder( 217 memory=encoder_embedding, memory_lengths=token_len 218 ) 219 220 spec_pred_postnet = self.postnet(mel_spec=spec_pred_dec) 221 222 if not self.calculate_loss and not self.training: 223 return spec_pred_dec, spec_pred_postnet, gate_pred, alignments, pred_length 224 225 return spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments 226 227 @typecheck( 228 input_types={"tokens": NeuralType(('B', 'T'), EmbeddedTextType())}, 229 output_types={"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType())}, 230 ) 231 def generate_spectrogram(self, *, tokens): 232 self.eval() 233 self.calculate_loss = False 234 token_len = torch.tensor([len(i) for i in tokens]).to(self.device) 235 tensors = self(tokens=tokens, token_len=token_len) 236 spectrogram_pred = tensors[1] 237 238 if spectrogram_pred.shape[0] > 1: 239 # Silence all frames past the predicted end 240 mask = ~get_mask_from_lengths(tensors[-1]) 241 mask = mask.expand(spectrogram_pred.shape[1], mask.size(0), mask.size(1)) 242 mask = mask.permute(1, 0, 2) 243 spectrogram_pred.data.masked_fill_(mask, self.pad_value) 244 245 return spectrogram_pred 246 247 def training_step(self, batch, batch_idx): 248 audio, audio_len, tokens, token_len = batch 249 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, _ = self.forward( 250 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 251 ) 252 253 loss, _ = self.loss( 254 spec_pred_dec=spec_pred_dec, 255 spec_pred_postnet=spec_pred_postnet, 256 gate_pred=gate_pred, 257 spec_target=spec_target, 258 spec_target_len=spec_target_len, 259 pad_value=self.pad_value, 260 ) 261 262 output = { 263 'loss': loss, 264 'progress_bar': {'training_loss': loss}, 265 'log': {'loss': loss}, 266 } 267 return output 268 269 def validation_step(self, batch, batch_idx): 270 audio, audio_len, tokens, token_len = batch 271 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments = self.forward( 272 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 273 ) 274 275 loss, gate_target = self.loss( 276 spec_pred_dec=spec_pred_dec, 277 spec_pred_postnet=spec_pred_postnet, 278 gate_pred=gate_pred, 279 spec_target=spec_target, 280 spec_target_len=spec_target_len, 281 pad_value=self.pad_value, 282 ) 283 loss = { 284 "val_loss": loss, 285 "mel_target": spec_target, 286 "mel_postnet": spec_pred_postnet, 287 "gate": gate_pred, 288 "gate_target": gate_target, 289 "alignments": alignments, 290 } 291 self.validation_step_outputs.append(loss) 292 return loss 293 294 def on_validation_epoch_end(self): 295 if self.logger is not None and self.logger.experiment is not None: 296 logger = self.logger.experiment 297 for logger in self.trainer.loggers: 298 if isinstance(logger, TensorBoardLogger): 299 logger = logger.experiment 300 break 301 if isinstance(logger, TensorBoardLogger): 302 tacotron2_log_to_tb_func( 303 logger, 304 self.validation_step_outputs[0].values(), 305 self.global_step, 306 tag="val", 307 log_images=True, 308 add_audio=False, 309 ) 310 elif isinstance(logger, WandbLogger): 311 tacotron2_log_to_wandb_func( 312 logger, 313 self.validation_step_outputs[0].values(), 314 self.global_step, 315 tag="val", 316 log_images=True, 317 add_audio=False, 318 ) 319 avg_loss = torch.stack( 320 [x['val_loss'] for x in self.validation_step_outputs] 321 ).mean() # This reduces across batches, not workers! 322 self.log('val_loss', avg_loss) 323 self.validation_step_outputs.clear() # free memory 324 325 def _setup_normalizer(self, cfg): 326 if "text_normalizer" in cfg: 327 normalizer_kwargs = {} 328 329 if "whitelist" in cfg.text_normalizer: 330 normalizer_kwargs["whitelist"] = self.register_artifact( 331 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 332 ) 333 334 try: 335 import nemo_text_processing 336 337 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 338 except Exception as e: 339 logging.error(e) 340 raise ImportError( 341 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 342 ) 343 344 self.text_normalizer_call = self.normalizer.normalize 345 if "text_normalizer_call_kwargs" in cfg: 346 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 347 348 def _setup_tokenizer(self, cfg): 349 text_tokenizer_kwargs = {} 350 if "g2p" in cfg.text_tokenizer and cfg.text_tokenizer.g2p is not None: 351 # for backward compatibility 352 if ( 353 self._is_model_being_restored() 354 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 355 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 356 ): 357 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 358 cfg.text_tokenizer.g2p["_target_"] 359 ) 360 361 g2p_kwargs = {} 362 363 if "phoneme_dict" in cfg.text_tokenizer.g2p: 364 g2p_kwargs["phoneme_dict"] = self.register_artifact( 365 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 366 ) 367 368 if "heteronyms" in cfg.text_tokenizer.g2p: 369 g2p_kwargs["heteronyms"] = self.register_artifact( 370 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 371 ) 372 373 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 374 375 self.tokenizer = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 376 377 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 378 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 379 raise ValueError(f"No dataset for {name}") 380 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 381 raise ValueError(f"No dataloder_params for {name}") 382 if shuffle_should_be: 383 if 'shuffle' not in cfg.dataloader_params: 384 logging.warning( 385 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 386 "config. Manually setting to True" 387 ) 388 with open_dict(cfg.dataloader_params): 389 cfg.dataloader_params.shuffle = True 390 elif not cfg.dataloader_params.shuffle: 391 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 392 elif not shuffle_should_be and cfg.dataloader_params.shuffle: 393 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 394 395 dataset = instantiate( 396 cfg.dataset, 397 text_normalizer=self.normalizer, 398 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 399 text_tokenizer=self.tokenizer, 400 ) 401 402 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 403 404 def setup_training_data(self, cfg): 405 self._train_dl = self.__setup_dataloader_from_config(cfg) 406 407 def setup_validation_data(self, cfg): 408 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="validation") 409 410 @classmethod 411 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 412 """ 413 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 414 Returns: 415 List of available pre-trained models. 416 """ 417 list_of_models = [] 418 model = PretrainedModelInfo( 419 pretrained_model_name="tts_en_tacotron2", 420 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_tacotron2/versions/1.10.0/files/tts_en_tacotron2.nemo", 421 description="This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.", 422 class_=cls, 423 aliases=["Tacotron2-22050Hz"], 424 ) 425 list_of_models.append(model) 426 return list_of_models 427 [end of nemo/collections/tts/models/tacotron2.py] [start of nemo/core/config/modelPT.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Dict, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.core import config 21 from nemo.core.classes.dataset import DatasetConfig 22 from nemo.utils import exp_manager 23 24 25 @dataclass 26 class SchedConfig: 27 name: str = MISSING 28 min_lr: float = 0.0 29 last_epoch: int = -1 30 31 32 @dataclass 33 class OptimConfig: 34 name: str = MISSING 35 sched: Optional[SchedConfig] = None 36 37 38 @dataclass 39 class ModelConfig: 40 """ 41 Model component inside ModelPT 42 """ 43 44 # ... 45 train_ds: Optional[DatasetConfig] = None 46 validation_ds: Optional[DatasetConfig] = None 47 test_ds: Optional[DatasetConfig] = None 48 optim: Optional[OptimConfig] = None 49 50 51 @dataclass 52 class HydraConfig: 53 run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) 54 job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) 55 56 57 @dataclass 58 class NemoConfig: 59 name: str = MISSING 60 model: ModelConfig = MISSING 61 trainer: config.TrainerConfig = config.TrainerConfig( 62 strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' 63 ) 64 exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() 65 hydra: HydraConfig = HydraConfig() 66 67 68 class ModelConfigBuilder: 69 def __init__(self, model_cfg: ModelConfig): 70 """ 71 Base class for any Model Config Builder. 72 73 A Model Config Builder is a utility class that accepts a ModelConfig dataclass, 74 and via a set of utility methods (that are implemented by the subclassed ModelConfigBuilder), 75 builds a finalized ModelConfig that can be supplied to a NemoModel dataclass as 76 the `model` component. 77 78 Subclasses *must* implement the private method `_finalize_cfg`. 79 Inside this method, they must update `self.model_cfg` with all interdependent config 80 options that need to be set (either updated by user explicitly or with their default value). 81 82 The updated model config must then be preserved in `self.model_cfg`. 83 84 Example: 85 # Create the config builder 86 config_builder = <subclass>ModelConfigBuilder() 87 88 # Update the components of the config that are modifiable 89 config_builder.set_X(X) 90 config_builder.set_Y(Y) 91 92 # Create a "finalized" config dataclass that will contain all the updates 93 # that were specified by the builder 94 model_config = config_builder.build() 95 96 # Use model config as is (or further update values), then create a new Model 97 model = nemo.<domain>.models.<ModelName>Model(cfg=model_config, trainer=Trainer()) 98 99 Supported build methods: 100 - set_train_ds: All model configs can accept a subclass of `DatasetConfig` as their 101 training config. Subclasses can override this method to enable auto-complete 102 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 103 104 - set_validation_ds: All model configs can accept a subclass of `DatasetConfig` as their 105 validation config. Subclasses can override this method to enable auto-complete 106 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 107 108 - set_test_ds: All model configs can accept a subclass of `DatasetConfig` as their 109 test config. Subclasses can override this method to enable auto-complete 110 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 111 112 - set_optim: A build method that supports changes to the Optimizer (and optionally, 113 the Scheduler) used for training the model. The function accepts two inputs - 114 115 `cfg`: A subclass of `OptimizerParams` - any OptimizerParams subclass can be used, 116 in order to select an appropriate Optimizer. Examples: AdamParams. 117 118 `sched_cfg`: A subclass of `SchedulerParams` - any SchedulerParams subclass can be used, 119 in order to select an appropriate Scheduler. Examples: CosineAnnealingParams. 120 Note that this argument is optional. 121 122 - build(): The method which should return a "finalized" ModelConfig dataclass. 123 Subclasses *should* always override this method, and update the signature 124 of this method with the return type of the Dataclass, so that it enables 125 autocomplete for the user. 126 127 Example: 128 def build(self) -> EncDecCTCConfig: 129 return super().build() 130 131 Any additional build methods must be added by subclasses of ModelConfigBuilder. 132 133 Args: 134 model_cfg: 135 """ 136 self.model_cfg = model_cfg 137 self.train_ds_cfg = None 138 self.validation_ds_cfg = None 139 self.test_ds_cfg = None 140 self.optim_cfg = None 141 142 def set_train_ds(self, cfg: Optional[DatasetConfig] = None): 143 self.model_cfg.train_ds = cfg 144 145 def set_validation_ds(self, cfg: Optional[DatasetConfig] = None): 146 self.model_cfg.validation_ds = cfg 147 148 def set_test_ds(self, cfg: Optional[DatasetConfig] = None): 149 self.model_cfg.test_ds = cfg 150 151 def set_optim(self, cfg: config.OptimizerParams, sched_cfg: Optional[config.SchedulerParams] = None): 152 @dataclass 153 class WrappedOptimConfig(OptimConfig, cfg.__class__): 154 pass 155 156 # Setup optim 157 optim_name = cfg.__class__.__name__.replace("Params", "").lower() 158 wrapped_cfg = WrappedOptimConfig(name=optim_name, sched=None, **vars(cfg)) 159 160 if sched_cfg is not None: 161 162 @dataclass 163 class WrappedSchedConfig(SchedConfig, sched_cfg.__class__): 164 pass 165 166 # Setup scheduler 167 sched_name = sched_cfg.__class__.__name__.replace("Params", "") 168 wrapped_sched_cfg = WrappedSchedConfig(name=sched_name, **vars(sched_cfg)) 169 170 wrapped_cfg.sched = wrapped_sched_cfg 171 172 self.model_cfg.optim = wrapped_cfg 173 174 def _finalize_cfg(self): 175 raise NotImplementedError() 176 177 def build(self) -> ModelConfig: 178 # validate config 179 self._finalize_cfg() 180 181 return self.model_cfg 182 [end of nemo/core/config/modelPT.py] [start of nemo/utils/exp_manager.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import glob 16 import os 17 import subprocess 18 import sys 19 import time 20 import warnings 21 from dataclasses import dataclass 22 from datetime import timedelta 23 from pathlib import Path 24 from shutil import copy, move 25 from typing import Any, Dict, List, Optional, Tuple, Union 26 27 import pytorch_lightning 28 import torch 29 from hydra.core.hydra_config import HydraConfig 30 from hydra.utils import get_original_cwd 31 from omegaconf import DictConfig, OmegaConf, open_dict 32 from pytorch_lightning.callbacks import Callback, ModelCheckpoint 33 from pytorch_lightning.callbacks.early_stopping import EarlyStopping 34 from pytorch_lightning.callbacks.timer import Interval, Timer 35 from pytorch_lightning.loggers import MLFlowLogger, TensorBoardLogger, WandbLogger 36 from pytorch_lightning.loops import _TrainingEpochLoop 37 from pytorch_lightning.strategies.ddp import DDPStrategy 38 39 from nemo.collections.common.callbacks import EMA 40 from nemo.constants import NEMO_ENV_VARNAME_TESTING, NEMO_ENV_VARNAME_VERSION 41 from nemo.utils import logging, timers 42 from nemo.utils.app_state import AppState 43 from nemo.utils.callbacks import NeMoModelCheckpoint, PreemptionCallback 44 from nemo.utils.env_var_parsing import get_envbool 45 from nemo.utils.exceptions import NeMoBaseException 46 from nemo.utils.get_rank import is_global_rank_zero 47 from nemo.utils.lightning_logger_patch import add_filehandlers_to_pl_logger 48 from nemo.utils.loggers import ClearMLLogger, ClearMLParams, DLLogger, DLLoggerParams, MLFlowParams 49 from nemo.utils.model_utils import uninject_model_parallel_rank 50 51 52 class NotFoundError(NeMoBaseException): 53 """ Raised when a file or folder is not found""" 54 55 56 class LoggerMisconfigurationError(NeMoBaseException): 57 """ Raised when a mismatch between trainer.logger and exp_manager occurs""" 58 59 def __init__(self, message): 60 message = ( 61 message 62 + " You can disable lighning's trainer from creating a logger by passing logger=False to its constructor." 63 ) 64 super().__init__(message) 65 66 67 class CheckpointMisconfigurationError(NeMoBaseException): 68 """ Raised when a mismatch between trainer.callbacks and exp_manager occurs""" 69 70 71 @dataclass 72 class EarlyStoppingParams: 73 monitor: str = "val_loss" # The metric that early stopping should consider. 74 mode: str = "min" # inform early stopping whether to look for increase or decrease in monitor. 75 min_delta: float = 0.001 # smallest change to consider as improvement. 76 patience: int = 10 # how many (continuous) validation cycles to wait with no improvement and stopping training. 77 verbose: bool = True 78 strict: bool = True 79 check_finite: bool = True 80 stopping_threshold: Optional[float] = None 81 divergence_threshold: Optional[float] = None 82 check_on_train_epoch_end: Optional[bool] = None 83 log_rank_zero_only: bool = False 84 85 86 @dataclass 87 class CallbackParams: 88 filepath: Optional[str] = None # Deprecated 89 dirpath: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 90 filename: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 91 monitor: Optional[str] = "val_loss" 92 verbose: Optional[bool] = True 93 save_last: Optional[bool] = True 94 save_top_k: Optional[int] = 3 95 save_weights_only: Optional[bool] = False 96 mode: Optional[str] = "min" 97 auto_insert_metric_name: bool = True 98 every_n_epochs: Optional[int] = 1 99 every_n_train_steps: Optional[int] = None 100 train_time_interval: Optional[str] = None 101 prefix: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 102 postfix: str = ".nemo" 103 save_best_model: bool = False 104 always_save_nemo: bool = False 105 save_nemo_on_train_end: Optional[bool] = True # Whether to automatically save .nemo file durin on_train_end hook 106 model_parallel_size: Optional[int] = None # tensor parallel size * pipeline parallel size 107 save_on_train_epoch_end: Optional[bool] = False # Save after training, not after validation 108 109 110 @dataclass 111 class StepTimingParams: 112 reduction: Optional[str] = "mean" 113 # if True torch.cuda.synchronize() is called on start/stop 114 sync_cuda: Optional[bool] = False 115 # if positive, defines the size of a sliding window for computing mean 116 buffer_size: Optional[int] = 1 117 118 119 @dataclass 120 class EMAParams: 121 enable: Optional[bool] = False 122 decay: Optional[float] = 0.999 123 cpu_offload: Optional[bool] = False 124 validate_original_weights: Optional[bool] = False 125 every_n_steps: int = 1 126 127 128 @dataclass 129 class ExpManagerConfig: 130 """Experiment Manager config for validation of passed arguments. 131 """ 132 133 # Log dir creation parameters 134 explicit_log_dir: Optional[str] = None 135 exp_dir: Optional[str] = None 136 name: Optional[str] = None 137 version: Optional[str] = None 138 use_datetime_version: Optional[bool] = True 139 resume_if_exists: Optional[bool] = False 140 resume_past_end: Optional[bool] = False 141 resume_ignore_no_checkpoint: Optional[bool] = False 142 resume_from_checkpoint: Optional[str] = None 143 # Logging parameters 144 create_tensorboard_logger: Optional[bool] = True 145 summary_writer_kwargs: Optional[Dict[Any, Any]] = None 146 create_wandb_logger: Optional[bool] = False 147 wandb_logger_kwargs: Optional[Dict[Any, Any]] = None 148 create_mlflow_logger: Optional[bool] = False 149 mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() 150 create_dllogger_logger: Optional[bool] = False 151 dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() 152 create_clearml_logger: Optional[bool] = False 153 clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() 154 # Checkpointing parameters 155 create_checkpoint_callback: Optional[bool] = True 156 checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() 157 create_early_stopping_callback: Optional[bool] = False 158 early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() 159 create_preemption_callback: Optional[bool] = True 160 # Additional exp_manager arguments 161 files_to_copy: Optional[List[str]] = None 162 # logs timing of train/val/test steps 163 log_step_timing: Optional[bool] = True 164 step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() 165 # Configures creation of log files for different ranks 166 log_local_rank_0_only: Optional[bool] = False 167 log_global_rank_0_only: Optional[bool] = False 168 # disable initial validation when resuming from a checkpoint saved during validation 169 disable_validation_on_resume: Optional[bool] = True 170 ema: Optional[EMAParams] = EMAParams() 171 # Wall clock time limit 172 max_time_per_run: Optional[str] = None 173 # time to sleep non 0 ranks during initialization 174 seconds_to_sleep: float = 5 175 176 177 class TimingCallback(Callback): 178 """ 179 Logs execution time of train/val/test steps 180 """ 181 182 def __init__(self, timer_kwargs={}): 183 self.timer = timers.NamedTimer(**timer_kwargs) 184 185 def _on_batch_start(self, name): 186 # reset only if we do not return mean of a sliding window 187 if self.timer.buffer_size <= 0: 188 self.timer.reset(name) 189 190 self.timer.start(name) 191 192 def _on_batch_end(self, name, pl_module): 193 self.timer.stop(name) 194 # Set the `batch_size=1` as WAR for `dataloader_iter`, which is not used for any metric 195 pl_module.log( 196 name + ' in s', 197 self.timer[name], 198 on_step=True, 199 on_epoch=False, 200 batch_size=1, 201 prog_bar=(name == "train_step_timing"), 202 ) 203 204 def on_train_batch_start(self, trainer, pl_module, batch, batch_idx): 205 self._on_batch_start("train_step_timing") 206 207 def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): 208 self._on_batch_end("train_step_timing", pl_module) 209 210 def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 211 self._on_batch_start("validation_step_timing") 212 213 def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 214 self._on_batch_end("validation_step_timing", pl_module) 215 216 def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 217 self._on_batch_start("test_step_timing") 218 219 def on_test_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 220 self._on_batch_end("test_step_timing", pl_module) 221 222 def on_before_backward(self, trainer, pl_module, loss): 223 self._on_batch_start("train_backward_timing") 224 225 def on_after_backward(self, trainer, pl_module): 226 self._on_batch_end("train_backward_timing", pl_module) 227 228 229 def exp_manager(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None) -> Optional[Path]: 230 """ 231 exp_manager is a helper function used to manage folders for experiments. It follows the pytorch lightning paradigm 232 of exp_dir/model_or_experiment_name/version. If the lightning trainer has a logger, exp_manager will get exp_dir, 233 name, and version from the logger. Otherwise it will use the exp_dir and name arguments to create the logging 234 directory. exp_manager also allows for explicit folder creation via explicit_log_dir. 235 236 The version can be a datetime string or an integer. Datestime version can be disabled if use_datetime_version is set 237 to False. It optionally creates TensorBoardLogger, WandBLogger, DLLogger, MLFlowLogger, ClearMLLogger, 238 ModelCheckpoint objects from pytorch lightning. 239 It copies sys.argv, and git information if available to the logging directory. It creates a log file for each 240 process to log their output into. 241 242 exp_manager additionally has a resume feature (resume_if_exists) which can be used to continuing training from 243 the constructed log_dir. When you need to continue the training repeatedly (like on a cluster which you need 244 multiple consecutive jobs), you need to avoid creating the version folders. Therefore from v1.0.0, when 245 resume_if_exists is set to True, creating the version folders is ignored. 246 247 Args: 248 trainer (pytorch_lightning.Trainer): The lightning trainer. 249 cfg (DictConfig, dict): Can have the following keys: 250 251 - explicit_log_dir (str, Path): Can be used to override exp_dir/name/version folder creation. Defaults to 252 None, which will use exp_dir, name, and version to construct the logging directory. 253 - exp_dir (str, Path): The base directory to create the logging directory. Defaults to None, which logs to 254 ./nemo_experiments. 255 - name (str): The name of the experiment. Defaults to None which turns into "default" via name = name or 256 "default". 257 - version (str): The version of the experiment. Defaults to None which uses either a datetime string or 258 lightning's TensorboardLogger system of using version_{int}. 259 - use_datetime_version (bool): Whether to use a datetime string for version. Defaults to True. 260 - resume_if_exists (bool): Whether this experiment is resuming from a previous run. If True, it sets 261 trainer._checkpoint_connector._ckpt_path so that the trainer should auto-resume. exp_manager will move files 262 under log_dir to log_dir/run_{int}. Defaults to False. From v1.0.0, when resume_if_exists is True, 263 we would not create version folders to make it easier to find the log folder for next runs. 264 - resume_past_end (bool): exp_manager errors out if resume_if_exists is True and a checkpoint matching 265 ``*end.ckpt`` indicating a previous training run fully completed. This behaviour can be disabled, in which 266 case the ``*end.ckpt`` will be loaded by setting resume_past_end to True. Defaults to False. 267 - resume_ignore_no_checkpoint (bool): exp_manager errors out if resume_if_exists is True and no checkpoint 268 could be found. This behaviour can be disabled, in which case exp_manager will print a message and 269 continue without restoring, by setting resume_ignore_no_checkpoint to True. Defaults to False. 270 - resume_from_checkpoint (str): Can be used to specify a path to a specific checkpoint file to load from. This will 271 override any checkpoint found when resume_if_exists is True. Defaults to None. 272 - create_tensorboard_logger (bool): Whether to create a tensorboard logger and attach it to the pytorch 273 lightning trainer. Defaults to True. 274 - summary_writer_kwargs (dict): A dictionary of kwargs that can be passed to lightning's TensorboardLogger 275 class. Note that log_dir is passed by exp_manager and cannot exist in this dict. Defaults to None. 276 - create_wandb_logger (bool): Whether to create a Weights and Baises logger and attach it to the pytorch 277 lightning trainer. Defaults to False. 278 - wandb_logger_kwargs (dict): A dictionary of kwargs that can be passed to lightning's WandBLogger 279 class. Note that name and project are required parameters if create_wandb_logger is True. 280 Defaults to None. 281 - create_mlflow_logger (bool): Whether to create an MLFlow logger and attach it to the pytorch lightning 282 training. Defaults to False 283 - mlflow_logger_kwargs (dict): optional parameters for the MLFlow logger 284 - create_dllogger_logger (bool): Whether to create an DLLogger logger and attach it to the pytorch lightning 285 training. Defaults to False 286 - dllogger_logger_kwargs (dict): optional parameters for the DLLogger logger 287 - create_clearml_logger (bool): Whether to create an ClearML logger and attach it to the pytorch lightning 288 training. Defaults to False 289 - clearml_logger_kwargs (dict): optional parameters for the ClearML logger 290 - create_checkpoint_callback (bool): Whether to create a ModelCheckpoint callback and attach it to the 291 pytorch lightning trainer. The ModelCheckpoint saves the top 3 models with the best "val_loss", the most 292 recent checkpoint under ``*last.ckpt``, and the final checkpoint after training completes under ``*end.ckpt``. 293 Defaults to True. 294 - create_early_stopping_callback (bool): Flag to decide if early stopping should be used to stop training. Default is False. 295 See EarlyStoppingParams dataclass above. 296 - create_preemption_callback (bool): Flag to decide whether to enable preemption callback to save checkpoints and exit training 297 immediately upon preemption. Default is True. 298 - files_to_copy (list): A list of files to copy to the experiment logging directory. Defaults to None which 299 copies no files. 300 - log_local_rank_0_only (bool): Whether to only create log files for local rank 0. Defaults to False. 301 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 302 - log_global_rank_0_only (bool): Whether to only create log files for global rank 0. Defaults to False. 303 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 304 - max_time (str): The maximum wall clock time *per run*. This is intended to be used on clusters where you want 305 a checkpoint to be saved after this specified time and be able to resume from that checkpoint. Defaults to None. 306 - seconds_to_sleep (float): seconds to sleep non rank 0 processes for. Used to give enough time for rank 0 to initialize 307 308 returns: 309 log_dir (Path): The final logging directory where logging files are saved. Usually the concatenation of 310 exp_dir, name, and version. 311 """ 312 # Add rank information to logger 313 # Note: trainer.global_rank and trainer.is_global_zero are not set until trainer.fit, so have to hack around it 314 local_rank = int(os.environ.get("LOCAL_RANK", 0)) 315 global_rank = trainer.node_rank * trainer.num_devices + local_rank 316 logging.rank = global_rank 317 318 if cfg is None: 319 logging.error("exp_manager did not receive a cfg argument. It will be disabled.") 320 return 321 if trainer.fast_dev_run: 322 logging.info("Trainer was called with fast_dev_run. exp_manager will return without any functionality.") 323 return 324 325 # Ensure passed cfg is compliant with ExpManagerConfig 326 schema = OmegaConf.structured(ExpManagerConfig) 327 if isinstance(cfg, dict): 328 cfg = OmegaConf.create(cfg) 329 elif not isinstance(cfg, DictConfig): 330 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 331 cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True)) 332 cfg = OmegaConf.merge(schema, cfg) 333 334 error_checks(trainer, cfg) # Ensures that trainer options are compliant with NeMo and exp_manager arguments 335 336 log_dir, exp_dir, name, version = get_log_dir( 337 trainer=trainer, 338 exp_dir=cfg.exp_dir, 339 name=cfg.name, 340 version=cfg.version, 341 explicit_log_dir=cfg.explicit_log_dir, 342 use_datetime_version=cfg.use_datetime_version, 343 resume_if_exists=cfg.resume_if_exists, 344 ) 345 346 check_resume( 347 trainer, 348 log_dir, 349 cfg.resume_if_exists, 350 cfg.resume_past_end, 351 cfg.resume_ignore_no_checkpoint, 352 cfg.checkpoint_callback_params.dirpath, 353 cfg.resume_from_checkpoint, 354 ) 355 356 checkpoint_name = name 357 # If name returned from get_log_dir is "", use cfg.name for checkpointing 358 if checkpoint_name is None or checkpoint_name == '': 359 checkpoint_name = cfg.name or "default" 360 361 # Set mlflow name if it's not set, before the main name is erased 362 if cfg.create_mlflow_logger and (not cfg.mlflow_logger_kwargs.get("experiment_name", None)): 363 cfg.mlflow_logger_kwargs.experiment_name = cfg.name 364 logging.warning( 365 'mlflow logger specified but no experiment name set. Using the same as Tensorboard: %s', 366 cfg.mlflow_logger_kwargs.experiment_name, 367 ) 368 369 cfg.name = name # Used for configure_loggers so that the log_dir is properly set even if name is "" 370 cfg.version = version 371 372 # update app_state with log_dir, exp_dir, etc 373 app_state = AppState() 374 app_state.log_dir = log_dir 375 app_state.exp_dir = exp_dir 376 app_state.name = name 377 app_state.version = version 378 app_state.checkpoint_name = checkpoint_name 379 app_state.create_checkpoint_callback = cfg.create_checkpoint_callback 380 app_state.checkpoint_callback_params = cfg.checkpoint_callback_params 381 382 # Create the logging directory if it does not exist 383 os.makedirs(log_dir, exist_ok=True) # Cannot limit creation to global zero as all ranks write to own log file 384 logging.info(f'Experiments will be logged at {log_dir}') 385 trainer._default_root_dir = log_dir 386 387 if cfg.log_local_rank_0_only is True and cfg.log_global_rank_0_only is True: 388 raise ValueError( 389 f"Cannot set both log_local_rank_0_only and log_global_rank_0_only to True. Please set either one or neither." 390 ) 391 392 # This is set if the env var NEMO_TESTING is set to True. 393 nemo_testing = get_envbool(NEMO_ENV_VARNAME_TESTING, False) 394 395 # Handle logging to file 396 log_file = log_dir / f'nemo_log_globalrank-{global_rank}_localrank-{local_rank}.txt' 397 if cfg.log_local_rank_0_only is True and not nemo_testing: 398 if local_rank == 0: 399 logging.add_file_handler(log_file) 400 elif cfg.log_global_rank_0_only is True and not nemo_testing: 401 if global_rank == 0: 402 logging.add_file_handler(log_file) 403 else: 404 # Logs on all ranks. 405 logging.add_file_handler(log_file) 406 407 # For some reason, LearningRateLogger requires trainer to have a logger. Safer to create logger on all ranks 408 # not just global rank 0. 409 if ( 410 cfg.create_tensorboard_logger 411 or cfg.create_wandb_logger 412 or cfg.create_mlflow_logger 413 or cfg.create_dllogger_logger 414 or cfg.create_clearml_logger 415 ): 416 configure_loggers( 417 trainer, 418 exp_dir, 419 log_dir, 420 cfg.name, 421 cfg.version, 422 cfg.checkpoint_callback_params, 423 cfg.create_tensorboard_logger, 424 cfg.summary_writer_kwargs, 425 cfg.create_wandb_logger, 426 cfg.wandb_logger_kwargs, 427 cfg.create_mlflow_logger, 428 cfg.mlflow_logger_kwargs, 429 cfg.create_dllogger_logger, 430 cfg.dllogger_logger_kwargs, 431 cfg.create_clearml_logger, 432 cfg.clearml_logger_kwargs, 433 ) 434 435 # add loggers timing callbacks 436 if cfg.log_step_timing: 437 timing_callback = TimingCallback(timer_kwargs=cfg.step_timing_kwargs or {}) 438 trainer.callbacks.insert(0, timing_callback) 439 440 if cfg.ema.enable: 441 ema_callback = EMA( 442 decay=cfg.ema.decay, 443 validate_original_weights=cfg.ema.validate_original_weights, 444 cpu_offload=cfg.ema.cpu_offload, 445 every_n_steps=cfg.ema.every_n_steps, 446 ) 447 trainer.callbacks.append(ema_callback) 448 449 if cfg.create_early_stopping_callback: 450 early_stop_callback = EarlyStopping(**cfg.early_stopping_callback_params) 451 trainer.callbacks.append(early_stop_callback) 452 453 if cfg.create_checkpoint_callback: 454 configure_checkpointing( 455 trainer, 456 log_dir, 457 checkpoint_name, 458 cfg.resume_if_exists, 459 cfg.checkpoint_callback_params, 460 cfg.create_preemption_callback, 461 ) 462 463 if cfg.disable_validation_on_resume: 464 # extend training loop to skip initial validation when resuming from checkpoint 465 configure_no_restart_validation_training_loop(trainer) 466 # Setup a stateless timer for use on clusters. 467 if cfg.max_time_per_run is not None: 468 found_ptl_timer = False 469 for idx, callback in enumerate(trainer.callbacks): 470 if isinstance(callback, Timer): 471 # NOTE: PTL does not expose a `trainer.max_time`. By the time we are in this function, PTL has already setup a timer if the user specifies `trainer.max_time` so best we can do is replace that. 472 # Working: If only `trainer.max_time` is set - it behaves as a normal PTL timer. If only `exp_manager.max_time_per_run` is set - it behaves as a StateLessTimer. If both are set, it also behaves as a StateLessTimer. 473 logging.warning( 474 f'Found a PTL Timer callback, replacing with a StatelessTimer callback. This will happen if you set trainer.max_time as well as exp_manager.max_time_per_run.' 475 ) 476 trainer.callbacks[idx] = StatelessTimer(cfg.max_time_per_run) 477 found_ptl_timer = True 478 break 479 480 if not found_ptl_timer: 481 trainer.max_time = cfg.max_time_per_run 482 trainer.callbacks.append(StatelessTimer(cfg.max_time_per_run)) 483 484 if is_global_rank_zero(): 485 # Move files_to_copy to folder and add git information if present 486 if cfg.files_to_copy: 487 for _file in cfg.files_to_copy: 488 copy(Path(_file), log_dir) 489 490 # Create files for cmd args and git info 491 with open(log_dir / 'cmd-args.log', 'w', encoding='utf-8') as _file: 492 _file.write(" ".join(sys.argv)) 493 494 # Try to get git hash 495 git_repo, git_hash = get_git_hash() 496 if git_repo: 497 with open(log_dir / 'git-info.log', 'w', encoding='utf-8') as _file: 498 _file.write(f'commit hash: {git_hash}') 499 _file.write(get_git_diff()) 500 501 # Add err_file logging to global_rank zero 502 logging.add_err_file_handler(log_dir / 'nemo_error_log.txt') 503 504 # Add lightning file logging to global_rank zero 505 add_filehandlers_to_pl_logger(log_dir / 'lightning_logs.txt', log_dir / 'nemo_error_log.txt') 506 507 elif trainer.num_nodes * trainer.num_devices > 1: 508 # sleep other ranks so rank 0 can finish 509 # doing the initialization such as moving files 510 time.sleep(cfg.seconds_to_sleep) 511 512 return log_dir 513 514 515 def error_checks(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None): 516 """ 517 Checks that the passed trainer is compliant with NeMo and exp_manager's passed configuration. Checks that: 518 - Throws error when hydra has changed the working directory. This causes issues with lightning's DDP 519 - Throws error when trainer has loggers defined but create_tensorboard_logger or create_wandB_logger 520 or create_mlflow_logger or create_dllogger_logger is True 521 - Prints error messages when 1) run on multi-node and not Slurm, and 2) run on multi-gpu without DDP 522 """ 523 if HydraConfig.initialized() and get_original_cwd() != os.getcwd(): 524 raise ValueError( 525 "Hydra changed the working directory. This interferes with ExpManger's functionality. Please pass " 526 "hydra.run.dir=. to your python script." 527 ) 528 if trainer.logger is not None and ( 529 cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger 530 ): 531 raise LoggerMisconfigurationError( 532 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and either " 533 f"create_tensorboard_logger: {cfg.create_tensorboard_logger} or create_wandb_logger: " 534 f"{cfg.create_wandb_logger} or create_mlflow_logger: {cfg.create_mlflow_logger}" 535 f"or create_dllogger_logger: {cfg.create_mlflow_logger} was set to True. " 536 "These can only be used if trainer does not already have a logger." 537 ) 538 if trainer.num_nodes > 1 and not check_slurm(trainer): 539 logging.error( 540 "You are running multi-node training without SLURM handling the processes." 541 " Please note that this is not tested in NeMo and could result in errors." 542 ) 543 if trainer.num_devices > 1 and not isinstance(trainer.strategy, DDPStrategy): 544 logging.error( 545 "You are running multi-gpu without ddp.Please note that this is not tested in NeMo and could result in " 546 "errors." 547 ) 548 549 550 def check_resume( 551 trainer: 'pytorch_lightning.Trainer', 552 log_dir: str, 553 resume_if_exists: bool = False, 554 resume_past_end: bool = False, 555 resume_ignore_no_checkpoint: bool = False, 556 dirpath: str = None, 557 resume_from_checkpoint: str = None, 558 ): 559 """Checks that resume=True was used correctly with the arguments pass to exp_manager. Sets 560 trainer._checkpoint_connector._ckpt_path as necessary. 561 562 Returns: 563 log_dir (Path): The log_dir 564 exp_dir (str): The base exp_dir without name nor version 565 name (str): The name of the experiment 566 version (str): The version of the experiment 567 568 Raises: 569 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 570 ValueError: If resume is True, and there were more than 1 checkpoint could found. 571 """ 572 573 if not log_dir: 574 raise ValueError(f"Resuming requires the log_dir {log_dir} to be passed to exp_manager") 575 576 checkpoint = None 577 if resume_from_checkpoint: 578 checkpoint = resume_from_checkpoint 579 if resume_if_exists: 580 # Use <log_dir>/checkpoints/ unless `dirpath` is set 581 checkpoint_dir = Path(dirpath) if dirpath else Path(Path(log_dir) / "checkpoints") 582 583 # when using distributed checkpointing, checkpoint_dir is a directory of directories 584 # we check for this here 585 dist_checkpoints = [d for d in list(checkpoint_dir.glob("*")) if d.is_dir()] 586 end_dist_checkpoints = [d for d in dist_checkpoints if d.match("*end")] 587 last_dist_checkpoints = [d for d in dist_checkpoints if d.match("*last")] 588 589 end_checkpoints = end_dist_checkpoints if end_dist_checkpoints else list(checkpoint_dir.rglob("*end.ckpt")) 590 last_checkpoints = last_dist_checkpoints if last_dist_checkpoints else list(checkpoint_dir.rglob("*last.ckpt")) 591 592 if not checkpoint_dir.exists() or (not len(end_checkpoints) > 0 and not len(last_checkpoints) > 0): 593 if resume_ignore_no_checkpoint: 594 warn = f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. " 595 if checkpoint is None: 596 warn += "Training from scratch." 597 elif checkpoint == resume_from_checkpoint: 598 warn += f"Training from {resume_from_checkpoint}." 599 logging.warning(warn) 600 else: 601 raise NotFoundError( 602 f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. Cannot resume." 603 ) 604 elif len(end_checkpoints) > 0: 605 if resume_past_end: 606 if len(end_checkpoints) > 1: 607 if 'mp_rank' in str(end_checkpoints[0]): 608 checkpoint = end_checkpoints[0] 609 else: 610 raise ValueError(f"Multiple checkpoints {end_checkpoints} that matches *end.ckpt.") 611 else: 612 raise ValueError( 613 f"Found {end_checkpoints[0]} indicating that the last training run has already completed." 614 ) 615 elif len(last_checkpoints) > 1: 616 if 'mp_rank' in str(last_checkpoints[0]) or 'tp_rank' in str(last_checkpoints[0]): 617 checkpoint = last_checkpoints[0] 618 checkpoint = uninject_model_parallel_rank(checkpoint) 619 else: 620 raise ValueError(f"Multiple checkpoints {last_checkpoints} that matches *last.ckpt.") 621 else: 622 checkpoint = last_checkpoints[0] 623 624 # PTL 2.0 supports ckpt_path instead of resume_from_checkpoint as the trainer flag 625 if checkpoint is not None: 626 trainer.ckpt_path = str(checkpoint) 627 logging.info(f'Resuming training from checkpoint: {trainer.ckpt_path}') 628 629 if is_global_rank_zero(): 630 # Check to see if any files exist that need to be moved 631 files_to_move = [] 632 if Path(log_dir).exists(): 633 for child in Path(log_dir).iterdir(): 634 if child.is_file(): 635 files_to_move.append(child) 636 637 if len(files_to_move) > 0: 638 # Move old files to a new folder 639 other_run_dirs = Path(log_dir).glob("run_*") 640 run_count = 0 641 for fold in other_run_dirs: 642 if fold.is_dir(): 643 run_count += 1 644 new_run_dir = Path(Path(log_dir) / f"run_{run_count}") 645 new_run_dir.mkdir() 646 for _file in files_to_move: 647 move(str(_file), str(new_run_dir)) 648 649 650 def check_explicit_log_dir( 651 trainer: 'pytorch_lightning.Trainer', explicit_log_dir: Union[Path, str], exp_dir: str, name: str, version: str 652 ) -> Tuple[Path, str, str, str]: 653 """ Checks that the passed arguments are compatible with explicit_log_dir. 654 655 Returns: 656 log_dir (Path): the log_dir 657 exp_dir (str): the base exp_dir without name nor version 658 name (str): The name of the experiment 659 version (str): The version of the experiment 660 661 Raise: 662 LoggerMisconfigurationError 663 """ 664 if trainer.logger is not None: 665 raise LoggerMisconfigurationError( 666 "The pytorch lightning trainer that was passed to exp_manager contained a logger and explicit_log_dir: " 667 f"{explicit_log_dir} was pass to exp_manager. Please remove the logger from the lightning trainer." 668 ) 669 # Checking only (explicit_log_dir) vs (exp_dir and version). 670 # The `name` will be used as the actual name of checkpoint/archive. 671 if exp_dir or version: 672 logging.error( 673 f"exp_manager received explicit_log_dir: {explicit_log_dir} and at least one of exp_dir: {exp_dir}, " 674 f"or version: {version}. Please note that exp_dir, name, and version will be ignored." 675 ) 676 if is_global_rank_zero() and Path(explicit_log_dir).exists(): 677 logging.warning(f"Exp_manager is logging to {explicit_log_dir}, but it already exists.") 678 return Path(explicit_log_dir), str(explicit_log_dir), "", "" 679 680 681 def get_log_dir( 682 trainer: 'pytorch_lightning.Trainer', 683 exp_dir: str = None, 684 name: str = None, 685 version: str = None, 686 explicit_log_dir: str = None, 687 use_datetime_version: bool = True, 688 resume_if_exists: bool = False, 689 ) -> Tuple[Path, str, str, str]: 690 """ 691 Obtains the log_dir used for exp_manager. 692 693 Returns: 694 log_dir (Path): the log_dir 695 exp_dir (str): the base exp_dir without name nor version 696 name (str): The name of the experiment 697 version (str): The version of the experiment 698 explicit_log_dir (str): The explicit path to the log folder. Defaults to False. 699 use_datetime_version (bool): Uses date and time as the version of the log folder. Defaults to True. 700 resume_if_exists (bool): if resume_if_exists of the exp_manager's config is enabled or not. When enabled, the 701 version folders would not get created. 702 703 Raise: 704 LoggerMisconfigurationError: If trainer is incompatible with arguments 705 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 706 ValueError: If resume is True, and there were more than 1 checkpoint could found. 707 """ 708 if explicit_log_dir: # If explicit log_dir was passed, short circuit 709 return check_explicit_log_dir(trainer, explicit_log_dir, exp_dir, name, version) 710 711 # Default exp_dir to ./nemo_experiments if None was passed 712 _exp_dir = exp_dir 713 if exp_dir is None: 714 _exp_dir = str(Path.cwd() / 'nemo_experiments') 715 716 # If the user has already defined a logger for the trainer, use the logger defaults for logging directory 717 if trainer.logger is not None: 718 if trainer.logger.save_dir: 719 if exp_dir: 720 raise LoggerMisconfigurationError( 721 "The pytorch lightning trainer that was passed to exp_manager contained a logger, the logger's " 722 f"save_dir was not None, and exp_dir ({exp_dir}) was not None. If trainer.logger.save_dir " 723 "exists, exp_manager will use trainer.logger.save_dir as the logging directory and exp_dir " 724 "must be None." 725 ) 726 _exp_dir = trainer.logger.save_dir 727 if name: 728 raise LoggerMisconfigurationError( 729 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and name: " 730 f"{name} was also passed to exp_manager. If the trainer contains a " 731 "logger, exp_manager will use trainer.logger.name, and name passed to exp_manager must be None." 732 ) 733 name = trainer.logger.name 734 version = f"version_{trainer.logger.version}" 735 # Use user-defined exp_dir, project_name, exp_name, and versioning options 736 else: 737 name = name or "default" 738 version = version or os.environ.get(NEMO_ENV_VARNAME_VERSION, None) 739 740 if not version: 741 if resume_if_exists: 742 logging.warning( 743 "No version folders would be created under the log folder as 'resume_if_exists' is enabled." 744 ) 745 version = None 746 elif is_global_rank_zero(): 747 if use_datetime_version: 748 version = time.strftime('%Y-%m-%d_%H-%M-%S') 749 else: 750 tensorboard_logger = TensorBoardLogger(save_dir=Path(_exp_dir), name=name, version=version) 751 version = f"version_{tensorboard_logger.version}" 752 os.environ[NEMO_ENV_VARNAME_VERSION] = "" if version is None else version 753 754 log_dir = Path(_exp_dir) / Path(str(name)) / Path("" if version is None else str(version)) 755 return log_dir, str(_exp_dir), name, version 756 757 758 def get_git_hash(): 759 """ 760 Helper function that tries to get the commit hash if running inside a git folder 761 762 returns: 763 Bool: Whether the git subprocess ran without error 764 str: git subprocess output or error message 765 """ 766 try: 767 return ( 768 True, 769 subprocess.check_output(['git', 'rev-parse', 'HEAD'], stderr=subprocess.STDOUT).decode(), 770 ) 771 except subprocess.CalledProcessError as err: 772 return False, "{}\n".format(err.output.decode("utf-8")) 773 774 775 def get_git_diff(): 776 """ 777 Helper function that tries to get the git diff if running inside a git folder 778 779 returns: 780 Bool: Whether the git subprocess ran without error 781 str: git subprocess output or error message 782 """ 783 try: 784 return subprocess.check_output(['git', 'diff'], stderr=subprocess.STDOUT).decode() 785 except subprocess.CalledProcessError as err: 786 return "{}\n".format(err.output.decode("utf-8")) 787 788 789 def configure_loggers( 790 trainer: 'pytorch_lightning.Trainer', 791 exp_dir: [Path, str], 792 log_dir: [Path, str], 793 name: str, 794 version: str, 795 checkpoint_callback_params: dict, 796 create_tensorboard_logger: bool, 797 summary_writer_kwargs: dict, 798 create_wandb_logger: bool, 799 wandb_kwargs: dict, 800 create_mlflow_logger: bool, 801 mlflow_kwargs: dict, 802 create_dllogger_logger: bool, 803 dllogger_kwargs: dict, 804 create_clearml_logger: bool, 805 clearml_kwargs: dict, 806 ): 807 """ 808 Creates TensorboardLogger and/or WandBLogger / MLFlowLogger / DLlogger / ClearMLLogger and attach them to trainer. 809 Raises ValueError if summary_writer_kwargs or wandb_kwargs are misconfigured. 810 """ 811 # Potentially create tensorboard logger and/or WandBLogger / MLFlowLogger / DLLogger 812 logger_list = [] 813 if create_tensorboard_logger: 814 if summary_writer_kwargs is None: 815 summary_writer_kwargs = {} 816 elif "log_dir" in summary_writer_kwargs: 817 raise ValueError( 818 "You cannot pass `log_dir` as part of `summary_writer_kwargs`. `log_dir` is handled by lightning's " 819 "TensorBoardLogger logger." 820 ) 821 tensorboard_logger = TensorBoardLogger(save_dir=exp_dir, name=name, version=version, **summary_writer_kwargs) 822 logger_list.append(tensorboard_logger) 823 logging.info("TensorboardLogger has been set up") 824 825 if create_wandb_logger: 826 if wandb_kwargs is None: 827 wandb_kwargs = {} 828 if "name" not in wandb_kwargs and "project" not in wandb_kwargs: 829 raise ValueError("name and project are required for wandb_logger") 830 831 # Update the wandb save_dir 832 if wandb_kwargs.get('save_dir', None) is None: 833 wandb_kwargs['save_dir'] = exp_dir 834 os.makedirs(wandb_kwargs['save_dir'], exist_ok=True) 835 wandb_logger = WandbLogger(version=version, **wandb_kwargs) 836 837 logger_list.append(wandb_logger) 838 logging.info("WandBLogger has been set up") 839 840 if create_mlflow_logger: 841 mlflow_logger = MLFlowLogger(run_name=version, **mlflow_kwargs) 842 843 logger_list.append(mlflow_logger) 844 logging.info("MLFlowLogger has been set up") 845 846 if create_dllogger_logger: 847 dllogger_logger = DLLogger(**dllogger_kwargs) 848 849 logger_list.append(dllogger_logger) 850 logging.info("DLLogger has been set up") 851 852 if create_clearml_logger: 853 clearml_logger = ClearMLLogger( 854 clearml_cfg=clearml_kwargs, 855 log_dir=log_dir, 856 prefix=name, 857 save_best_model=checkpoint_callback_params.save_best_model, 858 ) 859 860 logger_list.append(clearml_logger) 861 logging.info("ClearMLLogger has been set up") 862 863 trainer._logger_connector.configure_logger(logger_list) 864 865 866 def configure_checkpointing( 867 trainer: 'pytorch_lightning.Trainer', 868 log_dir: Path, 869 name: str, 870 resume: bool, 871 params: 'DictConfig', 872 create_preemption_callback: bool, 873 ): 874 """ Adds ModelCheckpoint to trainer. Raises CheckpointMisconfigurationError if trainer already has a ModelCheckpoint 875 callback 876 """ 877 for callback in trainer.callbacks: 878 if isinstance(callback, ModelCheckpoint): 879 raise CheckpointMisconfigurationError( 880 "The pytorch lightning trainer that was passed to exp_manager contained a ModelCheckpoint " 881 "and create_checkpoint_callback was set to True. Please either set create_checkpoint_callback " 882 "to False, or remove ModelCheckpoint from the lightning trainer" 883 ) 884 # Create the callback and attach it to trainer 885 if "filepath" in params: 886 if params.filepath is not None: 887 logging.warning("filepath is deprecated. Please switch to dirpath and filename instead") 888 if params.dirpath is None: 889 params.dirpath = Path(params.filepath).parent 890 if params.filename is None: 891 params.filename = Path(params.filepath).name 892 with open_dict(params): 893 del params["filepath"] 894 if params.dirpath is None: 895 params.dirpath = Path(log_dir / 'checkpoints') 896 if params.filename is None: 897 params.filename = f'{name}--{{{params.monitor}:.4f}}-{{epoch}}' 898 if params.prefix is None: 899 params.prefix = name 900 NeMoModelCheckpoint.CHECKPOINT_NAME_LAST = params.filename + '-last' 901 902 logging.debug(params.dirpath) 903 logging.debug(params.filename) 904 logging.debug(params.prefix) 905 906 if "val" in params.monitor: 907 if ( 908 trainer.max_epochs is not None 909 and trainer.max_epochs != -1 910 and trainer.max_epochs < trainer.check_val_every_n_epoch 911 ): 912 logging.error( 913 "The checkpoint callback was told to monitor a validation value but trainer.max_epochs(" 914 f"{trainer.max_epochs}) was less than trainer.check_val_every_n_epoch({trainer.check_val_every_n_epoch}" 915 f"). It is very likely this run will fail with ModelCheckpoint(monitor='{params.monitor}') not found " 916 "in the returned metrics. Please ensure that validation is run within trainer.max_epochs." 917 ) 918 elif trainer.max_steps is not None and trainer.max_steps != -1: 919 logging.warning( 920 "The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to " 921 f"{trainer.max_steps}. Please ensure that max_steps will run for at least " 922 f"{trainer.check_val_every_n_epoch} epochs to ensure that checkpointing will not error out." 923 ) 924 925 checkpoint_callback = NeMoModelCheckpoint(n_resume=resume, **params) 926 checkpoint_callback.last_model_path = trainer.ckpt_path or "" 927 if 'mp_rank' in checkpoint_callback.last_model_path or 'tp_rank' in checkpoint_callback.last_model_path: 928 checkpoint_callback.last_model_path = uninject_model_parallel_rank(checkpoint_callback.last_model_path) 929 trainer.callbacks.append(checkpoint_callback) 930 if create_preemption_callback: 931 # Check if cuda is avialable as preemption is supported only on GPUs 932 if torch.cuda.is_available(): 933 ## By default PreemptionCallback handles SIGTERM. To handle other signals pass the signal in the call as below: 934 ## PreemptionCallback(checkpoint_callback, signal.SIGCHLD) 935 preemption_callback = PreemptionCallback(checkpoint_callback) 936 trainer.callbacks.append(preemption_callback) 937 else: 938 logging.info("Preemption is supported only on GPUs, disabling preemption") 939 940 941 def check_slurm(trainer): 942 try: 943 return trainer.accelerator_connector.is_slurm_managing_tasks 944 except AttributeError: 945 return False 946 947 948 class StatelessTimer(Timer): 949 """Extension of PTL timers to be per run.""" 950 951 def __init__(self, duration: timedelta = None, interval: str = Interval.step, verbose: bool = True,) -> None: 952 super().__init__(duration, interval, verbose) 953 954 # Override PTL Timer's state dict to not store elapsed time information so that we can restore and continue training. 955 def state_dict(self) -> Dict[str, Any]: 956 return {} 957 958 def load_state_dict(self, state_dict: Dict[str, Any]) -> None: 959 return 960 961 962 def configure_no_restart_validation_training_loop(trainer: pytorch_lightning.Trainer) -> None: 963 if type(trainer.fit_loop.epoch_loop) != _TrainingEpochLoop: 964 warnings.warn("Detected custom epoch loop. Skipping no validation on restart support.", UserWarning) 965 return 966 ## Pass trainer object to avoid trainer getting overwritten as None 967 loop = SkipResumeTrainingValidationLoop(trainer, trainer.min_steps, trainer.max_steps) 968 trainer.fit_loop.epoch_loop = loop 969 970 971 class SkipResumeTrainingValidationLoop(_TrainingEpochLoop): 972 """ 973 Extend the PTL Epoch loop to skip validating when resuming. 974 This happens when resuming a checkpoint that has already run validation, but loading restores 975 the training state before validation has run. 976 """ 977 978 def _should_check_val_fx(self) -> bool: 979 if self.restarting and self.global_step % self.trainer.val_check_batch == 0: 980 return False 981 return super()._should_check_val_fx() 982 983 984 def clean_exp_ckpt(exp_log_dir: Union[str, Path], remove_ckpt: bool = True, remove_nemo: bool = False): 985 """ 986 Helper method that removes Pytorch Lightning .ckpt files or NeMo .nemo files from the checkpoint directory 987 988 Args: 989 exp_log_dir: str path to the root directory of the current experiment. 990 remove_ckpt: bool, whether to remove all *.ckpt files in the checkpoints directory. 991 remove_nemo: bool, whether to remove all *.nemo files in the checkpoints directory. 992 """ 993 exp_log_dir = str(exp_log_dir) 994 995 if remove_ckpt: 996 logging.info("Deleting *.ckpt files ...") 997 ckpt_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.ckpt")) 998 for filepath in ckpt_files: 999 os.remove(filepath) 1000 logging.info(f"Deleted file : {filepath}") 1001 1002 if remove_nemo: 1003 logging.info("Deleting *.nemo files ...") 1004 nemo_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.nemo")) 1005 for filepath in nemo_files: 1006 os.remove(filepath) 1007 logging.info(f"Deleted file : {filepath}") 1008 [end of nemo/utils/exp_manager.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR model with CTC decoder. To evaluate a model with 19 # Transducer (RNN-T) decoder use another script 'scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py'. 20 # NeMo's beam search decoders are capable of using the KenLM's N-gram models 21 # to find the best candidates. This script supports both character level and BPE level 22 # encodings and models which is detected automatically from the type of the model. 23 # You may train the LM model with 'scripts/asr_language_modeling/ngram_lm/train_kenlm.py'. 24 25 # Config Help 26 27 To discover all arguments of the script, please run : 28 python eval_beamsearch_ngram.py --help 29 python eval_beamsearch_ngram.py --cfg job 30 31 # USAGE 32 33 python eval_beamsearch_ngram.py nemo_model_file=<path to the .nemo file of the model> \ 34 input_manifest=<path to the evaluation JSON manifest file> \ 35 kenlm_model_file=<path to the binary KenLM model> \ 36 beam_width=[<list of the beam widths, separated with commas>] \ 37 beam_alpha=[<list of the beam alphas, separated with commas>] \ 38 beam_beta=[<list of the beam betas, separated with commas>] \ 39 preds_output_folder=<optional folder to store the predictions> \ 40 probs_cache_file=null \ 41 decoding_mode=beamsearch_ngram 42 ... 43 44 45 # Grid Search for Hyper parameters 46 47 For grid search, you can provide a list of arguments as follows - 48 49 beam_width=[4,8,16,....] \ 50 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 51 beam_beta=[-1.0,-0.5,0.0,...,1.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 from dataclasses import dataclass, field, is_dataclass 64 from pathlib import Path 65 from typing import List, Optional 66 67 import editdistance 68 import numpy as np 69 import torch 70 from omegaconf import MISSING, OmegaConf 71 from sklearn.model_selection import ParameterGrid 72 from tqdm.auto import tqdm 73 74 import nemo.collections.asr as nemo_asr 75 from nemo.collections.asr.models import EncDecHybridRNNTCTCModel 76 from nemo.collections.asr.parts.submodules import ctc_beam_decoding 77 from nemo.collections.asr.parts.utils.transcribe_utils import PunctuationCapitalization, TextProcessingConfig 78 from nemo.core.config import hydra_runner 79 from nemo.utils import logging 80 81 # fmt: off 82 83 84 @dataclass 85 class EvalBeamSearchNGramConfig: 86 """ 87 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 88 """ 89 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 90 nemo_model_file: str = MISSING 91 92 # File paths 93 input_manifest: str = MISSING # The manifest file of the evaluation set 94 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 95 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 96 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 97 98 # Parameters for inference 99 acoustic_batch_size: int = 16 # The batch size to calculate log probabilities 100 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 101 device: str = "cuda" # The device to load the model onto to calculate log probabilities 102 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 103 104 # Beam Search hyperparameters 105 106 # The decoding scheme to be used for evaluation. 107 # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] 108 decoding_mode: str = "beamsearch_ngram" 109 110 beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding 111 beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding 112 beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding 113 114 decoding_strategy: str = "beam" 115 decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) 116 117 text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( 118 punctuation_marks = ".,?", 119 separate_punctuation = False, 120 do_lowercase = False, 121 rm_punctuation = False, 122 ) 123 # fmt: on 124 125 126 def beam_search_eval( 127 model: nemo_asr.models.ASRModel, 128 cfg: EvalBeamSearchNGramConfig, 129 all_probs: List[torch.Tensor], 130 target_transcripts: List[str], 131 preds_output_file: str = None, 132 lm_path: str = None, 133 beam_alpha: float = 1.0, 134 beam_beta: float = 0.0, 135 beam_width: int = 128, 136 beam_batch_size: int = 128, 137 progress_bar: bool = True, 138 punctuation_capitalization: PunctuationCapitalization = None, 139 ): 140 level = logging.getEffectiveLevel() 141 logging.setLevel(logging.CRITICAL) 142 # Reset config 143 model.change_decoding_strategy(None) 144 145 # Override the beam search config with current search candidate configuration 146 cfg.decoding.beam_size = beam_width 147 cfg.decoding.beam_alpha = beam_alpha 148 cfg.decoding.beam_beta = beam_beta 149 cfg.decoding.return_best_hypothesis = False 150 cfg.decoding.kenlm_path = cfg.kenlm_model_file 151 152 # Update model's decoding strategy config 153 model.cfg.decoding.strategy = cfg.decoding_strategy 154 model.cfg.decoding.beam = cfg.decoding 155 156 # Update model's decoding strategy 157 if isinstance(model, EncDecHybridRNNTCTCModel): 158 model.change_decoding_strategy(model.cfg.decoding, decoder_type='ctc') 159 decoding = model.ctc_decoding 160 else: 161 model.change_decoding_strategy(model.cfg.decoding) 162 decoding = model.decoding 163 logging.setLevel(level) 164 165 wer_dist_first = cer_dist_first = 0 166 wer_dist_best = cer_dist_best = 0 167 words_count = 0 168 chars_count = 0 169 sample_idx = 0 170 if preds_output_file: 171 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 172 173 if progress_bar: 174 it = tqdm( 175 range(int(np.ceil(len(all_probs) / beam_batch_size))), 176 desc=f"Beam search decoding with width={beam_width}, alpha={beam_alpha}, beta={beam_beta}", 177 ncols=120, 178 ) 179 else: 180 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 181 for batch_idx in it: 182 # disabling type checking 183 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 184 probs_lens = torch.tensor([prob.shape[0] for prob in probs_batch]) 185 with torch.no_grad(): 186 packed_batch = torch.zeros(len(probs_batch), max(probs_lens), probs_batch[0].shape[-1], device='cpu') 187 188 for prob_index in range(len(probs_batch)): 189 packed_batch[prob_index, : probs_lens[prob_index], :] = torch.tensor( 190 probs_batch[prob_index], device=packed_batch.device, dtype=packed_batch.dtype 191 ) 192 193 _, beams_batch = decoding.ctc_decoder_predictions_tensor( 194 packed_batch, decoder_lengths=probs_lens, return_hypotheses=True, 195 ) 196 197 for beams_idx, beams in enumerate(beams_batch): 198 target = target_transcripts[sample_idx + beams_idx] 199 target_split_w = target.split() 200 target_split_c = list(target) 201 words_count += len(target_split_w) 202 chars_count += len(target_split_c) 203 wer_dist_min = cer_dist_min = 10000 204 for candidate_idx, candidate in enumerate(beams): # type: (int, ctc_beam_decoding.rnnt_utils.Hypothesis) 205 pred_text = candidate.text 206 if cfg.text_processing.do_lowercase: 207 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 208 if cfg.text_processing.rm_punctuation: 209 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 210 if cfg.text_processing.separate_punctuation: 211 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 212 pred_split_w = pred_text.split() 213 wer_dist = editdistance.eval(target_split_w, pred_split_w) 214 pred_split_c = list(pred_text) 215 cer_dist = editdistance.eval(target_split_c, pred_split_c) 216 217 wer_dist_min = min(wer_dist_min, wer_dist) 218 cer_dist_min = min(cer_dist_min, cer_dist) 219 220 if candidate_idx == 0: 221 # first candidate 222 wer_dist_first += wer_dist 223 cer_dist_first += cer_dist 224 225 score = candidate.score 226 if preds_output_file: 227 out_file.write('{}\t{}\n'.format(pred_text, score)) 228 wer_dist_best += wer_dist_min 229 cer_dist_best += cer_dist_min 230 sample_idx += len(probs_batch) 231 232 if preds_output_file: 233 out_file.close() 234 logging.info(f"Stored the predictions of beam search decoding at '{preds_output_file}'.") 235 236 if lm_path: 237 logging.info( 238 'WER/CER with beam search decoding and N-gram model = {:.2%}/{:.2%}'.format( 239 wer_dist_first / words_count, cer_dist_first / chars_count 240 ) 241 ) 242 else: 243 logging.info( 244 'WER/CER with beam search decoding = {:.2%}/{:.2%}'.format( 245 wer_dist_first / words_count, cer_dist_first / chars_count 246 ) 247 ) 248 logging.info( 249 'Oracle WER/CER in candidates with perfect LM= {:.2%}/{:.2%}'.format( 250 wer_dist_best / words_count, cer_dist_best / chars_count 251 ) 252 ) 253 logging.info(f"=================================================================================") 254 255 return wer_dist_first / words_count, cer_dist_first / chars_count 256 257 258 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 259 def main(cfg: EvalBeamSearchNGramConfig): 260 logging.warning("This file will be renamed to eval_beamsearch_ngram_ctc.py in the future NeMo (1.21) release.") 261 if is_dataclass(cfg): 262 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 263 264 valid_decoding_modes = ["greedy", "beamsearch", "beamsearch_ngram"] 265 if cfg.decoding_mode not in valid_decoding_modes: 266 raise ValueError( 267 f"Given decoding_mode={cfg.decoding_mode} is invalid. Available options are :\n" f"{valid_decoding_modes}" 268 ) 269 270 if cfg.nemo_model_file.endswith('.nemo'): 271 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 272 else: 273 logging.warning( 274 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 275 ) 276 asr_model = nemo_asr.models.ASRModel.from_pretrained( 277 cfg.nemo_model_file, map_location=torch.device(cfg.device) 278 ) 279 280 target_transcripts = [] 281 manifest_dir = Path(cfg.input_manifest).parent 282 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 283 audio_file_paths = [] 284 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 285 data = json.loads(line) 286 audio_file = Path(data['audio_filepath']) 287 if not audio_file.is_file() and not audio_file.is_absolute(): 288 audio_file = manifest_dir / audio_file 289 target_transcripts.append(data['text']) 290 audio_file_paths.append(str(audio_file.absolute())) 291 292 punctuation_capitalization = PunctuationCapitalization(cfg.text_processing.punctuation_marks) 293 if cfg.text_processing.do_lowercase: 294 target_transcripts = punctuation_capitalization.do_lowercase(target_transcripts) 295 if cfg.text_processing.rm_punctuation: 296 target_transcripts = punctuation_capitalization.rm_punctuation(target_transcripts) 297 if cfg.text_processing.separate_punctuation: 298 target_transcripts = punctuation_capitalization.separate_punctuation(target_transcripts) 299 300 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 301 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 302 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 303 with open(cfg.probs_cache_file, 'rb') as probs_file: 304 all_probs = pickle.load(probs_file) 305 306 if len(all_probs) != len(audio_file_paths): 307 raise ValueError( 308 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 309 f"match the manifest file. You may need to delete the probabilities cached file." 310 ) 311 else: 312 313 @contextlib.contextmanager 314 def default_autocast(): 315 yield 316 317 if cfg.use_amp: 318 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 319 logging.info("AMP is enabled!\n") 320 autocast = torch.cuda.amp.autocast 321 322 else: 323 autocast = default_autocast 324 else: 325 326 autocast = default_autocast 327 328 with autocast(): 329 with torch.no_grad(): 330 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 331 asr_model.cur_decoder = 'ctc' 332 all_logits = asr_model.transcribe(audio_file_paths, batch_size=cfg.acoustic_batch_size, logprobs=True) 333 334 all_probs = all_logits 335 if cfg.probs_cache_file: 336 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 337 with open(cfg.probs_cache_file, 'wb') as f_dump: 338 pickle.dump(all_probs, f_dump) 339 340 wer_dist_greedy = 0 341 cer_dist_greedy = 0 342 words_count = 0 343 chars_count = 0 344 for batch_idx, probs in enumerate(all_probs): 345 preds = np.argmax(probs, axis=1) 346 preds_tensor = torch.tensor(preds, device='cpu').unsqueeze(0) 347 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 348 pred_text = asr_model.ctc_decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 349 else: 350 pred_text = asr_model._wer.decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 351 352 if cfg.text_processing.do_lowercase: 353 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 354 if cfg.text_processing.rm_punctuation: 355 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 356 if cfg.text_processing.separate_punctuation: 357 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 358 359 pred_split_w = pred_text.split() 360 target_split_w = target_transcripts[batch_idx].split() 361 pred_split_c = list(pred_text) 362 target_split_c = list(target_transcripts[batch_idx]) 363 364 wer_dist = editdistance.eval(target_split_w, pred_split_w) 365 cer_dist = editdistance.eval(target_split_c, pred_split_c) 366 367 wer_dist_greedy += wer_dist 368 cer_dist_greedy += cer_dist 369 words_count += len(target_split_w) 370 chars_count += len(target_split_c) 371 372 logging.info('Greedy WER/CER = {:.2%}/{:.2%}'.format(wer_dist_greedy / words_count, cer_dist_greedy / chars_count)) 373 374 asr_model = asr_model.to('cpu') 375 376 if cfg.decoding_mode == "beamsearch_ngram": 377 if not os.path.exists(cfg.kenlm_model_file): 378 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 379 lm_path = cfg.kenlm_model_file 380 else: 381 lm_path = None 382 383 # 'greedy' decoding_mode would skip the beam search decoding 384 if cfg.decoding_mode in ["beamsearch_ngram", "beamsearch"]: 385 if cfg.beam_width is None or cfg.beam_alpha is None or cfg.beam_beta is None: 386 raise ValueError("beam_width, beam_alpha and beam_beta are needed to perform beam search decoding.") 387 params = {'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'beam_beta': cfg.beam_beta} 388 hp_grid = ParameterGrid(params) 389 hp_grid = list(hp_grid) 390 391 best_wer_beam_size, best_cer_beam_size = None, None 392 best_wer_alpha, best_cer_alpha = None, None 393 best_wer_beta, best_cer_beta = None, None 394 best_wer, best_cer = 1e6, 1e6 395 396 logging.info(f"==============================Starting the beam search decoding===============================") 397 logging.info(f"Grid search size: {len(hp_grid)}") 398 logging.info(f"It may take some time...") 399 logging.info(f"==============================================================================================") 400 401 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 402 os.mkdir(cfg.preds_output_folder) 403 for hp in hp_grid: 404 if cfg.preds_output_folder: 405 preds_output_file = os.path.join( 406 cfg.preds_output_folder, 407 f"preds_out_width{hp['beam_width']}_alpha{hp['beam_alpha']}_beta{hp['beam_beta']}.tsv", 408 ) 409 else: 410 preds_output_file = None 411 412 candidate_wer, candidate_cer = beam_search_eval( 413 asr_model, 414 cfg, 415 all_probs=all_probs, 416 target_transcripts=target_transcripts, 417 preds_output_file=preds_output_file, 418 lm_path=lm_path, 419 beam_width=hp["beam_width"], 420 beam_alpha=hp["beam_alpha"], 421 beam_beta=hp["beam_beta"], 422 beam_batch_size=cfg.beam_batch_size, 423 progress_bar=True, 424 punctuation_capitalization=punctuation_capitalization, 425 ) 426 427 if candidate_cer < best_cer: 428 best_cer_beam_size = hp["beam_width"] 429 best_cer_alpha = hp["beam_alpha"] 430 best_cer_beta = hp["beam_beta"] 431 best_cer = candidate_cer 432 433 if candidate_wer < best_wer: 434 best_wer_beam_size = hp["beam_width"] 435 best_wer_alpha = hp["beam_alpha"] 436 best_wer_beta = hp["beam_beta"] 437 best_wer = candidate_wer 438 439 logging.info( 440 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 441 f'Beam alpha = {best_wer_alpha}, Beam beta = {best_wer_beta}' 442 ) 443 444 logging.info( 445 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 446 f'Beam alpha = {best_cer_alpha}, Beam beta = {best_cer_beta}' 447 ) 448 logging.info(f"=================================================================================") 449 450 451 if __name__ == '__main__': 452 main() 453 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR Transducer model. NeMo's beam search decoders are capable of using the 19 # KenLM's N-gram models to find the best candidates. This script supports both character level and BPE level 20 # encodings and models which is detected automatically from the type of the model. 21 # You may train the LM model with 'scripts/ngram_lm/train_kenlm.py'. 22 23 # Config Help 24 25 To discover all arguments of the script, please run : 26 python eval_beamsearch_ngram.py --help 27 python eval_beamsearch_ngram.py --cfg job 28 29 # USAGE 30 31 python eval_beamsearch_ngram_transducer.py nemo_model_file=<path to the .nemo file of the model> \ 32 input_manifest=<path to the evaluation JSON manifest file \ 33 kenlm_model_file=<path to the binary KenLM model> \ 34 beam_width=[<list of the beam widths, separated with commas>] \ 35 beam_alpha=[<list of the beam alphas, separated with commas>] \ 36 preds_output_folder=<optional folder to store the predictions> \ 37 probs_cache_file=null \ 38 decoding_strategy=<greedy_batch or maes decoding> 39 maes_prefix_alpha=[<list of the maes prefix alphas, separated with commas>] \ 40 maes_expansion_gamma=[<list of the maes expansion gammas, separated with commas>] \ 41 hat_subtract_ilm=<in case of HAT model: subtract internal LM or not> \ 42 hat_ilm_weight=[<in case of HAT model: list of the HAT internal LM weights, separated with commas>] \ 43 ... 44 45 46 # Grid Search for Hyper parameters 47 48 For grid search, you can provide a list of arguments as follows - 49 50 beam_width=[4,8,16,....] \ 51 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 import tempfile 64 from dataclasses import dataclass, field, is_dataclass 65 from pathlib import Path 66 from typing import List, Optional 67 68 import editdistance 69 import numpy as np 70 import torch 71 from omegaconf import MISSING, OmegaConf 72 from sklearn.model_selection import ParameterGrid 73 from tqdm.auto import tqdm 74 75 import nemo.collections.asr as nemo_asr 76 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding 77 from nemo.core.config import hydra_runner 78 from nemo.utils import logging 79 80 # fmt: off 81 82 83 @dataclass 84 class EvalBeamSearchNGramConfig: 85 """ 86 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 87 """ 88 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 89 nemo_model_file: str = MISSING 90 91 # File paths 92 input_manifest: str = MISSING # The manifest file of the evaluation set 93 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 94 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 95 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 96 97 # Parameters for inference 98 acoustic_batch_size: int = 128 # The batch size to calculate log probabilities 99 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 100 device: str = "cuda" # The device to load the model onto to calculate log probabilities 101 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 102 num_workers: int = 1 # Number of workers for DataLoader 103 104 # The decoding scheme to be used for evaluation 105 decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] 106 107 # Beam Search hyperparameters 108 beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding 109 beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding 110 111 maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding 112 maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding 113 114 # HAT related parameters (only for internal lm subtraction) 115 hat_subtract_ilm: bool = False 116 hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) 117 118 decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) 119 120 121 # fmt: on 122 123 124 def decoding_step( 125 model: nemo_asr.models.ASRModel, 126 cfg: EvalBeamSearchNGramConfig, 127 all_probs: List[torch.Tensor], 128 target_transcripts: List[str], 129 preds_output_file: str = None, 130 beam_batch_size: int = 128, 131 progress_bar: bool = True, 132 ): 133 level = logging.getEffectiveLevel() 134 logging.setLevel(logging.CRITICAL) 135 # Reset config 136 model.change_decoding_strategy(None) 137 138 cfg.decoding.hat_ilm_weight = cfg.decoding.hat_ilm_weight * cfg.hat_subtract_ilm 139 # Override the beam search config with current search candidate configuration 140 cfg.decoding.return_best_hypothesis = False 141 cfg.decoding.ngram_lm_model = cfg.kenlm_model_file 142 cfg.decoding.hat_subtract_ilm = cfg.hat_subtract_ilm 143 144 # Update model's decoding strategy config 145 model.cfg.decoding.strategy = cfg.decoding_strategy 146 model.cfg.decoding.beam = cfg.decoding 147 148 # Update model's decoding strategy 149 model.change_decoding_strategy(model.cfg.decoding) 150 logging.setLevel(level) 151 152 wer_dist_first = cer_dist_first = 0 153 wer_dist_best = cer_dist_best = 0 154 words_count = 0 155 chars_count = 0 156 sample_idx = 0 157 if preds_output_file: 158 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 159 160 if progress_bar: 161 if cfg.decoding_strategy == "greedy_batch": 162 description = "Greedy_batch decoding.." 163 else: 164 description = f"{cfg.decoding_strategy} decoding with bw={cfg.decoding.beam_size}, ba={cfg.decoding.ngram_lm_alpha}, ma={cfg.decoding.maes_prefix_alpha}, mg={cfg.decoding.maes_expansion_gamma}, hat_ilmw={cfg.decoding.hat_ilm_weight}" 165 it = tqdm(range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=description, ncols=120) 166 else: 167 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 168 for batch_idx in it: 169 # disabling type checking 170 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 171 probs_lens = torch.tensor([prob.shape[-1] for prob in probs_batch]) 172 with torch.no_grad(): 173 packed_batch = torch.zeros(len(probs_batch), probs_batch[0].shape[0], max(probs_lens), device='cpu') 174 175 for prob_index in range(len(probs_batch)): 176 packed_batch[prob_index, :, : probs_lens[prob_index]] = torch.tensor( 177 probs_batch[prob_index].unsqueeze(0), device=packed_batch.device, dtype=packed_batch.dtype 178 ) 179 best_hyp_batch, beams_batch = model.decoding.rnnt_decoder_predictions_tensor( 180 packed_batch, probs_lens, return_hypotheses=True, 181 ) 182 if cfg.decoding_strategy == "greedy_batch": 183 beams_batch = [[x] for x in best_hyp_batch] 184 185 for beams_idx, beams in enumerate(beams_batch): 186 target = target_transcripts[sample_idx + beams_idx] 187 target_split_w = target.split() 188 target_split_c = list(target) 189 words_count += len(target_split_w) 190 chars_count += len(target_split_c) 191 wer_dist_min = cer_dist_min = 10000 192 for candidate_idx, candidate in enumerate(beams): # type: (int, rnnt_beam_decoding.rnnt_utils.Hypothesis) 193 pred_text = candidate.text 194 pred_split_w = pred_text.split() 195 wer_dist = editdistance.eval(target_split_w, pred_split_w) 196 pred_split_c = list(pred_text) 197 cer_dist = editdistance.eval(target_split_c, pred_split_c) 198 199 wer_dist_min = min(wer_dist_min, wer_dist) 200 cer_dist_min = min(cer_dist_min, cer_dist) 201 202 if candidate_idx == 0: 203 # first candidate 204 wer_dist_first += wer_dist 205 cer_dist_first += cer_dist 206 207 score = candidate.score 208 if preds_output_file: 209 out_file.write('{}\t{}\n'.format(pred_text, score)) 210 wer_dist_best += wer_dist_min 211 cer_dist_best += cer_dist_min 212 sample_idx += len(probs_batch) 213 214 if cfg.decoding_strategy == "greedy_batch": 215 return wer_dist_first / words_count, cer_dist_first / chars_count 216 217 if preds_output_file: 218 out_file.close() 219 logging.info(f"Stored the predictions of {cfg.decoding_strategy} decoding at '{preds_output_file}'.") 220 221 if cfg.decoding.ngram_lm_model: 222 logging.info( 223 f"WER/CER with {cfg.decoding_strategy} decoding and N-gram model = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 224 ) 225 else: 226 logging.info( 227 f"WER/CER with {cfg.decoding_strategy} decoding = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 228 ) 229 logging.info( 230 f"Oracle WER/CER in candidates with perfect LM= {wer_dist_best / words_count:.2%}/{cer_dist_best / chars_count:.2%}" 231 ) 232 logging.info(f"=================================================================================") 233 234 return wer_dist_first / words_count, cer_dist_first / chars_count 235 236 237 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 238 def main(cfg: EvalBeamSearchNGramConfig): 239 if is_dataclass(cfg): 240 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 241 242 valid_decoding_strategis = ["greedy_batch", "beam", "tsd", "alsd", "maes"] 243 if cfg.decoding_strategy not in valid_decoding_strategis: 244 raise ValueError( 245 f"Given decoding_strategy={cfg.decoding_strategy} is invalid. Available options are :\n" 246 f"{valid_decoding_strategis}" 247 ) 248 249 if cfg.nemo_model_file.endswith('.nemo'): 250 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 251 else: 252 logging.warning( 253 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 254 ) 255 asr_model = nemo_asr.models.ASRModel.from_pretrained( 256 cfg.nemo_model_file, map_location=torch.device(cfg.device) 257 ) 258 259 if cfg.kenlm_model_file: 260 if not os.path.exists(cfg.kenlm_model_file): 261 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 262 if cfg.decoding_strategy != "maes": 263 raise ValueError(f"Decoding with kenlm model is supported only for maes decoding algorithm.") 264 lm_path = cfg.kenlm_model_file 265 else: 266 lm_path = None 267 cfg.beam_alpha = [0.0] 268 if cfg.hat_subtract_ilm: 269 assert lm_path, "kenlm must be set for hat internal lm subtraction" 270 271 if cfg.decoding_strategy != "maes": 272 cfg.maes_prefix_alpha, cfg.maes_expansion_gamma, cfg.hat_ilm_weight = [0], [0], [0] 273 274 target_transcripts = [] 275 manifest_dir = Path(cfg.input_manifest).parent 276 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 277 audio_file_paths = [] 278 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 279 data = json.loads(line) 280 audio_file = Path(data['audio_filepath']) 281 if not audio_file.is_file() and not audio_file.is_absolute(): 282 audio_file = manifest_dir / audio_file 283 target_transcripts.append(data['text']) 284 audio_file_paths.append(str(audio_file.absolute())) 285 286 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 287 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 288 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 289 with open(cfg.probs_cache_file, 'rb') as probs_file: 290 all_probs = pickle.load(probs_file) 291 292 if len(all_probs) != len(audio_file_paths): 293 raise ValueError( 294 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 295 f"match the manifest file. You may need to delete the probabilities cached file." 296 ) 297 else: 298 299 @contextlib.contextmanager 300 def default_autocast(): 301 yield 302 303 if cfg.use_amp: 304 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 305 logging.info("AMP is enabled!\n") 306 autocast = torch.cuda.amp.autocast 307 308 else: 309 autocast = default_autocast 310 else: 311 312 autocast = default_autocast 313 314 # manual calculation of encoder_embeddings 315 with autocast(): 316 with torch.no_grad(): 317 asr_model.eval() 318 asr_model.encoder.freeze() 319 device = next(asr_model.parameters()).device 320 all_probs = [] 321 with tempfile.TemporaryDirectory() as tmpdir: 322 with open(os.path.join(tmpdir, 'manifest.json'), 'w', encoding='utf-8') as fp: 323 for audio_file in audio_file_paths: 324 entry = {'audio_filepath': audio_file, 'duration': 100000, 'text': ''} 325 fp.write(json.dumps(entry) + '\n') 326 config = { 327 'paths2audio_files': audio_file_paths, 328 'batch_size': cfg.acoustic_batch_size, 329 'temp_dir': tmpdir, 330 'num_workers': cfg.num_workers, 331 'channel_selector': None, 332 'augmentor': None, 333 } 334 temporary_datalayer = asr_model._setup_transcribe_dataloader(config) 335 for test_batch in tqdm(temporary_datalayer, desc="Transcribing", disable=True): 336 encoded, encoded_len = asr_model.forward( 337 input_signal=test_batch[0].to(device), input_signal_length=test_batch[1].to(device) 338 ) 339 # dump encoder embeddings per file 340 for idx in range(encoded.shape[0]): 341 encoded_no_pad = encoded[idx, :, : encoded_len[idx]] 342 all_probs.append(encoded_no_pad) 343 344 if cfg.probs_cache_file: 345 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 346 with open(cfg.probs_cache_file, 'wb') as f_dump: 347 pickle.dump(all_probs, f_dump) 348 349 if cfg.decoding_strategy == "greedy_batch": 350 asr_model = asr_model.to('cpu') 351 candidate_wer, candidate_cer = decoding_step( 352 asr_model, 353 cfg, 354 all_probs=all_probs, 355 target_transcripts=target_transcripts, 356 beam_batch_size=cfg.beam_batch_size, 357 progress_bar=True, 358 ) 359 logging.info(f"Greedy batch WER/CER = {candidate_wer:.2%}/{candidate_cer:.2%}") 360 361 asr_model = asr_model.to('cpu') 362 363 # 'greedy_batch' decoding_strategy would skip the beam search decoding 364 if cfg.decoding_strategy in ["beam", "tsd", "alsd", "maes"]: 365 if cfg.beam_width is None or cfg.beam_alpha is None: 366 raise ValueError("beam_width and beam_alpha are needed to perform beam search decoding.") 367 params = { 368 'beam_width': cfg.beam_width, 369 'beam_alpha': cfg.beam_alpha, 370 'maes_prefix_alpha': cfg.maes_prefix_alpha, 371 'maes_expansion_gamma': cfg.maes_expansion_gamma, 372 'hat_ilm_weight': cfg.hat_ilm_weight, 373 } 374 hp_grid = ParameterGrid(params) 375 hp_grid = list(hp_grid) 376 377 best_wer_beam_size, best_cer_beam_size = None, None 378 best_wer_alpha, best_cer_alpha = None, None 379 best_wer, best_cer = 1e6, 1e6 380 381 logging.info( 382 f"==============================Starting the {cfg.decoding_strategy} decoding===============================" 383 ) 384 logging.info(f"Grid search size: {len(hp_grid)}") 385 logging.info(f"It may take some time...") 386 logging.info(f"==============================================================================================") 387 388 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 389 os.mkdir(cfg.preds_output_folder) 390 for hp in hp_grid: 391 if cfg.preds_output_folder: 392 results_file = f"preds_out_{cfg.decoding_strategy}_bw{hp['beam_width']}" 393 if cfg.decoding_strategy == "maes": 394 results_file = f"{results_file}_ma{hp['maes_prefix_alpha']}_mg{hp['maes_expansion_gamma']}" 395 if cfg.kenlm_model_file: 396 results_file = f"{results_file}_ba{hp['beam_alpha']}" 397 if cfg.hat_subtract_ilm: 398 results_file = f"{results_file}_hat_ilmw{hp['hat_ilm_weight']}" 399 preds_output_file = os.path.join(cfg.preds_output_folder, f"{results_file}.tsv") 400 else: 401 preds_output_file = None 402 403 cfg.decoding.beam_size = hp["beam_width"] 404 cfg.decoding.ngram_lm_alpha = hp["beam_alpha"] 405 cfg.decoding.maes_prefix_alpha = hp["maes_prefix_alpha"] 406 cfg.decoding.maes_expansion_gamma = hp["maes_expansion_gamma"] 407 cfg.decoding.hat_ilm_weight = hp["hat_ilm_weight"] 408 409 candidate_wer, candidate_cer = decoding_step( 410 asr_model, 411 cfg, 412 all_probs=all_probs, 413 target_transcripts=target_transcripts, 414 preds_output_file=preds_output_file, 415 beam_batch_size=cfg.beam_batch_size, 416 progress_bar=True, 417 ) 418 419 if candidate_cer < best_cer: 420 best_cer_beam_size = hp["beam_width"] 421 best_cer_alpha = hp["beam_alpha"] 422 best_cer_ma = hp["maes_prefix_alpha"] 423 best_cer_mg = hp["maes_expansion_gamma"] 424 best_cer_hat_ilm_weight = hp["hat_ilm_weight"] 425 best_cer = candidate_cer 426 427 if candidate_wer < best_wer: 428 best_wer_beam_size = hp["beam_width"] 429 best_wer_alpha = hp["beam_alpha"] 430 best_wer_ma = hp["maes_prefix_alpha"] 431 best_wer_ga = hp["maes_expansion_gamma"] 432 best_wer_hat_ilm_weight = hp["hat_ilm_weight"] 433 best_wer = candidate_wer 434 435 wer_hat_parameter = "" 436 if cfg.hat_subtract_ilm: 437 wer_hat_parameter = f"HAT ilm weight = {best_wer_hat_ilm_weight}, " 438 logging.info( 439 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 440 f'Beam alpha = {best_wer_alpha}, {wer_hat_parameter}' 441 f'maes_prefix_alpha = {best_wer_ma}, maes_expansion_gamma = {best_wer_ga} ' 442 ) 443 444 cer_hat_parameter = "" 445 if cfg.hat_subtract_ilm: 446 cer_hat_parameter = f"HAT ilm weight = {best_cer_hat_ilm_weight}" 447 logging.info( 448 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 449 f'Beam alpha = {best_cer_alpha}, {cer_hat_parameter} ' 450 f'maes_prefix_alpha = {best_cer_ma}, maes_expansion_gamma = {best_cer_mg}' 451 ) 452 logging.info(f"=================================================================================") 453 454 455 if __name__ == '__main__': 456 main() 457 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of scripts/confidence_ensembles/build_ensemble.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 This script provides a functionality to create confidence-based ensembles 17 from a collection of pretrained models. 18 19 For more details see the paper https://arxiv.org/abs/2306.15824 20 or tutorial in tutorials/asr/Confidence_Ensembles.ipynb 21 22 You would typically use this script by providing a yaml config file or overriding 23 default options from command line. 24 25 Usage examples: 26 27 1. Building an ensemble of two monolingual models with default settings (no confidence tuning). 28 29 python build_ensemble.py --config-path=. --config-name=ensemble_config.yaml 30 ensemble.0.model=stt_it_conformer_ctc_large 31 ensemble.0.training_manifest=<path to the Italian data of 100+ utterances (no transcription required)> 32 ensemble.1.model=stt_es_conformer_ctc_large 33 ensemble.1.training_manifest=<path to the Spanish data of 100+ utterances (no transcription required)> 34 output_path=<path to the desired location of the .nemo checkpoint> 35 36 You can have more than 2 models and can control transcription settings (e.g., batch size) 37 with ``transcription.<any argument of examples/asr/transcribe_speech.py>`` parameters. 38 39 2. If you want to get improved results, you can enable tuning of the confidence and logistic regression (LR) parameters. 40 E.g. 41 42 python build_ensemble.py 43 <all arguments like in the previous example> 44 ensemble.0.dev_manifest=<path to the dev data that's required for tuning> 45 ... 46 # IMPORTANT: see the note below if you use > 2 models! 47 ensemble.N.dev_manifest=<path to the dev data that's required for tuning> 48 tune_confidence=True # to allow confidence tuning. LR is tuned by default 49 50 As with any tuning, it is recommended to have reasonably large validation set for each model, 51 otherwise you might overfit to the validation data. 52 53 Note that if you add additional models (> 2) you will need to modify ensemble_config.yaml 54 or create a new one with added models in there. While it's theoretically possible to 55 fully override such parameters from commandline, hydra is very unfriendly for such 56 use-cases, so it's strongly recommended to be creating new configs. 57 58 3. If you want to precisely control tuning grid search, you can do that with 59 60 python build_ensemble.py 61 <all arguments as in the previous examples> 62 tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set 63 tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set 64 65 You can check the dataclasses in this file for the full list of supported 66 arguments and their default values. 67 """ 68 69 import atexit 70 71 # using default logging to be able to silence unnecessary messages from nemo 72 import logging 73 import os 74 import random 75 import sys 76 import tempfile 77 from copy import deepcopy 78 from dataclasses import dataclass 79 from pathlib import Path 80 from typing import Dict, List, Optional, Tuple 81 82 import joblib 83 import numpy as np 84 import pytorch_lightning as pl 85 from omegaconf import MISSING, DictConfig, OmegaConf 86 from sklearn.linear_model import LogisticRegression 87 from sklearn.metrics import confusion_matrix 88 from sklearn.pipeline import Pipeline, make_pipeline 89 from sklearn.preprocessing import StandardScaler 90 from tqdm import tqdm 91 92 from nemo.collections.asr.models.confidence_ensemble import ( 93 ConfidenceEnsembleModel, 94 ConfidenceSpec, 95 compute_confidence, 96 get_filtered_logprobs, 97 ) 98 from nemo.collections.asr.parts.utils.asr_confidence_utils import ( 99 ConfidenceConfig, 100 ConfidenceMeasureConfig, 101 get_confidence_aggregation_bank, 102 get_confidence_measure_bank, 103 ) 104 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 105 from nemo.core.config import hydra_runner 106 107 LOG = logging.getLogger(__file__) 108 109 # adding Python path. If not found, asking user to get the file 110 try: 111 sys.path.append(str(Path(__file__).parents[2] / "examples" / "asr")) 112 import transcribe_speech 113 except ImportError: 114 # if users run script normally from nemo repo, this shouldn't be triggered as 115 # we modify the path above. But if they downloaded the build_ensemble.py as 116 # an isolated script, we'd ask them to also download corresponding version 117 # of the transcribe_speech.py 118 print( 119 "Current script depends on 'examples/asr/transcribe_speech.py', but can't find it. " 120 "If it's not present, download it from the NeMo github manually and put inside this folder." 121 ) 122 123 124 @dataclass 125 class EnsembleConfig: 126 # .nemo path or pretrained name 127 model: str = MISSING 128 # path to the training data manifest (non-tarred) 129 training_manifest: str = MISSING 130 # specify to limit the number of training samples 131 # 100 is most likely enough, but setting higher default just in case 132 max_training_samples: int = 1000 133 # specify to provide dev data manifest for HP tuning 134 dev_manifest: Optional[str] = None 135 136 137 @dataclass 138 class TuneConfidenceConfig: 139 # important parameter, so should always be tuned 140 exclude_blank: Tuple[bool] = (True, False) 141 # prod is pretty much always worse, so not including by default 142 aggregation: Tuple[str] = ("mean", "min", "max") 143 # not including max prob, as there is always an entropy-based metric 144 # that's better but otherwise including everything 145 confidence_type: Tuple[str] = ( 146 "entropy_renyi_exp", 147 "entropy_renyi_lin", 148 "entropy_tsallis_exp", 149 "entropy_tsallis_lin", 150 "entropy_gibbs_lin", 151 "entropy_gibbs_exp", 152 ) 153 154 # TODO: currently it's not possible to efficiently tune temperature, as we always 155 # apply log-softmax in the decoder, so to try different values it will be required 156 # to rerun the decoding, which is very slow. To support this for one-off experiments 157 # it's possible to modify the code of CTC decoder / Transducer joint to 158 # remove log-softmax and then apply it directly in this script with the temperature 159 # 160 # Alternatively, one can run this script multiple times with different values of 161 # temperature and pick the best performing ensemble. Note that this will increase 162 # tuning time by the number of temperature values tried. On the other hand, 163 # the above approach is a lot more efficient and will only slightly increase 164 # the total tuning runtime. 165 166 # very important to tune for max prob, but for entropy metrics 1.0 is almost always best 167 # temperature: Tuple[float] = (1.0,) 168 169 # not that important, but can sometimes make a small difference 170 alpha: Tuple[float] = (0.25, 0.33, 0.5, 1.0) 171 172 def get_grid_size(self) -> int: 173 """Returns the total number of points in the search space.""" 174 if "max_prob" in self.confidence_type: 175 return ( 176 len(self.exclude_blank) 177 * len(self.aggregation) 178 * ((len(self.confidence_type) - 1) * len(self.alpha) + 1) 179 ) 180 return len(self.exclude_blank) * len(self.aggregation) * len(self.confidence_type) * len(self.alpha) 181 182 183 @dataclass 184 class TuneLogisticRegressionConfig: 185 # will have log-uniform grid over this range with that many points 186 # note that a value of 10000.0 (not regularization) is always added 187 C_num_points: int = 10 188 C_min: float = 0.0001 189 C_max: float = 10.0 190 191 # not too important 192 multi_class: Tuple[str] = ("ovr", "multinomial") 193 194 # should try to include weights directly if the data is too imbalanced 195 class_weight: Tuple = (None, "balanced") 196 197 # increase if getting many warnings that algorithm didn't converge 198 max_iter: int = 1000 199 200 201 @dataclass 202 class BuildEnsembleConfig: 203 # where to save the resulting ensemble model 204 output_path: str = MISSING 205 206 # each model specification 207 ensemble: List[EnsembleConfig] = MISSING 208 209 random_seed: int = 0 # for reproducibility 210 211 # default confidence, can override 212 confidence: ConfidenceConfig = ConfidenceConfig( 213 # we keep frame confidences and apply aggregation manually to get full-utterance confidence 214 preserve_frame_confidence=True, 215 exclude_blank=True, 216 aggregation="mean", 217 measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False 229 # used to specify what to tune over. By default runs tuning over some 230 # reasonalbe grid, so that it does not take forever. 231 # Can be changed as needed 232 tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() 233 234 # very fast to tune and can be important in case of imbalanced datasets 235 # will automatically set to False if dev data is not available 236 tune_logistic_regression: bool = True 237 tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() 238 239 def __post_init__(self): 240 """Checking that if any dev data is provided, all are provided. 241 242 Will also auto-set tune_logistic_regression to False if no dev data 243 is available. 244 245 If tune_confidence is set to True (user choice) and no dev data is 246 provided, will raise an error. 247 """ 248 num_dev_data = 0 249 for ensemble_cfg in self.ensemble: 250 num_dev_data += ensemble_cfg.dev_manifest is not None 251 if num_dev_data == 0: 252 if self.tune_confidence: 253 raise ValueError("tune_confidence is set to True, but no dev data is provided") 254 LOG.info("Setting tune_logistic_regression = False since no dev data is provided") 255 self.tune_logistic_regression = False 256 return 257 258 if num_dev_data < len(self.ensemble): 259 raise ValueError( 260 "Some ensemble configs specify dev data, but some don't. Either all have to specify it or none!" 261 ) 262 263 264 def calculate_score(features: np.ndarray, labels: np.ndarray, pipe: Pipeline) -> Tuple[float, np.ndarray]: 265 """Score is always calculated as mean of the per-class scores. 266 267 This is done to account for possible class imbalances. 268 269 Args: 270 features: numpy array of features of shape [N x D], where N is the 271 number of objects (typically a total number of utterances in 272 all datasets) and D is the total number of confidence scores 273 used to train the model (typically = number of models). 274 labels: numpy array of shape [N] contatining ground-truth model indices. 275 pipe: classification pipeline (currently, standardization + logistic 276 regression). 277 278 Returns: 279 tuple: score value in [0, 1] and full classification confusion matrix. 280 """ 281 predictions = pipe.predict(features) 282 conf_m = confusion_matrix(labels, predictions) 283 score = np.diag(conf_m).sum() / conf_m.sum() 284 return score, conf_m 285 286 287 def train_model_selection( 288 training_features: np.ndarray, 289 training_labels: np.ndarray, 290 dev_features: Optional[np.ndarray] = None, 291 dev_labels: Optional[np.ndarray] = None, 292 tune_lr: bool = False, 293 tune_lr_cfg: Optional[TuneLogisticRegressionConfig] = None, 294 verbose: bool = False, 295 ) -> Tuple[Pipeline, float]: 296 """Trains model selection block with an (optional) tuning of the parameters. 297 298 Returns a pipeline consisting of feature standardization and logistic 299 regression. If tune_lr is set to True, dev features/labels will be used 300 to tune the hyperparameters of the logistic regression with the grid 301 search that's defined via ``tune_lr_cfg``. 302 303 If no tuning is requested, uses the following parameters:: 304 305 best_pipe = make_pipeline( 306 StandardScaler(), 307 LogisticRegression( 308 multi_class="multinomial", 309 C=10000.0, 310 max_iter=1000, 311 class_weight="balanced", 312 ), 313 ) 314 315 Args: 316 training_features: numpy array of features of shape [N x D], where N is 317 the number of objects (typically a total number of utterances in 318 all training datasets) and D is the total number of confidence 319 scores used to train the model (typically = number of models). 320 training_labels: numpy array of shape [N] contatining ground-truth 321 model indices. 322 dev_features: same as training, but for the validation subset. 323 dev_labels: same as training, but for the validation subset. 324 tune_lr: controls whether tuning of LR hyperparameters is performed. 325 If set to True, it's required to also provide dev features/labels. 326 tune_lr_cfg: specifies what values of LR hyperparameters to try. 327 verbose: if True, will output final training/dev scores. 328 329 Returns: 330 tuple: trained model selection pipeline, best score (or -1 if no tuning 331 was done). 332 """ 333 if not tune_lr: 334 # default parameters: C=10000.0 disables regularization 335 best_pipe = make_pipeline( 336 StandardScaler(), 337 LogisticRegression(multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced"), 338 ) 339 max_score = -1 340 else: 341 C_pms = np.append( 342 np.exp(np.linspace(np.log(tune_lr_cfg.C_min), np.log(tune_lr_cfg.C_max), tune_lr_cfg.C_num_points)), 343 10000.0, 344 ) 345 max_score = 0 346 best_pipe = None 347 for class_weight in tune_lr_cfg.class_weight: 348 for multi_class in tune_lr_cfg.multi_class: 349 for C in C_pms: 350 pipe = make_pipeline( 351 StandardScaler(), 352 LogisticRegression( 353 multi_class=multi_class, C=C, max_iter=tune_lr_cfg.max_iter, class_weight=class_weight 354 ), 355 ) 356 pipe.fit(training_features, training_labels) 357 score, confusion = calculate_score(dev_features, dev_labels, pipe) 358 if score > max_score: 359 max_score = score 360 best_pipe = pipe 361 362 best_pipe.fit(training_features, training_labels) 363 if verbose: 364 accuracy, confusion = calculate_score(training_features, training_labels, best_pipe) 365 LOG.info("Training fit accuracy: %.4f", accuracy * 100.0) 366 LOG.info("Training confusion matrix:\n%s", str(confusion)) 367 if dev_features is not None and verbose: 368 accuracy, confusion = calculate_score(dev_features, dev_labels, best_pipe) 369 LOG.info("Dev fit accuracy: %.4f", accuracy * 100.0) 370 LOG.info("Dev confusion matrix:\n%s", str(confusion)) 371 372 return best_pipe, max_score 373 374 375 def subsample_manifest(manifest_file: str, max_samples: int) -> str: 376 """Will save a subsampled version of the manifest to the same folder. 377 378 Have to save to the same folder to support relative paths. 379 380 Args: 381 manifest_file: path to the manifest file that needs subsampling. 382 max_samples: how many samples to retain. Will randomly select that 383 many lines from the manifest. 384 385 Returns: 386 str: the path to the subsampled manifest file. 387 """ 388 with open(manifest_file, "rt", encoding="utf-8") as fin: 389 lines = fin.readlines() 390 if max_samples < len(lines): 391 lines = random.sample(lines, max_samples) 392 output_file = manifest_file + "-subsampled" 393 with open(output_file, "wt", encoding="utf-8") as fout: 394 fout.write("".join(lines)) 395 return output_file 396 397 398 def cleanup_subsampled_manifests(subsampled_manifests: List[str]): 399 """Removes all generated subsamples manifests.""" 400 for manifest in subsampled_manifests: 401 os.remove(manifest) 402 403 404 def compute_all_confidences( 405 hypothesis: Hypothesis, tune_confidence_cfg: TuneConfidenceConfig 406 ) -> Dict[ConfidenceSpec, float]: 407 """Computes a set of confidence scores from a given hypothesis. 408 409 Works with the output of both CTC and Transducer decoding. 410 411 Args: 412 hypothesis: generated hypothesis as returned from the transcribe 413 method of the ASR model. 414 tune_confidence_cfg: config specifying what confidence scores to 415 compute. 416 417 Returns: 418 dict: dictionary with confidenct spec -> confidence score mapping. 419 """ 420 conf_values = {} 421 422 for exclude_blank in tune_confidence_cfg.exclude_blank: 423 filtered_logprobs = get_filtered_logprobs(hypothesis, exclude_blank) 424 vocab_size = filtered_logprobs.shape[1] 425 for aggregation in tune_confidence_cfg.aggregation: 426 aggr_func = get_confidence_aggregation_bank()[aggregation] 427 for conf_type in tune_confidence_cfg.confidence_type: 428 conf_func = get_confidence_measure_bank()[conf_type] 429 if conf_type == "max_prob": # skipping alpha in this case 430 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=1.0)).cpu().item() 431 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, 1.0)] = conf_value 432 else: 433 for alpha in tune_confidence_cfg.alpha: 434 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=alpha)).cpu().item() 435 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, alpha)] = conf_value 436 437 return conf_values 438 439 440 def find_best_confidence( 441 train_confidences: List[List[Dict[ConfidenceSpec, float]]], 442 train_labels: List[int], 443 dev_confidences: List[List[Dict[ConfidenceSpec, float]]], 444 dev_labels: List[int], 445 tune_lr: bool, 446 tune_lr_config: TuneConfidenceConfig, 447 ) -> Tuple[ConfidenceConfig, Pipeline]: 448 """Finds the best confidence configuration for model selection. 449 450 Will loop over all values in the confidence dictionary and fit the LR 451 model (optionally tuning its HPs). The best performing confidence (on the 452 dev set) will be used for the final LR model. 453 454 Args: 455 train_confidences: this is an object of type 456 ``List[List[Dict[ConfidenceSpec, float]]]``. The shape of this 457 object is [M, N, S], where 458 M: number of models 459 N: number of utterances in all training sets 460 S: number of confidence scores to try 461 462 This argument will be used to construct np.array objects for each 463 of the confidence scores with the shape [M, N] 464 465 train_labels: ground-truth labels of the correct model for each data 466 points. This is a list of size [N] 467 dev_confidences: same as training, but for the validation subset. 468 dev_labels: same as training, but for the validation subset. 469 tune_lr: controls whether tuning of LR hyperparameters is performed. 470 tune_lr_cfg: specifies what values of LR hyperparameters to try. 471 472 Returns: 473 tuple: best confidence config, best model selection pipeline 474 """ 475 max_score = 0 476 best_pipe = None 477 best_conf_spec = None 478 LOG.info("Evaluation all confidences. Total grid size: %d", len(train_confidences[0][0].keys())) 479 for conf_spec in tqdm(train_confidences[0][0].keys()): 480 cur_train_confidences = [] 481 for model_confs in train_confidences: 482 cur_train_confidences.append([]) 483 for model_conf in model_confs: 484 cur_train_confidences[-1].append(model_conf[conf_spec]) 485 cur_dev_confidences = [] 486 for model_confs in dev_confidences: 487 cur_dev_confidences.append([]) 488 for model_conf in model_confs: 489 cur_dev_confidences[-1].append(model_conf[conf_spec]) 490 # transposing with zip(*list) 491 training_features = np.array(list(zip(*cur_train_confidences))) 492 training_labels = np.array(train_labels) 493 dev_features = np.array(list(zip(*cur_dev_confidences))) 494 dev_labels = np.array(dev_labels) 495 pipe, score = train_model_selection( 496 training_features, training_labels, dev_features, dev_labels, tune_lr, tune_lr_config, 497 ) 498 if max_score < score: 499 max_score = score 500 best_pipe = pipe 501 best_conf_spec = conf_spec 502 LOG.info("Found better parameters: %s. New score: %.4f", str(conf_spec), max_score) 503 504 return best_conf_spec.to_confidence_config(), best_pipe 505 506 507 @hydra_runner(config_name="BuildEnsembleConfig", schema=BuildEnsembleConfig) 508 def main(cfg: BuildEnsembleConfig): 509 # silencing all messages from nemo/ptl to avoid dumping tons of configs to the stdout 510 logging.getLogger('pytorch_lightning').setLevel(logging.CRITICAL) 511 logging.getLogger('nemo_logger').setLevel(logging.CRITICAL) 512 LOG.info(f'Build ensemble config:\n{OmegaConf.to_yaml(cfg)}') 513 514 # to ensure post init is called 515 cfg = BuildEnsembleConfig(**cfg) 516 517 pl.seed_everything(cfg.random_seed) 518 cfg.transcription.random_seed = None # seed is already applied 519 cfg.transcription.return_transcriptions = True 520 cfg.transcription.preserve_alignment = True 521 cfg.transcription.ctc_decoding.temperature = cfg.temperature 522 cfg.transcription.rnnt_decoding.temperature = cfg.temperature 523 # this ensures that generated output is after log-softmax for consistency with CTC 524 525 train_confidences = [] 526 dev_confidences = [] 527 train_labels = [] 528 dev_labels = [] 529 530 # registering clean-up function that will hold on to this list and 531 # should clean up even if there is partial error in some of the transcribe 532 # calls 533 subsampled_manifests = [] 534 atexit.register(cleanup_subsampled_manifests, subsampled_manifests) 535 536 # note that we loop over the same config. 537 # This is intentional, as we need to run all models on all datasets 538 # this loop will do the following things: 539 # 1. Goes through each model X each training dataset 540 # 2. Computes predictions by directly calling transcribe_speech.main 541 # 3. Converts transcription to the confidence score(s) as specified in the config 542 # 4. If dev sets are provided, computes the same for them 543 # 5. Creates a list of ground-truth model indices by mapping each model 544 # to its own training dataset as specified in the config. 545 # 6. After the loop, we either run tuning over all confidence scores or 546 # directly use a single score to fit logistic regression and save the 547 # final ensemble model. 548 for model_idx, model_cfg in enumerate(cfg.ensemble): 549 train_model_confidences = [] 550 dev_model_confidences = [] 551 for data_idx, data_cfg in enumerate(cfg.ensemble): 552 if model_idx == 0: # generating subsampled manifests only one time 553 subsampled_manifests.append( 554 subsample_manifest(data_cfg.training_manifest, data_cfg.max_training_samples) 555 ) 556 subsampled_manifest = subsampled_manifests[data_idx] 557 558 if model_cfg.model.endswith(".nemo"): 559 cfg.transcription.model_path = model_cfg.model 560 else: # assuming pretrained model 561 cfg.transcription.pretrained_name = model_cfg.model 562 563 cfg.transcription.dataset_manifest = subsampled_manifest 564 565 # training 566 with tempfile.NamedTemporaryFile() as output_file: 567 cfg.transcription.output_filename = output_file.name 568 LOG.info("Transcribing training dataset %d with model %d", data_idx, model_idx) 569 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 570 LOG.info("Generating confidence scores") 571 # TODO: parallelize this loop? 572 for transcription in tqdm(transcriptions): 573 if cfg.tune_confidence: 574 train_model_confidences.append( 575 compute_all_confidences(transcription, cfg.tune_confidence_config) 576 ) 577 else: 578 train_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 579 if model_idx == 0: # labels are the same for all models 580 train_labels.append(data_idx) 581 582 # optional dev 583 if data_cfg.dev_manifest is not None: 584 cfg.transcription.dataset_manifest = data_cfg.dev_manifest 585 with tempfile.NamedTemporaryFile() as output_file: 586 cfg.transcription.output_filename = output_file.name 587 LOG.info("Transcribing dev dataset %d with model %d", data_idx, model_idx) 588 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 589 LOG.info("Generating confidence scores") 590 for transcription in tqdm(transcriptions): 591 if cfg.tune_confidence: 592 dev_model_confidences.append( 593 compute_all_confidences(transcription, cfg.tune_confidence_config) 594 ) 595 else: 596 dev_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 597 if model_idx == 0: # labels are the same for all models 598 dev_labels.append(data_idx) 599 600 train_confidences.append(train_model_confidences) 601 if dev_model_confidences: 602 dev_confidences.append(dev_model_confidences) 603 604 if cfg.tune_confidence: 605 best_confidence, model_selection_block = find_best_confidence( 606 train_confidences, 607 train_labels, 608 dev_confidences, 609 dev_labels, 610 cfg.tune_logistic_regression, 611 cfg.tune_logistic_regression_config, 612 ) 613 else: 614 best_confidence = cfg.confidence 615 # transposing with zip(*list) 616 training_features = np.array(list(zip(*train_confidences))) 617 training_labels = np.array(train_labels) 618 if dev_confidences: 619 dev_features = np.array(list(zip(*dev_confidences))) 620 dev_labels = np.array(dev_labels) 621 else: 622 dev_features = None 623 dev_labels = None 624 model_selection_block, _ = train_model_selection( 625 training_features, 626 training_labels, 627 dev_features, 628 dev_labels, 629 cfg.tune_logistic_regression, 630 cfg.tune_logistic_regression_config, 631 verbose=True, 632 ) 633 634 with tempfile.TemporaryDirectory() as tmpdir: 635 model_selection_block_path = os.path.join(tmpdir, 'model_selection_block.pkl') 636 joblib.dump(model_selection_block, model_selection_block_path) 637 638 # creating ensemble checkpoint 639 ensemble_model = ConfidenceEnsembleModel( 640 cfg=DictConfig( 641 { 642 'model_selection_block': model_selection_block_path, 643 'confidence': best_confidence, 644 'temperature': cfg.temperature, 645 'load_models': [model_cfg.model for model_cfg in cfg.ensemble], 646 } 647 ), 648 trainer=None, 649 ) 650 ensemble_model.save_to(cfg.output_path) 651 652 653 if __name__ == '__main__': 654 main() 655 [end of scripts/confidence_ensembles/build_ensemble.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import json 16 import os 17 from dataclasses import dataclass, is_dataclass 18 from pathlib import Path 19 from typing import Optional 20 21 import pytorch_lightning as pl 22 import torch 23 from omegaconf import MISSING, OmegaConf 24 from sklearn.model_selection import ParameterGrid 25 26 from nemo.collections.asr.metrics.rnnt_wer import RNNTDecodingConfig 27 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 28 from nemo.collections.asr.models import ASRModel, EncDecRNNTModel 29 from nemo.collections.asr.parts.utils.asr_confidence_benchmarking_utils import ( 30 apply_confidence_parameters, 31 run_confidence_benchmark, 32 ) 33 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig 34 from nemo.core.config import hydra_runner 35 from nemo.utils import logging 36 37 """ 38 Get confidence metrics and curve plots for a given model, dataset, and confidence parameters. 39 40 # Arguments 41 model_path: Path to .nemo ASR checkpoint 42 pretrained_name: Name of pretrained ASR model (from NGC registry) 43 dataset_manifest: Path to dataset JSON manifest file (in NeMo format) 44 output_dir: Output directory to store a report and curve plot directories 45 46 batch_size: batch size during inference 47 num_workers: number of workers during inference 48 49 cuda: Optional int to enable or disable execution of model on certain CUDA device 50 amp: Bool to decide if Automatic Mixed Precision should be used during inference 51 audio_type: Str filetype of the audio. Supported = wav, flac, mp3 52 53 target_level: Word- or token-level confidence. Supported = word, token, auto (for computing both word and token) 54 confidence_cfg: Config with confidence parameters 55 grid_params: Dictionary with lists of parameters to iteratively benchmark on 56 57 # Usage 58 ASR model can be specified by either "model_path" or "pretrained_name". 59 Data for transcription are defined with "dataset_manifest". 60 Results are returned as a benchmark report and curve plots. 61 62 python benchmark_asr_confidence.py \ 63 model_path=null \ 64 pretrained_name=null \ 65 dataset_manifest="" \ 66 output_dir="" \ 67 batch_size=64 \ 68 num_workers=8 \ 69 cuda=0 \ 70 amp=True \ 71 target_level="word" \ 72 confidence_cfg.exclude_blank=False \ 73 'grid_params="{\"aggregation\": [\"min\", \"prod\"], \"alpha\": [0.33, 0.5]}"' 74 """ 75 76 77 def get_experiment_params(cfg): 78 """Get experiment parameters from a confidence config and generate the experiment name. 79 80 Returns: 81 List of experiment parameters. 82 String with the experiment name. 83 """ 84 blank = "no_blank" if cfg.exclude_blank else "blank" 85 aggregation = cfg.aggregation 86 method_name = cfg.measure_cfg.name 87 alpha = cfg.measure_cfg.alpha 88 if method_name == "entropy": 89 entropy_type = cfg.measure_cfg.entropy_type 90 entropy_norm = cfg.measure_cfg.entropy_norm 91 experiment_param_list = [ 92 aggregation, 93 str(cfg.exclude_blank), 94 method_name, 95 entropy_type, 96 entropy_norm, 97 str(alpha), 98 ] 99 experiment_str = "-".join([aggregation, blank, method_name, entropy_type, entropy_norm, str(alpha)]) 100 else: 101 experiment_param_list = [aggregation, str(cfg.exclude_blank), method_name, "-", "-", str(alpha)] 102 experiment_str = "-".join([aggregation, blank, method_name, str(alpha)]) 103 return experiment_param_list, experiment_str 104 105 106 @dataclass 107 class ConfidenceBenchmarkingConfig: 108 # Required configs 109 model_path: Optional[str] = None # Path to a .nemo file 110 pretrained_name: Optional[str] = None # Name of a pretrained model 111 dataset_manifest: str = MISSING 112 output_dir: str = MISSING 113 114 # General configs 115 batch_size: int = 32 116 num_workers: int = 4 117 118 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA 119 # device anyway, and do inference on CPU only if CUDA device is not found. 120 # If `cuda` is a negative number, inference will be on CPU only. 121 cuda: Optional[int] = None 122 amp: bool = False 123 audio_type: str = "wav" 124 125 # Confidence configs 126 target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) 127 confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) 128 grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on 129 130 131 @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) 132 def main(cfg: ConfidenceBenchmarkingConfig): 133 torch.set_grad_enabled(False) 134 135 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 136 137 if is_dataclass(cfg): 138 cfg = OmegaConf.structured(cfg) 139 140 if cfg.model_path is None and cfg.pretrained_name is None: 141 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None!") 142 143 # setup GPU 144 if cfg.cuda is None: 145 if torch.cuda.is_available(): 146 device = [0] # use 0th CUDA device 147 accelerator = 'gpu' 148 else: 149 device = 1 150 accelerator = 'cpu' 151 else: 152 device = [cfg.cuda] 153 accelerator = 'gpu' 154 155 map_location = torch.device('cuda:{}'.format(device[0]) if accelerator == 'gpu' else 'cpu') 156 157 # setup model 158 if cfg.model_path is not None: 159 # restore model from .nemo file path 160 model_cfg = ASRModel.restore_from(restore_path=cfg.model_path, return_config=True) 161 classpath = model_cfg.target # original class path 162 imported_class = model_utils.import_class_by_path(classpath) # type: ASRModel 163 logging.info(f"Restoring model : {imported_class.__name__}") 164 asr_model = imported_class.restore_from( 165 restore_path=cfg.model_path, map_location=map_location 166 ) # type: ASRModel 167 else: 168 # restore model by name 169 asr_model = ASRModel.from_pretrained( 170 model_name=cfg.pretrained_name, map_location=map_location 171 ) # type: ASRModel 172 173 trainer = pl.Trainer(devices=device, accelerator=accelerator) 174 asr_model.set_trainer(trainer) 175 asr_model = asr_model.eval() 176 177 # Check if ctc or rnnt model 178 is_rnnt = isinstance(asr_model, EncDecRNNTModel) 179 180 # Check that the model has the `change_decoding_strategy` method 181 if not hasattr(asr_model, 'change_decoding_strategy'): 182 raise RuntimeError("The asr_model you are using must have the `change_decoding_strategy` method.") 183 184 # get filenames and reference texts from manifest 185 filepaths = [] 186 reference_texts = [] 187 if os.stat(cfg.dataset_manifest).st_size == 0: 188 logging.error(f"The input dataset_manifest {cfg.dataset_manifest} is empty. Exiting!") 189 return None 190 manifest_dir = Path(cfg.dataset_manifest).parent 191 with open(cfg.dataset_manifest, 'r') as f: 192 for line in f: 193 item = json.loads(line) 194 audio_file = Path(item['audio_filepath']) 195 if not audio_file.is_file() and not audio_file.is_absolute(): 196 audio_file = manifest_dir / audio_file 197 filepaths.append(str(audio_file.absolute())) 198 reference_texts.append(item['text']) 199 200 # setup AMP (optional) 201 autocast = None 202 if cfg.amp and torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 203 logging.info("AMP enabled!\n") 204 autocast = torch.cuda.amp.autocast 205 206 # do grid-based benchmarking if grid_params is provided, otherwise a regular one 207 work_dir = Path(cfg.output_dir) 208 os.makedirs(work_dir, exist_ok=True) 209 report_legend = ( 210 ",".join( 211 [ 212 "model_type", 213 "aggregation", 214 "blank", 215 "method_name", 216 "entropy_type", 217 "entropy_norm", 218 "alpha", 219 "target_level", 220 "auc_roc", 221 "auc_pr", 222 "auc_nt", 223 "nce", 224 "ece", 225 "auc_yc", 226 "std_yc", 227 "max_yc", 228 ] 229 ) 230 + "\n" 231 ) 232 model_typename = "RNNT" if is_rnnt else "CTC" 233 report_file = work_dir / Path("report.csv") 234 if cfg.grid_params: 235 asr_model.change_decoding_strategy( 236 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 237 if is_rnnt 238 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 239 ) 240 params = json.loads(cfg.grid_params) 241 hp_grid = ParameterGrid(params) 242 hp_grid = list(hp_grid) 243 244 logging.info(f"==============================Running a benchmarking with grid search=========================") 245 logging.info(f"Grid search size: {len(hp_grid)}") 246 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directories near the file") 247 logging.info(f"==============================================================================================") 248 249 with open(report_file, "tw", encoding="utf-8") as f: 250 f.write(report_legend) 251 f.flush() 252 for i, hp in enumerate(hp_grid): 253 logging.info(f"Run # {i + 1}, grid: `{hp}`") 254 asr_model.change_decoding_strategy(apply_confidence_parameters(asr_model.cfg.decoding, hp)) 255 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 256 plot_dir = work_dir / Path(experiment_name) 257 results = run_confidence_benchmark( 258 asr_model, 259 cfg.target_level, 260 filepaths, 261 reference_texts, 262 cfg.batch_size, 263 cfg.num_workers, 264 plot_dir, 265 autocast, 266 ) 267 for level, result in results.items(): 268 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 269 f.flush() 270 else: 271 asr_model.change_decoding_strategy( 272 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 273 if is_rnnt 274 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 275 ) 276 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 277 plot_dir = work_dir / Path(experiment_name) 278 279 logging.info(f"==============================Running a single benchmarking===================================") 280 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directory `{plot_dir}`") 281 282 with open(report_file, "tw", encoding="utf-8") as f: 283 f.write(report_legend) 284 f.flush() 285 results = run_confidence_benchmark( 286 asr_model, 287 cfg.batch_size, 288 cfg.num_workers, 289 cfg.target_level, 290 filepaths, 291 reference_texts, 292 plot_dir, 293 autocast, 294 ) 295 for level, result in results.items(): 296 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 297 logging.info(f"===========================================Done===============================================") 298 299 300 if __name__ == '__main__': 301 main() 302 [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 """ 15 # This script converts an existing audio dataset with a manifest to 16 # a tarred and sharded audio dataset that can be read by the 17 # TarredAudioToTextDataLayer. 18 19 # Please make sure your audio_filepath DOES NOT CONTAIN '-sub'! 20 # Because we will use it to handle files which have duplicate filenames but with different offsets 21 # (see function create_shard for details) 22 23 24 # Bucketing can help to improve the training speed. You may use --buckets_num to specify the number of buckets. 25 # It creates multiple tarred datasets, one per bucket, based on the audio durations. 26 # The range of [min_duration, max_duration) is split into equal sized buckets. 27 # Recommend to use --sort_in_shards to speedup the training by reducing the paddings in the batches 28 # More info on how to use bucketing feature: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/datasets.html 29 30 # If valid NVIDIA DALI version is installed, will also generate the corresponding DALI index files that need to be 31 # supplied to the config in order to utilize webdataset for efficient large dataset handling. 32 # NOTE: DALI + Webdataset is NOT compatible with Bucketing support ! 33 34 # Usage: 35 1) Creating a new tarfile dataset 36 37 python convert_to_tarred_audio_dataset.py \ 38 --manifest_path=<path to the manifest file> \ 39 --target_dir=<path to output directory> \ 40 --num_shards=<number of tarfiles that will contain the audio> \ 41 --max_duration=<float representing maximum duration of audio samples> \ 42 --min_duration=<float representing minimum duration of audio samples> \ 43 --shuffle --shuffle_seed=1 \ 44 --sort_in_shards \ 45 --workers=-1 46 47 48 2) Concatenating more tarfiles to a pre-existing tarred dataset 49 50 python convert_to_tarred_audio_dataset.py \ 51 --manifest_path=<path to the tarred manifest file> \ 52 --metadata_path=<path to the metadata.yaml (or metadata_version_{X}.yaml) file> \ 53 --target_dir=<path to output directory where the original tarfiles are contained> \ 54 --max_duration=<float representing maximum duration of audio samples> \ 55 --min_duration=<float representing minimum duration of audio samples> \ 56 --shuffle --shuffle_seed=1 \ 57 --sort_in_shards \ 58 --workers=-1 \ 59 --concat_manifest_paths \ 60 <space separated paths to 1 or more manifest files to concatenate into the original tarred dataset> 61 62 3) Writing an empty metadata file 63 64 python convert_to_tarred_audio_dataset.py \ 65 --target_dir=<path to output directory> \ 66 # any other optional argument 67 --num_shards=8 \ 68 --max_duration=16.7 \ 69 --min_duration=0.01 \ 70 --shuffle \ 71 --workers=-1 \ 72 --sort_in_shards \ 73 --shuffle_seed=1 \ 74 --write_metadata 75 76 """ 77 import argparse 78 import copy 79 import json 80 import os 81 import random 82 import tarfile 83 from collections import defaultdict 84 from dataclasses import dataclass, field 85 from datetime import datetime 86 from typing import Any, List, Optional 87 88 from joblib import Parallel, delayed 89 from omegaconf import DictConfig, OmegaConf, open_dict 90 91 try: 92 import create_dali_tarred_dataset_index as dali_index 93 94 DALI_INDEX_SCRIPT_AVAILABLE = True 95 except (ImportError, ModuleNotFoundError, FileNotFoundError): 96 DALI_INDEX_SCRIPT_AVAILABLE = False 97 98 parser = argparse.ArgumentParser( 99 description="Convert an existing ASR dataset to tarballs compatible with TarredAudioToTextDataLayer." 100 ) 101 parser.add_argument( 102 "--manifest_path", default=None, type=str, required=False, help="Path to the existing dataset's manifest." 103 ) 104 105 parser.add_argument( 106 '--concat_manifest_paths', 107 nargs='+', 108 default=None, 109 type=str, 110 required=False, 111 help="Path to the additional dataset's manifests that will be concatenated with base dataset.", 112 ) 113 114 # Optional arguments 115 parser.add_argument( 116 "--target_dir", 117 default='./tarred', 118 type=str, 119 help="Target directory for resulting tarballs and manifest. Defaults to `./tarred`. Creates the path if necessary.", 120 ) 121 122 parser.add_argument( 123 "--metadata_path", required=False, default=None, type=str, help="Path to metadata file for the dataset.", 124 ) 125 126 parser.add_argument( 127 "--num_shards", 128 default=-1, 129 type=int, 130 help="Number of shards (tarballs) to create. Used for partitioning data among workers.", 131 ) 132 parser.add_argument( 133 '--max_duration', 134 default=None, 135 required=True, 136 type=float, 137 help='Maximum duration of audio clip in the dataset. By default, it is None and is required to be set.', 138 ) 139 parser.add_argument( 140 '--min_duration', 141 default=None, 142 type=float, 143 help='Minimum duration of audio clip in the dataset. By default, it is None and will not filter files.', 144 ) 145 parser.add_argument( 146 "--shuffle", 147 action='store_true', 148 help="Whether or not to randomly shuffle the samples in the manifest before tarring/sharding.", 149 ) 150 151 parser.add_argument( 152 "--keep_files_together", 153 action='store_true', 154 help="Whether or not to keep entries from the same file (but different offsets) together when sorting before tarring/sharding.", 155 ) 156 157 parser.add_argument( 158 "--sort_in_shards", 159 action='store_true', 160 help="Whether or not to sort samples inside the shards based on their duration.", 161 ) 162 163 parser.add_argument( 164 "--buckets_num", type=int, default=1, help="Number of buckets to create based on duration.", 165 ) 166 167 parser.add_argument("--shuffle_seed", type=int, default=None, help="Random seed for use if shuffling is enabled.") 168 parser.add_argument( 169 '--write_metadata', 170 action='store_true', 171 help=( 172 "Flag to write a blank metadata with the current call config. " 173 "Note that the metadata will not contain the number of shards, " 174 "and it must be filled out by the user." 175 ), 176 ) 177 parser.add_argument( 178 "--no_shard_manifests", 179 action='store_true', 180 help="Do not write sharded manifests along with the aggregated manifest.", 181 ) 182 parser.add_argument('--workers', type=int, default=1, help='Number of worker processes') 183 args = parser.parse_args() 184 185 186 @dataclass 187 class ASRTarredDatasetConfig: 188 num_shards: int = -1 189 shuffle: bool = False 190 max_duration: Optional[float] = None 191 min_duration: Optional[float] = None 192 shuffle_seed: Optional[int] = None 193 sort_in_shards: bool = True 194 shard_manifests: bool = True 195 keep_files_together: bool = False 196 197 198 @dataclass 199 class ASRTarredDatasetMetadata: 200 created_datetime: Optional[str] = None 201 version: int = 0 202 num_samples_per_shard: Optional[int] = None 203 is_concatenated_manifest: bool = False 204 205 dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() 206 history: Optional[List[Any]] = field(default_factory=lambda: []) 207 208 def __post_init__(self): 209 self.created_datetime = self.get_current_datetime() 210 211 def get_current_datetime(self): 212 return datetime.now().strftime("%m-%d-%Y %H-%M-%S") 213 214 @classmethod 215 def from_config(cls, config: DictConfig): 216 obj = cls() 217 obj.__dict__.update(**config) 218 return obj 219 220 @classmethod 221 def from_file(cls, filepath: str): 222 config = OmegaConf.load(filepath) 223 return ASRTarredDatasetMetadata.from_config(config=config) 224 225 226 class ASRTarredDatasetBuilder: 227 """ 228 Helper class that constructs a tarred dataset from scratch, or concatenates tarred datasets 229 together and constructs manifests for them. 230 """ 231 232 def __init__(self): 233 self.config = None 234 235 def configure(self, config: ASRTarredDatasetConfig): 236 """ 237 Sets the config generated from command line overrides. 238 239 Args: 240 config: ASRTarredDatasetConfig dataclass object. 241 """ 242 self.config = config # type: ASRTarredDatasetConfig 243 244 if self.config.num_shards < 0: 245 raise ValueError("`num_shards` must be > 0. Please fill in the metadata information correctly.") 246 247 def create_new_dataset(self, manifest_path: str, target_dir: str = "./tarred/", num_workers: int = 0): 248 """ 249 Creates a new tarred dataset from a given manifest file. 250 251 Args: 252 manifest_path: Path to the original ASR manifest. 253 target_dir: Output directory. 254 num_workers: Integer denoting number of parallel worker processes which will write tarfiles. 255 Defaults to 1 - which denotes sequential worker process. 256 257 Output: 258 Writes tarfiles, along with the tarred dataset compatible manifest file. 259 Also preserves a record of the metadata used to construct this tarred dataset. 260 """ 261 if self.config is None: 262 raise ValueError("Config has not been set. Please call `configure(config: ASRTarredDatasetConfig)`") 263 264 if manifest_path is None: 265 raise FileNotFoundError("Manifest filepath cannot be None !") 266 267 config = self.config # type: ASRTarredDatasetConfig 268 269 if not os.path.exists(target_dir): 270 os.makedirs(target_dir) 271 272 # Read the existing manifest 273 entries, total_duration, filtered_entries, filtered_duration = self._read_manifest(manifest_path, config) 274 275 if len(filtered_entries) > 0: 276 print(f"Filtered {len(filtered_entries)} files which amounts to {filtered_duration} seconds of audio.") 277 print( 278 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 279 ) 280 281 if len(entries) == 0: 282 print("No tarred dataset was created as there were 0 valid samples after filtering!") 283 return 284 if config.shuffle: 285 random.seed(config.shuffle_seed) 286 print("Shuffling...") 287 if config.keep_files_together: 288 filename_entries = defaultdict(list) 289 for ent in entries: 290 filename_entries[ent["audio_filepath"]].append(ent) 291 filenames = list(filename_entries.keys()) 292 random.shuffle(filenames) 293 shuffled_entries = [] 294 for filename in filenames: 295 shuffled_entries += filename_entries[filename] 296 entries = shuffled_entries 297 else: 298 random.shuffle(entries) 299 300 # Create shards and updated manifest entries 301 print(f"Number of samples added : {len(entries)}") 302 print(f"Remainder: {len(entries) % config.num_shards}") 303 304 start_indices = [] 305 end_indices = [] 306 # Build indices 307 for i in range(config.num_shards): 308 start_idx = (len(entries) // config.num_shards) * i 309 end_idx = start_idx + (len(entries) // config.num_shards) 310 print(f"Shard {i} has entries {start_idx} ~ {end_idx}") 311 files = set() 312 for ent_id in range(start_idx, end_idx): 313 files.add(entries[ent_id]["audio_filepath"]) 314 print(f"Shard {i} contains {len(files)} files") 315 if i == config.num_shards - 1: 316 # We discard in order to have the same number of entries per shard. 317 print(f"Have {len(entries) - end_idx} entries left over that will be discarded.") 318 319 start_indices.append(start_idx) 320 end_indices.append(end_idx) 321 322 manifest_folder, _ = os.path.split(manifest_path) 323 324 with Parallel(n_jobs=num_workers, verbose=config.num_shards) as parallel: 325 # Call parallel tarfile construction 326 new_entries_list = parallel( 327 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, i, manifest_folder) 328 for i, (start_idx, end_idx) in enumerate(zip(start_indices, end_indices)) 329 ) 330 331 if config.shard_manifests: 332 sharded_manifests_dir = target_dir + '/sharded_manifests' 333 if not os.path.exists(sharded_manifests_dir): 334 os.makedirs(sharded_manifests_dir) 335 336 for manifest in new_entries_list: 337 shard_id = manifest[0]['shard_id'] 338 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 339 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 340 for entry in manifest: 341 json.dump(entry, m2) 342 m2.write('\n') 343 344 # Flatten the list of list of entries to a list of entries 345 new_entries = [sample for manifest in new_entries_list for sample in manifest] 346 del new_entries_list 347 348 print("Total number of entries in manifest :", len(new_entries)) 349 350 # Write manifest 351 new_manifest_path = os.path.join(target_dir, 'tarred_audio_manifest.json') 352 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 353 for entry in new_entries: 354 json.dump(entry, m2) 355 m2.write('\n') 356 357 # Write metadata (default metadata for new datasets) 358 new_metadata_path = os.path.join(target_dir, 'metadata.yaml') 359 metadata = ASRTarredDatasetMetadata() 360 361 # Update metadata 362 metadata.dataset_config = config 363 metadata.num_samples_per_shard = len(new_entries) // config.num_shards 364 365 # Write metadata 366 metadata_yaml = OmegaConf.structured(metadata) 367 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 368 369 def create_concatenated_dataset( 370 self, 371 base_manifest_path: str, 372 manifest_paths: List[str], 373 metadata: ASRTarredDatasetMetadata, 374 target_dir: str = "./tarred_concatenated/", 375 num_workers: int = 1, 376 ): 377 """ 378 Creates new tarfiles in order to create a concatenated dataset, whose manifest contains the data for 379 both the original dataset as well as the new data submitted in manifest paths. 380 381 Args: 382 base_manifest_path: Path to the manifest file which contains the information for the original 383 tarred dataset (with flattened paths). 384 manifest_paths: List of one or more paths to manifest files that will be concatenated with above 385 base tarred dataset. 386 metadata: ASRTarredDatasetMetadata dataclass instance with overrides from command line. 387 target_dir: Output directory 388 389 Output: 390 Writes tarfiles which with indices mapping to a "concatenated" tarred dataset, 391 along with the tarred dataset compatible manifest file which includes information 392 about all the datasets that comprise the concatenated dataset. 393 394 Also preserves a record of the metadata used to construct this tarred dataset. 395 """ 396 if not os.path.exists(target_dir): 397 os.makedirs(target_dir) 398 399 if base_manifest_path is None: 400 raise FileNotFoundError("Base manifest filepath cannot be None !") 401 402 if manifest_paths is None or len(manifest_paths) == 0: 403 raise FileNotFoundError("List of additional manifest filepaths cannot be None !") 404 405 config = ASRTarredDatasetConfig(**(metadata.dataset_config)) 406 407 # Read the existing manifest (no filtering here) 408 base_entries, _, _, _ = self._read_manifest(base_manifest_path, config) 409 print(f"Read base manifest containing {len(base_entries)} samples.") 410 411 # Precompute number of samples per shard 412 if metadata.num_samples_per_shard is None: 413 num_samples_per_shard = len(base_entries) // config.num_shards 414 else: 415 num_samples_per_shard = metadata.num_samples_per_shard 416 417 print("Number of samples per shard :", num_samples_per_shard) 418 419 # Compute min and max duration and update config (if no metadata passed) 420 print(f"Selected max duration : {config.max_duration}") 421 print(f"Selected min duration : {config.min_duration}") 422 423 entries = [] 424 for new_manifest_idx in range(len(manifest_paths)): 425 new_entries, total_duration, filtered_new_entries, filtered_duration = self._read_manifest( 426 manifest_paths[new_manifest_idx], config 427 ) 428 429 if len(filtered_new_entries) > 0: 430 print( 431 f"Filtered {len(filtered_new_entries)} files which amounts to {filtered_duration:0.2f}" 432 f" seconds of audio from manifest {manifest_paths[new_manifest_idx]}." 433 ) 434 print( 435 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 436 ) 437 438 entries.extend(new_entries) 439 440 if len(entries) == 0: 441 print("No tarred dataset was created as there were 0 valid samples after filtering!") 442 return 443 444 if config.shuffle: 445 random.seed(config.shuffle_seed) 446 print("Shuffling...") 447 random.shuffle(entries) 448 449 # Drop last section of samples that cannot be added onto a chunk 450 drop_count = len(entries) % num_samples_per_shard 451 total_new_entries = len(entries) 452 entries = entries[:-drop_count] 453 454 print( 455 f"Dropping {drop_count} samples from total new samples {total_new_entries} since they cannot " 456 f"be added into a uniformly sized chunk." 457 ) 458 459 # Create shards and updated manifest entries 460 num_added_shards = len(entries) // num_samples_per_shard 461 462 print(f"Number of samples in base dataset : {len(base_entries)}") 463 print(f"Number of samples in additional datasets : {len(entries)}") 464 print(f"Number of added shards : {num_added_shards}") 465 print(f"Remainder: {len(entries) % num_samples_per_shard}") 466 467 start_indices = [] 468 end_indices = [] 469 shard_indices = [] 470 for i in range(num_added_shards): 471 start_idx = (len(entries) // num_added_shards) * i 472 end_idx = start_idx + (len(entries) // num_added_shards) 473 shard_idx = i + config.num_shards 474 print(f"Shard {shard_idx} has entries {start_idx + len(base_entries)} ~ {end_idx + len(base_entries)}") 475 476 start_indices.append(start_idx) 477 end_indices.append(end_idx) 478 shard_indices.append(shard_idx) 479 480 manifest_folder, _ = os.path.split(base_manifest_path) 481 482 with Parallel(n_jobs=num_workers, verbose=num_added_shards) as parallel: 483 # Call parallel tarfile construction 484 new_entries_list = parallel( 485 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, shard_idx, manifest_folder) 486 for i, (start_idx, end_idx, shard_idx) in enumerate(zip(start_indices, end_indices, shard_indices)) 487 ) 488 489 if config.shard_manifests: 490 sharded_manifests_dir = target_dir + '/sharded_manifests' 491 if not os.path.exists(sharded_manifests_dir): 492 os.makedirs(sharded_manifests_dir) 493 494 for manifest in new_entries_list: 495 shard_id = manifest[0]['shard_id'] 496 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 497 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 498 for entry in manifest: 499 json.dump(entry, m2) 500 m2.write('\n') 501 502 # Flatten the list of list of entries to a list of entries 503 new_entries = [sample for manifest in new_entries_list for sample in manifest] 504 del new_entries_list 505 506 # Write manifest 507 if metadata is None: 508 new_version = 1 # start with `1`, where `0` indicates the base manifest + dataset 509 else: 510 new_version = metadata.version + 1 511 512 print("Total number of entries in manifest :", len(base_entries) + len(new_entries)) 513 514 new_manifest_path = os.path.join(target_dir, f'tarred_audio_manifest_version_{new_version}.json') 515 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 516 # First write all the entries of base manifest 517 for entry in base_entries: 518 json.dump(entry, m2) 519 m2.write('\n') 520 521 # Finally write the new entries 522 for entry in new_entries: 523 json.dump(entry, m2) 524 m2.write('\n') 525 526 # Preserve historical metadata 527 base_metadata = metadata 528 529 # Write metadata (updated metadata for concatenated datasets) 530 new_metadata_path = os.path.join(target_dir, f'metadata_version_{new_version}.yaml') 531 metadata = ASRTarredDatasetMetadata() 532 533 # Update config 534 config.num_shards = config.num_shards + num_added_shards 535 536 # Update metadata 537 metadata.version = new_version 538 metadata.dataset_config = config 539 metadata.num_samples_per_shard = num_samples_per_shard 540 metadata.is_concatenated_manifest = True 541 metadata.created_datetime = metadata.get_current_datetime() 542 543 # Attach history 544 current_metadata = OmegaConf.structured(base_metadata.history) 545 metadata.history = current_metadata 546 547 # Write metadata 548 metadata_yaml = OmegaConf.structured(metadata) 549 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 550 551 def _read_manifest(self, manifest_path: str, config: ASRTarredDatasetConfig): 552 """Read and filters data from the manifest""" 553 # Read the existing manifest 554 entries = [] 555 total_duration = 0.0 556 filtered_entries = [] 557 filtered_duration = 0.0 558 with open(manifest_path, 'r', encoding='utf-8') as m: 559 for line in m: 560 entry = json.loads(line) 561 if (config.max_duration is None or entry['duration'] < config.max_duration) and ( 562 config.min_duration is None or entry['duration'] >= config.min_duration 563 ): 564 entries.append(entry) 565 total_duration += entry["duration"] 566 else: 567 filtered_entries.append(entry) 568 filtered_duration += entry['duration'] 569 570 return entries, total_duration, filtered_entries, filtered_duration 571 572 def _create_shard(self, entries, target_dir, shard_id, manifest_folder): 573 """Creates a tarball containing the audio files from `entries`. 574 """ 575 if self.config.sort_in_shards: 576 entries.sort(key=lambda x: x["duration"], reverse=False) 577 578 new_entries = [] 579 tar = tarfile.open(os.path.join(target_dir, f'audio_{shard_id}.tar'), mode='w', dereference=True) 580 581 count = dict() 582 for entry in entries: 583 # We squash the filename since we do not preserve directory structure of audio files in the tarball. 584 if os.path.exists(entry["audio_filepath"]): 585 audio_filepath = entry["audio_filepath"] 586 else: 587 audio_filepath = os.path.join(manifest_folder, entry["audio_filepath"]) 588 if not os.path.exists(audio_filepath): 589 raise FileNotFoundError(f"Could not find {entry['audio_filepath']}!") 590 591 base, ext = os.path.splitext(audio_filepath) 592 base = base.replace('/', '_') 593 # Need the following replacement as long as WebDataset splits on first period 594 base = base.replace('.', '_') 595 squashed_filename = f'{base}{ext}' 596 if squashed_filename not in count: 597 tar.add(audio_filepath, arcname=squashed_filename) 598 to_write = squashed_filename 599 count[squashed_filename] = 1 600 else: 601 to_write = base + "-sub" + str(count[squashed_filename]) + ext 602 count[squashed_filename] += 1 603 604 new_entry = { 605 'audio_filepath': to_write, 606 'duration': entry['duration'], 607 'shard_id': shard_id, # Keep shard ID for recordkeeping 608 } 609 610 if 'label' in entry: 611 new_entry['label'] = entry['label'] 612 613 if 'text' in entry: 614 new_entry['text'] = entry['text'] 615 616 if 'offset' in entry: 617 new_entry['offset'] = entry['offset'] 618 619 if 'lang' in entry: 620 new_entry['lang'] = entry['lang'] 621 622 new_entries.append(new_entry) 623 624 tar.close() 625 return new_entries 626 627 @classmethod 628 def setup_history(cls, base_metadata: ASRTarredDatasetMetadata, history: List[Any]): 629 if 'history' in base_metadata.keys(): 630 for history_val in base_metadata.history: 631 cls.setup_history(history_val, history) 632 633 if base_metadata is not None: 634 metadata_copy = copy.deepcopy(base_metadata) 635 with open_dict(metadata_copy): 636 metadata_copy.pop('history', None) 637 history.append(metadata_copy) 638 639 640 def main(): 641 if args.buckets_num > 1: 642 bucket_length = (args.max_duration - args.min_duration) / float(args.buckets_num) 643 for i in range(args.buckets_num): 644 min_duration = args.min_duration + i * bucket_length 645 max_duration = min_duration + bucket_length 646 if i == args.buckets_num - 1: 647 # add a small number to cover the samples with exactly duration of max_duration in the last bucket. 648 max_duration += 1e-5 649 target_dir = os.path.join(args.target_dir, f"bucket{i+1}") 650 print(f"Creating bucket {i+1} with min_duration={min_duration} and max_duration={max_duration} ...") 651 print(f"Results are being saved at: {target_dir}.") 652 create_tar_datasets(min_duration=min_duration, max_duration=max_duration, target_dir=target_dir) 653 print(f"Bucket {i+1} is created.") 654 else: 655 create_tar_datasets(min_duration=args.min_duration, max_duration=args.max_duration, target_dir=args.target_dir) 656 657 658 def create_tar_datasets(min_duration: float, max_duration: float, target_dir: str): 659 builder = ASRTarredDatasetBuilder() 660 661 shard_manifests = False if args.no_shard_manifests else True 662 663 if args.write_metadata: 664 metadata = ASRTarredDatasetMetadata() 665 dataset_cfg = ASRTarredDatasetConfig( 666 num_shards=args.num_shards, 667 shuffle=args.shuffle, 668 max_duration=max_duration, 669 min_duration=min_duration, 670 shuffle_seed=args.shuffle_seed, 671 sort_in_shards=args.sort_in_shards, 672 shard_manifests=shard_manifests, 673 keep_files_together=args.keep_files_together, 674 ) 675 metadata.dataset_config = dataset_cfg 676 677 output_path = os.path.join(target_dir, 'default_metadata.yaml') 678 OmegaConf.save(metadata, output_path, resolve=True) 679 print(f"Default metadata written to {output_path}") 680 exit(0) 681 682 if args.concat_manifest_paths is None or len(args.concat_manifest_paths) == 0: 683 print("Creating new tarred dataset ...") 684 685 # Create a tarred dataset from scratch 686 config = ASRTarredDatasetConfig( 687 num_shards=args.num_shards, 688 shuffle=args.shuffle, 689 max_duration=max_duration, 690 min_duration=min_duration, 691 shuffle_seed=args.shuffle_seed, 692 sort_in_shards=args.sort_in_shards, 693 shard_manifests=shard_manifests, 694 keep_files_together=args.keep_files_together, 695 ) 696 builder.configure(config) 697 builder.create_new_dataset(manifest_path=args.manifest_path, target_dir=target_dir, num_workers=args.workers) 698 699 else: 700 if args.buckets_num > 1: 701 raise ValueError("Concatenation feature does not support buckets_num > 1.") 702 print("Concatenating multiple tarred datasets ...") 703 704 # Implicitly update config from base details 705 if args.metadata_path is not None: 706 metadata = ASRTarredDatasetMetadata.from_file(args.metadata_path) 707 else: 708 raise ValueError("`metadata` yaml file path must be provided!") 709 710 # Preserve history 711 history = [] 712 builder.setup_history(OmegaConf.structured(metadata), history) 713 metadata.history = history 714 715 # Add command line overrides (everything other than num_shards) 716 metadata.dataset_config.max_duration = max_duration 717 metadata.dataset_config.min_duration = min_duration 718 metadata.dataset_config.shuffle = args.shuffle 719 metadata.dataset_config.shuffle_seed = args.shuffle_seed 720 metadata.dataset_config.sort_in_shards = args.sort_in_shards 721 metadata.dataset_config.shard_manifests = shard_manifests 722 723 builder.configure(metadata.dataset_config) 724 725 # Concatenate a tarred dataset onto a previous one 726 builder.create_concatenated_dataset( 727 base_manifest_path=args.manifest_path, 728 manifest_paths=args.concat_manifest_paths, 729 metadata=metadata, 730 target_dir=target_dir, 731 num_workers=args.workers, 732 ) 733 734 if DALI_INDEX_SCRIPT_AVAILABLE and dali_index.INDEX_CREATOR_AVAILABLE: 735 print("Constructing DALI Tarfile Index - ", target_dir) 736 index_config = dali_index.DALITarredIndexConfig(tar_dir=target_dir, workers=args.workers) 737 dali_index.main(index_config) 738 739 740 if __name__ == "__main__": 741 main() 742 [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of tools/nemo_forced_aligner/align.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import math 17 import os 18 from dataclasses import dataclass, field, is_dataclass 19 from pathlib import Path 20 from typing import List, Optional 21 22 import torch 23 from omegaconf import OmegaConf 24 from utils.data_prep import ( 25 add_t_start_end_to_utt_obj, 26 get_batch_starts_ends, 27 get_batch_variables, 28 get_manifest_lines_batch, 29 is_entry_in_all_lines, 30 is_entry_in_any_lines, 31 ) 32 from utils.make_ass_files import make_ass_files 33 from utils.make_ctm_files import make_ctm_files 34 from utils.make_output_manifest import write_manifest_out_line 35 from utils.viterbi_decoding import viterbi_decoding 36 37 from nemo.collections.asr.models.ctc_models import EncDecCTCModel 38 from nemo.collections.asr.models.hybrid_rnnt_ctc_models import EncDecHybridRNNTCTCModel 39 from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchASR 40 from nemo.collections.asr.parts.utils.transcribe_utils import setup_model 41 from nemo.core.config import hydra_runner 42 from nemo.utils import logging 43 44 """ 45 Align the utterances in manifest_filepath. 46 Results are saved in ctm files in output_dir. 47 48 Arguments: 49 pretrained_name: string specifying the name of a CTC NeMo ASR model which will be automatically downloaded 50 from NGC and used for generating the log-probs which we will use to do alignment. 51 Note: NFA can only use CTC models (not Transducer models) at the moment. 52 model_path: string specifying the local filepath to a CTC NeMo ASR model which will be used to generate the 53 log-probs which we will use to do alignment. 54 Note: NFA can only use CTC models (not Transducer models) at the moment. 55 Note: if a model_path is provided, it will override the pretrained_name. 56 manifest_filepath: filepath to the manifest of the data you want to align, 57 containing 'audio_filepath' and 'text' fields. 58 output_dir: the folder where output CTM files and new JSON manifest will be saved. 59 align_using_pred_text: if True, will transcribe the audio using the specified model and then use that transcription 60 as the reference text for the forced alignment. 61 transcribe_device: None, or a string specifying the device that will be used for generating log-probs (i.e. "transcribing"). 62 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 63 (otherwise will set it to 'cpu'). 64 viterbi_device: None, or string specifying the device that will be used for doing Viterbi decoding. 65 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 66 (otherwise will set it to 'cpu'). 67 batch_size: int specifying batch size that will be used for generating log-probs and doing Viterbi decoding. 68 use_local_attention: boolean flag specifying whether to try to use local attention for the ASR Model (will only 69 work if the ASR Model is a Conformer model). If local attention is used, we will set the local attention context 70 size to [64,64]. 71 additional_segment_grouping_separator: an optional string used to separate the text into smaller segments. 72 If this is not specified, then the whole text will be treated as a single segment. 73 remove_blank_tokens_from_ctm: a boolean denoting whether to remove <blank> tokens from token-level output CTMs. 74 audio_filepath_parts_in_utt_id: int specifying how many of the 'parts' of the audio_filepath 75 we will use (starting from the final part of the audio_filepath) to determine the 76 utt_id that will be used in the CTM files. Note also that any spaces that are present in the audio_filepath 77 will be replaced with dashes, so as not to change the number of space-separated elements in the 78 CTM files. 79 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 1 => utt_id will be "e1" 80 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 2 => utt_id will be "d_e1" 81 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 3 => utt_id will be "c_d_e1" 82 use_buffered_infer: False, if set True, using streaming to do get the logits for alignment 83 This flag is useful when aligning large audio file. 84 However, currently the chunk streaming inference does not support batch inference, 85 which means even you set batch_size > 1, it will only infer one by one instead of doing 86 the whole batch inference together. 87 chunk_len_in_secs: float chunk length in seconds 88 total_buffer_in_secs: float Length of buffer (chunk + left and right padding) in seconds 89 chunk_batch_size: int batch size for buffered chunk inference, 90 which will cut one audio into segments and do inference on chunk_batch_size segments at a time 91 92 simulate_cache_aware_streaming: False, if set True, using cache aware streaming to do get the logits for alignment 93 94 save_output_file_formats: List of strings specifying what type of output files to save (default: ["ctm", "ass"]) 95 ctm_file_config: CTMFileConfig to specify the configuration of the output CTM files 96 ass_file_config: ASSFileConfig to specify the configuration of the output ASS files 97 """ 98 99 100 @dataclass 101 class CTMFileConfig: 102 remove_blank_tokens: bool = False 103 # minimum duration (in seconds) for timestamps in the CTM.If any line in the CTM has a 104 # duration lower than this, it will be enlarged from the middle outwards until it 105 # meets the minimum_timestamp_duration, or reaches the beginning or end of the audio file. 106 # Note that this may cause timestamps to overlap. 107 minimum_timestamp_duration: float = 0 108 109 110 @dataclass 111 class ASSFileConfig: 112 fontsize: int = 20 113 vertical_alignment: str = "center" 114 # if resegment_text_to_fill_space is True, the ASS files will use new segments 115 # such that each segment will not take up more than (approximately) max_lines_per_segment 116 # when the ASS file is applied to a video 117 resegment_text_to_fill_space: bool = False 118 max_lines_per_segment: int = 2 119 text_already_spoken_rgb: List[int] = field(default_factory=lambda: [49, 46, 61]) # dark gray 120 text_being_spoken_rgb: List[int] = field(default_factory=lambda: [57, 171, 9]) # dark green 121 text_not_yet_spoken_rgb: List[int] = field(default_factory=lambda: [194, 193, 199]) # light gray 122 123 124 @dataclass 125 class AlignmentConfig: 126 # Required configs 127 pretrained_name: Optional[str] = None 128 model_path: Optional[str] = None 129 manifest_filepath: Optional[str] = None 130 output_dir: Optional[str] = None 131 132 # General configs 133 align_using_pred_text: bool = False 134 transcribe_device: Optional[str] = None 135 viterbi_device: Optional[str] = None 136 batch_size: int = 1 137 use_local_attention: bool = True 138 additional_segment_grouping_separator: Optional[str] = None 139 audio_filepath_parts_in_utt_id: int = 1 140 141 # Buffered chunked streaming configs 142 use_buffered_chunked_streaming: bool = False 143 chunk_len_in_secs: float = 1.6 144 total_buffer_in_secs: float = 4.0 145 chunk_batch_size: int = 32 146 147 # Cache aware streaming configs 148 simulate_cache_aware_streaming: Optional[bool] = False 149 150 # Output file configs 151 save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) 152 ctm_file_config: CTMFileConfig = CTMFileConfig() 153 ass_file_config: ASSFileConfig = ASSFileConfig() 154 155 156 @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) 157 def main(cfg: AlignmentConfig): 158 159 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 160 161 if is_dataclass(cfg): 162 cfg = OmegaConf.structured(cfg) 163 164 # Validate config 165 if cfg.model_path is None and cfg.pretrained_name is None: 166 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None") 167 168 if cfg.model_path is not None and cfg.pretrained_name is not None: 169 raise ValueError("One of cfg.model_path and cfg.pretrained_name must be None") 170 171 if cfg.manifest_filepath is None: 172 raise ValueError("cfg.manifest_filepath must be specified") 173 174 if cfg.output_dir is None: 175 raise ValueError("cfg.output_dir must be specified") 176 177 if cfg.batch_size < 1: 178 raise ValueError("cfg.batch_size cannot be zero or a negative number") 179 180 if cfg.additional_segment_grouping_separator == "" or cfg.additional_segment_grouping_separator == " ": 181 raise ValueError("cfg.additional_grouping_separator cannot be empty string or space character") 182 183 if cfg.ctm_file_config.minimum_timestamp_duration < 0: 184 raise ValueError("cfg.minimum_timestamp_duration cannot be a negative number") 185 186 if cfg.ass_file_config.vertical_alignment not in ["top", "center", "bottom"]: 187 raise ValueError("cfg.ass_file_config.vertical_alignment must be one of 'top', 'center' or 'bottom'") 188 189 for rgb_list in [ 190 cfg.ass_file_config.text_already_spoken_rgb, 191 cfg.ass_file_config.text_already_spoken_rgb, 192 cfg.ass_file_config.text_already_spoken_rgb, 193 ]: 194 if len(rgb_list) != 3: 195 raise ValueError( 196 "cfg.ass_file_config.text_already_spoken_rgb," 197 " cfg.ass_file_config.text_being_spoken_rgb," 198 " and cfg.ass_file_config.text_already_spoken_rgb all need to contain" 199 " exactly 3 elements." 200 ) 201 202 # Validate manifest contents 203 if not is_entry_in_all_lines(cfg.manifest_filepath, "audio_filepath"): 204 raise RuntimeError( 205 "At least one line in cfg.manifest_filepath does not contain an 'audio_filepath' entry. " 206 "All lines must contain an 'audio_filepath' entry." 207 ) 208 209 if cfg.align_using_pred_text: 210 if is_entry_in_any_lines(cfg.manifest_filepath, "pred_text"): 211 raise RuntimeError( 212 "Cannot specify cfg.align_using_pred_text=True when the manifest at cfg.manifest_filepath " 213 "contains 'pred_text' entries. This is because the audio will be transcribed and may produce " 214 "a different 'pred_text'. This may cause confusion." 215 ) 216 else: 217 if not is_entry_in_all_lines(cfg.manifest_filepath, "text"): 218 raise RuntimeError( 219 "At least one line in cfg.manifest_filepath does not contain a 'text' entry. " 220 "NFA requires all lines to contain a 'text' entry when cfg.align_using_pred_text=False." 221 ) 222 223 # init devices 224 if cfg.transcribe_device is None: 225 transcribe_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 226 else: 227 transcribe_device = torch.device(cfg.transcribe_device) 228 logging.info(f"Device to be used for transcription step (`transcribe_device`) is {transcribe_device}") 229 230 if cfg.viterbi_device is None: 231 viterbi_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 232 else: 233 viterbi_device = torch.device(cfg.viterbi_device) 234 logging.info(f"Device to be used for viterbi step (`viterbi_device`) is {viterbi_device}") 235 236 if transcribe_device.type == 'cuda' or viterbi_device.type == 'cuda': 237 logging.warning( 238 'One or both of transcribe_device and viterbi_device are GPUs. If you run into OOM errors ' 239 'it may help to change both devices to be the CPU.' 240 ) 241 242 # load model 243 model, _ = setup_model(cfg, transcribe_device) 244 model.eval() 245 246 if isinstance(model, EncDecHybridRNNTCTCModel): 247 model.change_decoding_strategy(decoder_type="ctc") 248 249 if cfg.use_local_attention: 250 logging.info( 251 "Flag use_local_attention is set to True => will try to use local attention for model if it allows it" 252 ) 253 model.change_attention_model(self_attention_model="rel_pos_local_attn", att_context_size=[64, 64]) 254 255 if not (isinstance(model, EncDecCTCModel) or isinstance(model, EncDecHybridRNNTCTCModel)): 256 raise NotImplementedError( 257 f"Model is not an instance of NeMo EncDecCTCModel or ENCDecHybridRNNTCTCModel." 258 " Currently only instances of these models are supported" 259 ) 260 261 if cfg.ctm_file_config.minimum_timestamp_duration > 0: 262 logging.warning( 263 f"cfg.ctm_file_config.minimum_timestamp_duration has been set to {cfg.ctm_file_config.minimum_timestamp_duration} seconds. " 264 "This may cause the alignments for some tokens/words/additional segments to be overlapping." 265 ) 266 267 buffered_chunk_params = {} 268 if cfg.use_buffered_chunked_streaming: 269 model_cfg = copy.deepcopy(model._cfg) 270 271 OmegaConf.set_struct(model_cfg.preprocessor, False) 272 # some changes for streaming scenario 273 model_cfg.preprocessor.dither = 0.0 274 model_cfg.preprocessor.pad_to = 0 275 276 if model_cfg.preprocessor.normalize != "per_feature": 277 logging.error( 278 "Only EncDecCTCModelBPE models trained with per_feature normalization are supported currently" 279 ) 280 # Disable config overwriting 281 OmegaConf.set_struct(model_cfg.preprocessor, True) 282 283 feature_stride = model_cfg.preprocessor['window_stride'] 284 model_stride_in_secs = feature_stride * cfg.model_downsample_factor 285 total_buffer = cfg.total_buffer_in_secs 286 chunk_len = float(cfg.chunk_len_in_secs) 287 tokens_per_chunk = math.ceil(chunk_len / model_stride_in_secs) 288 mid_delay = math.ceil((chunk_len + (total_buffer - chunk_len) / 2) / model_stride_in_secs) 289 logging.info(f"tokens_per_chunk is {tokens_per_chunk}, mid_delay is {mid_delay}") 290 291 model = FrameBatchASR( 292 asr_model=model, 293 frame_len=chunk_len, 294 total_buffer=cfg.total_buffer_in_secs, 295 batch_size=cfg.chunk_batch_size, 296 ) 297 buffered_chunk_params = { 298 "delay": mid_delay, 299 "model_stride_in_secs": model_stride_in_secs, 300 "tokens_per_chunk": tokens_per_chunk, 301 } 302 # get start and end line IDs of batches 303 starts, ends = get_batch_starts_ends(cfg.manifest_filepath, cfg.batch_size) 304 305 # init output_timestep_duration = None and we will calculate and update it during the first batch 306 output_timestep_duration = None 307 308 # init f_manifest_out 309 os.makedirs(cfg.output_dir, exist_ok=True) 310 tgt_manifest_name = str(Path(cfg.manifest_filepath).stem) + "_with_output_file_paths.json" 311 tgt_manifest_filepath = str(Path(cfg.output_dir) / tgt_manifest_name) 312 f_manifest_out = open(tgt_manifest_filepath, 'w') 313 314 # get alignment and save in CTM batch-by-batch 315 for start, end in zip(starts, ends): 316 manifest_lines_batch = get_manifest_lines_batch(cfg.manifest_filepath, start, end) 317 318 (log_probs_batch, y_batch, T_batch, U_batch, utt_obj_batch, output_timestep_duration,) = get_batch_variables( 319 manifest_lines_batch, 320 model, 321 cfg.additional_segment_grouping_separator, 322 cfg.align_using_pred_text, 323 cfg.audio_filepath_parts_in_utt_id, 324 output_timestep_duration, 325 cfg.simulate_cache_aware_streaming, 326 cfg.use_buffered_chunked_streaming, 327 buffered_chunk_params, 328 ) 329 330 alignments_batch = viterbi_decoding(log_probs_batch, y_batch, T_batch, U_batch, viterbi_device) 331 332 for utt_obj, alignment_utt in zip(utt_obj_batch, alignments_batch): 333 334 utt_obj = add_t_start_end_to_utt_obj(utt_obj, alignment_utt, output_timestep_duration) 335 336 if "ctm" in cfg.save_output_file_formats: 337 utt_obj = make_ctm_files(utt_obj, cfg.output_dir, cfg.ctm_file_config,) 338 339 if "ass" in cfg.save_output_file_formats: 340 utt_obj = make_ass_files(utt_obj, cfg.output_dir, cfg.ass_file_config) 341 342 write_manifest_out_line( 343 f_manifest_out, utt_obj, 344 ) 345 346 f_manifest_out.close() 347 348 return None 349 350 351 if __name__ == "__main__": 352 main() 353 [end of tools/nemo_forced_aligner/align.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
8a892b86186dbdf61803d75570cb5c58471e9dda
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-09-30T01:26:50Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,9 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,9 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): @@ -2217,7 +2219,9 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -181,7 +181,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() + measure_cfg: ConfidenceMeasureConfig = field(default_factory=lambda: ConfidenceMeasureConfig()) method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -110,7 +110,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
NVIDIA__NeMo-7616
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] 1 2 |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg 17 :target: https://badge.fury.io/py/nemo-toolkit 18 :alt: Release version 19 20 .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg 21 :target: https://badge.fury.io/py/nemo-toolkit 22 :alt: Python version 23 24 .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads 25 :target: https://pepy.tech/project/nemo-toolkit 26 :alt: PyPi total downloads 27 28 .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push 29 :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml 30 :alt: CodeQL 31 32 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 33 :target: https://github.com/psf/black 34 :alt: Code style: black 35 36 .. _main-readme: 37 38 **NVIDIA NeMo** 39 =============== 40 41 Introduction 42 ------------ 43 44 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), 45 text-to-speech synthesis (TTS), large language models (LLMs), and 46 natural language processing (NLP). 47 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) 48 and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 49 50 All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and 51 training is automatically scalable to 1000s of GPUs. 52 Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. 53 NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. 54 55 Getting started with NeMo is simple. 56 State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and 57 `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. 58 These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. 59 60 We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that 61 can all be run on `Google Colab <https://colab.research.google.com>`_. 62 63 For advanced users that want to train NeMo models from scratch or finetune existing NeMo models 64 we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. 65 66 For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. 67 The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ 68 which can be used to find the optimal model parallel configuration for training on a specific cluster. 69 70 Also see the two introductory videos below for a high level overview of NeMo. 71 72 * Developing State-Of-The-Art Conversational AI Models in Three Lines of Code. 73 * NVIDIA NeMo: Toolkit for Conversational AI at PyData Yerevan 2022. 74 75 |three_lines| |pydata| 76 77 .. |pydata| image:: https://img.youtube.com/vi/J-P6Sczmas8/maxres3.jpg 78 :target: https://www.youtube.com/embed/J-P6Sczmas8?mute=0&start=14&autoplay=0 79 :width: 600 80 :alt: Develop Conversational AI Models in 3 Lines 81 82 .. |three_lines| image:: https://img.youtube.com/vi/wBgpMf_KQVw/maxresdefault.jpg 83 :target: https://www.youtube.com/embed/wBgpMf_KQVw?mute=0&start=0&autoplay=0 84 :width: 600 85 :alt: Introduction at PyData@Yerevan 2022 86 87 Key Features 88 ------------ 89 90 * Speech processing 91 * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ 92 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 93 * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ 94 * Jasper, QuartzNet, CitriNet, ContextNet 95 * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer 96 * Squeezeformer-CTC and Squeezeformer-Transducer 97 * LSTM-Transducer (RNNT) and LSTM-CTC 98 * Supports the following decoders/losses: 99 * CTC 100 * Transducer/RNNT 101 * Hybrid Transducer/CTC 102 * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ 103 * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ 104 * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ 105 * Beam Search decoding 106 * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 107 * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ 108 * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) 109 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 110 * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ 111 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 112 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ 113 * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet 114 * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) 115 * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer 116 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 117 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 118 * Natural Language Processing 119 * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ 120 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ 121 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 122 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 123 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 124 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 125 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 126 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 127 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 128 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 129 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 130 * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ 131 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 132 * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ 133 * Text-to-Speech Synthesis (TTS): 134 * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 135 * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 136 * Vocoders: HiFiGAN, UnivNet, WaveGlow 137 * End-to-End Models: VITS 138 * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 139 * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ 140 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ 141 * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ 142 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 143 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 144 * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ 145 146 147 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 148 149 Requirements 150 ------------ 151 152 1) Python 3.10 or above 153 2) Pytorch 1.13.1 or above 154 3) NVIDIA GPU for training 155 156 Documentation 157 ------------- 158 159 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 160 :alt: Documentation Status 161 :scale: 100% 162 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 163 164 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 165 :alt: Documentation Status 166 :scale: 100% 167 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 168 169 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 170 | Version | Status | Description | 171 +=========+=============+==========================================================================================================================================+ 172 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 173 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 174 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 175 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 176 177 Tutorials 178 --------- 179 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 180 181 Getting help with NeMo 182 ---------------------- 183 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 184 185 186 Installation 187 ------------ 188 189 Conda 190 ~~~~~ 191 192 We recommend installing NeMo in a fresh Conda environment. 193 194 .. code-block:: bash 195 196 conda create --name nemo python==3.10.12 197 conda activate nemo 198 199 Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. 200 201 .. code-block:: bash 202 203 conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia 204 205 The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. 206 207 Pip 208 ~~~ 209 Use this installation mode if you want the latest released version. 210 211 .. code-block:: bash 212 213 apt-get update && apt-get install -y libsndfile1 ffmpeg 214 pip install Cython 215 pip install nemo_toolkit['all'] 216 217 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 218 219 Pip from source 220 ~~~~~~~~~~~~~~~ 221 Use this installation mode if you want the version from a particular GitHub branch (e.g main). 222 223 .. code-block:: bash 224 225 apt-get update && apt-get install -y libsndfile1 ffmpeg 226 pip install Cython 227 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 228 229 230 From source 231 ~~~~~~~~~~~ 232 Use this installation mode if you are contributing to NeMo. 233 234 .. code-block:: bash 235 236 apt-get update && apt-get install -y libsndfile1 ffmpeg 237 git clone https://github.com/NVIDIA/NeMo 238 cd NeMo 239 ./reinstall.sh 240 241 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 242 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 243 244 RNNT 245 ~~~~ 246 Note that RNNT requires numba to be installed from conda. 247 248 .. code-block:: bash 249 250 conda remove numba 251 pip uninstall numba 252 conda install -c conda-forge numba 253 254 NeMo Megatron 255 ~~~~~~~~~~~~~ 256 NeMo Megatron training requires NVIDIA Apex to be installed. 257 Install it manually if not using the NVIDIA PyTorch container. 258 259 To install Apex, run 260 261 .. code-block:: bash 262 263 git clone https://github.com/NVIDIA/apex.git 264 cd apex 265 git checkout 52e18c894223800cb611682dce27d88050edf1de 266 pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ 267 268 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. 269 270 While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. 271 This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 272 273 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: 274 275 .. code-block:: bash 276 277 conda install -c nvidia cuda-nvprof=11.8 278 279 packaging is also needed: 280 281 .. code-block:: bash 282 283 pip install packaging 284 285 With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. 286 287 288 Transformer Engine 289 ~~~~~~~~~~~~~~~~~~ 290 NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ 291 Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. 292 `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. 293 294 .. code-block:: bash 295 296 pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable 297 298 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. 299 300 Transformer Engine requires PyTorch to be built with CUDA 11.8. 301 302 303 Flash Attention 304 ~~~~~~~~~~~~~~~~~~~~ 305 Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. 306 307 .. code-block:: bash 308 309 pip install flash-attn 310 pip install triton==2.0.0.dev20221202 311 312 NLP inference UI 313 ~~~~~~~~~~~~~~~~~~~~ 314 To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. 315 316 .. code-block:: bash 317 318 pip install gradio==3.34.0 319 320 NeMo Text Processing 321 ~~~~~~~~~~~~~~~~~~~~ 322 NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. 323 324 Docker containers: 325 ~~~~~~~~~~~~~~~~~~ 326 We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. 327 328 To use built container, please run 329 330 .. code-block:: bash 331 332 docker pull nvcr.io/nvidia/nemo:23.06 333 334 To build a nemo container with Dockerfile from a branch, please run 335 336 .. code-block:: bash 337 338 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 339 340 341 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. 342 343 .. code-block:: bash 344 345 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 346 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 347 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 348 349 Examples 350 -------- 351 352 Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 353 354 355 Contributing 356 ------------ 357 358 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 359 360 Publications 361 ------------ 362 363 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 364 365 License 366 ------- 367 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 368 [end of README.rst] [start of examples/asr/experimental/k2/align_speech_parallel.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 # Based on examples/asr/transcribe_speech_parallel.py 17 # ASR alignment with multi-GPU/multi-node support for large datasets 18 # It supports both tarred and non-tarred datasets 19 # Arguments 20 # model: path to a nemo/PTL checkpoint file or name of a pretrained model 21 # predict_ds: config of the dataset/dataloader 22 # aligner_args: aligner config 23 # output_path: path to store the predictions 24 # model_stride: model downsampling factor, 8 for Citrinet models and 4 for Conformer models 25 # 26 # Results of each GPU/worker is written into a file named 'predictions_{rank}.json, and aggregated results of all workers are written into 'predictions_all.json' 27 28 Example for non-tarred datasets: 29 30 python align_speech_parallel.py \ 31 model=stt_en_conformer_ctc_large \ 32 predict_ds.manifest_filepath=/dataset/manifest_file.json \ 33 predict_ds.batch_size=16 \ 34 output_path=/tmp/ 35 36 Example for tarred datasets: 37 38 python align_speech_parallel.py \ 39 predict_ds.is_tarred=true \ 40 predict_ds.manifest_filepath=/tarred_dataset/tarred_audio_manifest.json \ 41 predict_ds.tarred_audio_filepaths=/tarred_dataset/audio__OP_0..127_CL_.tar \ 42 ... 43 44 By default the trainer uses all the GPUs available and default precision is FP32. 45 By setting the trainer config you may control these configs. For example to do the predictions with AMP on just two GPUs: 46 47 python align_speech_parallel.py \ 48 trainer.precision=16 \ 49 trainer.gpus=2 \ 50 ... 51 52 You may control the dataloader's config by setting the predict_ds: 53 54 python align_speech_parallel.py \ 55 predict_ds.num_workers=8 \ 56 predict_ds.min_duration=2.0 \ 57 predict_ds.sample_rate=16000 \ 58 model=stt_en_conformer_ctc_small \ 59 ... 60 61 You may control the aligner's config by setting the aligner_args: 62 aligner_args.alignment_type=argmax \ 63 aligner_args.word_output=False \ 64 aligner_args.cpu_decoding=True \ 65 aligner_args.decode_batch_size=8 \ 66 aligner_args.ctc_cfg.prob_suppress_index=-1 \ 67 aligner_args.ctc_cfg.prob_suppress_value=0.5 \ 68 aligner_args.rnnt_cfg.predictor_window_size=10 \ 69 aligner_args.decoder_module_cfg.intersect_pruned=true \ 70 aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ 71 ... 72 73 """ 74 75 76 import os 77 from dataclasses import dataclass, is_dataclass 78 from typing import Optional 79 80 import pytorch_lightning as ptl 81 import torch 82 from omegaconf import MISSING, OmegaConf 83 84 from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter 85 from nemo.collections.asr.models import ASRModel 86 from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig 87 from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig 88 from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel 89 from nemo.core.config import TrainerConfig, hydra_runner 90 from nemo.utils import logging 91 from nemo.utils.get_rank import is_global_rank_zero 92 93 94 @dataclass 95 class ParallelAlignmentConfig: 96 model: Optional[str] = None # name 97 predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) 98 aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() 99 output_path: str = MISSING 100 model_stride: int = 8 101 102 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") 103 104 # there arguments will be ignored 105 return_predictions: bool = False 106 use_cer: bool = False 107 108 109 def match_train_config(predict_ds, train_ds): 110 # It copies the important configurations from the train dataset of the model 111 # into the predict_ds to be used for prediction. It is needed to match the training configurations. 112 if train_ds is None: 113 return 114 115 predict_ds.sample_rate = train_ds.get("sample_rate", 16000) 116 cfg_name_list = [ 117 "int_values", 118 "use_start_end_token", 119 "blank_index", 120 "unk_index", 121 "normalize", 122 "parser", 123 "eos_id", 124 "bos_id", 125 "pad_id", 126 ] 127 128 if is_dataclass(predict_ds): 129 predict_ds = OmegaConf.structured(predict_ds) 130 for cfg_name in cfg_name_list: 131 if hasattr(train_ds, cfg_name): 132 setattr(predict_ds, cfg_name, getattr(train_ds, cfg_name)) 133 134 return predict_ds 135 136 137 @hydra_runner(config_name="AlignmentConfig", schema=ParallelAlignmentConfig) 138 def main(cfg: ParallelAlignmentConfig): 139 if cfg.model.endswith(".nemo"): 140 logging.info("Attempting to initialize from .nemo file") 141 model = ASRModel.restore_from(restore_path=cfg.model, map_location="cpu") 142 elif cfg.model.endswith(".ckpt"): 143 logging.info("Attempting to initialize from .ckpt file") 144 model = ASRModel.load_from_checkpoint(checkpoint_path=cfg.model, map_location="cpu") 145 else: 146 logging.info( 147 "Attempting to initialize from a pretrained model as the model name does not have the extension of .nemo or .ckpt" 148 ) 149 model = ASRModel.from_pretrained(model_name=cfg.model, map_location="cpu") 150 151 trainer = ptl.Trainer(**cfg.trainer) 152 153 cfg.predict_ds.return_sample_id = True 154 cfg.return_predictions = False 155 cfg.use_cer = False 156 cfg.predict_ds = match_train_config(predict_ds=cfg.predict_ds, train_ds=model._cfg.train_ds) 157 data_loader = model._setup_dataloader_from_config(cfg.predict_ds) 158 159 os.makedirs(cfg.output_path, exist_ok=True) 160 # trainer.global_rank is not valid before predict() is called. Need this hack to find the correct global_rank. 161 global_rank = trainer.node_rank * trainer.num_devices + int(os.environ.get("LOCAL_RANK", 0)) 162 output_file = os.path.join(cfg.output_path, f"predictions_{global_rank}.json") 163 output_ctm_dir = os.path.join(cfg.output_path, "ctm") 164 predictor_writer = ASRCTMPredictionWriter( 165 dataset=data_loader.dataset, 166 output_file=output_file, 167 output_ctm_dir=output_ctm_dir, 168 time_per_frame=cfg.model_stride * model._cfg.preprocessor['window_stride'], 169 ) 170 trainer.callbacks.extend([predictor_writer]) 171 172 aligner_wrapper = AlignerWrapperModel(model=model, cfg=cfg.aligner_args) 173 trainer.predict(model=aligner_wrapper, dataloaders=data_loader, return_predictions=cfg.return_predictions) 174 samples_num = predictor_writer.close_output_file() 175 176 logging.info( 177 f"Prediction on rank {global_rank} is done for {samples_num} samples and results are stored in {output_file}." 178 ) 179 180 if torch.distributed.is_initialized(): 181 torch.distributed.barrier() 182 183 samples_num = 0 184 if is_global_rank_zero(): 185 output_file = os.path.join(cfg.output_path, f"predictions_all.json") 186 logging.info(f"Prediction files are being aggregated in {output_file}.") 187 with open(output_file, 'tw', encoding="utf-8") as outf: 188 for rank in range(trainer.world_size): 189 input_file = os.path.join(cfg.output_path, f"predictions_{rank}.json") 190 with open(input_file, 'r', encoding="utf-8") as inpf: 191 lines = inpf.readlines() 192 samples_num += len(lines) 193 outf.writelines(lines) 194 logging.info( 195 f"Prediction is done for {samples_num} samples in total on all workers and results are aggregated in {output_file}." 196 ) 197 198 199 if __name__ == '__main__': 200 main() 201 [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import re 17 from abc import abstractmethod 18 from dataclasses import dataclass, is_dataclass 19 from typing import Callable, Dict, List, Optional, Tuple, Union 20 21 import editdistance 22 import numpy as np 23 import torch 24 from omegaconf import OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.metrics.wer import move_dimension_to_the_front 28 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding as beam_decode 29 from nemo.collections.asr.parts.submodules import rnnt_greedy_decoding as greedy_decode 30 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 31 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 32 from nemo.utils import logging 33 34 __all__ = ['RNNTDecoding', 'RNNTWER'] 35 36 37 class AbstractRNNTDecoding(ConfidenceMixin): 38 """ 39 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 40 41 Args: 42 decoding_cfg: A dict-like object which contains the following key-value pairs. 43 strategy: str value which represents the type of decoding that can occur. 44 Possible values are : 45 - greedy, greedy_batch (for greedy decoding). 46 - beam, tsd, alsd (for beam search decoding). 47 48 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 49 tokens as well as the decoded string. Default is False in order to avoid double decoding 50 unless required. 51 52 preserve_alignments: Bool flag which preserves the history of logprobs generated during 53 decoding (sample / batched). When set to true, the Hypothesis will contain 54 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 55 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 56 57 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 58 with the `return_hypotheses` flag set to True. 59 60 The length of the list corresponds to the Acoustic Length (T). 61 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 62 U is the number of target tokens for the current timestep Ti. 63 64 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 65 word based timestamp mapping the output log-probabilities to discrete intervals of timestamps. 66 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 67 68 rnnt_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 69 Can take the following values - "char" for character/subword time stamps, "word" for word level 70 time stamps and "all" (default), for both character level and word level time stamps. 71 72 word_seperator: Str token representing the seperator between words. 73 74 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 75 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 76 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of ints. 77 78 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 79 scores. In order to obtain hypotheses with confidence scores, please utilize 80 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 81 82 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 83 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 84 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 85 86 The length of the list corresponds to the Acoustic Length (T). 87 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 88 U is the number of target tokens for the current timestep Ti. 89 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 90 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 91 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 92 93 The length of the list corresponds to the number of recognized tokens. 94 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 95 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 96 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 97 98 The length of the list corresponds to the number of recognized words. 99 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 100 from the `token_confidence`. 101 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 102 Valid options are `mean`, `min`, `max`, `prod`. 103 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 104 confidence scores. 105 106 name: The method name (str). 107 Supported values: 108 - 'max_prob' for using the maximum token probability as a confidence. 109 - 'entropy' for using a normalized entropy of a log-likelihood vector. 110 111 entropy_type: Which type of entropy to use (str). 112 Used if confidence_method_cfg.name is set to `entropy`. 113 Supported values: 114 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 115 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 116 Note that for this entropy, the alpha should comply the following inequality: 117 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 118 where V is the model vocabulary size. 119 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 120 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 121 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 122 More: https://en.wikipedia.org/wiki/Tsallis_entropy 123 - 'renyi' for the Rรฉnyi entropy. 124 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 125 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 126 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 127 128 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 129 When the alpha equals one, scaling is not applied to 'max_prob', 130 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 131 132 entropy_norm: A mapping of the entropy value to the interval [0,1]. 133 Supported values: 134 - 'lin' for using the linear mapping. 135 - 'exp' for using exponential mapping with linear shift. 136 137 The config may further contain the following sub-dictionaries: 138 "greedy": 139 max_symbols: int, describing the maximum number of target tokens to decode per 140 timestep during greedy decoding. Setting to larger values allows longer sentences 141 to be decoded, at the cost of increased execution time. 142 preserve_frame_confidence: Same as above, overrides above value. 143 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 144 145 "beam": 146 beam_size: int, defining the beam size for beam search. Must be >= 1. 147 If beam_size == 1, will perform cached greedy search. This might be slightly different 148 results compared to the greedy search above. 149 150 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 151 Set to True by default. 152 153 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 154 hypotheses after beam search has concluded. This flag is set by default. 155 156 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 157 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 158 at increased cost to execution time. 159 160 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 161 If an integer is provided, it can decode sequences of that particular maximum length. 162 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 163 where seq_len is the length of the acoustic model output (T). 164 165 NOTE: 166 If a float is provided, it can be greater than 1! 167 By default, a float of 2.0 is used so that a target sequence can be at most twice 168 as long as the acoustic model output length T. 169 170 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 171 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 172 173 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 174 in order to reduce expensive beam search cost later. int >= 0. 175 176 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 177 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 178 and affects the speed of inference since large values will perform large beam search in the next step. 179 180 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 181 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 182 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 183 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 184 expansion apart from the "most likely" candidate. 185 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 186 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 187 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 188 tuned on a validation set. 189 190 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 191 192 decoder: The Decoder/Prediction network module. 193 joint: The Joint network module. 194 blank_id: The id of the RNNT blank token. 195 """ 196 197 def __init__(self, decoding_cfg, decoder, joint, blank_id: int): 198 super(AbstractRNNTDecoding, self).__init__() 199 200 # Convert dataclass to config object 201 if is_dataclass(decoding_cfg): 202 decoding_cfg = OmegaConf.structured(decoding_cfg) 203 204 self.cfg = decoding_cfg 205 self.blank_id = blank_id 206 self.num_extra_outputs = joint.num_extra_outputs 207 self.big_blank_durations = self.cfg.get("big_blank_durations", None) 208 self.durations = self.cfg.get("durations", None) 209 self.compute_hypothesis_token_set = self.cfg.get("compute_hypothesis_token_set", False) 210 self.compute_langs = decoding_cfg.get('compute_langs', False) 211 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 212 self.joint_fused_batch_size = self.cfg.get('fused_batch_size', None) 213 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 214 self.word_seperator = self.cfg.get('word_seperator', ' ') 215 216 if self.durations is not None: # this means it's a TDT model. 217 if blank_id == 0: 218 raise ValueError("blank_id must equal len(non_blank_vocabs) for TDT models") 219 if self.big_blank_durations is not None: 220 raise ValueError("duration and big_blank_durations can't both be not None") 221 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 222 raise ValueError("currently only greedy and greedy_batch inference is supported for TDT models") 223 224 if self.big_blank_durations is not None: # this means it's a multi-blank model. 225 if blank_id == 0: 226 raise ValueError("blank_id must equal len(vocabs) for multi-blank RNN-T models") 227 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 228 raise ValueError( 229 "currently only greedy and greedy_batch inference is supported for multi-blank models" 230 ) 231 232 possible_strategies = ['greedy', 'greedy_batch', 'beam', 'tsd', 'alsd', 'maes'] 233 if self.cfg.strategy not in possible_strategies: 234 raise ValueError(f"Decoding strategy must be one of {possible_strategies}") 235 236 # Update preserve alignments 237 if self.preserve_alignments is None: 238 if self.cfg.strategy in ['greedy', 'greedy_batch']: 239 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 240 241 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 242 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 243 244 # Update compute timestamps 245 if self.compute_timestamps is None: 246 if self.cfg.strategy in ['greedy', 'greedy_batch']: 247 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 248 249 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 250 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 251 252 # Test if alignments are being preserved for RNNT 253 if self.compute_timestamps is True and self.preserve_alignments is False: 254 raise ValueError("If `compute_timesteps` flag is set, then `preserve_alignments` flag must also be set.") 255 256 # initialize confidence-related fields 257 self._init_confidence(self.cfg.get('confidence_cfg', None)) 258 259 # Confidence estimation is not implemented for these strategies 260 if ( 261 not self.preserve_frame_confidence 262 and self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes'] 263 and self.cfg.beam.get('preserve_frame_confidence', False) 264 ): 265 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 266 267 if self.cfg.strategy == 'greedy': 268 if self.big_blank_durations is None: 269 if self.durations is None: 270 self.decoding = greedy_decode.GreedyRNNTInfer( 271 decoder_model=decoder, 272 joint_model=joint, 273 blank_index=self.blank_id, 274 max_symbols_per_step=( 275 self.cfg.greedy.get('max_symbols', None) 276 or self.cfg.greedy.get('max_symbols_per_step', None) 277 ), 278 preserve_alignments=self.preserve_alignments, 279 preserve_frame_confidence=self.preserve_frame_confidence, 280 confidence_method_cfg=self.confidence_method_cfg, 281 ) 282 else: 283 self.decoding = greedy_decode.GreedyTDTInfer( 284 decoder_model=decoder, 285 joint_model=joint, 286 blank_index=self.blank_id, 287 durations=self.durations, 288 max_symbols_per_step=( 289 self.cfg.greedy.get('max_symbols', None) 290 or self.cfg.greedy.get('max_symbols_per_step', None) 291 ), 292 preserve_alignments=self.preserve_alignments, 293 preserve_frame_confidence=self.preserve_frame_confidence, 294 confidence_method_cfg=self.confidence_method_cfg, 295 ) 296 else: 297 self.decoding = greedy_decode.GreedyMultiblankRNNTInfer( 298 decoder_model=decoder, 299 joint_model=joint, 300 blank_index=self.blank_id, 301 big_blank_durations=self.big_blank_durations, 302 max_symbols_per_step=( 303 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 304 ), 305 preserve_alignments=self.preserve_alignments, 306 preserve_frame_confidence=self.preserve_frame_confidence, 307 confidence_method_cfg=self.confidence_method_cfg, 308 ) 309 310 elif self.cfg.strategy == 'greedy_batch': 311 if self.big_blank_durations is None: 312 if self.durations is None: 313 self.decoding = greedy_decode.GreedyBatchedRNNTInfer( 314 decoder_model=decoder, 315 joint_model=joint, 316 blank_index=self.blank_id, 317 max_symbols_per_step=( 318 self.cfg.greedy.get('max_symbols', None) 319 or self.cfg.greedy.get('max_symbols_per_step', None) 320 ), 321 preserve_alignments=self.preserve_alignments, 322 preserve_frame_confidence=self.preserve_frame_confidence, 323 confidence_method_cfg=self.confidence_method_cfg, 324 ) 325 else: 326 self.decoding = greedy_decode.GreedyBatchedTDTInfer( 327 decoder_model=decoder, 328 joint_model=joint, 329 blank_index=self.blank_id, 330 durations=self.durations, 331 max_symbols_per_step=( 332 self.cfg.greedy.get('max_symbols', None) 333 or self.cfg.greedy.get('max_symbols_per_step', None) 334 ), 335 preserve_alignments=self.preserve_alignments, 336 preserve_frame_confidence=self.preserve_frame_confidence, 337 confidence_method_cfg=self.confidence_method_cfg, 338 ) 339 340 else: 341 self.decoding = greedy_decode.GreedyBatchedMultiblankRNNTInfer( 342 decoder_model=decoder, 343 joint_model=joint, 344 blank_index=self.blank_id, 345 big_blank_durations=self.big_blank_durations, 346 max_symbols_per_step=( 347 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 348 ), 349 preserve_alignments=self.preserve_alignments, 350 preserve_frame_confidence=self.preserve_frame_confidence, 351 confidence_method_cfg=self.confidence_method_cfg, 352 ) 353 354 elif self.cfg.strategy == 'beam': 355 356 self.decoding = beam_decode.BeamRNNTInfer( 357 decoder_model=decoder, 358 joint_model=joint, 359 beam_size=self.cfg.beam.beam_size, 360 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 361 search_type='default', 362 score_norm=self.cfg.beam.get('score_norm', True), 363 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 364 preserve_alignments=self.preserve_alignments, 365 ) 366 367 elif self.cfg.strategy == 'tsd': 368 369 self.decoding = beam_decode.BeamRNNTInfer( 370 decoder_model=decoder, 371 joint_model=joint, 372 beam_size=self.cfg.beam.beam_size, 373 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 374 search_type='tsd', 375 score_norm=self.cfg.beam.get('score_norm', True), 376 tsd_max_sym_exp_per_step=self.cfg.beam.get('tsd_max_sym_exp', 10), 377 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 378 preserve_alignments=self.preserve_alignments, 379 ) 380 381 elif self.cfg.strategy == 'alsd': 382 383 self.decoding = beam_decode.BeamRNNTInfer( 384 decoder_model=decoder, 385 joint_model=joint, 386 beam_size=self.cfg.beam.beam_size, 387 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 388 search_type='alsd', 389 score_norm=self.cfg.beam.get('score_norm', True), 390 alsd_max_target_len=self.cfg.beam.get('alsd_max_target_len', 2), 391 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 392 preserve_alignments=self.preserve_alignments, 393 ) 394 395 elif self.cfg.strategy == 'maes': 396 397 self.decoding = beam_decode.BeamRNNTInfer( 398 decoder_model=decoder, 399 joint_model=joint, 400 beam_size=self.cfg.beam.beam_size, 401 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 402 search_type='maes', 403 score_norm=self.cfg.beam.get('score_norm', True), 404 maes_num_steps=self.cfg.beam.get('maes_num_steps', 2), 405 maes_prefix_alpha=self.cfg.beam.get('maes_prefix_alpha', 1), 406 maes_expansion_gamma=self.cfg.beam.get('maes_expansion_gamma', 2.3), 407 maes_expansion_beta=self.cfg.beam.get('maes_expansion_beta', 2.0), 408 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 409 preserve_alignments=self.preserve_alignments, 410 ngram_lm_model=self.cfg.beam.get('ngram_lm_model', None), 411 ngram_lm_alpha=self.cfg.beam.get('ngram_lm_alpha', 0.0), 412 hat_subtract_ilm=self.cfg.beam.get('hat_subtract_ilm', False), 413 hat_ilm_weight=self.cfg.beam.get('hat_ilm_weight', 0.0), 414 ) 415 416 else: 417 418 raise ValueError( 419 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 420 f"but was provided {self.cfg.strategy}" 421 ) 422 423 # Update the joint fused batch size or disable it entirely if needed. 424 self.update_joint_fused_batch_size() 425 426 def rnnt_decoder_predictions_tensor( 427 self, 428 encoder_output: torch.Tensor, 429 encoded_lengths: torch.Tensor, 430 return_hypotheses: bool = False, 431 partial_hypotheses: Optional[List[Hypothesis]] = None, 432 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 433 """ 434 Decode an encoder output by autoregressive decoding of the Decoder+Joint networks. 435 436 Args: 437 encoder_output: torch.Tensor of shape [B, D, T]. 438 encoded_lengths: torch.Tensor containing lengths of the padded encoder outputs. Shape [B]. 439 return_hypotheses: bool. If set to True it will return list of Hypothesis or NBestHypotheses 440 441 Returns: 442 If `return_best_hypothesis` is set: 443 A tuple (hypotheses, None): 444 hypotheses - list of Hypothesis (best hypothesis per sample). 445 Look at rnnt_utils.Hypothesis for more information. 446 447 If `return_best_hypothesis` is not set: 448 A tuple(hypotheses, all_hypotheses) 449 hypotheses - list of Hypothesis (best hypothesis per sample). 450 Look at rnnt_utils.Hypothesis for more information. 451 all_hypotheses - list of NBestHypotheses. Each NBestHypotheses further contains a sorted 452 list of all the hypotheses of the model per sample. 453 Look at rnnt_utils.NBestHypotheses for more information. 454 """ 455 # Compute hypotheses 456 with torch.inference_mode(): 457 hypotheses_list = self.decoding( 458 encoder_output=encoder_output, encoded_lengths=encoded_lengths, partial_hypotheses=partial_hypotheses 459 ) # type: [List[Hypothesis]] 460 461 # extract the hypotheses 462 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 463 464 prediction_list = hypotheses_list 465 466 if isinstance(prediction_list[0], NBestHypotheses): 467 hypotheses = [] 468 all_hypotheses = [] 469 470 for nbest_hyp in prediction_list: # type: NBestHypotheses 471 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 472 decoded_hyps = self.decode_hypothesis(n_hyps) # type: List[str] 473 474 # If computing timestamps 475 if self.compute_timestamps is True: 476 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 477 for hyp_idx in range(len(decoded_hyps)): 478 decoded_hyps[hyp_idx] = self.compute_rnnt_timestamps(decoded_hyps[hyp_idx], timestamp_type) 479 480 hypotheses.append(decoded_hyps[0]) # best hypothesis 481 all_hypotheses.append(decoded_hyps) 482 483 if return_hypotheses: 484 return hypotheses, all_hypotheses 485 486 best_hyp_text = [h.text for h in hypotheses] 487 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 488 return best_hyp_text, all_hyp_text 489 490 else: 491 hypotheses = self.decode_hypothesis(prediction_list) # type: List[str] 492 493 # If computing timestamps 494 if self.compute_timestamps is True: 495 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 496 for hyp_idx in range(len(hypotheses)): 497 hypotheses[hyp_idx] = self.compute_rnnt_timestamps(hypotheses[hyp_idx], timestamp_type) 498 499 if return_hypotheses: 500 # greedy decoding, can get high-level confidence scores 501 if self.preserve_frame_confidence and ( 502 self.preserve_word_confidence or self.preserve_token_confidence 503 ): 504 hypotheses = self.compute_confidence(hypotheses) 505 return hypotheses, None 506 507 best_hyp_text = [h.text for h in hypotheses] 508 return best_hyp_text, None 509 510 def decode_hypothesis(self, hypotheses_list: List[Hypothesis]) -> List[Union[Hypothesis, NBestHypotheses]]: 511 """ 512 Decode a list of hypotheses into a list of strings. 513 514 Args: 515 hypotheses_list: List of Hypothesis. 516 517 Returns: 518 A list of strings. 519 """ 520 for ind in range(len(hypotheses_list)): 521 # Extract the integer encoded hypothesis 522 prediction = hypotheses_list[ind].y_sequence 523 524 if type(prediction) != list: 525 prediction = prediction.tolist() 526 527 # RNN-T sample level is already preprocessed by implicit RNNT decoding 528 # Simply remove any blank and possibly big blank tokens 529 if self.big_blank_durations is not None: # multi-blank RNNT 530 num_extra_outputs = len(self.big_blank_durations) 531 prediction = [p for p in prediction if p < self.blank_id - num_extra_outputs] 532 elif self.durations is not None: # TDT model. 533 prediction = [p for p in prediction if p < self.blank_id] 534 else: # standard RNN-T 535 prediction = [p for p in prediction if p != self.blank_id] 536 537 # De-tokenize the integer tokens; if not computing timestamps 538 if self.compute_timestamps is True: 539 # keep the original predictions, wrap with the number of repetitions per token and alignments 540 # this is done so that `rnnt_decoder_predictions_tensor()` can process this hypothesis 541 # in order to compute exact time stamps. 542 alignments = copy.deepcopy(hypotheses_list[ind].alignments) 543 token_repetitions = [1] * len(alignments) # preserve number of repetitions per token 544 hypothesis = (prediction, alignments, token_repetitions) 545 else: 546 hypothesis = self.decode_tokens_to_str(prediction) 547 548 # TODO: remove 549 # collapse leading spaces before . , ? for PC models 550 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 551 552 if self.compute_hypothesis_token_set: 553 hypotheses_list[ind].tokens = self.decode_ids_to_tokens(prediction) 554 555 # De-tokenize the integer tokens 556 hypotheses_list[ind].text = hypothesis 557 558 return hypotheses_list 559 560 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 561 """ 562 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 563 Assumes that `frame_confidence` is present in the hypotheses. 564 565 Args: 566 hypotheses_list: List of Hypothesis. 567 568 Returns: 569 A list of hypotheses with high-level confidence scores. 570 """ 571 if self.exclude_blank_from_confidence: 572 for hyp in hypotheses_list: 573 hyp.token_confidence = hyp.non_blank_frame_confidence 574 else: 575 for hyp in hypotheses_list: 576 offset = 0 577 token_confidence = [] 578 if len(hyp.timestep) > 0: 579 for ts, te in zip(hyp.timestep, hyp.timestep[1:] + [len(hyp.frame_confidence)]): 580 if ts != te: 581 # <blank> tokens are considered to belong to the last non-blank token, if any. 582 token_confidence.append( 583 self._aggregate_confidence( 584 [hyp.frame_confidence[ts][offset]] 585 + [fc[0] for fc in hyp.frame_confidence[ts + 1 : te]] 586 ) 587 ) 588 offset = 0 589 else: 590 token_confidence.append(hyp.frame_confidence[ts][offset]) 591 offset += 1 592 hyp.token_confidence = token_confidence 593 if self.preserve_word_confidence: 594 for hyp in hypotheses_list: 595 hyp.word_confidence = self._aggregate_token_confidence(hyp) 596 return hypotheses_list 597 598 @abstractmethod 599 def decode_tokens_to_str(self, tokens: List[int]) -> str: 600 """ 601 Implemented by subclass in order to decoder a token id list into a string. 602 603 Args: 604 tokens: List of int representing the token ids. 605 606 Returns: 607 A decoded string. 608 """ 609 raise NotImplementedError() 610 611 @abstractmethod 612 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 613 """ 614 Implemented by subclass in order to decode a token id list into a token list. 615 A token list is the string representation of each token id. 616 617 Args: 618 tokens: List of int representing the token ids. 619 620 Returns: 621 A list of decoded tokens. 622 """ 623 raise NotImplementedError() 624 625 @abstractmethod 626 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 627 """ 628 Implemented by subclass in order to 629 compute the most likely language ID (LID) string given the tokens. 630 631 Args: 632 tokens: List of int representing the token ids. 633 634 Returns: 635 A decoded LID string. 636 """ 637 raise NotImplementedError() 638 639 @abstractmethod 640 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 641 """ 642 Implemented by subclass in order to 643 decode a token id list into language ID (LID) list. 644 645 Args: 646 tokens: List of int representing the token ids. 647 648 Returns: 649 A list of decoded LIDS. 650 """ 651 raise NotImplementedError() 652 653 def update_joint_fused_batch_size(self): 654 if self.joint_fused_batch_size is None: 655 # do nothing and let the Joint itself handle setting up of the fused batch 656 return 657 658 if not hasattr(self.decoding.joint, 'set_fused_batch_size'): 659 logging.warning( 660 "The joint module does not have `set_fused_batch_size(int)` as a setter function.\n" 661 "Ignoring update of joint fused batch size." 662 ) 663 return 664 665 if not hasattr(self.decoding.joint, 'set_fuse_loss_wer'): 666 logging.warning( 667 "The joint module does not have `set_fuse_loss_wer(bool, RNNTLoss, RNNTWER)` " 668 "as a setter function.\n" 669 "Ignoring update of joint fused batch size." 670 ) 671 return 672 673 if self.joint_fused_batch_size > 0: 674 self.decoding.joint.set_fused_batch_size(self.joint_fused_batch_size) 675 else: 676 logging.info("Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.") 677 self.decoding.joint.set_fuse_loss_wer(False) 678 679 def compute_rnnt_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 680 assert timestamp_type in ['char', 'word', 'all'] 681 682 # Unpack the temporary storage 683 decoded_prediction, alignments, token_repetitions = hypothesis.text 684 685 # Retrieve offsets 686 char_offsets = word_offsets = None 687 char_offsets = self._compute_offsets(hypothesis, token_repetitions, self.blank_id) 688 689 # finally, set the flattened decoded predictions to text field for later text decoding 690 hypothesis.text = decoded_prediction 691 692 # Assert number of offsets and hypothesis tokens are 1:1 match. 693 num_flattened_tokens = 0 694 for t in range(len(char_offsets)): 695 # Subtract one here for the extra RNNT BLANK token emitted to designate "End of timestep" 696 num_flattened_tokens += len(char_offsets[t]['char']) - 1 697 698 if num_flattened_tokens != len(hypothesis.text): 699 raise ValueError( 700 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 701 " have to be of the same length, but are: " 702 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 703 f" {len(hypothesis.text)}" 704 ) 705 706 encoded_char_offsets = copy.deepcopy(char_offsets) 707 708 # Correctly process the token ids to chars/subwords. 709 for i, offsets in enumerate(char_offsets): 710 decoded_chars = [] 711 for char in offsets['char'][:-1]: # ignore the RNNT Blank token at end of every timestep with -1 subset 712 decoded_chars.append(self.decode_tokens_to_str([int(char)])) 713 char_offsets[i]["char"] = decoded_chars 714 715 # detect char vs subword models 716 lens = [] 717 for v in char_offsets: 718 tokens = v["char"] 719 # each token may be either 1 unicode token or multiple unicode token 720 # for character based models, only 1 token is used 721 # for subword, more than one token can be used. 722 # Computing max, then summing up total lens is a test to check for char vs subword 723 # For char models, len(lens) == sum(lens) 724 # but this is violated for subword models. 725 max_len = max(len(c) for c in tokens) 726 lens.append(max_len) 727 728 # array of one or more chars implies subword based model with multiple char emitted per TxU step (via subword) 729 if sum(lens) > len(lens): 730 text_type = 'subword' 731 else: 732 # full array of ones implies character based model with 1 char emitted per TxU step 733 text_type = 'char' 734 735 # retrieve word offsets from character offsets 736 word_offsets = None 737 if timestamp_type in ['word', 'all']: 738 if text_type == 'char': 739 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 740 else: 741 # utilize the copy of char offsets with the correct integer ids for tokens 742 # so as to avoid tokenize -> detokenize -> compare -> merge steps. 743 word_offsets = self._get_word_offsets_subwords_sentencepiece( 744 encoded_char_offsets, 745 hypothesis, 746 decode_ids_to_tokens=self.decode_ids_to_tokens, 747 decode_tokens_to_str=self.decode_tokens_to_str, 748 ) 749 750 # attach results 751 if len(hypothesis.timestep) > 0: 752 timestep_info = hypothesis.timestep 753 else: 754 timestep_info = [] 755 756 # Setup defaults 757 hypothesis.timestep = {"timestep": timestep_info} 758 759 # Add char / subword time stamps 760 if char_offsets is not None and timestamp_type in ['char', 'all']: 761 hypothesis.timestep['char'] = char_offsets 762 763 # Add word time stamps 764 if word_offsets is not None and timestamp_type in ['word', 'all']: 765 hypothesis.timestep['word'] = word_offsets 766 767 # Convert the flattened token indices to text 768 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 769 770 return hypothesis 771 772 @staticmethod 773 def _compute_offsets( 774 hypothesis: Hypothesis, token_repetitions: List[int], rnnt_token: int 775 ) -> List[Dict[str, Union[str, int]]]: 776 """ 777 Utility method that calculates the indidual time indices where a token starts and ends. 778 779 Args: 780 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 781 emitted at every time step after rnnt collapse. 782 token_repetitions: A list of ints representing the number of repetitions of each emitted token. 783 rnnt_token: The integer of the rnnt blank token used during rnnt collapse. 784 785 Returns: 786 787 """ 788 start_index = 0 789 790 # If the exact timestep information is available, utilize the 1st non-rnnt blank token timestep 791 # as the start index. 792 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 793 start_index = max(0, hypothesis.timestep[0] - 1) 794 795 # Construct the start and end indices brackets 796 end_indices = np.asarray(token_repetitions).cumsum() 797 start_indices = np.concatenate(([start_index], end_indices[:-1])) 798 799 # Process the TxU dangling alignment tensor, containing pairs of (logits, label) 800 alignment_labels = [al_logits_labels for al_logits_labels in hypothesis.text[1]] 801 for t in range(len(alignment_labels)): 802 for u in range(len(alignment_labels[t])): 803 alignment_labels[t][u] = alignment_labels[t][u][1] # pick label from (logit, label) tuple 804 805 # Merge the results per token into a list of dictionaries 806 offsets = [ 807 {"char": a, "start_offset": s, "end_offset": e} 808 for a, s, e in zip(alignment_labels, start_indices, end_indices) 809 ] 810 811 # Filter out RNNT token (blank at [t][0] position). This is because blank can only occur at end of a 812 # time step for RNNT, so if 0th token is blank, then that timestep is skipped. 813 offsets = list(filter(lambda offsets: offsets["char"][0] != rnnt_token, offsets)) 814 return offsets 815 816 @staticmethod 817 def _get_word_offsets_chars( 818 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 819 ) -> Dict[str, Union[str, float]]: 820 """ 821 Utility method which constructs word time stamps out of character time stamps. 822 823 References: 824 This code is a port of the Hugging Face code for word time stamp construction. 825 826 Args: 827 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 828 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 829 830 Returns: 831 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 832 "end_offset". 833 """ 834 word_offsets = [] 835 836 last_state = "SPACE" 837 word = "" 838 start_offset = 0 839 end_offset = 0 840 for i, offset in enumerate(offsets): 841 chars = offset["char"] 842 for char in chars: 843 state = "SPACE" if char == word_delimiter_char else "WORD" 844 845 if state == last_state: 846 # If we are in the same state as before, we simply repeat what we've done before 847 end_offset = offset["end_offset"] 848 word += char 849 else: 850 # Switching state 851 if state == "SPACE": 852 # Finishing a word 853 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 854 else: 855 # Starting a new word 856 start_offset = offset["start_offset"] 857 end_offset = offset["end_offset"] 858 word = char 859 860 last_state = state 861 862 if last_state == "WORD": 863 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 864 865 return word_offsets 866 867 @staticmethod 868 def _get_word_offsets_subwords_sentencepiece( 869 offsets: Dict[str, Union[str, float]], 870 hypothesis: Hypothesis, 871 decode_ids_to_tokens: Callable[[List[int]], str], 872 decode_tokens_to_str: Callable[[List[int]], str], 873 ) -> Dict[str, Union[str, float]]: 874 """ 875 Utility method which constructs word time stamps out of sub-word time stamps. 876 877 **Note**: Only supports Sentencepiece based tokenizers ! 878 879 Args: 880 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 881 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 882 after rnnt collapse. 883 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 884 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 885 886 Returns: 887 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 888 "end_offset". 889 """ 890 word_offsets = [] 891 built_token = [] 892 previous_token_index = 0 893 # For every offset token 894 for i, offset in enumerate(offsets): 895 # For every subword token in offset token list (ignoring the RNNT Blank token at the end) 896 for char in offset['char'][:-1]: 897 char = int(char) 898 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if built_token: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 # This should only be done when these arrays contain more than one element. 931 if offsets and word_offsets: 932 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 933 934 # If there are any remaining tokens left, inject them all into the final word offset. 935 # The start offset of this token is the start time of the next token to process. 936 # The end offset of this token is the end time of the last token from offsets. 937 # Note that built_token is a flat list; but offsets contains a nested list which 938 # may have different dimensionality. 939 # As such, we can't rely on the length of the list of built_token to index offsets. 940 if built_token: 941 # start from the previous token index as this hasn't been committed to word_offsets yet 942 # if we still have content in built_token 943 start_offset = offsets[previous_token_index]["start_offset"] 944 word_offsets.append( 945 { 946 "word": decode_tokens_to_str(built_token), 947 "start_offset": start_offset, 948 "end_offset": offsets[-1]["end_offset"], 949 } 950 ) 951 built_token.clear() 952 953 return word_offsets 954 955 956 class RNNTDecoding(AbstractRNNTDecoding): 957 """ 958 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 959 960 Args: 961 decoding_cfg: A dict-like object which contains the following key-value pairs. 962 strategy: str value which represents the type of decoding that can occur. 963 Possible values are : 964 - greedy, greedy_batch (for greedy decoding). 965 - beam, tsd, alsd (for beam search decoding). 966 967 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 968 tokens as well as the decoded string. Default is False in order to avoid double decoding 969 unless required. 970 971 preserve_alignments: Bool flag which preserves the history of logprobs generated during 972 decoding (sample / batched). When set to true, the Hypothesis will contain 973 the non-null value for `logprobs` in it. Here, `alignments` is a List of List of 974 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 975 976 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 977 with the `return_hypotheses` flag set to True. 978 979 The length of the list corresponds to the Acoustic Length (T). 980 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 981 U is the number of target tokens for the current timestep Ti. 982 983 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 984 scores. In order to obtain hypotheses with confidence scores, please utilize 985 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 986 987 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 988 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 989 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 990 991 The length of the list corresponds to the Acoustic Length (T). 992 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 993 U is the number of target tokens for the current timestep Ti. 994 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 995 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 996 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 997 998 The length of the list corresponds to the number of recognized tokens. 999 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1000 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1001 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1002 1003 The length of the list corresponds to the number of recognized words. 1004 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1005 from the `token_confidence`. 1006 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1007 Valid options are `mean`, `min`, `max`, `prod`. 1008 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1009 confidence scores. 1010 1011 name: The method name (str). 1012 Supported values: 1013 - 'max_prob' for using the maximum token probability as a confidence. 1014 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1015 1016 entropy_type: Which type of entropy to use (str). 1017 Used if confidence_method_cfg.name is set to `entropy`. 1018 Supported values: 1019 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1020 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1021 Note that for this entropy, the alpha should comply the following inequality: 1022 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1023 where V is the model vocabulary size. 1024 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1025 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1026 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1027 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1028 - 'renyi' for the Rรฉnyi entropy. 1029 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1030 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1031 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1032 1033 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1034 When the alpha equals one, scaling is not applied to 'max_prob', 1035 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1036 1037 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1038 Supported values: 1039 - 'lin' for using the linear mapping. 1040 - 'exp' for using exponential mapping with linear shift. 1041 1042 The config may further contain the following sub-dictionaries: 1043 "greedy": 1044 max_symbols: int, describing the maximum number of target tokens to decode per 1045 timestep during greedy decoding. Setting to larger values allows longer sentences 1046 to be decoded, at the cost of increased execution time. 1047 1048 preserve_frame_confidence: Same as above, overrides above value. 1049 1050 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 1051 1052 "beam": 1053 beam_size: int, defining the beam size for beam search. Must be >= 1. 1054 If beam_size == 1, will perform cached greedy search. This might be slightly different 1055 results compared to the greedy search above. 1056 1057 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 1058 Set to True by default. 1059 1060 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1061 hypotheses after beam search has concluded. This flag is set by default. 1062 1063 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 1064 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 1065 at increased cost to execution time. 1066 1067 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 1068 If an integer is provided, it can decode sequences of that particular maximum length. 1069 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 1070 where seq_len is the length of the acoustic model output (T). 1071 1072 NOTE: 1073 If a float is provided, it can be greater than 1! 1074 By default, a float of 2.0 is used so that a target sequence can be at most twice 1075 as long as the acoustic model output length T. 1076 1077 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 1078 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 1079 1080 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 1081 in order to reduce expensive beam search cost later. int >= 0. 1082 1083 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 1084 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 1085 and affects the speed of inference since large values will perform large beam search in the next step. 1086 1087 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 1088 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 1089 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 1090 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 1091 expansion apart from the "most likely" candidate. 1092 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 1093 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 1094 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 1095 tuned on a validation set. 1096 1097 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 1098 1099 decoder: The Decoder/Prediction network module. 1100 joint: The Joint network module. 1101 vocabulary: The vocabulary (excluding the RNNT blank token) which will be used for decoding. 1102 """ 1103 1104 def __init__( 1105 self, decoding_cfg, decoder, joint, vocabulary, 1106 ): 1107 # we need to ensure blank is the last token in the vocab for the case of RNNT and Multi-blank RNNT. 1108 blank_id = len(vocabulary) + joint.num_extra_outputs 1109 1110 if hasattr(decoding_cfg, 'model_type') and decoding_cfg.model_type == 'tdt': 1111 blank_id = len(vocabulary) 1112 1113 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1114 1115 super(RNNTDecoding, self).__init__( 1116 decoding_cfg=decoding_cfg, decoder=decoder, joint=joint, blank_id=blank_id, 1117 ) 1118 1119 if isinstance(self.decoding, beam_decode.BeamRNNTInfer): 1120 self.decoding.set_decoding_type('char') 1121 1122 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1123 """ 1124 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1125 1126 Args: 1127 hypothesis: Hypothesis 1128 1129 Returns: 1130 A list of word-level confidence scores. 1131 """ 1132 return self._aggregate_token_confidence_chars(hypothesis.words, hypothesis.token_confidence) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c < self.blank_id - self.num_extra_outputs] 1159 return token_list 1160 1161 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 1162 """ 1163 Compute the most likely language ID (LID) string given the tokens. 1164 1165 Args: 1166 tokens: List of int representing the token ids. 1167 1168 Returns: 1169 A decoded LID string. 1170 """ 1171 lang = self.tokenizer.ids_to_lang(tokens) 1172 return lang 1173 1174 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 1175 """ 1176 Decode a token id list into language ID (LID) list. 1177 1178 Args: 1179 tokens: List of int representing the token ids. 1180 1181 Returns: 1182 A list of decoded LIDS. 1183 """ 1184 lang_list = self.tokenizer.ids_to_text_and_langs(tokens) 1185 return lang_list 1186 1187 1188 class RNNTWER(Metric): 1189 """ 1190 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. 1191 When doing distributed training/evaluation the result of res=WER(predictions, targets, target_lengths) calls 1192 will be all-reduced between all workers using SUM operations. 1193 Here contains two numbers res=[wer_numerator, wer_denominator]. WER=wer_numerator/wer_denominator. 1194 1195 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. 1196 Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1197 1198 Example: 1199 def validation_step(self, batch, batch_idx): 1200 ... 1201 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1202 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1203 return self.val_outputs 1204 1205 def on_validation_epoch_end(self): 1206 ... 1207 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1208 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1209 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1210 self.val_outputs.clear() # free memory 1211 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1212 1213 Args: 1214 decoding: RNNTDecoding object that will perform autoregressive decoding of the RNNT model. 1215 batch_dim_index: Index of the batch dimension. 1216 use_cer: Whether to use Character Error Rate isntead of Word Error Rate. 1217 log_prediction: Whether to log a single decoded sample per call. 1218 1219 Returns: 1220 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenshtein's 1221 distances for all prediction - reference pairs, total number of words in all references. 1222 """ 1223 1224 full_state_update = True 1225 1226 def __init__( 1227 self, decoding: RNNTDecoding, batch_dim_index=0, use_cer=False, log_prediction=True, dist_sync_on_step=False 1228 ): 1229 super(RNNTWER, self).__init__(dist_sync_on_step=dist_sync_on_step) 1230 self.decoding = decoding 1231 self.batch_dim_index = batch_dim_index 1232 self.use_cer = use_cer 1233 self.log_prediction = log_prediction 1234 self.blank_id = self.decoding.blank_id 1235 self.labels_map = self.decoding.labels_map 1236 1237 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1238 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1239 1240 def update( 1241 self, 1242 encoder_output: torch.Tensor, 1243 encoded_lengths: torch.Tensor, 1244 targets: torch.Tensor, 1245 target_lengths: torch.Tensor, 1246 ) -> torch.Tensor: 1247 words = 0 1248 scores = 0 1249 references = [] 1250 with torch.no_grad(): 1251 # prediction_cpu_tensor = tensors[0].long().cpu() 1252 targets_cpu_tensor = targets.long().cpu() 1253 targets_cpu_tensor = move_dimension_to_the_front(targets_cpu_tensor, self.batch_dim_index) 1254 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1255 1256 # iterate over batch 1257 for ind in range(targets_cpu_tensor.shape[0]): 1258 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1259 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1260 1261 reference = self.decoding.decode_tokens_to_str(target) 1262 references.append(reference) 1263 1264 hypotheses, _ = self.decoding.rnnt_decoder_predictions_tensor(encoder_output, encoded_lengths) 1265 1266 if self.log_prediction: 1267 logging.info(f"\n") 1268 logging.info(f"reference :{references[0]}") 1269 logging.info(f"predicted :{hypotheses[0]}") 1270 1271 for h, r in zip(hypotheses, references): 1272 if self.use_cer: 1273 h_list = list(h) 1274 r_list = list(r) 1275 else: 1276 h_list = h.split() 1277 r_list = r.split() 1278 words += len(r_list) 1279 # Compute Levenshtein's distance 1280 scores += editdistance.eval(h_list, r_list) 1281 1282 self.scores += torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1283 self.words += torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1284 # return torch.tensor([scores, words]).to(predictions.device) 1285 1286 def compute(self): 1287 wer = self.scores.float() / self.words 1288 return wer, self.scores.detach(), self.words.detach() 1289 1290 1291 @dataclass 1292 class RNNTDecodingConfig: 1293 model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" 1294 strategy: str = "greedy_batch" 1295 1296 compute_hypothesis_token_set: bool = False 1297 1298 # preserve decoding alignments 1299 preserve_alignments: Optional[bool] = None 1300 1301 # confidence config 1302 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 1307 # compute RNNT time stamps 1308 compute_timestamps: Optional[bool] = None 1309 1310 # compute language IDs 1311 compute_langs: bool = False 1312 1313 # token representing word seperator 1314 word_seperator: str = " " 1315 1316 # type of timestamps to calculate 1317 rnnt_timestamp_type: str = "all" # can be char, word or all for both 1318 1319 # greedy decoding config 1320 greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() 1321 1322 # beam decoding config 1323 beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) 1324 1325 # can be used to change temperature for decoding 1326 temperature: float = 1.0 1327 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/metrics/wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import re 16 from abc import abstractmethod 17 from dataclasses import dataclass, is_dataclass 18 from typing import Callable, Dict, List, Optional, Tuple, Union 19 20 import editdistance 21 import jiwer 22 import numpy as np 23 import torch 24 from omegaconf import DictConfig, OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding 28 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 29 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 30 from nemo.utils import logging, logging_mode 31 32 __all__ = ['word_error_rate', 'word_error_rate_detail', 'WER', 'move_dimension_to_the_front'] 33 34 35 def word_error_rate(hypotheses: List[str], references: List[str], use_cer=False) -> float: 36 """ 37 Computes Average Word Error rate between two texts represented as 38 corresponding lists of string. 39 40 Hypotheses and references must have same length. 41 42 Args: 43 hypotheses (list): list of hypotheses 44 references(list) : list of references 45 use_cer (bool): set True to enable cer 46 47 Returns: 48 wer (float): average word error rate 49 """ 50 scores = 0 51 words = 0 52 if len(hypotheses) != len(references): 53 raise ValueError( 54 "In word error rate calculation, hypotheses and reference" 55 " lists must have the same number of elements. But I got:" 56 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 57 ) 58 for h, r in zip(hypotheses, references): 59 if use_cer: 60 h_list = list(h) 61 r_list = list(r) 62 else: 63 h_list = h.split() 64 r_list = r.split() 65 words += len(r_list) 66 # May deprecate using editdistance in future release for here and rest of codebase 67 # once we confirm jiwer is reliable. 68 scores += editdistance.eval(h_list, r_list) 69 if words != 0: 70 wer = 1.0 * scores / words 71 else: 72 wer = float('inf') 73 return wer 74 75 76 def word_error_rate_detail( 77 hypotheses: List[str], references: List[str], use_cer=False 78 ) -> Tuple[float, int, float, float, float]: 79 """ 80 Computes Average Word Error Rate with details (insertion rate, deletion rate, substitution rate) 81 between two texts represented as corresponding lists of string. 82 83 Hypotheses and references must have same length. 84 85 Args: 86 hypotheses (list): list of hypotheses 87 references(list) : list of references 88 use_cer (bool): set True to enable cer 89 90 Returns: 91 wer (float): average word error rate 92 words (int): Total number of words/charactors of given reference texts 93 ins_rate (float): average insertion error rate 94 del_rate (float): average deletion error rate 95 sub_rate (float): average substitution error rate 96 """ 97 scores = 0 98 words = 0 99 ops_count = {'substitutions': 0, 'insertions': 0, 'deletions': 0} 100 101 if len(hypotheses) != len(references): 102 raise ValueError( 103 "In word error rate calculation, hypotheses and reference" 104 " lists must have the same number of elements. But I got:" 105 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 106 ) 107 108 for h, r in zip(hypotheses, references): 109 if use_cer: 110 h_list = list(h) 111 r_list = list(r) 112 else: 113 h_list = h.split() 114 r_list = r.split() 115 116 # To get rid of the issue that jiwer does not allow empty string 117 if len(r_list) == 0: 118 if len(h_list) != 0: 119 errors = len(h_list) 120 ops_count['insertions'] += errors 121 else: 122 errors = 0 123 else: 124 if use_cer: 125 measures = jiwer.cer(r, h, return_dict=True) 126 else: 127 measures = jiwer.compute_measures(r, h) 128 129 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 130 ops_count['insertions'] += measures['insertions'] 131 ops_count['deletions'] += measures['deletions'] 132 ops_count['substitutions'] += measures['substitutions'] 133 134 scores += errors 135 words += len(r_list) 136 137 if words != 0: 138 wer = 1.0 * scores / words 139 ins_rate = 1.0 * ops_count['insertions'] / words 140 del_rate = 1.0 * ops_count['deletions'] / words 141 sub_rate = 1.0 * ops_count['substitutions'] / words 142 else: 143 wer, ins_rate, del_rate, sub_rate = float('inf'), float('inf'), float('inf'), float('inf') 144 145 return wer, words, ins_rate, del_rate, sub_rate 146 147 148 def word_error_rate_per_utt(hypotheses: List[str], references: List[str], use_cer=False) -> Tuple[List[float], float]: 149 """ 150 Computes Word Error Rate per utterance and the average WER 151 between two texts represented as corresponding lists of string. 152 153 Hypotheses and references must have same length. 154 155 Args: 156 hypotheses (list): list of hypotheses 157 references(list) : list of references 158 use_cer (bool): set True to enable cer 159 160 Returns: 161 wer_per_utt (List[float]): word error rate per utterance 162 avg_wer (float): average word error rate 163 """ 164 scores = 0 165 words = 0 166 wer_per_utt = [] 167 168 if len(hypotheses) != len(references): 169 raise ValueError( 170 "In word error rate calculation, hypotheses and reference" 171 " lists must have the same number of elements. But I got:" 172 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 173 ) 174 175 for h, r in zip(hypotheses, references): 176 if use_cer: 177 h_list = list(h) 178 r_list = list(r) 179 else: 180 h_list = h.split() 181 r_list = r.split() 182 183 # To get rid of the issue that jiwer does not allow empty string 184 if len(r_list) == 0: 185 if len(h_list) != 0: 186 errors = len(h_list) 187 wer_per_utt.append(float('inf')) 188 else: 189 if use_cer: 190 measures = jiwer.cer(r, h, return_dict=True) 191 er = measures['cer'] 192 else: 193 measures = jiwer.compute_measures(r, h) 194 er = measures['wer'] 195 196 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 197 wer_per_utt.append(er) 198 199 scores += errors 200 words += len(r_list) 201 202 if words != 0: 203 avg_wer = 1.0 * scores / words 204 else: 205 avg_wer = float('inf') 206 207 return wer_per_utt, avg_wer 208 209 210 def move_dimension_to_the_front(tensor, dim_index): 211 all_dims = list(range(tensor.ndim)) 212 return tensor.permute(*([dim_index] + all_dims[:dim_index] + all_dims[dim_index + 1 :])) 213 214 215 class AbstractCTCDecoding(ConfidenceMixin): 216 """ 217 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs. 218 219 Args: 220 decoding_cfg: A dict-like object which contains the following key-value pairs. 221 strategy: str value which represents the type of decoding that can occur. 222 Possible values are : 223 - greedy (for greedy decoding). 224 - beam (for DeepSpeed KenLM based decoding). 225 226 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 227 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 228 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 229 230 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 231 Can take the following values - "char" for character/subword time stamps, "word" for word level 232 time stamps and "all" (default), for both character level and word level time stamps. 233 234 word_seperator: Str token representing the seperator between words. 235 236 preserve_alignments: Bool flag which preserves the history of logprobs generated during 237 decoding (sample / batched). When set to true, the Hypothesis will contain 238 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 239 240 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 241 scores. In order to obtain hypotheses with confidence scores, please utilize 242 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 243 244 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 245 generated during decoding. When set to true, the Hypothesis will contain 246 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 247 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 248 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 249 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 250 251 The length of the list corresponds to the number of recognized tokens. 252 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 253 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 254 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 255 256 The length of the list corresponds to the number of recognized words. 257 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 258 from the `token_confidence`. 259 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 260 Valid options are `mean`, `min`, `max`, `prod`. 261 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 262 confidence scores. 263 264 name: The method name (str). 265 Supported values: 266 - 'max_prob' for using the maximum token probability as a confidence. 267 - 'entropy' for using a normalized entropy of a log-likelihood vector. 268 269 entropy_type: Which type of entropy to use (str). 270 Used if confidence_method_cfg.name is set to `entropy`. 271 Supported values: 272 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 273 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 274 Note that for this entropy, the alpha should comply the following inequality: 275 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 276 where V is the model vocabulary size. 277 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 278 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 279 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 280 More: https://en.wikipedia.org/wiki/Tsallis_entropy 281 - 'renyi' for the Rรฉnyi entropy. 282 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 285 286 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 287 When the alpha equals one, scaling is not applied to 'max_prob', 288 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 289 290 entropy_norm: A mapping of the entropy value to the interval [0,1]. 291 Supported values: 292 - 'lin' for using the linear mapping. 293 - 'exp' for using exponential mapping with linear shift. 294 295 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 296 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 297 298 The config may further contain the following sub-dictionaries: 299 "greedy": 300 preserve_alignments: Same as above, overrides above value. 301 compute_timestamps: Same as above, overrides above value. 302 preserve_frame_confidence: Same as above, overrides above value. 303 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 304 305 "beam": 306 beam_size: int, defining the beam size for beam search. Must be >= 1. 307 If beam_size == 1, will perform cached greedy search. This might be slightly different 308 results compared to the greedy search above. 309 310 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 311 hypotheses after beam search has concluded. This flag is set by default. 312 313 beam_alpha: float, the strength of the Language model on the final score of a token. 314 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 315 316 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 317 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 318 319 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 320 If the path is invalid (file is not found at path), will raise a deferred error at the moment 321 of calculation of beam search, so that users may update / change the decoding strategy 322 to point to the correct file. 323 324 blank_id: The id of the RNNT blank token. 325 """ 326 327 def __init__(self, decoding_cfg, blank_id: int): 328 super().__init__() 329 330 # Convert dataclas to config 331 if is_dataclass(decoding_cfg): 332 decoding_cfg = OmegaConf.structured(decoding_cfg) 333 334 if not isinstance(decoding_cfg, DictConfig): 335 decoding_cfg = OmegaConf.create(decoding_cfg) 336 337 OmegaConf.set_struct(decoding_cfg, False) 338 339 # update minimal config 340 minimal_cfg = ['greedy'] 341 for item in minimal_cfg: 342 if item not in decoding_cfg: 343 decoding_cfg[item] = OmegaConf.create({}) 344 345 self.cfg = decoding_cfg 346 self.blank_id = blank_id 347 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 348 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 349 self.batch_dim_index = self.cfg.get('batch_dim_index', 0) 350 self.word_seperator = self.cfg.get('word_seperator', ' ') 351 352 possible_strategies = ['greedy', 'beam', 'pyctcdecode', 'flashlight'] 353 if self.cfg.strategy not in possible_strategies: 354 raise ValueError(f"Decoding strategy must be one of {possible_strategies}. Given {self.cfg.strategy}") 355 356 # Update preserve alignments 357 if self.preserve_alignments is None: 358 if self.cfg.strategy in ['greedy']: 359 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 360 else: 361 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 362 363 # Update compute timestamps 364 if self.compute_timestamps is None: 365 if self.cfg.strategy in ['greedy']: 366 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 367 elif self.cfg.strategy in ['beam']: 368 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 369 370 # initialize confidence-related fields 371 self._init_confidence(self.cfg.get('confidence_cfg', None)) 372 373 # Confidence estimation is not implemented for strategies other than `greedy` 374 if ( 375 not self.preserve_frame_confidence 376 and self.cfg.strategy != 'greedy' 377 and self.cfg.beam.get('preserve_frame_confidence', False) 378 ): 379 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 380 381 # we need timestamps to extract non-blank per-frame confidence 382 if self.compute_timestamps is not None: 383 self.compute_timestamps |= self.preserve_frame_confidence 384 385 if self.cfg.strategy == 'greedy': 386 387 self.decoding = ctc_greedy_decoding.GreedyCTCInfer( 388 blank_id=self.blank_id, 389 preserve_alignments=self.preserve_alignments, 390 compute_timestamps=self.compute_timestamps, 391 preserve_frame_confidence=self.preserve_frame_confidence, 392 confidence_method_cfg=self.confidence_method_cfg, 393 ) 394 395 elif self.cfg.strategy == 'beam': 396 397 self.decoding = ctc_beam_decoding.BeamCTCInfer( 398 blank_id=blank_id, 399 beam_size=self.cfg.beam.get('beam_size', 1), 400 search_type='default', 401 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 402 preserve_alignments=self.preserve_alignments, 403 compute_timestamps=self.compute_timestamps, 404 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 405 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 406 kenlm_path=self.cfg.beam.get('kenlm_path', None), 407 ) 408 409 self.decoding.override_fold_consecutive_value = False 410 411 elif self.cfg.strategy == 'pyctcdecode': 412 413 self.decoding = ctc_beam_decoding.BeamCTCInfer( 414 blank_id=blank_id, 415 beam_size=self.cfg.beam.get('beam_size', 1), 416 search_type='pyctcdecode', 417 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 418 preserve_alignments=self.preserve_alignments, 419 compute_timestamps=self.compute_timestamps, 420 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 421 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 422 kenlm_path=self.cfg.beam.get('kenlm_path', None), 423 pyctcdecode_cfg=self.cfg.beam.get('pyctcdecode_cfg', None), 424 ) 425 426 self.decoding.override_fold_consecutive_value = False 427 428 elif self.cfg.strategy == 'flashlight': 429 430 self.decoding = ctc_beam_decoding.BeamCTCInfer( 431 blank_id=blank_id, 432 beam_size=self.cfg.beam.get('beam_size', 1), 433 search_type='flashlight', 434 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 435 preserve_alignments=self.preserve_alignments, 436 compute_timestamps=self.compute_timestamps, 437 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 438 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 439 kenlm_path=self.cfg.beam.get('kenlm_path', None), 440 flashlight_cfg=self.cfg.beam.get('flashlight_cfg', None), 441 ) 442 443 self.decoding.override_fold_consecutive_value = False 444 445 else: 446 raise ValueError( 447 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 448 f"but was provided {self.cfg.strategy}" 449 ) 450 451 def ctc_decoder_predictions_tensor( 452 self, 453 decoder_outputs: torch.Tensor, 454 decoder_lengths: torch.Tensor = None, 455 fold_consecutive: bool = True, 456 return_hypotheses: bool = False, 457 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 458 """ 459 Decodes a sequence of labels to words 460 461 Args: 462 decoder_outputs: An integer torch.Tensor of shape [Batch, Time, {Vocabulary}] (if ``batch_index_dim == 0``) or [Time, Batch] 463 (if ``batch_index_dim == 1``) of integer indices that correspond to the index of some character in the 464 label set. 465 decoder_lengths: Optional tensor of length `Batch` which contains the integer lengths 466 of the sequence in the padded `predictions` tensor. 467 fold_consecutive: Bool, determine whether to perform "ctc collapse", folding consecutive tokens 468 into a single token. 469 return_hypotheses: Bool flag whether to return just the decoding predictions of the model 470 or a Hypothesis object that holds information such as the decoded `text`, 471 the `alignment` of emited by the CTC Model, and the `length` of the sequence (if available). 472 May also contain the log-probabilities of the decoder (if this method is called via 473 transcribe()) 474 475 Returns: 476 Either a list of str which represent the CTC decoded strings per sample, 477 or a list of Hypothesis objects containing additional information. 478 """ 479 480 if isinstance(decoder_outputs, torch.Tensor): 481 decoder_outputs = move_dimension_to_the_front(decoder_outputs, self.batch_dim_index) 482 483 if ( 484 hasattr(self.decoding, 'override_fold_consecutive_value') 485 and self.decoding.override_fold_consecutive_value is not None 486 ): 487 logging.info( 488 f"Beam search requires that consecutive ctc tokens are not folded. \n" 489 f"Overriding provided value of `fold_consecutive` = {fold_consecutive} to " 490 f"{self.decoding.override_fold_consecutive_value}", 491 mode=logging_mode.ONCE, 492 ) 493 fold_consecutive = self.decoding.override_fold_consecutive_value 494 495 with torch.inference_mode(): 496 # Resolve the forward step of the decoding strategy 497 hypotheses_list = self.decoding( 498 decoder_output=decoder_outputs, decoder_lengths=decoder_lengths 499 ) # type: List[List[Hypothesis]] 500 501 # extract the hypotheses 502 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 503 504 if isinstance(hypotheses_list[0], NBestHypotheses): 505 hypotheses = [] 506 all_hypotheses = [] 507 508 for nbest_hyp in hypotheses_list: # type: NBestHypotheses 509 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 510 decoded_hyps = self.decode_hypothesis( 511 n_hyps, fold_consecutive 512 ) # type: List[Union[Hypothesis, NBestHypotheses]] 513 514 # If computing timestamps 515 if self.compute_timestamps is True: 516 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 517 for hyp_idx in range(len(decoded_hyps)): 518 decoded_hyps[hyp_idx] = self.compute_ctc_timestamps(decoded_hyps[hyp_idx], timestamp_type) 519 520 hypotheses.append(decoded_hyps[0]) # best hypothesis 521 all_hypotheses.append(decoded_hyps) 522 523 if return_hypotheses: 524 return hypotheses, all_hypotheses 525 526 best_hyp_text = [h.text for h in hypotheses] 527 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 528 return best_hyp_text, all_hyp_text 529 530 else: 531 hypotheses = self.decode_hypothesis( 532 hypotheses_list, fold_consecutive 533 ) # type: List[Union[Hypothesis, NBestHypotheses]] 534 535 # If computing timestamps 536 if self.compute_timestamps is True: 537 # greedy decoding, can get high-level confidence scores 538 if return_hypotheses and (self.preserve_word_confidence or self.preserve_token_confidence): 539 hypotheses = self.compute_confidence(hypotheses) 540 else: 541 # remove unused token_repetitions from Hypothesis.text 542 for hyp in hypotheses: 543 hyp.text = hyp.text[:2] 544 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 545 for hyp_idx in range(len(hypotheses)): 546 hypotheses[hyp_idx] = self.compute_ctc_timestamps(hypotheses[hyp_idx], timestamp_type) 547 548 if return_hypotheses: 549 return hypotheses, None 550 551 best_hyp_text = [h.text for h in hypotheses] 552 return best_hyp_text, None 553 554 def decode_hypothesis( 555 self, hypotheses_list: List[Hypothesis], fold_consecutive: bool 556 ) -> List[Union[Hypothesis, NBestHypotheses]]: 557 """ 558 Decode a list of hypotheses into a list of strings. 559 560 Args: 561 hypotheses_list: List of Hypothesis. 562 fold_consecutive: Whether to collapse the ctc blank tokens or not. 563 564 Returns: 565 A list of strings. 566 """ 567 for ind in range(len(hypotheses_list)): 568 # Extract the integer encoded hypothesis 569 hyp = hypotheses_list[ind] 570 prediction = hyp.y_sequence 571 predictions_len = hyp.length if hyp.length > 0 else None 572 573 if fold_consecutive: 574 if type(prediction) != list: 575 prediction = prediction.numpy().tolist() 576 577 if predictions_len is not None: 578 prediction = prediction[:predictions_len] 579 580 # CTC decoding procedure 581 decoded_prediction = [] 582 token_lengths = [] # preserve token lengths 583 token_repetitions = [] # preserve number of repetitions per token 584 585 previous = self.blank_id 586 last_length = 0 587 last_repetition = 1 588 589 for pidx, p in enumerate(prediction): 590 if (p != previous or previous == self.blank_id) and p != self.blank_id: 591 decoded_prediction.append(p) 592 593 token_lengths.append(pidx - last_length) 594 last_length = pidx 595 token_repetitions.append(last_repetition) 596 last_repetition = 1 597 598 if p == previous and previous != self.blank_id: 599 last_repetition += 1 600 601 previous = p 602 603 if len(token_repetitions) > 0: 604 token_repetitions = token_repetitions[1:] + [last_repetition] 605 606 else: 607 if predictions_len is not None: 608 prediction = prediction[:predictions_len] 609 decoded_prediction = prediction[prediction != self.blank_id].tolist() 610 token_lengths = [1] * len(decoded_prediction) # preserve number of repetitions per token 611 token_repetitions = [1] * len(decoded_prediction) # preserve number of repetitions per token 612 613 # De-tokenize the integer tokens; if not computing timestamps 614 if self.compute_timestamps is True: 615 # keep the original predictions, wrap with the number of repetitions per token 616 # this is done so that `ctc_decoder_predictions_tensor()` can process this hypothesis 617 # in order to compute exact time stamps. 618 hypothesis = (decoded_prediction, token_lengths, token_repetitions) 619 else: 620 hypothesis = self.decode_tokens_to_str(decoded_prediction) 621 622 # TODO: remove 623 # collapse leading spaces before . , ? for PC models 624 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 625 626 # Preserve this wrapped hypothesis or decoded text tokens. 627 hypotheses_list[ind].text = hypothesis 628 629 return hypotheses_list 630 631 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 632 """ 633 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 634 Assumes that `frame_confidence` is present in the hypotheses. 635 636 Args: 637 hypotheses_list: List of Hypothesis. 638 639 Returns: 640 A list of hypotheses with high-level confidence scores. 641 """ 642 for hyp in hypotheses_list: 643 if not isinstance(hyp.text, tuple) or len(hyp.text) != 3: 644 # the method must have been called in the wrong place 645 raise ValueError( 646 """Wrong format of the `text` attribute of a hypothesis.\n 647 Expected: (decoded_prediction, token_repetitions)\n 648 The method invocation is expected between .decode_hypothesis() and .compute_ctc_timestamps()""" 649 ) 650 token_repetitions = hyp.text[2] 651 hyp.text = hyp.text[:2] 652 token_confidence = [] 653 if self.exclude_blank_from_confidence: 654 non_blank_frame_confidence = hyp.non_blank_frame_confidence 655 i = 0 656 for tr in token_repetitions: 657 # token repetition can be zero 658 j = i + tr 659 token_confidence.append(self._aggregate_confidence(non_blank_frame_confidence[i:j])) 660 i = j 661 else: 662 # <blank> tokens are considered to belong to the last non-blank token, if any. 663 token_lengths = hyp.text[1] 664 if len(token_lengths) > 0: 665 ts = token_lengths[0] 666 for tl in token_lengths[1:] + [len(hyp.frame_confidence)]: 667 token_confidence.append(self._aggregate_confidence(hyp.frame_confidence[ts : ts + tl])) 668 ts += tl 669 hyp.token_confidence = token_confidence 670 if self.preserve_word_confidence: 671 for hyp in hypotheses_list: 672 hyp.word_confidence = self._aggregate_token_confidence(hyp) 673 return hypotheses_list 674 675 @abstractmethod 676 def decode_tokens_to_str(self, tokens: List[int]) -> str: 677 """ 678 Implemented by subclass in order to decoder a token id list into a string. 679 680 Args: 681 tokens: List of int representing the token ids. 682 683 Returns: 684 A decoded string. 685 """ 686 raise NotImplementedError() 687 688 @abstractmethod 689 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 690 """ 691 Implemented by subclass in order to decode a token id list into a token list. 692 A token list is the string representation of each token id. 693 694 Args: 695 tokens: List of int representing the token ids. 696 697 Returns: 698 A list of decoded tokens. 699 """ 700 raise NotImplementedError() 701 702 def compute_ctc_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 703 """ 704 Method to compute time stamps at char/subword, and word level given some hypothesis. 705 Requires the input hypothesis to contain a `text` field that is the tuple. The tuple contains - 706 the ctc collapsed integer ids, and the number of repetitions of each token. 707 708 Args: 709 hypothesis: A Hypothesis object, with a wrapped `text` field. 710 The `text` field must contain a tuple with two values - 711 The ctc collapsed integer ids 712 A list of integers that represents the number of repetitions per token. 713 timestamp_type: A str value that represents the type of time stamp calculated. 714 Can be one of "char", "word" or "all" 715 716 Returns: 717 A Hypothesis object with a modified `timestep` value, which is now a dictionary containing 718 the time stamp information. 719 """ 720 assert timestamp_type in ['char', 'word', 'all'] 721 722 # Unpack the temporary storage, and set the decoded predictions 723 decoded_prediction, token_lengths = hypothesis.text 724 hypothesis.text = decoded_prediction 725 726 # Retrieve offsets 727 char_offsets = word_offsets = None 728 char_offsets = self._compute_offsets(hypothesis, token_lengths, self.blank_id) 729 730 # Assert number of offsets and hypothesis tokens are 1:1 match. 731 if len(char_offsets) != len(hypothesis.text): 732 raise ValueError( 733 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 734 " have to be of the same length, but are: " 735 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 736 f" {len(hypothesis.text)}" 737 ) 738 739 # Correctly process the token ids to chars/subwords. 740 for i, char in enumerate(hypothesis.text): 741 char_offsets[i]["char"] = self.decode_tokens_to_str([char]) 742 743 # detect char vs subword models 744 lens = [len(list(v["char"])) > 1 for v in char_offsets] 745 if any(lens): 746 text_type = 'subword' 747 else: 748 text_type = 'char' 749 750 # retrieve word offsets from character offsets 751 word_offsets = None 752 if timestamp_type in ['word', 'all']: 753 if text_type == 'char': 754 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 755 else: 756 word_offsets = self._get_word_offsets_subwords_sentencepiece( 757 char_offsets, 758 hypothesis, 759 decode_ids_to_tokens=self.decode_ids_to_tokens, 760 decode_tokens_to_str=self.decode_tokens_to_str, 761 ) 762 763 # attach results 764 if len(hypothesis.timestep) > 0: 765 timestep_info = hypothesis.timestep 766 else: 767 timestep_info = [] 768 769 # Setup defaults 770 hypothesis.timestep = {"timestep": timestep_info} 771 772 # Add char / subword time stamps 773 if char_offsets is not None and timestamp_type in ['char', 'all']: 774 hypothesis.timestep['char'] = char_offsets 775 776 # Add word time stamps 777 if word_offsets is not None and timestamp_type in ['word', 'all']: 778 hypothesis.timestep['word'] = word_offsets 779 780 # Convert the token indices to text 781 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 782 783 return hypothesis 784 785 @staticmethod 786 def _compute_offsets( 787 hypothesis: Hypothesis, token_lengths: List[int], ctc_token: int 788 ) -> List[Dict[str, Union[str, int]]]: 789 """ 790 Utility method that calculates the indidual time indices where a token starts and ends. 791 792 Args: 793 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 794 emitted at every time step after ctc collapse. 795 token_lengths: A list of ints representing the lengths of each emitted token. 796 ctc_token: The integer of the ctc blank token used during ctc collapse. 797 798 Returns: 799 800 """ 801 start_index = 0 802 803 # If the exact timestep information is available, utilize the 1st non-ctc blank token timestep 804 # as the start index. 805 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 806 start_index = max(0, hypothesis.timestep[0] - 1) 807 808 # Construct the start and end indices brackets 809 end_indices = np.asarray(token_lengths).cumsum() 810 start_indices = np.concatenate(([start_index], end_indices[:-1])) 811 812 # Merge the results per token into a list of dictionaries 813 offsets = [ 814 {"char": t, "start_offset": s, "end_offset": e} 815 for t, s, e in zip(hypothesis.text, start_indices, end_indices) 816 ] 817 818 # Filter out CTC token 819 offsets = list(filter(lambda offsets: offsets["char"] != ctc_token, offsets)) 820 return offsets 821 822 @staticmethod 823 def _get_word_offsets_chars( 824 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 825 ) -> Dict[str, Union[str, float]]: 826 """ 827 Utility method which constructs word time stamps out of character time stamps. 828 829 References: 830 This code is a port of the Hugging Face code for word time stamp construction. 831 832 Args: 833 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 834 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 835 836 Returns: 837 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 838 "end_offset". 839 """ 840 word_offsets = [] 841 842 last_state = "SPACE" 843 word = "" 844 start_offset = 0 845 end_offset = 0 846 for i, offset in enumerate(offsets): 847 char = offset["char"] 848 state = "SPACE" if char == word_delimiter_char else "WORD" 849 850 if state == last_state: 851 # If we are in the same state as before, we simply repeat what we've done before 852 end_offset = offset["end_offset"] 853 word += char 854 else: 855 # Switching state 856 if state == "SPACE": 857 # Finishing a word 858 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 859 else: 860 # Starting a new word 861 start_offset = offset["start_offset"] 862 end_offset = offset["end_offset"] 863 word = char 864 865 last_state = state 866 if last_state == "WORD": 867 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 868 869 return word_offsets 870 871 @staticmethod 872 def _get_word_offsets_subwords_sentencepiece( 873 offsets: Dict[str, Union[str, float]], 874 hypothesis: Hypothesis, 875 decode_ids_to_tokens: Callable[[List[int]], str], 876 decode_tokens_to_str: Callable[[List[int]], str], 877 ) -> Dict[str, Union[str, float]]: 878 """ 879 Utility method which constructs word time stamps out of sub-word time stamps. 880 881 **Note**: Only supports Sentencepiece based tokenizers ! 882 883 Args: 884 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 885 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 886 after ctc collapse. 887 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 888 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 889 890 Returns: 891 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 892 "end_offset". 893 """ 894 word_offsets = [] 895 built_token = [] 896 previous_token_index = 0 897 # For every collapsed sub-word token 898 for i, char in enumerate(hypothesis.text): 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if len(built_token) > 0: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 if len(word_offsets) == 0: 931 # alaptev: sometimes word_offsets can be empty 932 if len(built_token) > 0: 933 word_offsets.append( 934 { 935 "word": decode_tokens_to_str(built_token), 936 "start_offset": offsets[0]["start_offset"], 937 "end_offset": offsets[-1]["end_offset"], 938 } 939 ) 940 built_token.clear() 941 else: 942 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 943 944 # If there are any remaining tokens left, inject them all into the final word offset. 945 # Note: The start offset of this token is the start time of the first token inside build_token. 946 # Note: The end offset of this token is the end time of the last token inside build_token 947 if len(built_token) > 0: 948 word_offsets.append( 949 { 950 "word": decode_tokens_to_str(built_token), 951 "start_offset": offsets[-(len(built_token))]["start_offset"], 952 "end_offset": offsets[-1]["end_offset"], 953 } 954 ) 955 built_token.clear() 956 957 return word_offsets 958 959 @property 960 def preserve_alignments(self): 961 return self._preserve_alignments 962 963 @preserve_alignments.setter 964 def preserve_alignments(self, value): 965 self._preserve_alignments = value 966 967 if hasattr(self, 'decoding'): 968 self.decoding.preserve_alignments = value 969 970 @property 971 def compute_timestamps(self): 972 return self._compute_timestamps 973 974 @compute_timestamps.setter 975 def compute_timestamps(self, value): 976 self._compute_timestamps = value 977 978 if hasattr(self, 'decoding'): 979 self.decoding.compute_timestamps = value 980 981 @property 982 def preserve_frame_confidence(self): 983 return self._preserve_frame_confidence 984 985 @preserve_frame_confidence.setter 986 def preserve_frame_confidence(self, value): 987 self._preserve_frame_confidence = value 988 989 if hasattr(self, 'decoding'): 990 self.decoding.preserve_frame_confidence = value 991 992 993 class CTCDecoding(AbstractCTCDecoding): 994 """ 995 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs for character 996 based models. 997 998 Args: 999 decoding_cfg: A dict-like object which contains the following key-value pairs. 1000 strategy: str value which represents the type of decoding that can occur. 1001 Possible values are : 1002 - greedy (for greedy decoding). 1003 - beam (for DeepSpeed KenLM based decoding). 1004 1005 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 1006 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 1007 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 1008 1009 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 1010 Can take the following values - "char" for character/subword time stamps, "word" for word level 1011 time stamps and "all" (default), for both character level and word level time stamps. 1012 1013 word_seperator: Str token representing the seperator between words. 1014 1015 preserve_alignments: Bool flag which preserves the history of logprobs generated during 1016 decoding (sample / batched). When set to true, the Hypothesis will contain 1017 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 1018 1019 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 1020 scores. In order to obtain hypotheses with confidence scores, please utilize 1021 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 1022 1023 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 1024 generated during decoding. When set to true, the Hypothesis will contain 1025 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 1026 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 1027 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1028 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 1029 1030 The length of the list corresponds to the number of recognized tokens. 1031 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1032 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1033 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1034 1035 The length of the list corresponds to the number of recognized words. 1036 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1037 from the `token_confidence`. 1038 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1039 Valid options are `mean`, `min`, `max`, `prod`. 1040 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1041 confidence scores. 1042 1043 name: The method name (str). 1044 Supported values: 1045 - 'max_prob' for using the maximum token probability as a confidence. 1046 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1047 1048 entropy_type: Which type of entropy to use (str). 1049 Used if confidence_method_cfg.name is set to `entropy`. 1050 Supported values: 1051 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1052 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1053 Note that for this entropy, the alpha should comply the following inequality: 1054 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1055 where V is the model vocabulary size. 1056 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1057 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1058 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1059 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1060 - 'renyi' for the Rรฉnyi entropy. 1061 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1062 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1063 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1064 1065 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1066 When the alpha equals one, scaling is not applied to 'max_prob', 1067 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1068 1069 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1070 Supported values: 1071 - 'lin' for using the linear mapping. 1072 - 'exp' for using exponential mapping with linear shift. 1073 1074 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 1075 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 1076 1077 The config may further contain the following sub-dictionaries: 1078 "greedy": 1079 preserve_alignments: Same as above, overrides above value. 1080 compute_timestamps: Same as above, overrides above value. 1081 preserve_frame_confidence: Same as above, overrides above value. 1082 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 1083 1084 "beam": 1085 beam_size: int, defining the beam size for beam search. Must be >= 1. 1086 If beam_size == 1, will perform cached greedy search. This might be slightly different 1087 results compared to the greedy search above. 1088 1089 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1090 hypotheses after beam search has concluded. This flag is set by default. 1091 1092 beam_alpha: float, the strength of the Language model on the final score of a token. 1093 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1094 1095 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 1096 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1097 1098 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 1099 If the path is invalid (file is not found at path), will raise a deferred error at the moment 1100 of calculation of beam search, so that users may update / change the decoding strategy 1101 to point to the correct file. 1102 1103 blank_id: The id of the RNNT blank token. 1104 """ 1105 1106 def __init__( 1107 self, decoding_cfg, vocabulary, 1108 ): 1109 blank_id = len(vocabulary) 1110 self.vocabulary = vocabulary 1111 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1112 1113 super().__init__(decoding_cfg=decoding_cfg, blank_id=blank_id) 1114 1115 # Finalize Beam Search Decoding framework 1116 if isinstance(self.decoding, ctc_beam_decoding.AbstractBeamCTCInfer): 1117 self.decoding.set_vocabulary(self.vocabulary) 1118 self.decoding.set_decoding_type('char') 1119 1120 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1121 """ 1122 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1123 1124 Args: 1125 hypothesis: Hypothesis 1126 1127 Returns: 1128 A list of word-level confidence scores. 1129 """ 1130 return self._aggregate_token_confidence_chars( 1131 self.decode_tokens_to_str(hypothesis.text[0]).split(), hypothesis.token_confidence 1132 ) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c != self.blank_id] 1159 return token_list 1160 1161 1162 class WER(Metric): 1163 """ 1164 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference 1165 texts. When doing distributed training/evaluation the result of ``res=WER(predictions, targets, target_lengths)`` 1166 calls will be all-reduced between all workers using SUM operations. Here ``res`` contains three numbers 1167 ``res=[wer, total_levenstein_distance, total_number_of_words]``. 1168 1169 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step 1170 results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1171 1172 Example: 1173 def validation_step(self, batch, batch_idx): 1174 ... 1175 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1176 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1177 return self.val_outputs 1178 1179 def on_validation_epoch_end(self): 1180 ... 1181 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1182 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1183 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1184 self.val_outputs.clear() # free memory 1185 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1186 1187 Args: 1188 decoding: An instance of CTCDecoding. 1189 use_cer: Whether to use Character Error Rate instead of Word Error Rate. 1190 log_prediction: Whether to log a single decoded sample per call. 1191 fold_consecutive: Whether repeated consecutive characters should be folded into one when decoding. 1192 1193 Returns: 1194 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenstein's 1195 distances for all prediction - reference pairs, total number of words in all references. 1196 """ 1197 1198 full_state_update: bool = True 1199 1200 def __init__( 1201 self, 1202 decoding: CTCDecoding, 1203 use_cer=False, 1204 log_prediction=True, 1205 fold_consecutive=True, 1206 dist_sync_on_step=False, 1207 ): 1208 super().__init__(dist_sync_on_step=dist_sync_on_step) 1209 1210 self.decoding = decoding 1211 self.use_cer = use_cer 1212 self.log_prediction = log_prediction 1213 self.fold_consecutive = fold_consecutive 1214 1215 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1216 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1217 1218 def update( 1219 self, 1220 predictions: torch.Tensor, 1221 targets: torch.Tensor, 1222 target_lengths: torch.Tensor, 1223 predictions_lengths: torch.Tensor = None, 1224 ): 1225 """ 1226 Updates metric state. 1227 Args: 1228 predictions: an integer torch.Tensor of shape ``[Batch, Time, {Vocabulary}]`` (if ``batch_dim_index == 0``) or 1229 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1230 targets: an integer torch.Tensor of shape ``[Batch, Time]`` (if ``batch_dim_index == 0``) or 1231 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1232 target_lengths: an integer torch.Tensor of shape ``[Batch]`` 1233 predictions_lengths: an integer torch.Tensor of shape ``[Batch]`` 1234 """ 1235 words = 0 1236 scores = 0 1237 references = [] 1238 with torch.no_grad(): 1239 # prediction_cpu_tensor = tensors[0].long().cpu() 1240 targets_cpu_tensor = targets.long().cpu() 1241 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1242 1243 # iterate over batch 1244 for ind in range(targets_cpu_tensor.shape[0]): 1245 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1246 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1247 reference = self.decoding.decode_tokens_to_str(target) 1248 references.append(reference) 1249 1250 hypotheses, _ = self.decoding.ctc_decoder_predictions_tensor( 1251 predictions, predictions_lengths, fold_consecutive=self.fold_consecutive 1252 ) 1253 1254 if self.log_prediction: 1255 logging.info(f"\n") 1256 logging.info(f"reference:{references[0]}") 1257 logging.info(f"predicted:{hypotheses[0]}") 1258 1259 for h, r in zip(hypotheses, references): 1260 if self.use_cer: 1261 h_list = list(h) 1262 r_list = list(r) 1263 else: 1264 h_list = h.split() 1265 r_list = r.split() 1266 words += len(r_list) 1267 # Compute Levenstein's distance 1268 scores += editdistance.eval(h_list, r_list) 1269 1270 self.scores = torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1271 self.words = torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1272 # return torch.tensor([scores, words]).to(predictions.device) 1273 1274 def compute(self): 1275 scores = self.scores.detach().float() 1276 words = self.words.detach().float() 1277 return scores / words, scores, words 1278 1279 1280 @dataclass 1281 class CTCDecodingConfig: 1282 strategy: str = "greedy" 1283 1284 # preserve decoding alignments 1285 preserve_alignments: Optional[bool] = None 1286 1287 # compute ctc time stamps 1288 compute_timestamps: Optional[bool] = None 1289 1290 # token representing word seperator 1291 word_seperator: str = " " 1292 1293 # type of timestamps to calculate 1294 ctc_timestamp_type: str = "all" # can be char, word or all for both 1295 1296 # batch dimension 1297 batch_dim_index: int = 0 1298 1299 # greedy decoding config 1300 greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() 1301 1302 # beam decoding config 1303 beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) 1304 1305 # confidence config 1306 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1307 1308 # can be used to change temperature for decoding 1309 temperature: float = 1.0 1310 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/asr/models/configs/aligner_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig 18 19 20 @dataclass 21 class AlignerCTCConfig: 22 prob_suppress_index: int = -1 23 prob_suppress_value: float = 1.0 24 25 26 @dataclass 27 class AlignerRNNTConfig: 28 predictor_window_size: int = 0 29 predictor_step_size: int = 1 30 31 32 @dataclass 33 class AlignerWrapperModelConfig: 34 alignment_type: str = "forced" 35 word_output: bool = True 36 cpu_decoding: bool = False 37 decode_batch_size: int = 0 38 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() 39 rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() 40 41 42 @dataclass 43 class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): 44 decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() 45 [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 22 from nemo.collections.asr.modules.audio_preprocessing import ( 23 AudioToMelSpectrogramPreprocessorConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class ASRDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[Any] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[Any] = None 40 tarred_shard_strategy: str = "scatter" 41 shard_manifests: bool = False 42 shuffle_n: int = 0 43 44 # Optional 45 int_values: Optional[int] = None 46 augmentor: Optional[Dict[str, Any]] = None 47 max_duration: Optional[float] = None 48 min_duration: Optional[float] = None 49 max_utts: int = 0 50 blank_index: int = -1 51 unk_index: int = -1 52 normalize: bool = False 53 trim: bool = True 54 parser: Optional[str] = 'en' 55 eos_id: Optional[int] = None 56 bos_id: Optional[int] = None 57 pad_id: int = 0 58 use_start_end_token: bool = False 59 return_sample_id: Optional[bool] = False 60 61 # bucketing params 62 bucketing_strategy: str = "synced_randomized" 63 bucketing_batch_size: Optional[Any] = None 64 bucketing_weights: Optional[List[int]] = None 65 66 67 @dataclass 68 class EncDecCTCConfig(model_cfg.ModelConfig): 69 # Model global arguments 70 sample_rate: int = 16000 71 repeat: int = 1 72 dropout: float = 0.0 73 separable: bool = False 74 labels: List[str] = MISSING 75 76 # Dataset configs 77 train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) 78 validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 79 test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 80 81 # Optimizer / Scheduler config 82 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 83 84 # Model component configs 85 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 86 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 87 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 88 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 89 decoding: CTCDecodingConfig = CTCDecodingConfig() 90 91 92 @dataclass 93 class EncDecCTCModelConfig(model_cfg.NemoConfig): 94 model: EncDecCTCConfig = EncDecCTCConfig() 95 96 97 @dataclass 98 class CacheAwareStreamingConfig: 99 chunk_size: int = 0 # the size of each chunk at each step, it can be a list of two integers to specify different chunk sizes for the first step and others 100 shift_size: int = 0 # the size of the shift in each step, it can be a list of two integers to specify different shift sizes for the first step and others 101 102 cache_drop_size: int = 0 # the number of steps to drop from the cache 103 last_channel_cache_size: int = 0 # the size of the needed cache for last channel layers 104 105 valid_out_len: int = 0 # the number of the steps in the final output which are valid (have the same value as in the offline mode) 106 107 pre_encode_cache_size: int = 0 # the size of the needed cache for the pre-encoding part of the model to avoid caching inside the pre-encoding layers 108 drop_extra_pre_encoded: int = 0 # the number of steps to get dropped after the pre-encoding layer 109 110 last_channel_num: int = 0 # number of the last channel layers (like MHA layers) which need caching in the model 111 last_time_num: int = 0 # number of the last time layers (like convolutions) which need caching in the model 112 [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderClassificationConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class EncDecClassificationDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[str] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[str] = None 40 tarred_shard_strategy: str = "scatter" 41 shuffle_n: int = 0 42 43 # Optional 44 int_values: Optional[int] = None 45 augmentor: Optional[Dict[str, Any]] = None 46 max_duration: Optional[float] = None 47 min_duration: Optional[float] = None 48 cal_labels_occurrence: Optional[bool] = False 49 50 # VAD Optional 51 vad_stream: Optional[bool] = None 52 window_length_in_sec: float = 0.31 53 shift_length_in_sec: float = 0.01 54 normalize_audio: bool = False 55 is_regression_task: bool = False 56 57 # bucketing params 58 bucketing_strategy: str = "synced_randomized" 59 bucketing_batch_size: Optional[Any] = None 60 bucketing_weights: Optional[List[int]] = None 61 62 63 @dataclass 64 class EncDecClassificationConfig(model_cfg.ModelConfig): 65 # Model global arguments 66 sample_rate: int = 16000 67 repeat: int = 1 68 dropout: float = 0.0 69 separable: bool = True 70 kernel_size_factor: float = 1.0 71 labels: List[str] = MISSING 72 timesteps: int = MISSING 73 74 # Dataset configs 75 train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 76 manifest_filepath=None, shuffle=True, trim_silence=False 77 ) 78 validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 79 manifest_filepath=None, shuffle=False 80 ) 81 test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 82 manifest_filepath=None, shuffle=False 83 ) 84 85 # Optimizer / Scheduler config 86 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 87 88 # Model component configs 89 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() 90 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 91 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 92 audio_length=timesteps 93 ) 94 95 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 96 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 97 98 99 @dataclass 100 class EncDecClassificationModelConfig(model_cfg.NemoConfig): 101 model: EncDecClassificationConfig = EncDecClassificationConfig() 102 [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import asdict, dataclass 16 from typing import Any, Dict, Optional, Tuple, Union 17 18 19 @dataclass 20 class DiarizerComponentConfig: 21 """Dataclass to imitate HydraConfig dict when accessing parameters.""" 22 23 def get(self, name: str, default: Optional[Any] = None): 24 return getattr(self, name, default) 25 26 def __iter__(self): 27 for key in asdict(self): 28 yield key 29 30 def dict(self) -> Dict: 31 return asdict(self) 32 33 34 @dataclass 35 class ASRDiarizerCTCDecoderParams: 36 pretrained_language_model: Optional[str] = None # KenLM model file: .arpa model file or .bin binary file. 37 beam_width: int = 32 38 alpha: float = 0.5 39 beta: float = 2.5 40 41 42 @dataclass 43 class ASRRealigningLMParams: 44 # Provide a KenLM language model in .arpa format. 45 arpa_language_model: Optional[str] = None 46 # Min number of words for the left context. 47 min_number_of_words: int = 3 48 # Max number of words for the right context. 49 max_number_of_words: int = 10 50 # The threshold for the difference between two log probability values from two hypotheses. 51 logprob_diff_threshold: float = 1.2 52 53 54 @dataclass 55 class ASRDiarizerParams(DiarizerComponentConfig): 56 # if True, speech segmentation for diarization is based on word-timestamps from ASR inference. 57 asr_based_vad: bool = False 58 # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD. 59 asr_based_vad_threshold: float = 1.0 60 # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null. 61 asr_batch_size: Optional[int] = None 62 # Native decoder delay. null is recommended to use the default values for each ASR model. 63 decoder_delay_in_sec: Optional[float] = None 64 # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. 65 word_ts_anchor_offset: Optional[float] = None 66 # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. 67 word_ts_anchor_pos: str = "start" 68 # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. 69 fix_word_ts_with_VAD: bool = False 70 # If True, use colored text to distinguish speakers in the output transcript. 71 colored_text: bool = False 72 # If True, the start and end time of each speaker turn is printed in the output transcript. 73 print_time: bool = True 74 # If True, the output transcript breaks the line to fix the line width (default is 90 chars) 75 break_lines: bool = False 76 77 78 @dataclass 79 class ASRDiarizerConfig(DiarizerComponentConfig): 80 model_path: Optional[str] = "stt_en_conformer_ctc_large" 81 parameters: ASRDiarizerParams = ASRDiarizerParams() 82 ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() 83 realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() 84 85 86 @dataclass 87 class VADParams(DiarizerComponentConfig): 88 window_length_in_sec: float = 0.15 # Window length in sec for VAD context input 89 shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction 90 smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) 91 overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter 92 onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech 93 offset: float = 0.1 # Offset threshold for detecting the end of a speech 94 pad_onset: float = 0.1 # Adding durations before each speech segment 95 pad_offset: float = 0 # Adding durations after each speech segment 96 min_duration_on: float = 0 # Threshold for small non_speech deletion 97 min_duration_off: float = 0.2 # Threshold for short speech segment deletion 98 filter_speech_first: bool = True 99 100 101 @dataclass 102 class VADConfig(DiarizerComponentConfig): 103 model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name 104 external_vad_manifest: Optional[str] = None 105 parameters: VADParams = VADParams() 106 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): 110 # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] 111 window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) 112 # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] 113 shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) 114 # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] 115 multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) 116 # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. 117 save_embeddings: bool = True 118 119 120 @dataclass 121 class SpeakerEmbeddingsConfig(DiarizerComponentConfig): 122 # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) 123 model_path: Optional[str] = None 124 parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() 125 126 127 @dataclass 128 class ClusteringParams(DiarizerComponentConfig): 129 # If True, use num of speakers value provided in manifest file. 130 oracle_num_speakers: bool = False 131 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. 132 max_num_speakers: int = 8 133 # If the number of segments is lower than this number, enhanced speaker counting is activated. 134 enhanced_count_thres: int = 80 135 # Determines the range of p-value search: 0 < p <= max_rp_threshold. 136 max_rp_threshold: float = 0.25 137 # The higher the number, the more values will be examined with more time. 138 sparse_search_volume: int = 30 139 # If True, take a majority vote on multiple p-values to estimate the number of speakers. 140 maj_vote_spk_count: bool = False 141 142 143 @dataclass 144 class ClusteringConfig(DiarizerComponentConfig): 145 parameters: ClusteringParams = ClusteringParams() 146 147 148 @dataclass 149 class MSDDParams(DiarizerComponentConfig): 150 # If True, use speaker embedding model in checkpoint, else provided speaker embedding model in config will be used. 151 use_speaker_model_from_ckpt: bool = True 152 # Batch size for MSDD inference. 153 infer_batch_size: int = 25 154 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. 155 sigmoid_threshold: Tuple[float] = (0.7,) 156 # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. 157 seq_eval_mode: bool = False 158 # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. 159 split_infer: bool = True 160 # The length of split short sequence when split_infer is True. 161 diar_window_length: int = 50 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): 174 manifest_filepath: Optional[str] = None 175 out_dir: Optional[str] = None 176 oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps 177 collar: float = 0.25 # Collar value for scoring 178 ignore_overlap: bool = True # Consider or ignore overlap segments while scoring 179 vad: VADConfig = VADConfig() 180 speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() 181 clustering: ClusteringConfig = ClusteringConfig() 182 msdd_model: MSDDConfig = MSDDConfig() 183 asr: ASRDiarizerConfig = ASRDiarizerConfig() 184 185 186 @dataclass 187 class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): 188 diarizer: DiarizerConfig = DiarizerConfig() 189 device: str = "cpu" 190 verbose: bool = False 191 batch_size: int = 64 192 num_workers: int = 1 193 sample_rate: int = 16000 194 name: str = "" 195 196 @classmethod 197 def init_config(cls, diar_model_path: str, vad_model_path: str, map_location: str, verbose: bool): 198 return NeuralDiarizerInferenceConfig( 199 DiarizerConfig( 200 vad=VADConfig(model_path=vad_model_path), msdd_model=MSDDConfig(model_path=diar_model_path), 201 ), 202 device=map_location, 203 verbose=verbose, 204 ) 205 [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig 18 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig 19 from nemo.core.config.modelPT import NemoConfig 20 21 22 @dataclass 23 class GraphModuleConfig: 24 criterion_type: str = "ml" 25 loss_type: str = "ctc" 26 split_batch_size: int = 0 27 dec_type: str = "topo" 28 transcribe_training: bool = True 29 backend_cfg: BackendConfig = BackendConfig() 30 31 32 @dataclass 33 class EncDecK2SeqConfig(EncDecCTCConfig): 34 graph_module_cfg: GraphModuleConfig = GraphModuleConfig() 35 36 37 @dataclass 38 class EncDecK2SeqModelConfig(NemoConfig): 39 model: EncDecK2SeqConfig = EncDecK2SeqConfig() 40 [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import classification_models_config as clf_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ( 27 ConvASRDecoderClassificationConfig, 28 ConvASREncoderConfig, 29 JasperEncoderConfig, 30 ) 31 from nemo.core.config import modelPT as model_cfg 32 33 34 # fmt: off 35 def matchboxnet_3x1x64(): 36 config = [ 37 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 38 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 39 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 40 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 41 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 42 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 43 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 44 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 45 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 46 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 47 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 48 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 49 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 50 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 51 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 52 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 53 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 54 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 55 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 56 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 57 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 58 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 59 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 60 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 61 ] 62 return config 63 64 65 def matchboxnet_3x1x64_vad(): 66 config = [ 67 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 68 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 69 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 70 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 71 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 72 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 73 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 74 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 75 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 76 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 77 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 78 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 79 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 80 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 81 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 82 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 83 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 84 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 85 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 86 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 87 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 88 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 89 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 90 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 91 ] 92 return config 93 94 95 # fmt: on 96 97 98 @dataclass 99 class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): 100 # Model global arguments 101 sample_rate: int = 16000 102 repeat: int = 1 103 dropout: float = 0.0 104 separable: bool = True 105 kernel_size_factor: float = 1.0 106 timesteps: int = 128 107 labels: List[str] = MISSING 108 109 # Dataset configs 110 train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 111 manifest_filepath=None, shuffle=True, trim_silence=False 112 ) 113 validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 114 manifest_filepath=None, shuffle=False 115 ) 116 test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 117 manifest_filepath=None, shuffle=False 118 ) 119 120 # Optimizer / Scheduler config 121 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 122 123 # Model general component configs 124 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) 125 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( 126 freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 127 ) 128 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 129 audio_length=128 130 ) 131 132 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 133 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 134 135 136 @dataclass 137 class MatchboxNetVADModelConfig(MatchboxNetModelConfig): 138 timesteps: int = 64 139 labels: List[str] = field(default_factory=lambda: ['background', 'speech']) 140 141 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = None 142 143 144 class EncDecClassificationModelConfigBuilder(model_cfg.ModelConfigBuilder): 145 VALID_CONFIGS = ['matchboxnet_3x1x64', 'matchboxnet_3x1x64_vad'] 146 147 def __init__(self, name: str = 'matchboxnet_3x1x64', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 148 if name not in EncDecClassificationModelConfigBuilder.VALID_CONFIGS: 149 raise ValueError("`name` must be one of : \n" f"{EncDecClassificationModelConfigBuilder.VALID_CONFIGS}") 150 151 self.name = name 152 153 if 'matchboxnet_3x1x64_vad' in name: 154 if encoder_cfg_func is None: 155 encoder_cfg_func = matchboxnet_3x1x64_vad 156 157 model_cfg = MatchboxNetVADModelConfig( 158 repeat=1, 159 separable=True, 160 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 161 decoder=ConvASRDecoderClassificationConfig(), 162 ) 163 164 elif 'matchboxnet_3x1x64' in name: 165 if encoder_cfg_func is None: 166 encoder_cfg_func = matchboxnet_3x1x64 167 168 model_cfg = MatchboxNetModelConfig( 169 repeat=1, 170 separable=False, 171 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 172 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 173 decoder=ConvASRDecoderClassificationConfig(), 174 ) 175 176 else: 177 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 178 179 super(EncDecClassificationModelConfigBuilder, self).__init__(model_cfg) 180 self.model_cfg: clf_cfg.EncDecClassificationConfig = model_cfg # enable type hinting 181 182 def set_labels(self, labels: List[str]): 183 self.model_cfg.labels = labels 184 185 def set_separable(self, separable: bool): 186 self.model_cfg.separable = separable 187 188 def set_repeat(self, repeat: int): 189 self.model_cfg.repeat = repeat 190 191 def set_sample_rate(self, sample_rate: int): 192 self.model_cfg.sample_rate = sample_rate 193 194 def set_dropout(self, dropout: float = 0.0): 195 self.model_cfg.dropout = dropout 196 197 def set_timesteps(self, timesteps: int): 198 self.model_cfg.timesteps = timesteps 199 200 def set_is_regression_task(self, is_regression_task: bool): 201 self.model_cfg.is_regression_task = is_regression_task 202 203 # Note: Autocomplete for users wont work without these overrides 204 # But practically it is not needed since python will infer at runtime 205 206 # def set_train_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 207 # super().set_train_ds(cfg) 208 # 209 # def set_validation_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 210 # super().set_validation_ds(cfg) 211 # 212 # def set_test_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 213 # super().set_test_ds(cfg) 214 215 def _finalize_cfg(self): 216 # propagate labels 217 self.model_cfg.train_ds.labels = self.model_cfg.labels 218 self.model_cfg.validation_ds.labels = self.model_cfg.labels 219 self.model_cfg.test_ds.labels = self.model_cfg.labels 220 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 221 222 # propagate num classes 223 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 224 225 # propagate sample rate 226 self.model_cfg.sample_rate = self.model_cfg.sample_rate 227 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 228 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 229 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 230 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 231 232 # propagate filters 233 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 234 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 235 236 # propagate timeteps 237 if self.model_cfg.crop_or_pad_augment is not None: 238 self.model_cfg.crop_or_pad_augment.audio_length = self.model_cfg.timesteps 239 240 # propagate separable 241 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 242 layer.separable = self.model_cfg.separable 243 244 # propagate repeat 245 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 246 layer.repeat = self.model_cfg.repeat 247 248 # propagate dropout 249 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 250 layer.dropout = self.model_cfg.dropout 251 252 def build(self) -> clf_cfg.EncDecClassificationConfig: 253 return super().build() 254 [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import asr_models_config as ctc_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMelSpectrogramPreprocessorConfig, 23 SpectrogramAugmentationConfig, 24 ) 25 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig, JasperEncoderConfig 26 from nemo.core.config import modelPT as model_cfg 27 28 29 # fmt: off 30 def qn_15x5(): 31 config = [ 32 JasperEncoderConfig(filters=256, repeat=1, kernel=[33], stride=[2], dilation=[1], dropout=0.0, 33 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 34 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 35 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 36 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 37 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 38 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 39 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 40 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 41 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 42 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 43 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 44 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 45 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 46 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 47 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 48 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 49 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 50 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 51 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 52 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 53 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 54 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 55 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 56 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 57 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 58 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 59 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 60 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 61 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 62 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 63 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 64 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 65 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 66 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 67 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 68 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 69 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 70 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 71 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 72 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 73 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 74 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 75 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 76 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 77 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 78 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 79 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 80 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 81 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 82 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 83 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 84 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 85 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 86 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 87 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 88 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 89 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 90 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 91 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 92 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 93 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 94 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 95 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 96 JasperEncoderConfig(filters=512, repeat=1, kernel=[87], stride=[1], dilation=[2], dropout=0.0, 97 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 98 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 99 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 100 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 101 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 102 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 103 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 104 ] 105 return config 106 107 108 def jasper_10x5_dr(): 109 config = [ 110 JasperEncoderConfig(filters=256, repeat=1, kernel=[11], stride=[2], dilation=[1], dropout=0.2, 111 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 112 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 113 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 114 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 115 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 116 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 117 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 118 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 119 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 120 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 121 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 122 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 123 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 124 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 125 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 126 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 127 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 128 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 129 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 130 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 131 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 132 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 133 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 134 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 135 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 136 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 137 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 138 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 139 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 140 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 141 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 142 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 143 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 144 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 145 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 146 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 147 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 148 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 149 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 150 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 151 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 152 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 153 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 154 JasperEncoderConfig(filters=896, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.4, 155 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 156 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 157 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 158 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.4, 159 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 160 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 161 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 162 ] 163 return config 164 # fmt: on 165 166 167 @dataclass 168 class JasperModelConfig(ctc_cfg.EncDecCTCConfig): 169 # Model global arguments 170 sample_rate: int = 16000 171 repeat: int = 1 172 dropout: float = 0.0 173 separable: bool = False 174 labels: List[str] = MISSING 175 176 # Dataset configs 177 train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( 178 manifest_filepath=None, shuffle=True, trim_silence=True 179 ) 180 validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 181 test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 182 183 # Optimizer / Scheduler config 184 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 185 186 # Model general component configs 187 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 188 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 189 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 190 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 191 192 193 @dataclass 194 class QuartzNetModelConfig(JasperModelConfig): 195 separable: bool = True 196 197 198 class EncDecCTCModelConfigBuilder(model_cfg.ModelConfigBuilder): 199 VALID_CONFIGS = ['quartznet_15x5', 'quartznet_15x5_zh', 'jasper_10x5dr'] 200 201 def __init__(self, name: str = 'quartznet_15x5', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 202 if name not in EncDecCTCModelConfigBuilder.VALID_CONFIGS: 203 raise ValueError("`name` must be one of : \n" f"{EncDecCTCModelConfigBuilder.VALID_CONFIGS}") 204 205 self.name = name 206 207 if 'quartznet_15x5' in name: 208 if encoder_cfg_func is None: 209 encoder_cfg_func = qn_15x5 210 211 model_cfg = QuartzNetModelConfig( 212 repeat=5, 213 separable=True, 214 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 215 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 216 decoder=ConvASRDecoderConfig(), 217 ) 218 219 elif 'jasper_10x5' in name: 220 if encoder_cfg_func is None: 221 encoder_cfg_func = jasper_10x5_dr 222 223 model_cfg = JasperModelConfig( 224 repeat=5, 225 separable=False, 226 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 227 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 228 decoder=ConvASRDecoderConfig(), 229 ) 230 231 else: 232 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 233 234 super(EncDecCTCModelConfigBuilder, self).__init__(model_cfg) 235 self.model_cfg: ctc_cfg.EncDecCTCConfig = model_cfg # enable type hinting 236 237 if 'zh' in name: 238 self.set_dataset_normalize(normalize=False) 239 240 def set_labels(self, labels: List[str]): 241 self.model_cfg.labels = labels 242 243 def set_separable(self, separable: bool): 244 self.model_cfg.separable = separable 245 246 def set_repeat(self, repeat: int): 247 self.model_cfg.repeat = repeat 248 249 def set_sample_rate(self, sample_rate: int): 250 self.model_cfg.sample_rate = sample_rate 251 252 def set_dropout(self, dropout: float = 0.0): 253 self.model_cfg.dropout = dropout 254 255 def set_dataset_normalize(self, normalize: bool): 256 self.model_cfg.train_ds.normalize = normalize 257 self.model_cfg.validation_ds.normalize = normalize 258 self.model_cfg.test_ds.normalize = normalize 259 260 # Note: Autocomplete for users wont work without these overrides 261 # But practically it is not needed since python will infer at runtime 262 263 # def set_train_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 264 # super().set_train_ds(cfg) 265 # 266 # def set_validation_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 267 # super().set_validation_ds(cfg) 268 # 269 # def set_test_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 270 # super().set_test_ds(cfg) 271 272 def _finalize_cfg(self): 273 # propagate labels 274 self.model_cfg.train_ds.labels = self.model_cfg.labels 275 self.model_cfg.validation_ds.labels = self.model_cfg.labels 276 self.model_cfg.test_ds.labels = self.model_cfg.labels 277 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 278 279 # propagate num classes 280 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 281 282 # propagate sample rate 283 self.model_cfg.sample_rate = self.model_cfg.sample_rate 284 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 285 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 286 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 287 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 288 289 # propagate filters 290 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 291 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 292 293 # propagate separable 294 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 295 layer.separable = self.model_cfg.separable 296 297 # propagate repeat 298 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 299 layer.repeat = self.model_cfg.repeat 300 301 # propagate dropout 302 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 303 layer.dropout = self.model_cfg.dropout 304 305 def build(self) -> ctc_cfg.EncDecCTCConfig: 306 return super().build() 307 [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import random 17 from abc import ABC, abstractmethod 18 from dataclasses import dataclass 19 from typing import Any, Dict, Optional, Tuple 20 21 import torch 22 from packaging import version 23 24 from nemo.collections.asr.parts.numba.spec_augment import SpecAugmentNumba, spec_augment_launch_heuristics 25 from nemo.collections.asr.parts.preprocessing.features import ( 26 FilterbankFeatures, 27 FilterbankFeaturesTA, 28 make_seq_mask_like, 29 ) 30 from nemo.collections.asr.parts.submodules.spectr_augment import SpecAugment, SpecCutout 31 from nemo.core.classes import Exportable, NeuralModule, typecheck 32 from nemo.core.neural_types import ( 33 AudioSignal, 34 LengthsType, 35 MelSpectrogramType, 36 MFCCSpectrogramType, 37 NeuralType, 38 SpectrogramType, 39 ) 40 from nemo.core.utils import numba_utils 41 from nemo.core.utils.numba_utils import __NUMBA_MINIMUM_VERSION__ 42 from nemo.utils import logging 43 44 try: 45 import torchaudio 46 import torchaudio.functional 47 import torchaudio.transforms 48 49 TORCHAUDIO_VERSION = version.parse(torchaudio.__version__) 50 TORCHAUDIO_VERSION_MIN = version.parse('0.5') 51 52 HAVE_TORCHAUDIO = True 53 except ModuleNotFoundError: 54 HAVE_TORCHAUDIO = False 55 56 __all__ = [ 57 'AudioToMelSpectrogramPreprocessor', 58 'AudioToSpectrogram', 59 'SpectrogramToAudio', 60 'AudioToMFCCPreprocessor', 61 'SpectrogramAugmentation', 62 'MaskedPatchAugmentation', 63 'CropOrPadSpectrogramAugmentation', 64 ] 65 66 67 class AudioPreprocessor(NeuralModule, ABC): 68 """ 69 An interface for Neural Modules that performs audio pre-processing, 70 transforming the wav files to features. 71 """ 72 73 def __init__(self, win_length, hop_length): 74 super().__init__() 75 76 self.win_length = win_length 77 self.hop_length = hop_length 78 79 self.torch_windows = { 80 'hann': torch.hann_window, 81 'hamming': torch.hamming_window, 82 'blackman': torch.blackman_window, 83 'bartlett': torch.bartlett_window, 84 'ones': torch.ones, 85 None: torch.ones, 86 } 87 88 @typecheck() 89 @torch.no_grad() 90 def forward(self, input_signal, length): 91 processed_signal, processed_length = self.get_features(input_signal, length) 92 93 return processed_signal, processed_length 94 95 @abstractmethod 96 def get_features(self, input_signal, length): 97 # Called by forward(). Subclasses should implement this. 98 pass 99 100 101 class AudioToMelSpectrogramPreprocessor(AudioPreprocessor, Exportable): 102 """Featurizer module that converts wavs to mel spectrograms. 103 104 Args: 105 sample_rate (int): Sample rate of the input audio data. 106 Defaults to 16000 107 window_size (float): Size of window for fft in seconds 108 Defaults to 0.02 109 window_stride (float): Stride of window for fft in seconds 110 Defaults to 0.01 111 n_window_size (int): Size of window for fft in samples 112 Defaults to None. Use one of window_size or n_window_size. 113 n_window_stride (int): Stride of window for fft in samples 114 Defaults to None. Use one of window_stride or n_window_stride. 115 window (str): Windowing function for fft. can be one of ['hann', 116 'hamming', 'blackman', 'bartlett'] 117 Defaults to "hann" 118 normalize (str): Can be one of ['per_feature', 'all_features']; all 119 other options disable feature normalization. 'all_features' 120 normalizes the entire spectrogram to be mean 0 with std 1. 121 'pre_features' normalizes per channel / freq instead. 122 Defaults to "per_feature" 123 n_fft (int): Length of FT window. If None, it uses the smallest power 124 of 2 that is larger than n_window_size. 125 Defaults to None 126 preemph (float): Amount of pre emphasis to add to audio. Can be 127 disabled by passing None. 128 Defaults to 0.97 129 features (int): Number of mel spectrogram freq bins to output. 130 Defaults to 64 131 lowfreq (int): Lower bound on mel basis in Hz. 132 Defaults to 0 133 highfreq (int): Lower bound on mel basis in Hz. 134 Defaults to None 135 log (bool): Log features. 136 Defaults to True 137 log_zero_guard_type(str): Need to avoid taking the log of zero. There 138 are two options: "add" or "clamp". 139 Defaults to "add". 140 log_zero_guard_value(float, or str): Add or clamp requires the number 141 to add with or clamp to. log_zero_guard_value can either be a float 142 or "tiny" or "eps". torch.finfo is used if "tiny" or "eps" is 143 passed. 144 Defaults to 2**-24. 145 dither (float): Amount of white-noise dithering. 146 Defaults to 1e-5 147 pad_to (int): Ensures that the output size of the time dimension is 148 a multiple of pad_to. 149 Defaults to 16 150 frame_splicing (int): Defaults to 1 151 exact_pad (bool): If True, sets stft center to False and adds padding, such that num_frames = audio_length 152 // hop_length. Defaults to False. 153 pad_value (float): The value that shorter mels are padded with. 154 Defaults to 0 155 mag_power (float): The power that the linear spectrogram is raised to 156 prior to multiplication with mel basis. 157 Defaults to 2 for a power spec 158 rng : Random number generator 159 nb_augmentation_prob (float) : Probability with which narrowband augmentation would be applied to 160 samples in the batch. 161 Defaults to 0.0 162 nb_max_freq (int) : Frequency above which all frequencies will be masked for narrowband augmentation. 163 Defaults to 4000 164 use_torchaudio: Whether to use the `torchaudio` implementation. 165 mel_norm: Normalization used for mel filterbank weights. 166 Defaults to 'slaney' (area normalization) 167 stft_exact_pad: Deprecated argument, kept for compatibility with older checkpoints. 168 stft_conv: Deprecated argument, kept for compatibility with older checkpoints. 169 """ 170 171 def save_to(self, save_path: str): 172 pass 173 174 @classmethod 175 def restore_from(cls, restore_path: str): 176 pass 177 178 @property 179 def input_types(self): 180 """Returns definitions of module input ports. 181 """ 182 return { 183 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 184 "length": NeuralType( 185 tuple('B'), LengthsType() 186 ), # Please note that length should be in samples not seconds. 187 } 188 189 @property 190 def output_types(self): 191 """Returns definitions of module output ports. 192 193 processed_signal: 194 0: AxisType(BatchTag) 195 1: AxisType(MelSpectrogramSignalTag) 196 2: AxisType(ProcessedTimeTag) 197 processed_length: 198 0: AxisType(BatchTag) 199 """ 200 return { 201 "processed_signal": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 202 "processed_length": NeuralType(tuple('B'), LengthsType()), 203 } 204 205 def __init__( 206 self, 207 sample_rate=16000, 208 window_size=0.02, 209 window_stride=0.01, 210 n_window_size=None, 211 n_window_stride=None, 212 window="hann", 213 normalize="per_feature", 214 n_fft=None, 215 preemph=0.97, 216 features=64, 217 lowfreq=0, 218 highfreq=None, 219 log=True, 220 log_zero_guard_type="add", 221 log_zero_guard_value=2 ** -24, 222 dither=1e-5, 223 pad_to=16, 224 frame_splicing=1, 225 exact_pad=False, 226 pad_value=0, 227 mag_power=2.0, 228 rng=None, 229 nb_augmentation_prob=0.0, 230 nb_max_freq=4000, 231 use_torchaudio: bool = False, 232 mel_norm="slaney", 233 stft_exact_pad=False, # Deprecated arguments; kept for config compatibility 234 stft_conv=False, # Deprecated arguments; kept for config compatibility 235 ): 236 super().__init__(n_window_size, n_window_stride) 237 238 self._sample_rate = sample_rate 239 if window_size and n_window_size: 240 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 241 if window_stride and n_window_stride: 242 raise ValueError( 243 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 244 ) 245 if window_size: 246 n_window_size = int(window_size * self._sample_rate) 247 if window_stride: 248 n_window_stride = int(window_stride * self._sample_rate) 249 250 # Given the long and similar argument list, point to the class and instantiate it by reference 251 if not use_torchaudio: 252 featurizer_class = FilterbankFeatures 253 else: 254 featurizer_class = FilterbankFeaturesTA 255 self.featurizer = featurizer_class( 256 sample_rate=self._sample_rate, 257 n_window_size=n_window_size, 258 n_window_stride=n_window_stride, 259 window=window, 260 normalize=normalize, 261 n_fft=n_fft, 262 preemph=preemph, 263 nfilt=features, 264 lowfreq=lowfreq, 265 highfreq=highfreq, 266 log=log, 267 log_zero_guard_type=log_zero_guard_type, 268 log_zero_guard_value=log_zero_guard_value, 269 dither=dither, 270 pad_to=pad_to, 271 frame_splicing=frame_splicing, 272 exact_pad=exact_pad, 273 pad_value=pad_value, 274 mag_power=mag_power, 275 rng=rng, 276 nb_augmentation_prob=nb_augmentation_prob, 277 nb_max_freq=nb_max_freq, 278 mel_norm=mel_norm, 279 stft_exact_pad=stft_exact_pad, # Deprecated arguments; kept for config compatibility 280 stft_conv=stft_conv, # Deprecated arguments; kept for config compatibility 281 ) 282 283 def input_example(self, max_batch: int = 8, max_dim: int = 32000, min_length: int = 200): 284 batch_size = torch.randint(low=1, high=max_batch, size=[1]).item() 285 max_length = torch.randint(low=min_length, high=max_dim, size=[1]).item() 286 signals = torch.rand(size=[batch_size, max_length]) * 2 - 1 287 lengths = torch.randint(low=min_length, high=max_dim, size=[batch_size]) 288 lengths[0] = max_length 289 return signals, lengths 290 291 def get_features(self, input_signal, length): 292 return self.featurizer(input_signal, length) 293 294 @property 295 def filter_banks(self): 296 return self.featurizer.filter_banks 297 298 299 class AudioToMFCCPreprocessor(AudioPreprocessor): 300 """Preprocessor that converts wavs to MFCCs. 301 Uses torchaudio.transforms.MFCC. 302 303 Args: 304 sample_rate: The sample rate of the audio. 305 Defaults to 16000. 306 window_size: Size of window for fft in seconds. Used to calculate the 307 win_length arg for mel spectrogram. 308 Defaults to 0.02 309 window_stride: Stride of window for fft in seconds. Used to caculate 310 the hop_length arg for mel spect. 311 Defaults to 0.01 312 n_window_size: Size of window for fft in samples 313 Defaults to None. Use one of window_size or n_window_size. 314 n_window_stride: Stride of window for fft in samples 315 Defaults to None. Use one of window_stride or n_window_stride. 316 window: Windowing function for fft. can be one of ['hann', 317 'hamming', 'blackman', 'bartlett', 'none', 'null']. 318 Defaults to 'hann' 319 n_fft: Length of FT window. If None, it uses the smallest power of 2 320 that is larger than n_window_size. 321 Defaults to None 322 lowfreq (int): Lower bound on mel basis in Hz. 323 Defaults to 0 324 highfreq (int): Lower bound on mel basis in Hz. 325 Defaults to None 326 n_mels: Number of mel filterbanks. 327 Defaults to 64 328 n_mfcc: Number of coefficients to retain 329 Defaults to 64 330 dct_type: Type of discrete cosine transform to use 331 norm: Type of norm to use 332 log: Whether to use log-mel spectrograms instead of db-scaled. 333 Defaults to True. 334 """ 335 336 @property 337 def input_types(self): 338 """Returns definitions of module input ports. 339 """ 340 return { 341 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 342 "length": NeuralType(tuple('B'), LengthsType()), 343 } 344 345 @property 346 def output_types(self): 347 """Returns definitions of module output ports. 348 """ 349 return { 350 "processed_signal": NeuralType(('B', 'D', 'T'), MFCCSpectrogramType()), 351 "processed_length": NeuralType(tuple('B'), LengthsType()), 352 } 353 354 def save_to(self, save_path: str): 355 pass 356 357 @classmethod 358 def restore_from(cls, restore_path: str): 359 pass 360 361 def __init__( 362 self, 363 sample_rate=16000, 364 window_size=0.02, 365 window_stride=0.01, 366 n_window_size=None, 367 n_window_stride=None, 368 window='hann', 369 n_fft=None, 370 lowfreq=0.0, 371 highfreq=None, 372 n_mels=64, 373 n_mfcc=64, 374 dct_type=2, 375 norm='ortho', 376 log=True, 377 ): 378 self._sample_rate = sample_rate 379 if not HAVE_TORCHAUDIO: 380 logging.error('Could not import torchaudio. Some features might not work.') 381 382 raise ModuleNotFoundError( 383 "torchaudio is not installed but is necessary for " 384 "AudioToMFCCPreprocessor. We recommend you try " 385 "building it from source for the PyTorch version you have." 386 ) 387 if window_size and n_window_size: 388 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 389 if window_stride and n_window_stride: 390 raise ValueError( 391 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 392 ) 393 # Get win_length (n_window_size) and hop_length (n_window_stride) 394 if window_size: 395 n_window_size = int(window_size * self._sample_rate) 396 if window_stride: 397 n_window_stride = int(window_stride * self._sample_rate) 398 399 super().__init__(n_window_size, n_window_stride) 400 401 mel_kwargs = {} 402 403 mel_kwargs['f_min'] = lowfreq 404 mel_kwargs['f_max'] = highfreq 405 mel_kwargs['n_mels'] = n_mels 406 407 mel_kwargs['n_fft'] = n_fft or 2 ** math.ceil(math.log2(n_window_size)) 408 409 mel_kwargs['win_length'] = n_window_size 410 mel_kwargs['hop_length'] = n_window_stride 411 412 # Set window_fn. None defaults to torch.ones. 413 window_fn = self.torch_windows.get(window, None) 414 if window_fn is None: 415 raise ValueError( 416 f"Window argument for AudioProcessor is invalid: {window}." 417 f"For no window function, use 'ones' or None." 418 ) 419 mel_kwargs['window_fn'] = window_fn 420 421 # Use torchaudio's implementation of MFCCs as featurizer 422 self.featurizer = torchaudio.transforms.MFCC( 423 sample_rate=self._sample_rate, 424 n_mfcc=n_mfcc, 425 dct_type=dct_type, 426 norm=norm, 427 log_mels=log, 428 melkwargs=mel_kwargs, 429 ) 430 431 def get_features(self, input_signal, length): 432 features = self.featurizer(input_signal) 433 seq_len = torch.ceil(length.to(torch.float32) / self.hop_length).to(dtype=torch.long) 434 return features, seq_len 435 436 437 class SpectrogramAugmentation(NeuralModule): 438 """ 439 Performs time and freq cuts in one of two ways. 440 SpecAugment zeroes out vertical and horizontal sections as described in 441 SpecAugment (https://arxiv.org/abs/1904.08779). Arguments for use with 442 SpecAugment are `freq_masks`, `time_masks`, `freq_width`, and `time_width`. 443 SpecCutout zeroes out rectangulars as described in Cutout 444 (https://arxiv.org/abs/1708.04552). Arguments for use with Cutout are 445 `rect_masks`, `rect_freq`, and `rect_time`. 446 447 Args: 448 freq_masks (int): how many frequency segments should be cut. 449 Defaults to 0. 450 time_masks (int): how many time segments should be cut 451 Defaults to 0. 452 freq_width (int): maximum number of frequencies to be cut in one 453 segment. 454 Defaults to 10. 455 time_width (int): maximum number of time steps to be cut in one 456 segment 457 Defaults to 10. 458 rect_masks (int): how many rectangular masks should be cut 459 Defaults to 0. 460 rect_freq (int): maximum size of cut rectangles along the frequency 461 dimension 462 Defaults to 5. 463 rect_time (int): maximum size of cut rectangles along the time 464 dimension 465 Defaults to 25. 466 """ 467 468 @property 469 def input_types(self): 470 """Returns definitions of module input types 471 """ 472 return { 473 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 474 "length": NeuralType(tuple('B'), LengthsType()), 475 } 476 477 @property 478 def output_types(self): 479 """Returns definitions of module output types 480 """ 481 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 482 483 def __init__( 484 self, 485 freq_masks=0, 486 time_masks=0, 487 freq_width=10, 488 time_width=10, 489 rect_masks=0, 490 rect_time=5, 491 rect_freq=20, 492 rng=None, 493 mask_value=0.0, 494 use_numba_spec_augment: bool = True, 495 ): 496 super().__init__() 497 498 if rect_masks > 0: 499 self.spec_cutout = SpecCutout(rect_masks=rect_masks, rect_time=rect_time, rect_freq=rect_freq, rng=rng,) 500 # self.spec_cutout.to(self._device) 501 else: 502 self.spec_cutout = lambda input_spec: input_spec 503 if freq_masks + time_masks > 0: 504 self.spec_augment = SpecAugment( 505 freq_masks=freq_masks, 506 time_masks=time_masks, 507 freq_width=freq_width, 508 time_width=time_width, 509 rng=rng, 510 mask_value=mask_value, 511 ) 512 else: 513 self.spec_augment = lambda input_spec, length: input_spec 514 515 # Check if numba is supported, and use a Numba kernel if it is 516 if use_numba_spec_augment and numba_utils.numba_cuda_is_supported(__NUMBA_MINIMUM_VERSION__): 517 logging.info('Numba CUDA SpecAugment kernel is being used') 518 self.spec_augment_numba = SpecAugmentNumba( 519 freq_masks=freq_masks, 520 time_masks=time_masks, 521 freq_width=freq_width, 522 time_width=time_width, 523 rng=rng, 524 mask_value=mask_value, 525 ) 526 else: 527 self.spec_augment_numba = None 528 529 @typecheck() 530 def forward(self, input_spec, length): 531 augmented_spec = self.spec_cutout(input_spec=input_spec) 532 533 # To run the Numba kernel, correct numba version is required as well as 534 # tensor must be on GPU and length must be provided 535 if self.spec_augment_numba is not None and spec_augment_launch_heuristics(augmented_spec, length): 536 augmented_spec = self.spec_augment_numba(input_spec=augmented_spec, length=length) 537 else: 538 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 539 return augmented_spec 540 541 542 class MaskedPatchAugmentation(NeuralModule): 543 """ 544 Zeroes out fixed size time patches of the spectrogram. 545 All samples in batch are guaranteed to have the same amount of masked time steps. 546 Optionally also performs frequency masking in the same way as SpecAugment. 547 Args: 548 patch_size (int): up to how many time steps does one patch consist of. 549 Defaults to 48. 550 mask_patches (float): how many patches should be masked in each sample. 551 if >= 1., interpreted as number of patches (after converting to int) 552 if <1., interpreted as fraction of total tokens to be masked (number of patches is rounded up) 553 Defaults to 10. 554 freq_masks (int): how many frequency segments should be cut. 555 Defaults to 0. 556 freq_width (int): maximum number of frequencies to be cut in a segment. 557 Defaults to 0. 558 """ 559 560 @property 561 def input_types(self): 562 """Returns definitions of module input types 563 """ 564 return { 565 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 566 "length": NeuralType(tuple('B'), LengthsType()), 567 } 568 569 @property 570 def output_types(self): 571 """Returns definitions of module output types 572 """ 573 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 574 575 def __init__( 576 self, patch_size: int = 48, mask_patches: float = 10.0, freq_masks: int = 0, freq_width: int = 0, 577 ): 578 super().__init__() 579 self.patch_size = patch_size 580 if mask_patches >= 1: 581 self.mask_patches = int(mask_patches) 582 elif mask_patches >= 0: 583 self._mask_fraction = mask_patches 584 self.mask_patches = None 585 else: 586 raise ValueError('mask_patches cannot be negative') 587 588 if freq_masks > 0: 589 self.spec_augment = SpecAugment(freq_masks=freq_masks, time_masks=0, freq_width=freq_width, time_width=0,) 590 else: 591 self.spec_augment = None 592 593 @typecheck() 594 def forward(self, input_spec, length): 595 augmented_spec = input_spec 596 597 min_len = torch.min(length) 598 599 if self.mask_patches is None: 600 # masking specified as fraction 601 len_fraction = int(min_len * self._mask_fraction) 602 mask_patches = len_fraction // self.patch_size + int(len_fraction % self.patch_size != 0) 603 else: 604 mask_patches = self.mask_patches 605 606 if min_len < self.patch_size * mask_patches: 607 mask_patches = min_len // self.patch_size 608 609 for idx in range(input_spec.shape[0]): 610 cur_len = length[idx] 611 patches = range(cur_len // self.patch_size) 612 masked_patches = random.sample(patches, mask_patches) 613 614 for mp in masked_patches: 615 augmented_spec[idx, :, mp * self.patch_size : (mp + 1) * self.patch_size] = 0.0 616 617 if self.spec_augment is not None: 618 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 619 620 return augmented_spec 621 622 623 class CropOrPadSpectrogramAugmentation(NeuralModule): 624 """ 625 Pad or Crop the incoming Spectrogram to a certain shape. 626 627 Args: 628 audio_length (int): the final number of timesteps that is required. 629 The signal will be either padded or cropped temporally to this 630 size. 631 """ 632 633 def __init__(self, audio_length): 634 super(CropOrPadSpectrogramAugmentation, self).__init__() 635 self.audio_length = audio_length 636 637 @typecheck() 638 @torch.no_grad() 639 def forward(self, input_signal, length): 640 image = input_signal 641 num_images = image.shape[0] 642 643 audio_length = self.audio_length 644 image_len = image.shape[-1] 645 646 # Crop long signal 647 if image_len > audio_length: # randomly slice 648 cutout_images = [] 649 offset = torch.randint(low=0, high=image_len - audio_length + 1, size=[num_images]) 650 651 for idx, offset in enumerate(offset): 652 cutout_images.append(image[idx : idx + 1, :, offset : offset + audio_length]) 653 654 image = torch.cat(cutout_images, dim=0) 655 del cutout_images 656 657 else: # symmetrically pad short signal with zeros 658 pad_left = (audio_length - image_len) // 2 659 pad_right = (audio_length - image_len) // 2 660 661 if (audio_length - image_len) % 2 == 1: 662 pad_right += 1 663 664 image = torch.nn.functional.pad(image, [pad_left, pad_right], mode="constant", value=0) 665 666 # Replace dynamic length sequences with static number of timesteps 667 length = (length * 0) + audio_length 668 669 return image, length 670 671 @property 672 def input_types(self): 673 """Returns definitions of module output ports. 674 """ 675 return { 676 "input_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 677 "length": NeuralType(tuple('B'), LengthsType()), 678 } 679 680 @property 681 def output_types(self): 682 """Returns definitions of module output ports. 683 """ 684 return { 685 "processed_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 686 "processed_length": NeuralType(tuple('B'), LengthsType()), 687 } 688 689 def save_to(self, save_path: str): 690 pass 691 692 @classmethod 693 def restore_from(cls, restore_path: str): 694 pass 695 696 697 class AudioToSpectrogram(NeuralModule): 698 """Transform a batch of input multi-channel signals into a batch of 699 STFT-based spectrograms. 700 701 Args: 702 fft_length: length of FFT 703 hop_length: length of hops/shifts of the sliding window 704 power: exponent for magnitude spectrogram. Default `None` will 705 return a complex-valued spectrogram 706 """ 707 708 def __init__(self, fft_length: int, hop_length: int, power: Optional[float] = None): 709 if not HAVE_TORCHAUDIO: 710 logging.error('Could not import torchaudio. Some features might not work.') 711 712 raise ModuleNotFoundError( 713 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 714 ) 715 716 super().__init__() 717 718 # For now, assume FFT length is divisible by two 719 if fft_length % 2 != 0: 720 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 721 722 self.stft = torchaudio.transforms.Spectrogram( 723 n_fft=fft_length, hop_length=hop_length, power=power, pad_mode='constant' 724 ) 725 726 # number of subbands 727 self.F = fft_length // 2 + 1 728 729 @property 730 def num_subbands(self) -> int: 731 return self.F 732 733 @property 734 def input_types(self) -> Dict[str, NeuralType]: 735 """Returns definitions of module output ports. 736 """ 737 return { 738 "input": NeuralType(('B', 'C', 'T'), AudioSignal()), 739 "input_length": NeuralType(('B',), LengthsType(), optional=True), 740 } 741 742 @property 743 def output_types(self) -> Dict[str, NeuralType]: 744 """Returns definitions of module output ports. 745 """ 746 return { 747 "output": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 748 "output_length": NeuralType(('B',), LengthsType()), 749 } 750 751 @typecheck() 752 def forward( 753 self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None 754 ) -> Tuple[torch.Tensor, torch.Tensor]: 755 """Convert a batch of C-channel input signals 756 into a batch of complex-valued spectrograms. 757 758 Args: 759 input: Time-domain input signal with C channels, shape (B, C, T) 760 input_length: Length of valid entries along the time dimension, shape (B,) 761 762 Returns: 763 Output spectrogram with F subbands and N time frames, shape (B, C, F, N) 764 and output length with shape (B,). 765 """ 766 B, T = input.size(0), input.size(-1) 767 input = input.view(B, -1, T) 768 769 # STFT output (B, C, F, N) 770 with torch.cuda.amp.autocast(enabled=False): 771 output = self.stft(input.float()) 772 773 if input_length is not None: 774 # Mask padded frames 775 output_length = self.get_output_length(input_length=input_length) 776 777 length_mask: torch.Tensor = make_seq_mask_like( 778 lengths=output_length, like=output, time_dim=-1, valid_ones=False 779 ) 780 output = output.masked_fill(length_mask, 0.0) 781 else: 782 # Assume all frames are valid for all examples in the batch 783 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 784 785 return output, output_length 786 787 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 788 """Get length of valid frames for the output. 789 790 Args: 791 input_length: number of valid samples, shape (B,) 792 793 Returns: 794 Number of valid frames, shape (B,) 795 """ 796 output_length = input_length.div(self.stft.hop_length, rounding_mode='floor').add(1).long() 797 return output_length 798 799 800 class SpectrogramToAudio(NeuralModule): 801 """Transform a batch of input multi-channel spectrograms into a batch of 802 time-domain multi-channel signals. 803 804 Args: 805 fft_length: length of FFT 806 hop_length: length of hops/shifts of the sliding window 807 power: exponent for magnitude spectrogram. Default `None` will 808 return a complex-valued spectrogram 809 """ 810 811 def __init__(self, fft_length: int, hop_length: int): 812 if not HAVE_TORCHAUDIO: 813 logging.error('Could not import torchaudio. Some features might not work.') 814 815 raise ModuleNotFoundError( 816 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 817 ) 818 819 super().__init__() 820 821 # For now, assume FFT length is divisible by two 822 if fft_length % 2 != 0: 823 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 824 825 self.istft = torchaudio.transforms.InverseSpectrogram( 826 n_fft=fft_length, hop_length=hop_length, pad_mode='constant' 827 ) 828 829 self.F = fft_length // 2 + 1 830 831 @property 832 def num_subbands(self) -> int: 833 return self.F 834 835 @property 836 def input_types(self) -> Dict[str, NeuralType]: 837 """Returns definitions of module output ports. 838 """ 839 return { 840 "input": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 841 "input_length": NeuralType(('B',), LengthsType(), optional=True), 842 } 843 844 @property 845 def output_types(self) -> Dict[str, NeuralType]: 846 """Returns definitions of module output ports. 847 """ 848 return { 849 "output": NeuralType(('B', 'C', 'T'), AudioSignal()), 850 "output_length": NeuralType(('B',), LengthsType()), 851 } 852 853 @typecheck() 854 def forward(self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None) -> torch.Tensor: 855 """Convert input complex-valued spectrogram to a time-domain 856 signal. Multi-channel IO is supported. 857 858 Args: 859 input: Input spectrogram for C channels, shape (B, C, F, N) 860 input_length: Length of valid entries along the time dimension, shape (B,) 861 862 Returns: 863 Time-domain signal with T time-domain samples and C channels, (B, C, T) 864 and output length with shape (B,). 865 """ 866 B, F, N = input.size(0), input.size(-2), input.size(-1) 867 assert F == self.F, f'Number of subbands F={F} not matching self.F={self.F}' 868 input = input.view(B, -1, F, N) 869 870 # iSTFT output (B, C, T) 871 with torch.cuda.amp.autocast(enabled=False): 872 output = self.istft(input.cfloat()) 873 874 if input_length is not None: 875 # Mask padded samples 876 output_length = self.get_output_length(input_length=input_length) 877 878 length_mask: torch.Tensor = make_seq_mask_like( 879 lengths=output_length, like=output, time_dim=-1, valid_ones=False 880 ) 881 output = output.masked_fill(length_mask, 0.0) 882 else: 883 # Assume all frames are valid for all examples in the batch 884 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 885 886 return output, output_length 887 888 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 889 """Get length of valid samples for the output. 890 891 Args: 892 input_length: number of valid frames, shape (B,) 893 894 Returns: 895 Number of valid samples, shape (B,) 896 """ 897 output_length = input_length.sub(1).mul(self.istft.hop_length).long() 898 return output_length 899 900 901 @dataclass 902 class AudioToMelSpectrogramPreprocessorConfig: 903 _target_: str = "nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor" 904 sample_rate: int = 16000 905 window_size: float = 0.02 906 window_stride: float = 0.01 907 n_window_size: Optional[int] = None 908 n_window_stride: Optional[int] = None 909 window: str = "hann" 910 normalize: str = "per_feature" 911 n_fft: Optional[int] = None 912 preemph: float = 0.97 913 features: int = 64 914 lowfreq: int = 0 915 highfreq: Optional[int] = None 916 log: bool = True 917 log_zero_guard_type: str = "add" 918 log_zero_guard_value: float = 2 ** -24 919 dither: float = 1e-5 920 pad_to: int = 16 921 frame_splicing: int = 1 922 exact_pad: bool = False 923 pad_value: int = 0 924 mag_power: float = 2.0 925 rng: Optional[str] = None 926 nb_augmentation_prob: float = 0.0 927 nb_max_freq: int = 4000 928 use_torchaudio: bool = False 929 mel_norm: str = "slaney" 930 stft_exact_pad: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 931 stft_conv: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 932 933 934 @dataclass 935 class AudioToMFCCPreprocessorConfig: 936 _target_: str = 'nemo.collections.asr.modules.AudioToMFCCPreprocessor' 937 sample_rate: int = 16000 938 window_size: float = 0.02 939 window_stride: float = 0.01 940 n_window_size: Optional[int] = None 941 n_window_stride: Optional[int] = None 942 window: str = 'hann' 943 n_fft: Optional[int] = None 944 lowfreq: Optional[float] = 0.0 945 highfreq: Optional[float] = None 946 n_mels: int = 64 947 n_mfcc: int = 64 948 dct_type: int = 2 949 norm: str = 'ortho' 950 log: bool = True 951 952 953 @dataclass 954 class SpectrogramAugmentationConfig: 955 _target_: str = "nemo.collections.asr.modules.SpectrogramAugmentation" 956 freq_masks: int = 0 957 time_masks: int = 0 958 freq_width: int = 0 959 time_width: Optional[Any] = 0 960 rect_masks: int = 0 961 rect_time: int = 0 962 rect_freq: int = 0 963 mask_value: float = 0 964 rng: Optional[Any] = None # random.Random() type 965 use_numba_spec_augment: bool = True 966 967 968 @dataclass 969 class CropOrPadSpectrogramAugmentationConfig: 970 audio_length: int 971 _target_: str = "nemo.collections.asr.modules.CropOrPadSpectrogramAugmentation" 972 973 974 @dataclass 975 class MaskedPatchAugmentationConfig: 976 patch_size: int = 48 977 mask_patches: float = 10.0 978 freq_masks: int = 0 979 freq_width: int = 0 980 _target_: str = "nemo.collections.asr.modules.MaskedPatchAugmentation" 981 [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/parts/k2/classes.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from abc import ABC 16 from dataclasses import dataclass 17 from typing import Any, Optional, Tuple 18 19 import torch 20 from omegaconf import DictConfig 21 22 from nemo.utils import logging 23 24 25 @dataclass 26 class GraphIntersectDenseConfig: 27 """Graph dense intersection config. 28 """ 29 30 search_beam: float = 20.0 31 output_beam: float = 10.0 32 min_active_states: int = 30 33 max_active_states: int = 10000 34 35 36 @dataclass 37 class GraphModuleConfig: 38 """Config for graph modules. 39 Typically used with graph losses and decoders. 40 """ 41 42 topo_type: str = "default" 43 topo_with_self_loops: bool = True 44 token_lm: Optional[Any] = None 45 intersect_pruned: bool = False 46 intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() 47 boost_coeff: float = 0.0 48 predictor_window_size: int = 0 49 predictor_step_size: int = 1 50 51 52 class ASRK2Mixin(ABC): 53 """k2 Mixin class that simplifies the construction of various models with k2-based losses. 54 55 It does the following: 56 - Sets up the graph loss and decoder (methods _init_k2 and update_k2_modules). 57 - Registers external graphs, if needed. 58 - Augments forward(...) with optional graph decoding to get accurate predictions. 59 """ 60 61 def _init_k2(self): 62 """ 63 k2-related initialization implementation. 64 65 This method is expected to run after the __init__ which sets self._cfg 66 self._cfg is expected to have the attribute graph_module_cfg 67 """ 68 if not hasattr(self, "_cfg"): 69 raise ValueError("self._cfg must be set before calling _init_k2().") 70 if not hasattr(self._cfg, "graph_module_cfg") or self._cfg.graph_module_cfg is None: 71 raise ValueError("self._cfg.graph_module_cfg must be set and cannot be None.") 72 self.graph_module_cfg = self._cfg.graph_module_cfg 73 74 # register token_lm for MAPLoss 75 criterion_type = self.graph_module_cfg.get("criterion_type", "ml") 76 self.use_graph_lm = criterion_type == "map" 77 if self.use_graph_lm: 78 token_lm_path = self.graph_module_cfg.backend_cfg.get("token_lm", None) 79 if token_lm_path is None: 80 raise ValueError( 81 f"graph_module_cfg.backend_cfg.token_lm is empty. It must be set for criterion_type == `{criterion_type}`" 82 ) 83 token_lm_path = self.register_artifact('graph_module_cfg.backend_cfg.token_lm', token_lm_path) 84 self.graph_module_cfg.backend_cfg["token_lm"] = token_lm_path 85 86 self.update_k2_modules(self.graph_module_cfg) 87 88 def update_k2_modules(self, input_cfg: DictConfig): 89 """ 90 Helper function to initialize or update k2 loss and transcribe_decoder. 91 92 Args: 93 input_cfg: DictConfig to take new parameters from. Schema is expected as in 94 nemo.collections.asr.models.configs.k2_sequence_models_config.GraphModuleConfig 95 """ 96 del self.loss 97 if hasattr(self, "transcribe_decoder"): 98 del self.transcribe_decoder 99 100 if hasattr(self, "joint"): 101 # RNNT 102 num_classes = self.joint.num_classes_with_blank - 1 103 else: 104 # CTC, MMI, ... 105 num_classes = self.decoder.num_classes_with_blank - 1 106 remove_consecutive = input_cfg.backend_cfg.get("topo_with_self_loops", True) and input_cfg.backend_cfg.get( 107 "topo_type", "default" 108 ) not in ["forced_blank", "identity",] 109 self._wer.remove_consecutive = remove_consecutive 110 111 from nemo.collections.asr.losses.lattice_losses import LatticeLoss 112 113 self.loss = LatticeLoss( 114 num_classes=num_classes, 115 reduction=self._cfg.get("ctc_reduction", "mean_batch"), 116 backend="k2", 117 criterion_type=input_cfg.get("criterion_type", "ml"), 118 loss_type=input_cfg.get("loss_type", "ctc"), 119 split_batch_size=input_cfg.get("split_batch_size", 0), 120 graph_module_cfg=input_cfg.backend_cfg, 121 ) 122 123 criterion_type = self.loss.criterion_type 124 self.use_graph_lm = criterion_type == "map" 125 transcribe_training = input_cfg.get("transcribe_training", False) 126 if transcribe_training and criterion_type == "ml": 127 logging.warning( 128 f"""You do not need to use transcribe_training=`{transcribe_training}` 129 with criterion_type=`{criterion_type}`. transcribe_training will be set to False.""" 130 ) 131 transcribe_training = False 132 self.transcribe_training = transcribe_training 133 if self.use_graph_lm: 134 from nemo.collections.asr.modules.graph_decoder import ViterbiDecoderWithGraph 135 136 self.transcribe_decoder = ViterbiDecoderWithGraph( 137 num_classes=num_classes, 138 backend="k2", 139 dec_type="token_lm", 140 return_type="1best", 141 return_ilabels=True, 142 output_aligned=True, 143 split_batch_size=input_cfg.get("split_batch_size", 0), 144 graph_module_cfg=input_cfg.backend_cfg, 145 ) 146 147 def _forward_k2_post_processing( 148 self, log_probs: torch.Tensor, encoded_length: torch.Tensor, greedy_predictions: torch.Tensor 149 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: 150 """ 151 k2-related post-processing parf of .forward() 152 153 Args: 154 log_probs: The log probabilities tensor of shape [B, T, D]. 155 encoded_length: The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 156 greedy_predictions: The greedy token predictions of the model of shape [B, T] 157 158 Returns: 159 A tuple of 3 elements - 160 1) The log probabilities tensor of shape [B, T, D]. 161 2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 162 3) The greedy token predictions of the model of shape [B, T] (via argmax) 163 """ 164 # greedy_predictions from .forward() are incorrect for criterion_type=`map` 165 # getting correct greedy_predictions, if needed 166 if self.use_graph_lm and (not self.training or self.transcribe_training): 167 greedy_predictions, encoded_length, _ = self.transcribe_decoder.forward( 168 log_probs=log_probs, log_probs_length=encoded_length 169 ) 170 return log_probs, encoded_length, greedy_predictions 171 [end of nemo/collections/asr/parts/k2/classes.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from dataclasses import dataclass 17 from typing import Any, Optional 18 19 import torch 20 from torch import nn as nn 21 22 from nemo.collections.asr.parts.submodules import multi_head_attention as mha 23 from nemo.collections.common.parts import adapter_modules 24 from nemo.core.classes.mixins import adapter_mixin_strategies 25 26 27 class MHAResidualAddAdapterStrategy(adapter_mixin_strategies.ResidualAddAdapterStrategy): 28 """ 29 An implementation of residual addition of an adapter module with its input for the MHA Adapters. 30 """ 31 32 def forward(self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin'): 33 """ 34 A basic strategy, comprising of a residual connection over the input, after forward pass by 35 the underlying adapter. Additional work is done to pack and unpack the dictionary of inputs and outputs. 36 37 Note: The `value` tensor is added to the output of the attention adapter as the residual connection. 38 39 Args: 40 input: A dictionary of multiple input arguments for the adapter module. 41 `query`, `key`, `value`: Original output tensor of the module, or the output of the 42 previous adapter (if more than one adapters are enabled). 43 `mask`: Attention mask. 44 `pos_emb`: Optional positional embedding for relative encoding. 45 adapter: The adapter module that is currently required to perform the forward pass. 46 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 47 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 48 49 Returns: 50 The result tensor, after one of the active adapters has finished its forward passes. 51 """ 52 out = self.compute_output(input, adapter, module=module) 53 54 # If not in training mode, or probability of stochastic depth is 0, skip step. 55 p = self.stochastic_depth 56 if not module.training or p == 0.0: 57 pass 58 else: 59 out = self.apply_stochastic_depth(out, input['value'], adapter, module=module) 60 61 # Return the residual connection output = input + adapter(input) 62 result = input['value'] + out 63 64 # If l2_lambda is activated, register the loss value 65 self.compute_auxiliary_losses(result, input['value'], adapter, module=module) 66 67 return result 68 69 def compute_output( 70 self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin' 71 ) -> torch.Tensor: 72 """ 73 Compute the output of a single adapter to some input. 74 75 Args: 76 input: Original output tensor of the module, or the output of the previous adapter (if more than 77 one adapters are enabled). 78 adapter: The adapter module that is currently required to perform the forward pass. 79 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 80 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 81 82 Returns: 83 The result tensor, after one of the active adapters has finished its forward passes. 84 """ 85 if isinstance(input, (list, tuple)): 86 out = adapter(*input) 87 elif isinstance(input, dict): 88 out = adapter(**input) 89 else: 90 out = adapter(input) 91 return out 92 93 94 @dataclass 95 class MHAResidualAddAdapterStrategyConfig(adapter_mixin_strategies.ResidualAddAdapterStrategyConfig): 96 _target_: str = "{0}.{1}".format( 97 MHAResidualAddAdapterStrategy.__module__, MHAResidualAddAdapterStrategy.__name__ 98 ) # mandatory field 99 100 101 class MultiHeadAttentionAdapter(mha.MultiHeadAttention, adapter_modules.AdapterModuleUtil): 102 """Multi-Head Attention layer of Transformer. 103 Args: 104 n_head (int): number of heads 105 n_feat (int): size of the features 106 dropout_rate (float): dropout rate 107 proj_dim (int, optional): Optional integer value for projection before computing attention. 108 If None, then there is no projection (equivalent to proj_dim = n_feat). 109 If > 0, then will project the n_feat to proj_dim before calculating attention. 110 If <0, then will equal n_head, so that each head has a projected dimension of 1. 111 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 112 """ 113 114 def __init__( 115 self, 116 n_head: int, 117 n_feat: int, 118 dropout_rate: float, 119 proj_dim: Optional[int] = None, 120 adapter_strategy: MHAResidualAddAdapterStrategy = None, 121 ): 122 super().__init__(n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, max_cache_len=0) 123 124 self.pre_norm = nn.LayerNorm(n_feat) 125 126 # Set the projection dim to number of heads automatically 127 if proj_dim is not None and proj_dim < 1: 128 proj_dim = n_head 129 130 self.proj_dim = proj_dim 131 132 # Recompute weights for projection dim 133 if self.proj_dim is not None: 134 if self.proj_dim % n_head != 0: 135 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 136 137 self.d_k = self.proj_dim // n_head 138 self.s_d_k = math.sqrt(self.d_k) 139 self.linear_q = nn.Linear(n_feat, self.proj_dim) 140 self.linear_k = nn.Linear(n_feat, self.proj_dim) 141 self.linear_v = nn.Linear(n_feat, self.proj_dim) 142 self.linear_out = nn.Linear(self.proj_dim, n_feat) 143 144 # Setup adapter strategy 145 self.setup_adapter_strategy(adapter_strategy) 146 147 # reset parameters for Q to be identity operation 148 self.reset_parameters() 149 150 def forward(self, query, key, value, mask, pos_emb=None, cache=None): 151 """Compute 'Scaled Dot Product Attention'. 152 Args: 153 query (torch.Tensor): (batch, time1, size) 154 key (torch.Tensor): (batch, time2, size) 155 value(torch.Tensor): (batch, time2, size) 156 mask (torch.Tensor): (batch, time1, time2) 157 cache (torch.Tensor) : (batch, time_cache, size) 158 159 returns: 160 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 161 cache (torch.Tensor) : (batch, time_cache_next, size) 162 """ 163 # Need to perform duplicate computations as at this point the tensors have been 164 # separated by the adapter forward 165 query = self.pre_norm(query) 166 key = self.pre_norm(key) 167 value = self.pre_norm(value) 168 169 return super().forward(query, key, value, mask, pos_emb, cache=cache) 170 171 def reset_parameters(self): 172 with torch.no_grad(): 173 nn.init.zeros_(self.linear_out.weight) 174 nn.init.zeros_(self.linear_out.bias) 175 176 def get_default_strategy_config(self) -> 'dataclass': 177 return MHAResidualAddAdapterStrategyConfig() 178 179 180 @dataclass 181 class MultiHeadAttentionAdapterConfig: 182 n_head: int 183 n_feat: int 184 dropout_rate: float = 0.0 185 proj_dim: Optional[int] = None 186 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 187 _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) 188 189 190 class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): 191 """Multi-Head Attention layer of Transformer-XL with support of relative positional encoding. 192 Paper: https://arxiv.org/abs/1901.02860 193 Args: 194 n_head (int): number of heads 195 n_feat (int): size of the features 196 dropout_rate (float): dropout rate 197 proj_dim (int, optional): Optional integer value for projection before computing attention. 198 If None, then there is no projection (equivalent to proj_dim = n_feat). 199 If > 0, then will project the n_feat to proj_dim before calculating attention. 200 If <0, then will equal n_head, so that each head has a projected dimension of 1. 201 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 202 """ 203 204 def __init__( 205 self, 206 n_head: int, 207 n_feat: int, 208 dropout_rate: float, 209 proj_dim: Optional[int] = None, 210 adapter_strategy: MHAResidualAddAdapterStrategyConfig = None, 211 ): 212 super().__init__( 213 n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, pos_bias_u=None, pos_bias_v=None, max_cache_len=0 214 ) 215 216 self.pre_norm = nn.LayerNorm(n_feat) 217 218 # Set the projection dim to number of heads automatically 219 if proj_dim is not None and proj_dim < 1: 220 proj_dim = n_head 221 222 self.proj_dim = proj_dim 223 224 # Recompute weights for projection dim 225 if self.proj_dim is not None: 226 if self.proj_dim % n_head != 0: 227 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 228 229 self.d_k = self.proj_dim // n_head 230 self.s_d_k = math.sqrt(self.d_k) 231 self.linear_q = nn.Linear(n_feat, self.proj_dim) 232 self.linear_k = nn.Linear(n_feat, self.proj_dim) 233 self.linear_v = nn.Linear(n_feat, self.proj_dim) 234 self.linear_out = nn.Linear(self.proj_dim, n_feat) 235 self.linear_pos = nn.Linear(n_feat, self.proj_dim, bias=False) 236 self.pos_bias_u = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 237 self.pos_bias_v = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 238 239 # Setup adapter strategy 240 self.setup_adapter_strategy(adapter_strategy) 241 242 # reset parameters for Q to be identity operation 243 self.reset_parameters() 244 245 def forward(self, query, key, value, mask, pos_emb, cache=None): 246 """Compute 'Scaled Dot Product Attention' with rel. positional encoding. 247 Args: 248 query (torch.Tensor): (batch, time1, size) 249 key (torch.Tensor): (batch, time2, size) 250 value(torch.Tensor): (batch, time2, size) 251 mask (torch.Tensor): (batch, time1, time2) 252 pos_emb (torch.Tensor) : (batch, time1, size) 253 cache (torch.Tensor) : (batch, time_cache, size) 254 Returns: 255 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 256 cache_next (torch.Tensor) : (batch, time_cache_next, size) 257 """ 258 # Need to perform duplicate computations as at this point the tensors have been 259 # separated by the adapter forward 260 query = self.pre_norm(query) 261 key = self.pre_norm(key) 262 value = self.pre_norm(value) 263 264 return super().forward(query, key, value, mask, pos_emb, cache=cache) 265 266 def reset_parameters(self): 267 with torch.no_grad(): 268 nn.init.zeros_(self.linear_out.weight) 269 nn.init.zeros_(self.linear_out.bias) 270 271 # NOTE: This exact procedure apparently highly important. 272 # Above operation is safe to do as self.linear_out.weight *= 0.0 (similar for bias) 273 # However: 274 # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 275 # For some reason at init sometimes it will cause the value of the tensor to become NaN 276 # All operations to compute matrix_ac and matrix_bd will then fail. 277 nn.init.zeros_(self.pos_bias_u) 278 nn.init.zeros_(self.pos_bias_v) 279 280 def get_default_strategy_config(self) -> 'dataclass': 281 return MHAResidualAddAdapterStrategyConfig() 282 283 284 @dataclass 285 class RelPositionMultiHeadAttentionAdapterConfig: 286 n_head: int 287 n_feat: int 288 dropout_rate: float = 0.0 289 proj_dim: Optional[int] = None 290 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 291 _target_: str = "{0}.{1}".format( 292 RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ 293 ) 294 295 296 class PositionalEncodingAdapter(mha.PositionalEncoding, adapter_modules.AdapterModuleUtil): 297 298 """ 299 Absolute positional embedding adapter. 300 301 .. note:: 302 303 Absolute positional embedding value is added to the input tensor *without residual connection* ! 304 Therefore, the input is changed, if you only require the positional embedding, drop the returned `x` ! 305 306 Args: 307 d_model (int): The input dimension of x. 308 max_len (int): The max sequence length. 309 xscale (float): The input scaling factor. Defaults to 1.0. 310 adapter_strategy (AbstractAdapterStrategy): By default, ReturnResultAdapterStrategyConfig. 311 An adapter composition function object. 312 NOTE: Since this is a positional encoding, it will not add a residual ! 313 """ 314 315 def __init__( 316 self, 317 d_model: int, 318 max_len: int = 5000, 319 xscale=1.0, 320 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 321 ): 322 323 super().__init__( 324 d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, 325 ) 326 327 # Setup adapter strategy 328 self.setup_adapter_strategy(adapter_strategy) 329 330 def get_default_strategy_config(self) -> 'dataclass': 331 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 332 333 334 @dataclass 335 class PositionalEncodingAdapterConfig: 336 d_model: int 337 max_len: int = 5000 338 xscale: float = 1.0 339 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 340 _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) 341 342 343 class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): 344 """ 345 Relative positional encoding for TransformerXL's layers 346 See : Appendix B in https://arxiv.org/abs/1901.02860 347 348 .. note:: 349 350 Relative positional embedding value is **not** added to the input tensor ! 351 Therefore, the input should be updated changed, if you only require the positional embedding, drop the returned `x` ! 352 353 Args: 354 d_model (int): embedding dim 355 max_len (int): maximum input length 356 xscale (bool): whether to scale the input by sqrt(d_model) 357 adapter_strategy: By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. 358 """ 359 360 def __init__( 361 self, 362 d_model: int, 363 max_len: int = 5000, 364 xscale=1.0, 365 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 366 ): 367 super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) 368 369 # Setup adapter strategy 370 self.setup_adapter_strategy(adapter_strategy) 371 372 def get_default_strategy_config(self) -> 'dataclass': 373 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 374 375 376 @dataclass 377 class RelPositionalEncodingAdapterConfig: 378 d_model: int 379 max_len: int = 5000 380 xscale: float = 1.0 381 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 382 _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) 383 [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import os 17 from dataclasses import dataclass 18 from typing import List, Optional, Tuple, Union 19 20 import torch 21 22 from nemo.collections.asr.parts.utils import rnnt_utils 23 from nemo.collections.common.tokenizers.tokenizer_spec import TokenizerSpec 24 from nemo.core.classes import Typing, typecheck 25 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 26 from nemo.utils import logging 27 28 DEFAULT_TOKEN_OFFSET = 100 29 30 31 def pack_hypotheses( 32 hypotheses: List[rnnt_utils.NBestHypotheses], logitlen: torch.Tensor, 33 ) -> List[rnnt_utils.NBestHypotheses]: 34 35 if logitlen is not None: 36 if hasattr(logitlen, 'cpu'): 37 logitlen_cpu = logitlen.to('cpu') 38 else: 39 logitlen_cpu = logitlen 40 41 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.NBestHypotheses 42 for candidate_idx, cand in enumerate(hyp.n_best_hypotheses): 43 cand.y_sequence = torch.tensor(cand.y_sequence, dtype=torch.long) 44 45 if logitlen is not None: 46 cand.length = logitlen_cpu[idx] 47 48 if cand.dec_state is not None: 49 cand.dec_state = _states_to_device(cand.dec_state) 50 51 return hypotheses 52 53 54 def _states_to_device(dec_state, device='cpu'): 55 if torch.is_tensor(dec_state): 56 dec_state = dec_state.to(device) 57 58 elif isinstance(dec_state, (list, tuple)): 59 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 60 61 return dec_state 62 63 64 class AbstractBeamCTCInfer(Typing): 65 """A beam CTC decoder. 66 67 Provides a common abstraction for sample level beam decoding. 68 69 Args: 70 blank_id: int, index of the blank token. Can be 0 or len(vocabulary). 71 beam_size: int, size of the beam used in the underlying beam search engine. 72 73 """ 74 75 @property 76 def input_types(self): 77 """Returns definitions of module input ports. 78 """ 79 return { 80 "decoder_output": NeuralType(('B', 'T', 'D'), LogprobsType()), 81 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 82 } 83 84 @property 85 def output_types(self): 86 """Returns definitions of module output ports. 87 """ 88 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 89 90 def __init__(self, blank_id: int, beam_size: int): 91 self.blank_id = blank_id 92 93 if beam_size < 1: 94 raise ValueError("Beam search size cannot be less than 1!") 95 96 self.beam_size = beam_size 97 98 # Variables set by corresponding setter methods 99 self.vocab = None 100 self.decoding_type = None 101 self.tokenizer = None 102 103 # Utility maps for vocabulary 104 self.vocab_index_map = None 105 self.index_vocab_map = None 106 107 # Internal variable, used to prevent double reduction of consecutive tokens (ctc collapse) 108 self.override_fold_consecutive_value = None 109 110 def set_vocabulary(self, vocab: List[str]): 111 """ 112 Set the vocabulary of the decoding framework. 113 114 Args: 115 vocab: List of str. Each token corresponds to its location in the vocabulary emitted by the model. 116 Note that this vocabulary must NOT contain the "BLANK" token. 117 """ 118 self.vocab = vocab 119 self.vocab_index_map = {v: i for i, v in enumerate(vocab)} 120 self.index_vocab_map = {i: v for i, v in enumerate(vocab)} 121 122 def set_decoding_type(self, decoding_type: str): 123 """ 124 Sets the decoding type of the framework. Can support either char or subword models. 125 126 Args: 127 decoding_type: Str corresponding to decoding type. Only supports "char" and "subword". 128 """ 129 decoding_type = decoding_type.lower() 130 supported_types = ['char', 'subword'] 131 132 if decoding_type not in supported_types: 133 raise ValueError( 134 f"Unsupported decoding type. Supported types = {supported_types}.\n" f"Given = {decoding_type}" 135 ) 136 137 self.decoding_type = decoding_type 138 139 def set_tokenizer(self, tokenizer: TokenizerSpec): 140 """ 141 Set the tokenizer of the decoding framework. 142 143 Args: 144 tokenizer: NeMo tokenizer object, which inherits from TokenizerSpec. 145 """ 146 self.tokenizer = tokenizer 147 148 @typecheck() 149 def forward( 150 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 151 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 152 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 153 Output token is generated auto-repressively. 154 155 Args: 156 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 157 decoder_lengths: list of int representing the length of each sequence 158 output sequence. 159 160 Returns: 161 packed list containing batch number of sentences (Hypotheses). 162 """ 163 raise NotImplementedError() 164 165 def __call__(self, *args, **kwargs): 166 return self.forward(*args, **kwargs) 167 168 169 class BeamCTCInfer(AbstractBeamCTCInfer): 170 """A greedy CTC decoder. 171 172 Provides a common abstraction for sample level and batch level greedy decoding. 173 174 Args: 175 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 176 preserve_alignments: Bool flag which preserves the history of logprobs generated during 177 decoding (sample / batched). When set to true, the Hypothesis will contain 178 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 179 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 180 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 181 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 182 183 """ 184 185 def __init__( 186 self, 187 blank_id: int, 188 beam_size: int, 189 search_type: str = "default", 190 return_best_hypothesis: bool = True, 191 preserve_alignments: bool = False, 192 compute_timestamps: bool = False, 193 beam_alpha: float = 1.0, 194 beam_beta: float = 0.0, 195 kenlm_path: str = None, 196 flashlight_cfg: Optional['FlashlightConfig'] = None, 197 pyctcdecode_cfg: Optional['PyCTCDecodeConfig'] = None, 198 ): 199 super().__init__(blank_id=blank_id, beam_size=beam_size) 200 201 self.search_type = search_type 202 self.return_best_hypothesis = return_best_hypothesis 203 self.preserve_alignments = preserve_alignments 204 self.compute_timestamps = compute_timestamps 205 206 if self.compute_timestamps: 207 raise ValueError(f"Currently this flag is not supported for beam search algorithms.") 208 209 self.vocab = None # This must be set by specific method by user before calling forward() ! 210 211 if search_type == "default" or search_type == "nemo": 212 self.search_algorithm = self.default_beam_search 213 elif search_type == "pyctcdecode": 214 self.search_algorithm = self._pyctcdecode_beam_search 215 elif search_type == "flashlight": 216 self.search_algorithm = self.flashlight_beam_search 217 else: 218 raise NotImplementedError( 219 f"The search type ({search_type}) supplied is not supported!\n" 220 f"Please use one of : (default, nemo, pyctcdecode)" 221 ) 222 223 # Log the beam search algorithm 224 logging.info(f"Beam search algorithm: {search_type}") 225 226 self.beam_alpha = beam_alpha 227 self.beam_beta = beam_beta 228 229 # Default beam search args 230 self.kenlm_path = kenlm_path 231 232 # PyCTCDecode params 233 if pyctcdecode_cfg is None: 234 pyctcdecode_cfg = PyCTCDecodeConfig() 235 self.pyctcdecode_cfg = pyctcdecode_cfg # type: PyCTCDecodeConfig 236 237 if flashlight_cfg is None: 238 flashlight_cfg = FlashlightConfig() 239 self.flashlight_cfg = flashlight_cfg 240 241 # Default beam search scorer functions 242 self.default_beam_scorer = None 243 self.pyctcdecode_beam_scorer = None 244 self.flashlight_beam_scorer = None 245 self.token_offset = 0 246 247 @typecheck() 248 def forward( 249 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 250 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 251 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 252 Output token is generated auto-repressively. 253 254 Args: 255 decoder_output: A tensor of size (batch, timesteps, features). 256 decoder_lengths: list of int representing the length of each sequence 257 output sequence. 258 259 Returns: 260 packed list containing batch number of sentences (Hypotheses). 261 """ 262 if self.vocab is None: 263 raise RuntimeError("Please set the vocabulary with `set_vocabulary()` before calling this function.") 264 265 if self.decoding_type is None: 266 raise ValueError("Please set the decoding type with `set_decoding_type()` before calling this function.") 267 268 with torch.no_grad(), torch.inference_mode(): 269 # Process each sequence independently 270 prediction_tensor = decoder_output 271 272 if prediction_tensor.ndim != 3: 273 raise ValueError( 274 f"`decoder_output` must be a tensor of shape [B, T, V] (log probs, float). " 275 f"Provided shape = {prediction_tensor.shape}" 276 ) 277 278 # determine type of input - logprobs or labels 279 out_len = decoder_lengths if decoder_lengths is not None else None 280 hypotheses = self.search_algorithm(prediction_tensor, out_len) 281 282 # Pack results into Hypotheses 283 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 284 285 # Pack the result 286 if self.return_best_hypothesis and isinstance(packed_result[0], rnnt_utils.NBestHypotheses): 287 packed_result = [res.n_best_hypotheses[0] for res in packed_result] # type: Hypothesis 288 289 return (packed_result,) 290 291 @torch.no_grad() 292 def default_beam_search( 293 self, x: torch.Tensor, out_len: torch.Tensor 294 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 295 """ 296 Open Seq2Seq Beam Search Algorithm (DeepSpeed) 297 298 Args: 299 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 300 and V is the vocabulary size. The tensor contains log-probabilities. 301 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 302 303 Returns: 304 A list of NBestHypotheses objects, one for each sequence in the batch. 305 """ 306 if self.compute_timestamps: 307 raise ValueError( 308 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 309 ) 310 311 if self.default_beam_scorer is None: 312 # Check for filepath 313 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 314 raise FileNotFoundError( 315 f"KenLM binary file not found at : {self.kenlm_path}. " 316 f"Please set a valid path in the decoding config." 317 ) 318 319 # perform token offset for subword models 320 if self.decoding_type == 'subword': 321 vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 322 else: 323 # char models 324 vocab = self.vocab 325 326 # Must import at runtime to avoid circular dependency due to module level import. 327 from nemo.collections.asr.modules.beam_search_decoder import BeamSearchDecoderWithLM 328 329 self.default_beam_scorer = BeamSearchDecoderWithLM( 330 vocab=vocab, 331 lm_path=self.kenlm_path, 332 beam_width=self.beam_size, 333 alpha=self.beam_alpha, 334 beta=self.beam_beta, 335 num_cpus=max(1, os.cpu_count()), 336 input_tensor=False, 337 ) 338 339 x = x.to('cpu') 340 341 with typecheck.disable_checks(): 342 data = [x[sample_id, : out_len[sample_id], :].softmax(dim=-1) for sample_id in range(len(x))] 343 beams_batch = self.default_beam_scorer.forward(log_probs=data, log_probs_length=None) 344 345 # For each sample in the batch 346 nbest_hypotheses = [] 347 for beams_idx, beams in enumerate(beams_batch): 348 # For each beam candidate / hypothesis in each sample 349 hypotheses = [] 350 for candidate_idx, candidate in enumerate(beams): 351 hypothesis = rnnt_utils.Hypothesis( 352 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 353 ) 354 355 # For subword encoding, NeMo will double encode the subword (multiple tokens) into a 356 # singular unicode id. In doing so, we preserve the semantic of the unicode token, and 357 # compress the size of the final KenLM ARPA / Binary file. 358 # In order to do double encoding, we shift the subword by some token offset. 359 # This step is ignored for character based models. 360 if self.decoding_type == 'subword': 361 pred_token_ids = [ord(c) - self.token_offset for c in candidate[1]] 362 else: 363 # Char models 364 pred_token_ids = [self.vocab_index_map[c] for c in candidate[1]] 365 366 # We preserve the token ids and the score for this hypothesis 367 hypothesis.y_sequence = pred_token_ids 368 hypothesis.score = candidate[0] 369 370 # If alignment must be preserved, we preserve a view of the output logprobs. 371 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 372 # require specific processing for each sample in the beam. 373 # This is done to preserve memory. 374 if self.preserve_alignments: 375 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 376 377 hypotheses.append(hypothesis) 378 379 # Wrap the result in NBestHypothesis. 380 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 381 nbest_hypotheses.append(hypotheses) 382 383 return nbest_hypotheses 384 385 @torch.no_grad() 386 def _pyctcdecode_beam_search( 387 self, x: torch.Tensor, out_len: torch.Tensor 388 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 389 """ 390 PyCTCDecode Beam Search Algorithm. Should support Char and Subword models. 391 392 Args: 393 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 394 and V is the vocabulary size. The tensor contains log-probabilities. 395 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 396 397 Returns: 398 A list of NBestHypotheses objects, one for each sequence in the batch. 399 """ 400 if self.compute_timestamps: 401 raise ValueError( 402 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 403 ) 404 405 try: 406 import pyctcdecode 407 except (ImportError, ModuleNotFoundError): 408 raise ImportError( 409 f"Could not load `pyctcdecode` library. Please install it from pip using :\n" 410 f"pip install --upgrade pyctcdecode" 411 ) 412 413 if self.pyctcdecode_beam_scorer is None: 414 self.pyctcdecode_beam_scorer = pyctcdecode.build_ctcdecoder( 415 labels=self.vocab, kenlm_model_path=self.kenlm_path, alpha=self.beam_alpha, beta=self.beam_beta 416 ) # type: pyctcdecode.BeamSearchDecoderCTC 417 418 x = x.to('cpu').numpy() 419 420 with typecheck.disable_checks(): 421 beams_batch = [] 422 for sample_id in range(len(x)): 423 logprobs = x[sample_id, : out_len[sample_id], :] 424 result = self.pyctcdecode_beam_scorer.decode_beams( 425 logprobs, 426 beam_width=self.beam_size, 427 beam_prune_logp=self.pyctcdecode_cfg.beam_prune_logp, 428 token_min_logp=self.pyctcdecode_cfg.token_min_logp, 429 prune_history=self.pyctcdecode_cfg.prune_history, 430 hotwords=self.pyctcdecode_cfg.hotwords, 431 hotword_weight=self.pyctcdecode_cfg.hotword_weight, 432 lm_start_state=None, 433 ) # Output format: text, last_lm_state, text_frames, logit_score, lm_score 434 beams_batch.append(result) 435 436 nbest_hypotheses = [] 437 for beams_idx, beams in enumerate(beams_batch): 438 hypotheses = [] 439 for candidate_idx, candidate in enumerate(beams): 440 # Candidate = (text, last_lm_state, text_frames, logit_score, lm_score) 441 hypothesis = rnnt_utils.Hypothesis( 442 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 443 ) 444 445 # TODO: Requires token ids to be returned rather than text. 446 if self.decoding_type == 'subword': 447 if self.tokenizer is None: 448 raise ValueError("Tokenizer must be provided for subword decoding. Use set_tokenizer().") 449 450 pred_token_ids = self.tokenizer.text_to_ids(candidate[0]) 451 else: 452 if self.vocab is None: 453 raise ValueError("Vocab must be provided for character decoding. Use set_vocab().") 454 455 chars = list(candidate[0]) 456 pred_token_ids = [self.vocab_index_map[c] for c in chars] 457 458 hypothesis.y_sequence = pred_token_ids 459 hypothesis.text = candidate[0] # text 460 hypothesis.score = candidate[4] # score 461 462 # Inject word level timestamps 463 hypothesis.timestep = candidate[2] # text_frames 464 465 if self.preserve_alignments: 466 hypothesis.alignments = torch.from_numpy(x[beams_idx][: out_len[beams_idx]]) 467 468 hypotheses.append(hypothesis) 469 470 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 471 nbest_hypotheses.append(hypotheses) 472 473 return nbest_hypotheses 474 475 @torch.no_grad() 476 def flashlight_beam_search( 477 self, x: torch.Tensor, out_len: torch.Tensor 478 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 479 """ 480 Flashlight Beam Search Algorithm. Should support Char and Subword models. 481 482 Args: 483 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 484 and V is the vocabulary size. The tensor contains log-probabilities. 485 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 486 487 Returns: 488 A list of NBestHypotheses objects, one for each sequence in the batch. 489 """ 490 if self.compute_timestamps: 491 raise ValueError( 492 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 493 ) 494 495 if self.flashlight_beam_scorer is None: 496 # Check for filepath 497 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 498 raise FileNotFoundError( 499 f"KenLM binary file not found at : {self.kenlm_path}. " 500 f"Please set a valid path in the decoding config." 501 ) 502 503 # perform token offset for subword models 504 # if self.decoding_type == 'subword': 505 # vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 506 # else: 507 # # char models 508 # vocab = self.vocab 509 510 # Must import at runtime to avoid circular dependency due to module level import. 511 from nemo.collections.asr.modules.flashlight_decoder import FlashLightKenLMBeamSearchDecoder 512 513 self.flashlight_beam_scorer = FlashLightKenLMBeamSearchDecoder( 514 lm_path=self.kenlm_path, 515 vocabulary=self.vocab, 516 tokenizer=self.tokenizer, 517 lexicon_path=self.flashlight_cfg.lexicon_path, 518 boost_path=self.flashlight_cfg.boost_path, 519 beam_size=self.beam_size, 520 beam_size_token=self.flashlight_cfg.beam_size_token, 521 beam_threshold=self.flashlight_cfg.beam_threshold, 522 lm_weight=self.beam_alpha, 523 word_score=self.beam_beta, 524 unk_weight=self.flashlight_cfg.unk_weight, 525 sil_weight=self.flashlight_cfg.sil_weight, 526 ) 527 528 x = x.to('cpu') 529 530 with typecheck.disable_checks(): 531 beams_batch = self.flashlight_beam_scorer.forward(log_probs=x) 532 533 # For each sample in the batch 534 nbest_hypotheses = [] 535 for beams_idx, beams in enumerate(beams_batch): 536 # For each beam candidate / hypothesis in each sample 537 hypotheses = [] 538 for candidate_idx, candidate in enumerate(beams): 539 hypothesis = rnnt_utils.Hypothesis( 540 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 541 ) 542 543 # We preserve the token ids and the score for this hypothesis 544 hypothesis.y_sequence = candidate['tokens'].tolist() 545 hypothesis.score = candidate['score'] 546 547 # If alignment must be preserved, we preserve a view of the output logprobs. 548 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 549 # require specific processing for each sample in the beam. 550 # This is done to preserve memory. 551 if self.preserve_alignments: 552 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 553 554 hypotheses.append(hypothesis) 555 556 # Wrap the result in NBestHypothesis. 557 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 558 nbest_hypotheses.append(hypotheses) 559 560 return nbest_hypotheses 561 562 def set_decoding_type(self, decoding_type: str): 563 super().set_decoding_type(decoding_type) 564 565 # Please check train_kenlm.py in scripts/asr_language_modeling/ to find out why we need 566 # TOKEN_OFFSET for BPE-based models 567 if self.decoding_type == 'subword': 568 self.token_offset = DEFAULT_TOKEN_OFFSET 569 570 571 @dataclass 572 class PyCTCDecodeConfig: 573 # These arguments cannot be imported from pyctcdecode (optional dependency) 574 # Therefore we copy the values explicitly 575 # Taken from pyctcdecode.constant 576 beam_prune_logp: float = -10.0 577 token_min_logp: float = -5.0 578 prune_history: bool = False 579 hotwords: Optional[List[str]] = None 580 hotword_weight: float = 10.0 581 582 583 @dataclass 584 class FlashlightConfig: 585 lexicon_path: Optional[str] = None 586 boost_path: Optional[str] = None 587 beam_size_token: int = 16 588 beam_threshold: float = 20.0 589 unk_weight: float = -math.inf 590 sil_weight: float = 0.0 591 592 593 @dataclass 594 class BeamCTCInferConfig: 595 beam_size: int 596 search_type: str = 'default' 597 preserve_alignments: bool = False 598 compute_timestamps: bool = False 599 return_best_hypothesis: bool = True 600 601 beam_alpha: float = 1.0 602 beam_beta: float = 0.0 603 kenlm_path: Optional[str] = None 604 605 flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() 606 pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() 607 [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import List, Optional 17 18 import torch 19 from omegaconf import DictConfig, OmegaConf 20 21 from nemo.collections.asr.parts.utils import rnnt_utils 22 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMethodConfig, ConfidenceMethodMixin 23 from nemo.core.classes import Typing, typecheck 24 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 25 from nemo.utils import logging 26 27 28 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 29 30 if logitlen is not None: 31 if hasattr(logitlen, 'cpu'): 32 logitlen_cpu = logitlen.to('cpu') 33 else: 34 logitlen_cpu = logitlen 35 36 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 37 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 38 39 if logitlen is not None: 40 hyp.length = logitlen_cpu[idx] 41 42 if hyp.dec_state is not None: 43 hyp.dec_state = _states_to_device(hyp.dec_state) 44 45 return hypotheses 46 47 48 def _states_to_device(dec_state, device='cpu'): 49 if torch.is_tensor(dec_state): 50 dec_state = dec_state.to(device) 51 52 elif isinstance(dec_state, (list, tuple)): 53 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 54 55 return dec_state 56 57 58 class GreedyCTCInfer(Typing, ConfidenceMethodMixin): 59 """A greedy CTC decoder. 60 61 Provides a common abstraction for sample level and batch level greedy decoding. 62 63 Args: 64 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 65 preserve_alignments: Bool flag which preserves the history of logprobs generated during 66 decoding (sample / batched). When set to true, the Hypothesis will contain 67 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 68 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 69 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 70 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 71 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 72 generated during decoding. When set to true, the Hypothesis will contain 73 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 74 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 75 confidence scores. 76 77 name: The method name (str). 78 Supported values: 79 - 'max_prob' for using the maximum token probability as a confidence. 80 - 'entropy' for using a normalized entropy of a log-likelihood vector. 81 82 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 83 Supported values: 84 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 85 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 86 Note that for this entropy, the alpha should comply the following inequality: 87 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 88 where V is the model vocabulary size. 89 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 90 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 91 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 92 More: https://en.wikipedia.org/wiki/Tsallis_entropy 93 - 'renyi' for the Rรฉnyi entropy. 94 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 95 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 96 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 97 98 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 99 When the alpha equals one, scaling is not applied to 'max_prob', 100 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 101 102 entropy_norm: A mapping of the entropy value to the interval [0,1]. 103 Supported values: 104 - 'lin' for using the linear mapping. 105 - 'exp' for using exponential mapping with linear shift. 106 107 """ 108 109 @property 110 def input_types(self): 111 """Returns definitions of module input ports. 112 """ 113 # Input can be of dimention - 114 # ('B', 'T', 'D') [Log probs] or ('B', 'T') [Labels] 115 116 return { 117 "decoder_output": NeuralType(None, LogprobsType()), 118 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 119 } 120 121 @property 122 def output_types(self): 123 """Returns definitions of module output ports. 124 """ 125 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 126 127 def __init__( 128 self, 129 blank_id: int, 130 preserve_alignments: bool = False, 131 compute_timestamps: bool = False, 132 preserve_frame_confidence: bool = False, 133 confidence_method_cfg: Optional[DictConfig] = None, 134 ): 135 super().__init__() 136 137 self.blank_id = blank_id 138 self.preserve_alignments = preserve_alignments 139 # we need timestamps to extract non-blank per-frame confidence 140 self.compute_timestamps = compute_timestamps | preserve_frame_confidence 141 self.preserve_frame_confidence = preserve_frame_confidence 142 143 # set confidence calculation method 144 self._init_confidence_method(confidence_method_cfg) 145 146 @typecheck() 147 def forward( 148 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 149 ): 150 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 151 Output token is generated auto-repressively. 152 153 Args: 154 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 155 decoder_lengths: list of int representing the length of each sequence 156 output sequence. 157 158 Returns: 159 packed list containing batch number of sentences (Hypotheses). 160 """ 161 with torch.inference_mode(): 162 hypotheses = [] 163 # Process each sequence independently 164 prediction_cpu_tensor = decoder_output.cpu() 165 166 if prediction_cpu_tensor.ndim < 2 or prediction_cpu_tensor.ndim > 3: 167 raise ValueError( 168 f"`decoder_output` must be a tensor of shape [B, T] (labels, int) or " 169 f"[B, T, V] (log probs, float). Provided shape = {prediction_cpu_tensor.shape}" 170 ) 171 172 # determine type of input - logprobs or labels 173 if prediction_cpu_tensor.ndim == 2: # labels 174 greedy_decode = self._greedy_decode_labels 175 else: 176 greedy_decode = self._greedy_decode_logprobs 177 178 for ind in range(prediction_cpu_tensor.shape[0]): 179 out_len = decoder_lengths[ind] if decoder_lengths is not None else None 180 hypothesis = greedy_decode(prediction_cpu_tensor[ind], out_len) 181 hypotheses.append(hypothesis) 182 183 # Pack results into Hypotheses 184 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 185 186 return (packed_result,) 187 188 @torch.no_grad() 189 def _greedy_decode_logprobs(self, x: torch.Tensor, out_len: torch.Tensor): 190 # x: [T, D] 191 # out_len: [seq_len] 192 193 # Initialize blank state and empty label set in Hypothesis 194 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 195 prediction = x.detach().cpu() 196 197 if out_len is not None: 198 prediction = prediction[:out_len] 199 200 prediction_logprobs, prediction_labels = prediction.max(dim=-1) 201 202 non_blank_ids = prediction_labels != self.blank_id 203 hypothesis.y_sequence = prediction_labels.numpy().tolist() 204 hypothesis.score = (prediction_logprobs[non_blank_ids]).sum() 205 206 if self.preserve_alignments: 207 # Preserve the logprobs, as well as labels after argmax 208 hypothesis.alignments = (prediction.clone(), prediction_labels.clone()) 209 210 if self.compute_timestamps: 211 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 212 213 if self.preserve_frame_confidence: 214 hypothesis.frame_confidence = self._get_confidence(prediction) 215 216 return hypothesis 217 218 @torch.no_grad() 219 def _greedy_decode_labels(self, x: torch.Tensor, out_len: torch.Tensor): 220 # x: [T] 221 # out_len: [seq_len] 222 223 # Initialize blank state and empty label set in Hypothesis 224 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 225 prediction_labels = x.detach().cpu() 226 227 if out_len is not None: 228 prediction_labels = prediction_labels[:out_len] 229 230 non_blank_ids = prediction_labels != self.blank_id 231 hypothesis.y_sequence = prediction_labels.numpy().tolist() 232 hypothesis.score = -1.0 233 234 if self.preserve_alignments: 235 raise ValueError("Requested for alignments, but predictions provided were labels, not log probabilities.") 236 237 if self.compute_timestamps: 238 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 239 240 if self.preserve_frame_confidence: 241 raise ValueError( 242 "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." 243 ) 244 245 return hypothesis 246 247 def __call__(self, *args, **kwargs): 248 return self.forward(*args, **kwargs) 249 250 251 @dataclass 252 class GreedyCTCInferConfig: 253 preserve_alignments: bool = False 254 compute_timestamps: bool = False 255 preserve_frame_confidence: bool = False 256 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 257 258 def __post_init__(self): 259 # OmegaConf.structured ensures that post_init check is always executed 260 self.confidence_method_cfg = OmegaConf.structured( 261 self.confidence_method_cfg 262 if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) 263 else ConfidenceMethodConfig(**self.confidence_method_cfg) 264 ) 265 [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 # Copyright 2017 Johns Hopkins University (Shinji Watanabe) 16 # 17 # Licensed under the Apache License, Version 2.0 (the "License"); 18 # you may not use this file except in compliance with the License. 19 # You may obtain a copy of the License at 20 # 21 # http://www.apache.org/licenses/LICENSE-2.0 22 # 23 # Unless required by applicable law or agreed to in writing, software 24 # distributed under the License is distributed on an "AS IS" BASIS, 25 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 26 # See the License for the specific language governing permissions and 27 # limitations under the License. 28 29 from dataclasses import dataclass 30 from typing import List, Optional, Tuple, Union 31 32 import numpy as np 33 import torch 34 from omegaconf import DictConfig, OmegaConf 35 36 from nemo.collections.asr.modules import rnnt_abstract 37 from nemo.collections.asr.parts.utils import rnnt_utils 38 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMethodConfig, ConfidenceMethodMixin 39 from nemo.collections.common.parts.rnn import label_collate 40 from nemo.core.classes import Typing, typecheck 41 from nemo.core.neural_types import AcousticEncodedRepresentation, ElementType, HypothesisType, LengthsType, NeuralType 42 from nemo.utils import logging 43 44 45 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 46 47 if hasattr(logitlen, 'cpu'): 48 logitlen_cpu = logitlen.to('cpu') 49 else: 50 logitlen_cpu = logitlen 51 52 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 53 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 54 hyp.length = logitlen_cpu[idx] 55 56 if hyp.dec_state is not None: 57 hyp.dec_state = _states_to_device(hyp.dec_state) 58 59 return hypotheses 60 61 62 def _states_to_device(dec_state, device='cpu'): 63 if torch.is_tensor(dec_state): 64 dec_state = dec_state.to(device) 65 66 elif isinstance(dec_state, (list, tuple)): 67 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 68 69 return dec_state 70 71 72 class _GreedyRNNTInfer(Typing, ConfidenceMethodMixin): 73 """A greedy transducer decoder. 74 75 Provides a common abstraction for sample level and batch level greedy decoding. 76 77 Args: 78 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 79 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 80 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 81 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 82 to a sequence in a single time step; if set to None then there is 83 no limit. 84 preserve_alignments: Bool flag which preserves the history of alignments generated during 85 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 86 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 87 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 88 89 The length of the list corresponds to the Acoustic Length (T). 90 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 91 U is the number of target tokens for the current timestep Ti. 92 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 93 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 94 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 95 96 The length of the list corresponds to the Acoustic Length (T). 97 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 98 U is the number of target tokens for the current timestep Ti. 99 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 100 confidence scores. 101 102 name: The method name (str). 103 Supported values: 104 - 'max_prob' for using the maximum token probability as a confidence. 105 - 'entropy' for using a normalized entropy of a log-likelihood vector. 106 107 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 108 Supported values: 109 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 110 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 111 Note that for this entropy, the alpha should comply the following inequality: 112 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 113 where V is the model vocabulary size. 114 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 115 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 116 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 117 More: https://en.wikipedia.org/wiki/Tsallis_entropy 118 - 'renyi' for the Rรฉnyi entropy. 119 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 120 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 121 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 122 123 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 124 When the alpha equals one, scaling is not applied to 'max_prob', 125 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 126 127 entropy_norm: A mapping of the entropy value to the interval [0,1]. 128 Supported values: 129 - 'lin' for using the linear mapping. 130 - 'exp' for using exponential mapping with linear shift. 131 """ 132 133 @property 134 def input_types(self): 135 """Returns definitions of module input ports. 136 """ 137 return { 138 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), 139 "encoded_lengths": NeuralType(tuple('B'), LengthsType()), 140 "partial_hypotheses": [NeuralType(elements_type=HypothesisType(), optional=True)], # must always be last 141 } 142 143 @property 144 def output_types(self): 145 """Returns definitions of module output ports. 146 """ 147 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 148 149 def __init__( 150 self, 151 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 152 joint_model: rnnt_abstract.AbstractRNNTJoint, 153 blank_index: int, 154 max_symbols_per_step: Optional[int] = None, 155 preserve_alignments: bool = False, 156 preserve_frame_confidence: bool = False, 157 confidence_method_cfg: Optional[DictConfig] = None, 158 ): 159 super().__init__() 160 self.decoder = decoder_model 161 self.joint = joint_model 162 163 self._blank_index = blank_index 164 self._SOS = blank_index # Start of single index 165 self.max_symbols = max_symbols_per_step 166 self.preserve_alignments = preserve_alignments 167 self.preserve_frame_confidence = preserve_frame_confidence 168 169 # set confidence calculation method 170 self._init_confidence_method(confidence_method_cfg) 171 172 def __call__(self, *args, **kwargs): 173 return self.forward(*args, **kwargs) 174 175 @torch.no_grad() 176 def _pred_step( 177 self, 178 label: Union[torch.Tensor, int], 179 hidden: Optional[torch.Tensor], 180 add_sos: bool = False, 181 batch_size: Optional[int] = None, 182 ) -> Tuple[torch.Tensor, torch.Tensor]: 183 """ 184 Common prediction step based on the AbstractRNNTDecoder implementation. 185 186 Args: 187 label: (int/torch.Tensor): Label or "Start-of-Signal" token. 188 hidden: (Optional torch.Tensor): RNN State vector 189 add_sos (bool): Whether to add a zero vector at the begging as "start of sentence" token. 190 batch_size: Batch size of the output tensor. 191 192 Returns: 193 g: (B, U, H) if add_sos is false, else (B, U + 1, H) 194 hid: (h, c) where h is the final sequence hidden state and c is 195 the final cell state: 196 h (tensor), shape (L, B, H) 197 c (tensor), shape (L, B, H) 198 """ 199 if isinstance(label, torch.Tensor): 200 # label: [batch, 1] 201 if label.dtype != torch.long: 202 label = label.long() 203 204 else: 205 # Label is an integer 206 if label == self._SOS: 207 return self.decoder.predict(None, hidden, add_sos=add_sos, batch_size=batch_size) 208 209 label = label_collate([[label]]) 210 211 # output: [B, 1, K] 212 return self.decoder.predict(label, hidden, add_sos=add_sos, batch_size=batch_size) 213 214 def _joint_step(self, enc, pred, log_normalize: Optional[bool] = None): 215 """ 216 Common joint step based on AbstractRNNTJoint implementation. 217 218 Args: 219 enc: Output of the Encoder model. A torch.Tensor of shape [B, 1, H1] 220 pred: Output of the Decoder model. A torch.Tensor of shape [B, 1, H2] 221 log_normalize: Whether to log normalize or not. None will log normalize only for CPU. 222 223 Returns: 224 logits of shape (B, T=1, U=1, V + 1) 225 """ 226 with torch.no_grad(): 227 logits = self.joint.joint(enc, pred) 228 229 if log_normalize is None: 230 if not logits.is_cuda: # Use log softmax only if on CPU 231 logits = logits.log_softmax(dim=len(logits.shape) - 1) 232 else: 233 if log_normalize: 234 logits = logits.log_softmax(dim=len(logits.shape) - 1) 235 236 return logits 237 238 239 class GreedyRNNTInfer(_GreedyRNNTInfer): 240 """A greedy transducer decoder. 241 242 Sequence level greedy decoding, performed auto-regressively. 243 244 Args: 245 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 246 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 247 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 248 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 249 to a sequence in a single time step; if set to None then there is 250 no limit. 251 preserve_alignments: Bool flag which preserves the history of alignments generated during 252 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 253 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 254 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 255 256 The length of the list corresponds to the Acoustic Length (T). 257 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 258 U is the number of target tokens for the current timestep Ti. 259 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 260 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 261 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 262 263 The length of the list corresponds to the Acoustic Length (T). 264 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 265 U is the number of target tokens for the current timestep Ti. 266 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 267 confidence scores. 268 269 name: The method name (str). 270 Supported values: 271 - 'max_prob' for using the maximum token probability as a confidence. 272 - 'entropy' for using a normalized entropy of a log-likelihood vector. 273 274 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 275 Supported values: 276 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 277 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 278 Note that for this entropy, the alpha should comply the following inequality: 279 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 280 where V is the model vocabulary size. 281 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 282 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/Tsallis_entropy 285 - 'renyi' for the Rรฉnyi entropy. 286 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 287 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 288 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 289 290 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 291 When the alpha equals one, scaling is not applied to 'max_prob', 292 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 293 294 entropy_norm: A mapping of the entropy value to the interval [0,1]. 295 Supported values: 296 - 'lin' for using the linear mapping. 297 - 'exp' for using exponential mapping with linear shift. 298 """ 299 300 def __init__( 301 self, 302 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 303 joint_model: rnnt_abstract.AbstractRNNTJoint, 304 blank_index: int, 305 max_symbols_per_step: Optional[int] = None, 306 preserve_alignments: bool = False, 307 preserve_frame_confidence: bool = False, 308 confidence_method_cfg: Optional[DictConfig] = None, 309 ): 310 super().__init__( 311 decoder_model=decoder_model, 312 joint_model=joint_model, 313 blank_index=blank_index, 314 max_symbols_per_step=max_symbols_per_step, 315 preserve_alignments=preserve_alignments, 316 preserve_frame_confidence=preserve_frame_confidence, 317 confidence_method_cfg=confidence_method_cfg, 318 ) 319 320 @typecheck() 321 def forward( 322 self, 323 encoder_output: torch.Tensor, 324 encoded_lengths: torch.Tensor, 325 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 326 ): 327 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 328 Output token is generated auto-regressively. 329 330 Args: 331 encoder_output: A tensor of size (batch, features, timesteps). 332 encoded_lengths: list of int representing the length of each sequence 333 output sequence. 334 335 Returns: 336 packed list containing batch number of sentences (Hypotheses). 337 """ 338 # Preserve decoder and joint training state 339 decoder_training_state = self.decoder.training 340 joint_training_state = self.joint.training 341 342 with torch.inference_mode(): 343 # Apply optional preprocessing 344 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 345 346 self.decoder.eval() 347 self.joint.eval() 348 349 hypotheses = [] 350 # Process each sequence independently 351 with self.decoder.as_frozen(), self.joint.as_frozen(): 352 for batch_idx in range(encoder_output.size(0)): 353 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 354 logitlen = encoded_lengths[batch_idx] 355 356 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 357 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 358 hypotheses.append(hypothesis) 359 360 # Pack results into Hypotheses 361 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 362 363 self.decoder.train(decoder_training_state) 364 self.joint.train(joint_training_state) 365 366 return (packed_result,) 367 368 @torch.no_grad() 369 def _greedy_decode( 370 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 371 ): 372 # x: [T, 1, D] 373 # out_len: [seq_len] 374 375 # Initialize blank state and empty label set in Hypothesis 376 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 377 378 if partial_hypotheses is not None: 379 hypothesis.last_token = partial_hypotheses.last_token 380 hypothesis.y_sequence = ( 381 partial_hypotheses.y_sequence.cpu().tolist() 382 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 383 else partial_hypotheses.y_sequence 384 ) 385 if partial_hypotheses.dec_state is not None: 386 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 387 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 388 389 if self.preserve_alignments: 390 # Alignments is a 2-dimensional dangling list representing T x U 391 hypothesis.alignments = [[]] 392 393 if self.preserve_frame_confidence: 394 hypothesis.frame_confidence = [[]] 395 396 # For timestep t in X_t 397 for time_idx in range(out_len): 398 # Extract encoder embedding at timestep t 399 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 400 f = x.narrow(dim=0, start=time_idx, length=1) 401 402 # Setup exit flags and counter 403 not_blank = True 404 symbols_added = 0 405 # While blank is not predicted, or we dont run out of max symbols per timestep 406 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 407 # In the first timestep, we initialize the network with RNNT Blank 408 # In later timesteps, we provide previous predicted label as input. 409 if hypothesis.last_token is None and hypothesis.dec_state is None: 410 last_label = self._SOS 411 else: 412 last_label = label_collate([[hypothesis.last_token]]) 413 414 # Perform prediction network and joint network steps. 415 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 416 # If preserving per-frame confidence, log_normalize must be true 417 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 418 0, 0, 0, : 419 ] 420 421 del g 422 423 # torch.max(0) op doesnt exist for FP 16. 424 if logp.dtype != torch.float32: 425 logp = logp.float() 426 427 # get index k, of max prob 428 v, k = logp.max(0) 429 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 430 431 if self.preserve_alignments: 432 # insert logprobs into last timestep 433 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 434 435 if self.preserve_frame_confidence: 436 # insert confidence into last timestep 437 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 438 439 del logp 440 441 # If blank token is predicted, exit inner loop, move onto next timestep t 442 if k == self._blank_index: 443 not_blank = False 444 else: 445 # Append token to label set, update RNN state. 446 hypothesis.y_sequence.append(k) 447 hypothesis.score += float(v) 448 hypothesis.timestep.append(time_idx) 449 hypothesis.dec_state = hidden_prime 450 hypothesis.last_token = k 451 452 # Increment token counter. 453 symbols_added += 1 454 455 if self.preserve_alignments: 456 # convert Ti-th logits into a torch array 457 hypothesis.alignments.append([]) # blank buffer for next timestep 458 459 if self.preserve_frame_confidence: 460 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 461 462 # Remove trailing empty list of Alignments 463 if self.preserve_alignments: 464 if len(hypothesis.alignments[-1]) == 0: 465 del hypothesis.alignments[-1] 466 467 # Remove trailing empty list of per-frame confidence 468 if self.preserve_frame_confidence: 469 if len(hypothesis.frame_confidence[-1]) == 0: 470 del hypothesis.frame_confidence[-1] 471 472 # Unpack the hidden states 473 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 474 475 return hypothesis 476 477 478 class GreedyBatchedRNNTInfer(_GreedyRNNTInfer): 479 """A batch level greedy transducer decoder. 480 481 Batch level greedy decoding, performed auto-regressively. 482 483 Args: 484 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 485 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 486 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 487 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 488 to a sequence in a single time step; if set to None then there is 489 no limit. 490 preserve_alignments: Bool flag which preserves the history of alignments generated during 491 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 492 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 493 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 494 495 The length of the list corresponds to the Acoustic Length (T). 496 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 497 U is the number of target tokens for the current timestep Ti. 498 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 499 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 500 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 501 502 The length of the list corresponds to the Acoustic Length (T). 503 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 504 U is the number of target tokens for the current timestep Ti. 505 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 506 confidence scores. 507 508 name: The method name (str). 509 Supported values: 510 - 'max_prob' for using the maximum token probability as a confidence. 511 - 'entropy' for using a normalized entropy of a log-likelihood vector. 512 513 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 514 Supported values: 515 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 516 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 517 Note that for this entropy, the alpha should comply the following inequality: 518 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 519 where V is the model vocabulary size. 520 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 521 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 522 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 523 More: https://en.wikipedia.org/wiki/Tsallis_entropy 524 - 'renyi' for the Rรฉnyi entropy. 525 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 526 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 527 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 528 529 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 530 When the alpha equals one, scaling is not applied to 'max_prob', 531 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 532 533 entropy_norm: A mapping of the entropy value to the interval [0,1]. 534 Supported values: 535 - 'lin' for using the linear mapping. 536 - 'exp' for using exponential mapping with linear shift. 537 """ 538 539 def __init__( 540 self, 541 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 542 joint_model: rnnt_abstract.AbstractRNNTJoint, 543 blank_index: int, 544 max_symbols_per_step: Optional[int] = None, 545 preserve_alignments: bool = False, 546 preserve_frame_confidence: bool = False, 547 confidence_method_cfg: Optional[DictConfig] = None, 548 ): 549 super().__init__( 550 decoder_model=decoder_model, 551 joint_model=joint_model, 552 blank_index=blank_index, 553 max_symbols_per_step=max_symbols_per_step, 554 preserve_alignments=preserve_alignments, 555 preserve_frame_confidence=preserve_frame_confidence, 556 confidence_method_cfg=confidence_method_cfg, 557 ) 558 559 # Depending on availability of `blank_as_pad` support 560 # switch between more efficient batch decoding technique 561 if self.decoder.blank_as_pad: 562 self._greedy_decode = self._greedy_decode_blank_as_pad 563 else: 564 self._greedy_decode = self._greedy_decode_masked 565 566 @typecheck() 567 def forward( 568 self, 569 encoder_output: torch.Tensor, 570 encoded_lengths: torch.Tensor, 571 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 572 ): 573 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 574 Output token is generated auto-regressively. 575 576 Args: 577 encoder_output: A tensor of size (batch, features, timesteps). 578 encoded_lengths: list of int representing the length of each sequence 579 output sequence. 580 581 Returns: 582 packed list containing batch number of sentences (Hypotheses). 583 """ 584 # Preserve decoder and joint training state 585 decoder_training_state = self.decoder.training 586 joint_training_state = self.joint.training 587 588 with torch.inference_mode(): 589 # Apply optional preprocessing 590 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 591 logitlen = encoded_lengths 592 593 self.decoder.eval() 594 self.joint.eval() 595 596 with self.decoder.as_frozen(), self.joint.as_frozen(): 597 inseq = encoder_output # [B, T, D] 598 hypotheses = self._greedy_decode( 599 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 600 ) 601 602 # Pack the hypotheses results 603 packed_result = pack_hypotheses(hypotheses, logitlen) 604 605 self.decoder.train(decoder_training_state) 606 self.joint.train(joint_training_state) 607 608 return (packed_result,) 609 610 def _greedy_decode_blank_as_pad( 611 self, 612 x: torch.Tensor, 613 out_len: torch.Tensor, 614 device: torch.device, 615 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 616 ): 617 if partial_hypotheses is not None: 618 raise NotImplementedError("`partial_hypotheses` support is not supported") 619 620 with torch.inference_mode(): 621 # x: [B, T, D] 622 # out_len: [B] 623 # device: torch.device 624 625 # Initialize list of Hypothesis 626 batchsize = x.shape[0] 627 hypotheses = [ 628 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 629 ] 630 631 # Initialize Hidden state matrix (shared by entire batch) 632 hidden = None 633 634 # If alignments need to be preserved, register a dangling list to hold the values 635 if self.preserve_alignments: 636 # alignments is a 3-dimensional dangling list representing B x T x U 637 for hyp in hypotheses: 638 hyp.alignments = [[]] 639 640 # If confidence scores need to be preserved, register a dangling list to hold the values 641 if self.preserve_frame_confidence: 642 # frame_confidence is a 3-dimensional dangling list representing B x T x U 643 for hyp in hypotheses: 644 hyp.frame_confidence = [[]] 645 646 # Last Label buffer + Last Label without blank buffer 647 # batch level equivalent of the last_label 648 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 649 650 # Mask buffers 651 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 652 653 # Get max sequence length 654 max_out_len = out_len.max() 655 for time_idx in range(max_out_len): 656 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 657 658 # Prepare t timestamp batch variables 659 not_blank = True 660 symbols_added = 0 661 662 # Reset blank mask 663 blank_mask.mul_(False) 664 665 # Update blank mask with time mask 666 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 667 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 668 blank_mask = time_idx >= out_len 669 # Start inner loop 670 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 671 # Batch prediction and joint network steps 672 # If very first prediction step, submit SOS tag (blank) to pred_step. 673 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 674 if time_idx == 0 and symbols_added == 0 and hidden is None: 675 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 676 else: 677 # Perform batch step prediction of decoder, getting new states and scores ("g") 678 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 679 680 # Batched joint step - Output = [B, V + 1] 681 # If preserving per-frame confidence, log_normalize must be true 682 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 683 :, 0, 0, : 684 ] 685 686 if logp.dtype != torch.float32: 687 logp = logp.float() 688 689 # Get index k, of max prob for batch 690 v, k = logp.max(1) 691 del g 692 693 # Update blank mask with current predicted blanks 694 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 695 k_is_blank = k == self._blank_index 696 blank_mask.bitwise_or_(k_is_blank) 697 all_blanks = torch.all(blank_mask) 698 699 del k_is_blank 700 701 # If preserving alignments, check if sequence length of sample has been reached 702 # before adding alignment 703 if self.preserve_alignments: 704 # Insert logprobs into last timestep per sample 705 logp_vals = logp.to('cpu') 706 logp_ids = logp_vals.max(1)[1] 707 for batch_idx, is_blank in enumerate(blank_mask): 708 # we only want to update non-blanks, unless we are at the last step in the loop where 709 # all elements produced blanks, otherwise there will be duplicate predictions 710 # saved in alignments 711 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 712 hypotheses[batch_idx].alignments[-1].append( 713 (logp_vals[batch_idx], logp_ids[batch_idx]) 714 ) 715 del logp_vals 716 717 # If preserving per-frame confidence, check if sequence length of sample has been reached 718 # before adding confidence scores 719 if self.preserve_frame_confidence: 720 # Insert probabilities into last timestep per sample 721 confidence = self._get_confidence(logp) 722 for batch_idx, is_blank in enumerate(blank_mask): 723 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 724 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 725 del logp 726 727 # If all samples predict / have predicted prior blanks, exit loop early 728 # This is equivalent to if single sample predicted k 729 if all_blanks: 730 not_blank = False 731 else: 732 # Collect batch indices where blanks occurred now/past 733 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 734 735 # Recover prior state for all samples which predicted blank now/past 736 if hidden is not None: 737 # LSTM has 2 states 738 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 739 740 elif len(blank_indices) > 0 and hidden is None: 741 # Reset state if there were some blank and other non-blank predictions in batch 742 # Original state is filled with zeros so we just multiply 743 # LSTM has 2 states 744 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 745 746 # Recover prior predicted label for all samples which predicted blank now/past 747 k[blank_indices] = last_label[blank_indices, 0] 748 749 # Update new label and hidden state for next iteration 750 last_label = k.clone().view(-1, 1) 751 hidden = hidden_prime 752 753 # Update predicted labels, accounting for time mask 754 # If blank was predicted even once, now or in the past, 755 # Force the current predicted label to also be blank 756 # This ensures that blanks propogate across all timesteps 757 # once they have occured (normally stopping condition of sample level loop). 758 for kidx, ki in enumerate(k): 759 if blank_mask[kidx] == 0: 760 hypotheses[kidx].y_sequence.append(ki) 761 hypotheses[kidx].timestep.append(time_idx) 762 hypotheses[kidx].score += float(v[kidx]) 763 symbols_added += 1 764 765 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 766 # Then preserve U at current timestep Ti 767 # Finally, forward the timestep history to Ti+1 for that sample 768 # All of this should only be done iff the current time index <= sample-level AM length. 769 # Otherwise ignore and move to next sample / next timestep. 770 if self.preserve_alignments: 771 772 # convert Ti-th logits into a torch array 773 for batch_idx in range(batchsize): 774 775 # this checks if current timestep <= sample-level AM length 776 # If current timestep > sample-level AM length, no alignments will be added 777 # Therefore the list of Uj alignments is empty here. 778 if len(hypotheses[batch_idx].alignments[-1]) > 0: 779 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 780 781 # Do the same if preserving per-frame confidence 782 if self.preserve_frame_confidence: 783 784 for batch_idx in range(batchsize): 785 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 786 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 787 788 # Remove trailing empty list of alignments at T_{am-len} x Uj 789 if self.preserve_alignments: 790 for batch_idx in range(batchsize): 791 if len(hypotheses[batch_idx].alignments[-1]) == 0: 792 del hypotheses[batch_idx].alignments[-1] 793 794 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 795 if self.preserve_frame_confidence: 796 for batch_idx in range(batchsize): 797 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 798 del hypotheses[batch_idx].frame_confidence[-1] 799 800 # Preserve states 801 for batch_idx in range(batchsize): 802 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 803 804 return hypotheses 805 806 def _greedy_decode_masked( 807 self, 808 x: torch.Tensor, 809 out_len: torch.Tensor, 810 device: torch.device, 811 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 812 ): 813 if partial_hypotheses is not None: 814 raise NotImplementedError("`partial_hypotheses` support is not supported") 815 816 # x: [B, T, D] 817 # out_len: [B] 818 # device: torch.device 819 820 # Initialize state 821 batchsize = x.shape[0] 822 hypotheses = [ 823 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 824 ] 825 826 # Initialize Hidden state matrix (shared by entire batch) 827 hidden = None 828 829 # If alignments need to be preserved, register a danling list to hold the values 830 if self.preserve_alignments: 831 # alignments is a 3-dimensional dangling list representing B x T x U 832 for hyp in hypotheses: 833 hyp.alignments = [[]] 834 else: 835 alignments = None 836 837 # If confidence scores need to be preserved, register a danling list to hold the values 838 if self.preserve_frame_confidence: 839 # frame_confidence is a 3-dimensional dangling list representing B x T x U 840 for hyp in hypotheses: 841 hyp.frame_confidence = [[]] 842 843 # Last Label buffer + Last Label without blank buffer 844 # batch level equivalent of the last_label 845 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 846 last_label_without_blank = last_label.clone() 847 848 # Mask buffers 849 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 850 851 # Get max sequence length 852 max_out_len = out_len.max() 853 854 with torch.inference_mode(): 855 for time_idx in range(max_out_len): 856 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 857 858 # Prepare t timestamp batch variables 859 not_blank = True 860 symbols_added = 0 861 862 # Reset blank mask 863 blank_mask.mul_(False) 864 865 # Update blank mask with time mask 866 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 867 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 868 blank_mask = time_idx >= out_len 869 870 # Start inner loop 871 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 872 # Batch prediction and joint network steps 873 # If very first prediction step, submit SOS tag (blank) to pred_step. 874 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 875 if time_idx == 0 and symbols_added == 0 and hidden is None: 876 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 877 else: 878 # Set a dummy label for the blank value 879 # This value will be overwritten by "blank" again the last label update below 880 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 881 last_label_without_blank_mask = last_label == self._blank_index 882 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 883 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 884 ~last_label_without_blank_mask 885 ] 886 887 # Perform batch step prediction of decoder, getting new states and scores ("g") 888 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 889 890 # Batched joint step - Output = [B, V + 1] 891 # If preserving per-frame confidence, log_normalize must be true 892 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 893 :, 0, 0, : 894 ] 895 896 if logp.dtype != torch.float32: 897 logp = logp.float() 898 899 # Get index k, of max prob for batch 900 v, k = logp.max(1) 901 del g 902 903 # Update blank mask with current predicted blanks 904 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 905 k_is_blank = k == self._blank_index 906 blank_mask.bitwise_or_(k_is_blank) 907 all_blanks = torch.all(blank_mask) 908 909 # If preserving alignments, check if sequence length of sample has been reached 910 # before adding alignment 911 if self.preserve_alignments: 912 # Insert logprobs into last timestep per sample 913 logp_vals = logp.to('cpu') 914 logp_ids = logp_vals.max(1)[1] 915 for batch_idx, is_blank in enumerate(blank_mask): 916 # we only want to update non-blanks, unless we are at the last step in the loop where 917 # all elements produced blanks, otherwise there will be duplicate predictions 918 # saved in alignments 919 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 920 hypotheses[batch_idx].alignments[-1].append( 921 (logp_vals[batch_idx], logp_ids[batch_idx]) 922 ) 923 924 del logp_vals 925 926 # If preserving per-frame confidence, check if sequence length of sample has been reached 927 # before adding confidence scores 928 if self.preserve_frame_confidence: 929 # Insert probabilities into last timestep per sample 930 confidence = self._get_confidence(logp) 931 for batch_idx, is_blank in enumerate(blank_mask): 932 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 933 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 934 del logp 935 936 # If all samples predict / have predicted prior blanks, exit loop early 937 # This is equivalent to if single sample predicted k 938 if blank_mask.all(): 939 not_blank = False 940 else: 941 # Collect batch indices where blanks occurred now/past 942 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 943 944 # Recover prior state for all samples which predicted blank now/past 945 if hidden is not None: 946 # LSTM has 2 states 947 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 948 949 elif len(blank_indices) > 0 and hidden is None: 950 # Reset state if there were some blank and other non-blank predictions in batch 951 # Original state is filled with zeros so we just multiply 952 # LSTM has 2 states 953 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 954 955 # Recover prior predicted label for all samples which predicted blank now/past 956 k[blank_indices] = last_label[blank_indices, 0] 957 958 # Update new label and hidden state for next iteration 959 last_label = k.view(-1, 1) 960 hidden = hidden_prime 961 962 # Update predicted labels, accounting for time mask 963 # If blank was predicted even once, now or in the past, 964 # Force the current predicted label to also be blank 965 # This ensures that blanks propogate across all timesteps 966 # once they have occured (normally stopping condition of sample level loop). 967 for kidx, ki in enumerate(k): 968 if blank_mask[kidx] == 0: 969 hypotheses[kidx].y_sequence.append(ki) 970 hypotheses[kidx].timestep.append(time_idx) 971 hypotheses[kidx].score += float(v[kidx]) 972 973 symbols_added += 1 974 975 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 976 # Then preserve U at current timestep Ti 977 # Finally, forward the timestep history to Ti+1 for that sample 978 # All of this should only be done iff the current time index <= sample-level AM length. 979 # Otherwise ignore and move to next sample / next timestep. 980 if self.preserve_alignments: 981 982 # convert Ti-th logits into a torch array 983 for batch_idx in range(batchsize): 984 985 # this checks if current timestep <= sample-level AM length 986 # If current timestep > sample-level AM length, no alignments will be added 987 # Therefore the list of Uj alignments is empty here. 988 if len(hypotheses[batch_idx].alignments[-1]) > 0: 989 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 990 991 # Do the same if preserving per-frame confidence 992 if self.preserve_frame_confidence: 993 994 for batch_idx in range(batchsize): 995 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 996 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 997 998 # Remove trailing empty list of alignments at T_{am-len} x Uj 999 if self.preserve_alignments: 1000 for batch_idx in range(batchsize): 1001 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1002 del hypotheses[batch_idx].alignments[-1] 1003 1004 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1005 if self.preserve_frame_confidence: 1006 for batch_idx in range(batchsize): 1007 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1008 del hypotheses[batch_idx].frame_confidence[-1] 1009 1010 # Preserve states 1011 for batch_idx in range(batchsize): 1012 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1013 1014 return hypotheses 1015 1016 1017 class ExportedModelGreedyBatchedRNNTInfer: 1018 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = None): 1019 self.encoder_model_path = encoder_model 1020 self.decoder_joint_model_path = decoder_joint_model 1021 self.max_symbols_per_step = max_symbols_per_step 1022 1023 # Will be populated at runtime 1024 self._blank_index = None 1025 1026 def __call__(self, audio_signal: torch.Tensor, length: torch.Tensor): 1027 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 1028 Output token is generated auto-regressively. 1029 1030 Args: 1031 encoder_output: A tensor of size (batch, features, timesteps). 1032 encoded_lengths: list of int representing the length of each sequence 1033 output sequence. 1034 1035 Returns: 1036 packed list containing batch number of sentences (Hypotheses). 1037 """ 1038 with torch.no_grad(): 1039 # Apply optional preprocessing 1040 encoder_output, encoded_lengths = self.run_encoder(audio_signal=audio_signal, length=length) 1041 1042 if torch.is_tensor(encoder_output): 1043 encoder_output = encoder_output.transpose(1, 2) 1044 else: 1045 encoder_output = encoder_output.transpose([0, 2, 1]) # (B, T, D) 1046 logitlen = encoded_lengths 1047 1048 inseq = encoder_output # [B, T, D] 1049 hypotheses, timestamps = self._greedy_decode(inseq, logitlen) 1050 1051 # Pack the hypotheses results 1052 packed_result = [rnnt_utils.Hypothesis(score=-1.0, y_sequence=[]) for _ in range(len(hypotheses))] 1053 for i in range(len(packed_result)): 1054 packed_result[i].y_sequence = torch.tensor(hypotheses[i], dtype=torch.long) 1055 packed_result[i].length = timestamps[i] 1056 1057 del hypotheses 1058 1059 return packed_result 1060 1061 def _greedy_decode(self, x, out_len): 1062 # x: [B, T, D] 1063 # out_len: [B] 1064 1065 # Initialize state 1066 batchsize = x.shape[0] 1067 hidden = self._get_initial_states(batchsize) 1068 target_lengths = torch.ones(batchsize, dtype=torch.int32) 1069 1070 # Output string buffer 1071 label = [[] for _ in range(batchsize)] 1072 timesteps = [[] for _ in range(batchsize)] 1073 1074 # Last Label buffer + Last Label without blank buffer 1075 # batch level equivalent of the last_label 1076 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long).numpy() 1077 if torch.is_tensor(x): 1078 last_label = torch.from_numpy(last_label).to(self.device) 1079 1080 # Mask buffers 1081 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool).numpy() 1082 1083 # Get max sequence length 1084 max_out_len = out_len.max() 1085 for time_idx in range(max_out_len): 1086 f = x[:, time_idx : time_idx + 1, :] # [B, 1, D] 1087 1088 if torch.is_tensor(f): 1089 f = f.transpose(1, 2) 1090 else: 1091 f = f.transpose([0, 2, 1]) 1092 1093 # Prepare t timestamp batch variables 1094 not_blank = True 1095 symbols_added = 0 1096 1097 # Reset blank mask 1098 blank_mask *= False 1099 1100 # Update blank mask with time mask 1101 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1102 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1103 blank_mask = time_idx >= out_len 1104 # Start inner loop 1105 while not_blank and (self.max_symbols_per_step is None or symbols_added < self.max_symbols_per_step): 1106 1107 # Batch prediction and joint network steps 1108 # If very first prediction step, submit SOS tag (blank) to pred_step. 1109 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1110 if time_idx == 0 and symbols_added == 0: 1111 g = torch.tensor([self._blank_index] * batchsize, dtype=torch.int32).view(-1, 1) 1112 else: 1113 if torch.is_tensor(last_label): 1114 g = last_label.type(torch.int32) 1115 else: 1116 g = last_label.astype(np.int32) 1117 1118 # Batched joint step - Output = [B, V + 1] 1119 joint_out, hidden_prime = self.run_decoder_joint(f, g, target_lengths, *hidden) 1120 logp, pred_lengths = joint_out 1121 logp = logp[:, 0, 0, :] 1122 1123 # Get index k, of max prob for batch 1124 if torch.is_tensor(logp): 1125 v, k = logp.max(1) 1126 else: 1127 k = np.argmax(logp, axis=1).astype(np.int32) 1128 1129 # Update blank mask with current predicted blanks 1130 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1131 k_is_blank = k == self._blank_index 1132 blank_mask |= k_is_blank 1133 1134 del k_is_blank 1135 del logp 1136 1137 # If all samples predict / have predicted prior blanks, exit loop early 1138 # This is equivalent to if single sample predicted k 1139 if blank_mask.all(): 1140 not_blank = False 1141 1142 else: 1143 # Collect batch indices where blanks occurred now/past 1144 if torch.is_tensor(blank_mask): 1145 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1146 else: 1147 blank_indices = blank_mask.astype(np.int32).nonzero() 1148 1149 if type(blank_indices) in (list, tuple): 1150 blank_indices = blank_indices[0] 1151 1152 # Recover prior state for all samples which predicted blank now/past 1153 if hidden is not None: 1154 # LSTM has 2 states 1155 for state_id in range(len(hidden)): 1156 hidden_prime[state_id][:, blank_indices, :] = hidden[state_id][:, blank_indices, :] 1157 1158 elif len(blank_indices) > 0 and hidden is None: 1159 # Reset state if there were some blank and other non-blank predictions in batch 1160 # Original state is filled with zeros so we just multiply 1161 # LSTM has 2 states 1162 for state_id in range(len(hidden_prime)): 1163 hidden_prime[state_id][:, blank_indices, :] *= 0.0 1164 1165 # Recover prior predicted label for all samples which predicted blank now/past 1166 k[blank_indices] = last_label[blank_indices, 0] 1167 1168 # Update new label and hidden state for next iteration 1169 if torch.is_tensor(k): 1170 last_label = k.clone().reshape(-1, 1) 1171 else: 1172 last_label = k.copy().reshape(-1, 1) 1173 hidden = hidden_prime 1174 1175 # Update predicted labels, accounting for time mask 1176 # If blank was predicted even once, now or in the past, 1177 # Force the current predicted label to also be blank 1178 # This ensures that blanks propogate across all timesteps 1179 # once they have occured (normally stopping condition of sample level loop). 1180 for kidx, ki in enumerate(k): 1181 if blank_mask[kidx] == 0: 1182 label[kidx].append(ki) 1183 timesteps[kidx].append(time_idx) 1184 1185 symbols_added += 1 1186 1187 return label, timesteps 1188 1189 def _setup_blank_index(self): 1190 raise NotImplementedError() 1191 1192 def run_encoder(self, audio_signal, length): 1193 raise NotImplementedError() 1194 1195 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1196 raise NotImplementedError() 1197 1198 def _get_initial_states(self, batchsize): 1199 raise NotImplementedError() 1200 1201 1202 class ONNXGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1203 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = 10): 1204 super().__init__( 1205 encoder_model=encoder_model, 1206 decoder_joint_model=decoder_joint_model, 1207 max_symbols_per_step=max_symbols_per_step, 1208 ) 1209 1210 try: 1211 import onnx 1212 import onnxruntime 1213 except (ModuleNotFoundError, ImportError): 1214 raise ImportError(f"`onnx` or `onnxruntime` could not be imported, please install the libraries.\n") 1215 1216 if torch.cuda.is_available(): 1217 # Try to use onnxruntime-gpu 1218 providers = ['TensorrtExecutionProvider', 'CUDAExecutionProvider'] 1219 else: 1220 # Fall back to CPU and onnxruntime-cpu 1221 providers = ['CPUExecutionProvider'] 1222 1223 onnx_session_opt = onnxruntime.SessionOptions() 1224 onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL 1225 1226 onnx_model = onnx.load(self.encoder_model_path) 1227 onnx.checker.check_model(onnx_model, full_check=True) 1228 self.encoder_model = onnx_model 1229 self.encoder = onnxruntime.InferenceSession( 1230 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1231 ) 1232 1233 onnx_model = onnx.load(self.decoder_joint_model_path) 1234 onnx.checker.check_model(onnx_model, full_check=True) 1235 self.decoder_joint_model = onnx_model 1236 self.decoder_joint = onnxruntime.InferenceSession( 1237 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1238 ) 1239 1240 logging.info("Successfully loaded encoder, decoder and joint onnx models !") 1241 1242 # Will be populated at runtime 1243 self._blank_index = None 1244 self.max_symbols_per_step = max_symbols_per_step 1245 1246 self._setup_encoder_input_output_keys() 1247 self._setup_decoder_joint_input_output_keys() 1248 self._setup_blank_index() 1249 1250 def _setup_encoder_input_output_keys(self): 1251 self.encoder_inputs = list(self.encoder_model.graph.input) 1252 self.encoder_outputs = list(self.encoder_model.graph.output) 1253 1254 def _setup_decoder_joint_input_output_keys(self): 1255 self.decoder_joint_inputs = list(self.decoder_joint_model.graph.input) 1256 self.decoder_joint_outputs = list(self.decoder_joint_model.graph.output) 1257 1258 def _setup_blank_index(self): 1259 # ASSUME: Single input with no time length information 1260 dynamic_dim = 257 1261 shapes = self.encoder_inputs[0].type.tensor_type.shape.dim 1262 ip_shape = [] 1263 for shape in shapes: 1264 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1265 ip_shape.append(dynamic_dim) # replace dynamic axes with constant 1266 else: 1267 ip_shape.append(int(shape.dim_value)) 1268 1269 enc_logits, encoded_length = self.run_encoder( 1270 audio_signal=torch.randn(*ip_shape), length=torch.randint(0, 1, size=(dynamic_dim,)) 1271 ) 1272 1273 # prepare states 1274 states = self._get_initial_states(batchsize=dynamic_dim) 1275 1276 # run decoder 1 step 1277 joint_out, states = self.run_decoder_joint(enc_logits, None, None, *states) 1278 log_probs, lengths = joint_out 1279 1280 self._blank_index = log_probs.shape[-1] - 1 # last token of vocab size is blank token 1281 logging.info( 1282 f"Enc-Dec-Joint step was evaluated, blank token id = {self._blank_index}; vocab size = {log_probs.shape[-1]}" 1283 ) 1284 1285 def run_encoder(self, audio_signal, length): 1286 if hasattr(audio_signal, 'cpu'): 1287 audio_signal = audio_signal.cpu().numpy() 1288 1289 if hasattr(length, 'cpu'): 1290 length = length.cpu().numpy() 1291 1292 ip = { 1293 self.encoder_inputs[0].name: audio_signal, 1294 self.encoder_inputs[1].name: length, 1295 } 1296 enc_out = self.encoder.run(None, ip) 1297 enc_out, encoded_length = enc_out # ASSUME: single output 1298 return enc_out, encoded_length 1299 1300 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1301 # ASSUME: Decoder is RNN Transducer 1302 if targets is None: 1303 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32) 1304 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32) 1305 1306 if hasattr(targets, 'cpu'): 1307 targets = targets.cpu().numpy() 1308 1309 if hasattr(target_length, 'cpu'): 1310 target_length = target_length.cpu().numpy() 1311 1312 ip = { 1313 self.decoder_joint_inputs[0].name: enc_logits, 1314 self.decoder_joint_inputs[1].name: targets, 1315 self.decoder_joint_inputs[2].name: target_length, 1316 } 1317 1318 num_states = 0 1319 if states is not None and len(states) > 0: 1320 num_states = len(states) 1321 for idx, state in enumerate(states): 1322 if hasattr(state, 'cpu'): 1323 state = state.cpu().numpy() 1324 1325 ip[self.decoder_joint_inputs[len(ip)].name] = state 1326 1327 dec_out = self.decoder_joint.run(None, ip) 1328 1329 # unpack dec output 1330 if num_states > 0: 1331 new_states = dec_out[-num_states:] 1332 dec_out = dec_out[:-num_states] 1333 else: 1334 new_states = None 1335 1336 return dec_out, new_states 1337 1338 def _get_initial_states(self, batchsize): 1339 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1340 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1341 num_states = len(input_state_nodes) 1342 if num_states == 0: 1343 return 1344 1345 input_states = [] 1346 for state_id in range(num_states): 1347 node = input_state_nodes[state_id] 1348 ip_shape = [] 1349 for shape_idx, shape in enumerate(node.type.tensor_type.shape.dim): 1350 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1351 ip_shape.append(batchsize) # replace dynamic axes with constant 1352 else: 1353 ip_shape.append(int(shape.dim_value)) 1354 1355 input_states.append(torch.zeros(*ip_shape)) 1356 1357 return input_states 1358 1359 1360 class TorchscriptGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1361 def __init__( 1362 self, 1363 encoder_model: str, 1364 decoder_joint_model: str, 1365 cfg: DictConfig, 1366 device: str, 1367 max_symbols_per_step: Optional[int] = 10, 1368 ): 1369 super().__init__( 1370 encoder_model=encoder_model, 1371 decoder_joint_model=decoder_joint_model, 1372 max_symbols_per_step=max_symbols_per_step, 1373 ) 1374 1375 self.cfg = cfg 1376 self.device = device 1377 1378 self.encoder = torch.jit.load(self.encoder_model_path, map_location=self.device) 1379 self.decoder_joint = torch.jit.load(self.decoder_joint_model_path, map_location=self.device) 1380 1381 logging.info("Successfully loaded encoder, decoder and joint torchscript models !") 1382 1383 # Will be populated at runtime 1384 self._blank_index = None 1385 self.max_symbols_per_step = max_symbols_per_step 1386 1387 self._setup_encoder_input_keys() 1388 self._setup_decoder_joint_input_keys() 1389 self._setup_blank_index() 1390 1391 def _setup_encoder_input_keys(self): 1392 arguments = self.encoder.forward.schema.arguments[1:] 1393 self.encoder_inputs = [arg for arg in arguments] 1394 1395 def _setup_decoder_joint_input_keys(self): 1396 arguments = self.decoder_joint.forward.schema.arguments[1:] 1397 self.decoder_joint_inputs = [arg for arg in arguments] 1398 1399 def _setup_blank_index(self): 1400 self._blank_index = len(self.cfg.joint.vocabulary) 1401 1402 logging.info(f"Blank token id = {self._blank_index}; vocab size = {len(self.cfg.joint.vocabulary) + 1}") 1403 1404 def run_encoder(self, audio_signal, length): 1405 enc_out = self.encoder(audio_signal, length) 1406 enc_out, encoded_length = enc_out # ASSUME: single output 1407 return enc_out, encoded_length 1408 1409 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1410 # ASSUME: Decoder is RNN Transducer 1411 if targets is None: 1412 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32, device=enc_logits.device) 1413 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32, device=enc_logits.device) 1414 1415 num_states = 0 1416 if states is not None and len(states) > 0: 1417 num_states = len(states) 1418 1419 dec_out = self.decoder_joint(enc_logits, targets, target_length, *states) 1420 1421 # unpack dec output 1422 if num_states > 0: 1423 new_states = dec_out[-num_states:] 1424 dec_out = dec_out[:-num_states] 1425 else: 1426 new_states = None 1427 1428 return dec_out, new_states 1429 1430 def _get_initial_states(self, batchsize): 1431 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1432 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1433 num_states = len(input_state_nodes) 1434 if num_states == 0: 1435 return 1436 1437 input_states = [] 1438 for state_id in range(num_states): 1439 # Hardcode shape size for LSTM (1 is for num layers in LSTM, which is flattened for export) 1440 ip_shape = [1, batchsize, self.cfg.model_defaults.pred_hidden] 1441 input_states.append(torch.zeros(*ip_shape, device=self.device)) 1442 1443 return input_states 1444 1445 1446 class GreedyMultiblankRNNTInfer(GreedyRNNTInfer): 1447 """A greedy transducer decoder for multi-blank RNN-T. 1448 1449 Sequence level greedy decoding, performed auto-regressively. 1450 1451 Args: 1452 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1453 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1454 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1455 big_blank_durations: a list containing durations for big blanks the model supports. 1456 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1457 to a sequence in a single time step; if set to None then there is 1458 no limit. 1459 preserve_alignments: Bool flag which preserves the history of alignments generated during 1460 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1461 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1462 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1463 The length of the list corresponds to the Acoustic Length (T). 1464 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1465 U is the number of target tokens for the current timestep Ti. 1466 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1467 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1468 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1469 The length of the list corresponds to the Acoustic Length (T). 1470 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1471 U is the number of target tokens for the current timestep Ti. 1472 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1473 confidence scores. 1474 1475 name: The method name (str). 1476 Supported values: 1477 - 'max_prob' for using the maximum token probability as a confidence. 1478 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1479 1480 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 1481 Supported values: 1482 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1483 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1484 Note that for this entropy, the alpha should comply the following inequality: 1485 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1486 where V is the model vocabulary size. 1487 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1488 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1489 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1490 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1491 - 'renyi' for the Rรฉnyi entropy. 1492 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1493 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1494 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1495 1496 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1497 When the alpha equals one, scaling is not applied to 'max_prob', 1498 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1499 1500 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1501 Supported values: 1502 - 'lin' for using the linear mapping. 1503 - 'exp' for using exponential mapping with linear shift. 1504 """ 1505 1506 def __init__( 1507 self, 1508 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1509 joint_model: rnnt_abstract.AbstractRNNTJoint, 1510 blank_index: int, 1511 big_blank_durations: list, 1512 max_symbols_per_step: Optional[int] = None, 1513 preserve_alignments: bool = False, 1514 preserve_frame_confidence: bool = False, 1515 confidence_method_cfg: Optional[DictConfig] = None, 1516 ): 1517 super().__init__( 1518 decoder_model=decoder_model, 1519 joint_model=joint_model, 1520 blank_index=blank_index, 1521 max_symbols_per_step=max_symbols_per_step, 1522 preserve_alignments=preserve_alignments, 1523 preserve_frame_confidence=preserve_frame_confidence, 1524 confidence_method_cfg=confidence_method_cfg, 1525 ) 1526 self.big_blank_durations = big_blank_durations 1527 self._SOS = blank_index - len(big_blank_durations) 1528 1529 @torch.no_grad() 1530 def _greedy_decode( 1531 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 1532 ): 1533 # x: [T, 1, D] 1534 # out_len: [seq_len] 1535 1536 # Initialize blank state and empty label set in Hypothesis 1537 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 1538 1539 if partial_hypotheses is not None: 1540 hypothesis.last_token = partial_hypotheses.last_token 1541 hypothesis.y_sequence = ( 1542 partial_hypotheses.y_sequence.cpu().tolist() 1543 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 1544 else partial_hypotheses.y_sequence 1545 ) 1546 if partial_hypotheses.dec_state is not None: 1547 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 1548 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 1549 1550 if self.preserve_alignments: 1551 # Alignments is a 2-dimensional dangling list representing T x U 1552 hypothesis.alignments = [[]] 1553 1554 if self.preserve_frame_confidence: 1555 hypothesis.frame_confidence = [[]] 1556 1557 # if this variable is > 1, it means the last emission was a big-blank and we need to skip frames. 1558 big_blank_duration = 1 1559 1560 # For timestep t in X_t 1561 for time_idx in range(out_len): 1562 if big_blank_duration > 1: 1563 # skip frames until big_blank_duration == 1. 1564 big_blank_duration -= 1 1565 continue 1566 # Extract encoder embedding at timestep t 1567 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 1568 f = x.narrow(dim=0, start=time_idx, length=1) 1569 1570 # Setup exit flags and counter 1571 not_blank = True 1572 symbols_added = 0 1573 1574 # While blank is not predicted, or we dont run out of max symbols per timestep 1575 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1576 # In the first timestep, we initialize the network with RNNT Blank 1577 # In later timesteps, we provide previous predicted label as input. 1578 if hypothesis.last_token is None and hypothesis.dec_state is None: 1579 last_label = self._SOS 1580 else: 1581 last_label = label_collate([[hypothesis.last_token]]) 1582 1583 # Perform prediction network and joint network steps. 1584 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 1585 # If preserving per-frame confidence, log_normalize must be true 1586 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1587 0, 0, 0, : 1588 ] 1589 1590 del g 1591 1592 # torch.max(0) op doesnt exist for FP 16. 1593 if logp.dtype != torch.float32: 1594 logp = logp.float() 1595 1596 # get index k, of max prob 1597 v, k = logp.max(0) 1598 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 1599 1600 # Note, we have non-blanks in the vocab first, followed by big blanks, and standard blank at last. 1601 # here we check if it's a big blank and if yes, set the duration variable. 1602 if k >= self._blank_index - len(self.big_blank_durations) and k < self._blank_index: 1603 big_blank_duration = self.big_blank_durations[self._blank_index - k - 1] 1604 1605 if self.preserve_alignments: 1606 # insert logprobs into last timestep 1607 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 1608 1609 if self.preserve_frame_confidence: 1610 # insert confidence into last timestep 1611 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 1612 1613 del logp 1614 1615 # If any type of blank token is predicted, exit inner loop, move onto next timestep t 1616 if k >= self._blank_index - len(self.big_blank_durations): 1617 not_blank = False 1618 else: 1619 # Append token to label set, update RNN state. 1620 hypothesis.y_sequence.append(k) 1621 hypothesis.score += float(v) 1622 hypothesis.timestep.append(time_idx) 1623 hypothesis.dec_state = hidden_prime 1624 hypothesis.last_token = k 1625 1626 # Increment token counter. 1627 symbols_added += 1 1628 1629 if self.preserve_alignments: 1630 # convert Ti-th logits into a torch array 1631 hypothesis.alignments.append([]) # blank buffer for next timestep 1632 1633 if self.preserve_frame_confidence: 1634 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 1635 1636 # Remove trailing empty list of Alignments 1637 if self.preserve_alignments: 1638 if len(hypothesis.alignments[-1]) == 0: 1639 del hypothesis.alignments[-1] 1640 1641 # Remove trailing empty list of per-frame confidence 1642 if self.preserve_frame_confidence: 1643 if len(hypothesis.frame_confidence[-1]) == 0: 1644 del hypothesis.frame_confidence[-1] 1645 1646 # Unpack the hidden states 1647 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 1648 1649 return hypothesis 1650 1651 1652 class GreedyBatchedMultiblankRNNTInfer(GreedyBatchedRNNTInfer): 1653 """A batch level greedy transducer decoder. 1654 Batch level greedy decoding, performed auto-regressively. 1655 Args: 1656 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1657 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1658 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1659 big_blank_durations: a list containing durations for big blanks the model supports. 1660 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1661 to a sequence in a single time step; if set to None then there is 1662 no limit. 1663 preserve_alignments: Bool flag which preserves the history of alignments generated during 1664 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1665 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1666 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1667 The length of the list corresponds to the Acoustic Length (T). 1668 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1669 U is the number of target tokens for the current timestep Ti. 1670 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1671 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1672 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1673 The length of the list corresponds to the Acoustic Length (T). 1674 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1675 U is the number of target tokens for the current timestep Ti. 1676 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1677 confidence scores. 1678 1679 name: The method name (str). 1680 Supported values: 1681 - 'max_prob' for using the maximum token probability as a confidence. 1682 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1683 1684 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 1685 Supported values: 1686 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1687 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1688 Note that for this entropy, the alpha should comply the following inequality: 1689 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1690 where V is the model vocabulary size. 1691 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1692 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1693 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1694 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1695 - 'renyi' for the Rรฉnyi entropy. 1696 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1697 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1698 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1699 1700 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1701 When the alpha equals one, scaling is not applied to 'max_prob', 1702 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1703 1704 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1705 Supported values: 1706 - 'lin' for using the linear mapping. 1707 - 'exp' for using exponential mapping with linear shift. 1708 """ 1709 1710 def __init__( 1711 self, 1712 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1713 joint_model: rnnt_abstract.AbstractRNNTJoint, 1714 blank_index: int, 1715 big_blank_durations: List[int], 1716 max_symbols_per_step: Optional[int] = None, 1717 preserve_alignments: bool = False, 1718 preserve_frame_confidence: bool = False, 1719 confidence_method_cfg: Optional[DictConfig] = None, 1720 ): 1721 super().__init__( 1722 decoder_model=decoder_model, 1723 joint_model=joint_model, 1724 blank_index=blank_index, 1725 max_symbols_per_step=max_symbols_per_step, 1726 preserve_alignments=preserve_alignments, 1727 preserve_frame_confidence=preserve_frame_confidence, 1728 confidence_method_cfg=confidence_method_cfg, 1729 ) 1730 self.big_blank_durations = big_blank_durations 1731 1732 # Depending on availability of `blank_as_pad` support 1733 # switch between more efficient batch decoding technique 1734 if self.decoder.blank_as_pad: 1735 self._greedy_decode = self._greedy_decode_blank_as_pad 1736 else: 1737 self._greedy_decode = self._greedy_decode_masked 1738 self._SOS = blank_index - len(big_blank_durations) 1739 1740 def _greedy_decode_blank_as_pad( 1741 self, 1742 x: torch.Tensor, 1743 out_len: torch.Tensor, 1744 device: torch.device, 1745 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1746 ): 1747 if partial_hypotheses is not None: 1748 raise NotImplementedError("`partial_hypotheses` support is not supported") 1749 1750 with torch.inference_mode(): 1751 # x: [B, T, D] 1752 # out_len: [B] 1753 # device: torch.device 1754 1755 # Initialize list of Hypothesis 1756 batchsize = x.shape[0] 1757 hypotheses = [ 1758 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1759 ] 1760 1761 # Initialize Hidden state matrix (shared by entire batch) 1762 hidden = None 1763 1764 # If alignments need to be preserved, register a danling list to hold the values 1765 if self.preserve_alignments: 1766 # alignments is a 3-dimensional dangling list representing B x T x U 1767 for hyp in hypotheses: 1768 hyp.alignments = [[]] 1769 1770 # If confidence scores need to be preserved, register a danling list to hold the values 1771 if self.preserve_frame_confidence: 1772 # frame_confidence is a 3-dimensional dangling list representing B x T x U 1773 for hyp in hypotheses: 1774 hyp.frame_confidence = [[]] 1775 1776 # Last Label buffer + Last Label without blank buffer 1777 # batch level equivalent of the last_label 1778 last_label = torch.full([batchsize, 1], fill_value=self._SOS, dtype=torch.long, device=device) 1779 1780 # this mask is true for if the emission is *any type* of blank. 1781 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 1782 1783 # Get max sequence length 1784 max_out_len = out_len.max() 1785 1786 # We have a mask for each big blank. A mask is "true" means: the previous emission is exactly the big-blank 1787 # with the corresponding duration, or has larger duration. E.g., for big_blank_mask for duration 2, it will 1788 # be set true if the previous emission was a big blank with duration 4, or 3 or 2; but false if prevoius 1789 # emission was a standard blank (with duration = 1). 1790 big_blank_masks = [torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device)] * len( 1791 self.big_blank_durations 1792 ) 1793 1794 # if this variable > 1, it means the previous emission was big-blank and we need to skip frames. 1795 big_blank_duration = 1 1796 1797 for time_idx in range(max_out_len): 1798 if big_blank_duration > 1: 1799 # skip frames until big_blank_duration == 1 1800 big_blank_duration -= 1 1801 continue 1802 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 1803 1804 # Prepare t timestamp batch variables 1805 not_blank = True 1806 symbols_added = 0 1807 1808 # Reset all blank masks 1809 blank_mask.mul_(False) 1810 for i in range(len(big_blank_masks)): 1811 big_blank_masks[i].mul_(False) 1812 1813 # Update blank mask with time mask 1814 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1815 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1816 blank_mask = time_idx >= out_len 1817 for i in range(len(big_blank_masks)): 1818 big_blank_masks[i] = time_idx >= out_len 1819 1820 # Start inner loop 1821 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1822 # Batch prediction and joint network steps 1823 # If very first prediction step, submit SOS tag (blank) to pred_step. 1824 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1825 if time_idx == 0 and symbols_added == 0 and hidden is None: 1826 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 1827 else: 1828 # Perform batch step prediction of decoder, getting new states and scores ("g") 1829 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 1830 1831 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 1832 # If preserving per-frame confidence, log_normalize must be true 1833 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1834 :, 0, 0, : 1835 ] 1836 1837 if logp.dtype != torch.float32: 1838 logp = logp.float() 1839 1840 # Get index k, of max prob for batch 1841 v, k = logp.max(1) 1842 del g 1843 1844 # Update blank mask with current predicted blanks 1845 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1846 k_is_blank = k >= self._blank_index - len(self.big_blank_durations) 1847 blank_mask.bitwise_or_(k_is_blank) 1848 1849 for i in range(len(big_blank_masks)): 1850 # using <= since as we mentioned before, the mask doesn't store exact matches. 1851 # instead, it is True when the predicted blank's duration is >= the duration that the 1852 # mask corresponds to. 1853 k_is_big_blank = k <= self._blank_index - 1 - i 1854 1855 # need to do a bitwise_and since it could also be a non-blank. 1856 k_is_big_blank.bitwise_and_(k_is_blank) 1857 big_blank_masks[i].bitwise_or_(k_is_big_blank) 1858 1859 del k_is_blank 1860 1861 # If preserving alignments, check if sequence length of sample has been reached 1862 # before adding alignment 1863 if self.preserve_alignments: 1864 # Insert logprobs into last timestep per sample 1865 logp_vals = logp.to('cpu') 1866 logp_ids = logp_vals.max(1)[1] 1867 for batch_idx in range(batchsize): 1868 if time_idx < out_len[batch_idx]: 1869 hypotheses[batch_idx].alignments[-1].append( 1870 (logp_vals[batch_idx], logp_ids[batch_idx]) 1871 ) 1872 del logp_vals 1873 1874 # If preserving per-frame confidence, check if sequence length of sample has been reached 1875 # before adding confidence scores 1876 if self.preserve_frame_confidence: 1877 # Insert probabilities into last timestep per sample 1878 confidence = self._get_confidence(logp) 1879 for batch_idx in range(batchsize): 1880 if time_idx < out_len[batch_idx]: 1881 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 1882 del logp 1883 1884 # If all samples predict / have predicted prior blanks, exit loop early 1885 # This is equivalent to if single sample predicted k 1886 if blank_mask.all(): 1887 not_blank = False 1888 else: 1889 # Collect batch indices where blanks occurred now/past 1890 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1891 1892 # Recover prior state for all samples which predicted blank now/past 1893 if hidden is not None: 1894 # LSTM has 2 states 1895 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 1896 1897 elif len(blank_indices) > 0 and hidden is None: 1898 # Reset state if there were some blank and other non-blank predictions in batch 1899 # Original state is filled with zeros so we just multiply 1900 # LSTM has 2 states 1901 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 1902 1903 # Recover prior predicted label for all samples which predicted blank now/past 1904 k[blank_indices] = last_label[blank_indices, 0] 1905 1906 # Update new label and hidden state for next iteration 1907 last_label = k.clone().view(-1, 1) 1908 hidden = hidden_prime 1909 1910 # Update predicted labels, accounting for time mask 1911 # If blank was predicted even once, now or in the past, 1912 # Force the current predicted label to also be blank 1913 # This ensures that blanks propogate across all timesteps 1914 # once they have occured (normally stopping condition of sample level loop). 1915 for kidx, ki in enumerate(k): 1916 if blank_mask[kidx] == 0: 1917 hypotheses[kidx].y_sequence.append(ki) 1918 hypotheses[kidx].timestep.append(time_idx) 1919 hypotheses[kidx].score += float(v[kidx]) 1920 1921 symbols_added += 1 1922 1923 for i in range(len(big_blank_masks) + 1): 1924 # The task here is find the shortest blank duration of all batches. 1925 # so we start from the shortest blank duration and go up, 1926 # and stop once we found the duration whose corresponding mask isn't all True. 1927 if i == len(big_blank_masks) or not big_blank_masks[i].all(): 1928 big_blank_duration = self.big_blank_durations[i - 1] if i > 0 else 1 1929 break 1930 1931 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 1932 # Then preserve U at current timestep Ti 1933 # Finally, forward the timestep history to Ti+1 for that sample 1934 # All of this should only be done iff the current time index <= sample-level AM length. 1935 # Otherwise ignore and move to next sample / next timestep. 1936 if self.preserve_alignments: 1937 1938 # convert Ti-th logits into a torch array 1939 for batch_idx in range(batchsize): 1940 1941 # this checks if current timestep <= sample-level AM length 1942 # If current timestep > sample-level AM length, no alignments will be added 1943 # Therefore the list of Uj alignments is empty here. 1944 if len(hypotheses[batch_idx].alignments[-1]) > 0: 1945 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 1946 1947 # Do the same if preserving per-frame confidence 1948 if self.preserve_frame_confidence: 1949 1950 for batch_idx in range(batchsize): 1951 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 1952 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 1953 1954 # Remove trailing empty list of alignments at T_{am-len} x Uj 1955 if self.preserve_alignments: 1956 for batch_idx in range(batchsize): 1957 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1958 del hypotheses[batch_idx].alignments[-1] 1959 1960 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1961 if self.preserve_frame_confidence: 1962 for batch_idx in range(batchsize): 1963 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1964 del hypotheses[batch_idx].frame_confidence[-1] 1965 1966 # Preserve states 1967 for batch_idx in range(batchsize): 1968 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1969 1970 return hypotheses 1971 1972 def _greedy_decode_masked( 1973 self, 1974 x: torch.Tensor, 1975 out_len: torch.Tensor, 1976 device: torch.device, 1977 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1978 ): 1979 if partial_hypotheses is not None: 1980 raise NotImplementedError("`partial_hypotheses` support is not supported") 1981 1982 if self.big_blank_durations != [1] * len(self.big_blank_durations): 1983 raise NotImplementedError( 1984 "Efficient frame-skipping version for multi-blank masked decoding is not supported." 1985 ) 1986 1987 # x: [B, T, D] 1988 # out_len: [B] 1989 # device: torch.device 1990 1991 # Initialize state 1992 batchsize = x.shape[0] 1993 hypotheses = [ 1994 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1995 ] 1996 1997 # Initialize Hidden state matrix (shared by entire batch) 1998 hidden = None 1999 2000 # If alignments need to be preserved, register a danling list to hold the values 2001 if self.preserve_alignments: 2002 # alignments is a 3-dimensional dangling list representing B x T x U 2003 for hyp in hypotheses: 2004 hyp.alignments = [[]] 2005 else: 2006 hyp.alignments = None 2007 2008 # If confidence scores need to be preserved, register a danling list to hold the values 2009 if self.preserve_frame_confidence: 2010 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2011 for hyp in hypotheses: 2012 hyp.frame_confidence = [[]] 2013 2014 # Last Label buffer + Last Label without blank buffer 2015 # batch level equivalent of the last_label 2016 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2017 last_label_without_blank = last_label.clone() 2018 2019 # Mask buffers 2020 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2021 2022 # Get max sequence length 2023 max_out_len = out_len.max() 2024 2025 with torch.inference_mode(): 2026 for time_idx in range(max_out_len): 2027 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2028 2029 # Prepare t timestamp batch variables 2030 not_blank = True 2031 symbols_added = 0 2032 2033 # Reset blank mask 2034 blank_mask.mul_(False) 2035 2036 # Update blank mask with time mask 2037 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2038 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2039 blank_mask = time_idx >= out_len 2040 2041 # Start inner loop 2042 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 2043 # Batch prediction and joint network steps 2044 # If very first prediction step, submit SOS tag (blank) to pred_step. 2045 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2046 if time_idx == 0 and symbols_added == 0 and hidden is None: 2047 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2048 else: 2049 # Set a dummy label for the blank value 2050 # This value will be overwritten by "blank" again the last label update below 2051 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 2052 last_label_without_blank_mask = last_label >= self._blank_index 2053 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 2054 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 2055 ~last_label_without_blank_mask 2056 ] 2057 2058 # Perform batch step prediction of decoder, getting new states and scores ("g") 2059 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 2060 2061 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2062 # If preserving per-frame confidence, log_normalize must be true 2063 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 2064 :, 0, 0, : 2065 ] 2066 2067 if logp.dtype != torch.float32: 2068 logp = logp.float() 2069 2070 # Get index k, of max prob for batch 2071 v, k = logp.max(1) 2072 del g 2073 2074 # Update blank mask with current predicted blanks 2075 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2076 k_is_blank = k == self._blank_index 2077 blank_mask.bitwise_or_(k_is_blank) 2078 2079 # If preserving alignments, check if sequence length of sample has been reached 2080 # before adding alignment 2081 if self.preserve_alignments: 2082 # Insert logprobs into last timestep per sample 2083 logp_vals = logp.to('cpu') 2084 logp_ids = logp_vals.max(1)[1] 2085 for batch_idx in range(batchsize): 2086 if time_idx < out_len[batch_idx]: 2087 hypotheses[batch_idx].alignments[-1].append( 2088 (logp_vals[batch_idx], logp_ids[batch_idx]) 2089 ) 2090 del logp_vals 2091 2092 # If preserving per-frame confidence, check if sequence length of sample has been reached 2093 # before adding confidence scores 2094 if self.preserve_frame_confidence: 2095 # Insert probabilities into last timestep per sample 2096 confidence = self._get_confidence(logp) 2097 for batch_idx in range(batchsize): 2098 if time_idx < out_len[batch_idx]: 2099 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 2100 del logp 2101 2102 # If all samples predict / have predicted prior blanks, exit loop early 2103 # This is equivalent to if single sample predicted k 2104 if blank_mask.all(): 2105 not_blank = False 2106 else: 2107 # Collect batch indices where blanks occurred now/past 2108 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2109 2110 # Recover prior state for all samples which predicted blank now/past 2111 if hidden is not None: 2112 # LSTM has 2 states 2113 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2114 2115 elif len(blank_indices) > 0 and hidden is None: 2116 # Reset state if there were some blank and other non-blank predictions in batch 2117 # Original state is filled with zeros so we just multiply 2118 # LSTM has 2 states 2119 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2120 2121 # Recover prior predicted label for all samples which predicted blank now/past 2122 k[blank_indices] = last_label[blank_indices, 0] 2123 2124 # Update new label and hidden state for next iteration 2125 last_label = k.view(-1, 1) 2126 hidden = hidden_prime 2127 2128 # Update predicted labels, accounting for time mask 2129 # If blank was predicted even once, now or in the past, 2130 # Force the current predicted label to also be blank 2131 # This ensures that blanks propogate across all timesteps 2132 # once they have occured (normally stopping condition of sample level loop). 2133 for kidx, ki in enumerate(k): 2134 if blank_mask[kidx] == 0: 2135 hypotheses[kidx].y_sequence.append(ki) 2136 hypotheses[kidx].timestep.append(time_idx) 2137 hypotheses[kidx].score += float(v[kidx]) 2138 2139 symbols_added += 1 2140 2141 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 2142 # Then preserve U at current timestep Ti 2143 # Finally, forward the timestep history to Ti+1 for that sample 2144 # All of this should only be done iff the current time index <= sample-level AM length. 2145 # Otherwise ignore and move to next sample / next timestep. 2146 if self.preserve_alignments: 2147 2148 # convert Ti-th logits into a torch array 2149 for batch_idx in range(batchsize): 2150 2151 # this checks if current timestep <= sample-level AM length 2152 # If current timestep > sample-level AM length, no alignments will be added 2153 # Therefore the list of Uj alignments is empty here. 2154 if len(hypotheses[batch_idx].alignments[-1]) > 0: 2155 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 2156 2157 # Do the same if preserving per-frame confidence 2158 if self.preserve_frame_confidence: 2159 2160 for batch_idx in range(batchsize): 2161 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 2162 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 2163 2164 # Remove trailing empty list of alignments at T_{am-len} x Uj 2165 if self.preserve_alignments: 2166 for batch_idx in range(batchsize): 2167 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2168 del hypotheses[batch_idx].alignments[-1] 2169 2170 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2171 if self.preserve_frame_confidence: 2172 for batch_idx in range(batchsize): 2173 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2174 del hypotheses[batch_idx].frame_confidence[-1] 2175 2176 # Preserve states 2177 for batch_idx in range(batchsize): 2178 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2179 2180 return hypotheses 2181 2182 2183 @dataclass 2184 class GreedyRNNTInferConfig: 2185 max_symbols_per_step: Optional[int] = 10 2186 preserve_alignments: bool = False 2187 preserve_frame_confidence: bool = False 2188 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 2189 2190 def __post_init__(self): 2191 # OmegaConf.structured ensures that post_init check is always executed 2192 self.confidence_method_cfg = OmegaConf.structured( 2193 self.confidence_method_cfg 2194 if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) 2195 else ConfidenceMethodConfig(**self.confidence_method_cfg) 2196 ) 2197 2198 2199 @dataclass 2200 class GreedyBatchedRNNTInferConfig: 2201 max_symbols_per_step: Optional[int] = 10 2202 preserve_alignments: bool = False 2203 preserve_frame_confidence: bool = False 2204 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 2205 2206 def __post_init__(self): 2207 # OmegaConf.structured ensures that post_init check is always executed 2208 self.confidence_method_cfg = OmegaConf.structured( 2209 self.confidence_method_cfg 2210 if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) 2211 else ConfidenceMethodConfig(**self.confidence_method_cfg) 2212 ) 2213 2214 2215 class GreedyTDTInfer(_GreedyRNNTInfer): 2216 """A greedy TDT decoder. 2217 2218 Sequence level greedy decoding, performed auto-regressively. 2219 2220 Args: 2221 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2222 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2223 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2224 durations: a list containing durations for TDT. 2225 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2226 to a sequence in a single time step; if set to None then there is 2227 no limit. 2228 preserve_alignments: Bool flag which preserves the history of alignments generated during 2229 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2230 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2231 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2232 The length of the list corresponds to the Acoustic Length (T). 2233 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2234 U is the number of target tokens for the current timestep Ti. 2235 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2236 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2237 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2238 The length of the list corresponds to the Acoustic Length (T). 2239 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2240 U is the number of target tokens for the current timestep Ti. 2241 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 2242 confidence scores. 2243 2244 name: The method name (str). 2245 Supported values: 2246 - 'max_prob' for using the maximum token probability as a confidence. 2247 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2248 2249 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 2250 Supported values: 2251 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2252 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2253 Note that for this entropy, the alpha should comply the following inequality: 2254 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2255 where V is the model vocabulary size. 2256 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2257 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2258 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2259 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2260 - 'renyi' for the Rรฉnyi entropy. 2261 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2262 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2263 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2264 2265 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2266 When the alpha equals one, scaling is not applied to 'max_prob', 2267 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2268 2269 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2270 Supported values: 2271 - 'lin' for using the linear mapping. 2272 - 'exp' for using exponential mapping with linear shift. 2273 """ 2274 2275 def __init__( 2276 self, 2277 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2278 joint_model: rnnt_abstract.AbstractRNNTJoint, 2279 blank_index: int, 2280 durations: list, 2281 max_symbols_per_step: Optional[int] = None, 2282 preserve_alignments: bool = False, 2283 preserve_frame_confidence: bool = False, 2284 confidence_method_cfg: Optional[DictConfig] = None, 2285 ): 2286 super().__init__( 2287 decoder_model=decoder_model, 2288 joint_model=joint_model, 2289 blank_index=blank_index, 2290 max_symbols_per_step=max_symbols_per_step, 2291 preserve_alignments=preserve_alignments, 2292 preserve_frame_confidence=preserve_frame_confidence, 2293 confidence_method_cfg=confidence_method_cfg, 2294 ) 2295 self.durations = durations 2296 2297 @typecheck() 2298 def forward( 2299 self, 2300 encoder_output: torch.Tensor, 2301 encoded_lengths: torch.Tensor, 2302 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2303 ): 2304 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2305 Output token is generated auto-regressively. 2306 Args: 2307 encoder_output: A tensor of size (batch, features, timesteps). 2308 encoded_lengths: list of int representing the length of each sequence 2309 output sequence. 2310 Returns: 2311 packed list containing batch number of sentences (Hypotheses). 2312 """ 2313 # Preserve decoder and joint training state 2314 decoder_training_state = self.decoder.training 2315 joint_training_state = self.joint.training 2316 2317 with torch.inference_mode(): 2318 # Apply optional preprocessing 2319 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2320 2321 self.decoder.eval() 2322 self.joint.eval() 2323 2324 hypotheses = [] 2325 # Process each sequence independently 2326 with self.decoder.as_frozen(), self.joint.as_frozen(): 2327 for batch_idx in range(encoder_output.size(0)): 2328 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 2329 logitlen = encoded_lengths[batch_idx] 2330 2331 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 2332 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 2333 hypotheses.append(hypothesis) 2334 2335 # Pack results into Hypotheses 2336 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 2337 2338 self.decoder.train(decoder_training_state) 2339 self.joint.train(joint_training_state) 2340 2341 return (packed_result,) 2342 2343 @torch.no_grad() 2344 def _greedy_decode( 2345 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 2346 ): 2347 # x: [T, 1, D] 2348 # out_len: [seq_len] 2349 2350 # Initialize blank state and empty label set in Hypothesis 2351 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 2352 2353 if partial_hypotheses is not None: 2354 hypothesis.last_token = partial_hypotheses.last_token 2355 hypothesis.y_sequence = ( 2356 partial_hypotheses.y_sequence.cpu().tolist() 2357 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 2358 else partial_hypotheses.y_sequence 2359 ) 2360 if partial_hypotheses.dec_state is not None: 2361 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 2362 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 2363 2364 if self.preserve_alignments: 2365 # Alignments is a 2-dimensional dangling list representing T x U 2366 hypothesis.alignments = [[]] 2367 2368 if self.preserve_frame_confidence: 2369 hypothesis.frame_confidence = [[]] 2370 2371 time_idx = 0 2372 while time_idx < out_len: 2373 # Extract encoder embedding at timestep t 2374 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 2375 f = x.narrow(dim=0, start=time_idx, length=1) 2376 2377 # Setup exit flags and counter 2378 not_blank = True 2379 symbols_added = 0 2380 2381 need_loop = True 2382 # While blank is not predicted, or we dont run out of max symbols per timestep 2383 while need_loop and (self.max_symbols is None or symbols_added < self.max_symbols): 2384 # In the first timestep, we initialize the network with RNNT Blank 2385 # In later timesteps, we provide previous predicted label as input. 2386 if hypothesis.last_token is None and hypothesis.dec_state is None: 2387 last_label = self._SOS 2388 else: 2389 last_label = label_collate([[hypothesis.last_token]]) 2390 2391 # Perform prediction network and joint network steps. 2392 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 2393 # If preserving per-frame confidence, log_normalize must be true 2394 logits = self._joint_step(f, g, log_normalize=False) 2395 logp = logits[0, 0, 0, : -len(self.durations)] 2396 if self.preserve_frame_confidence: 2397 logp = torch.log_softmax(logp, -1) 2398 2399 duration_logp = torch.log_softmax(logits[0, 0, 0, -len(self.durations) :], dim=-1) 2400 del g 2401 2402 # torch.max(0) op doesnt exist for FP 16. 2403 if logp.dtype != torch.float32: 2404 logp = logp.float() 2405 2406 # get index k, of max prob 2407 v, k = logp.max(0) 2408 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 2409 2410 d_v, d_k = duration_logp.max(0) 2411 d_k = d_k.item() 2412 2413 skip = self.durations[d_k] 2414 2415 if self.preserve_alignments: 2416 # insert logprobs into last timestep 2417 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 2418 2419 if self.preserve_frame_confidence: 2420 # insert confidence into last timestep 2421 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 2422 2423 del logp 2424 2425 # If blank token is predicted, exit inner loop, move onto next timestep t 2426 if k == self._blank_index: 2427 not_blank = False 2428 else: 2429 # Append token to label set, update RNN state. 2430 hypothesis.y_sequence.append(k) 2431 hypothesis.score += float(v) 2432 hypothesis.timestep.append(time_idx) 2433 hypothesis.dec_state = hidden_prime 2434 hypothesis.last_token = k 2435 2436 # Increment token counter. 2437 symbols_added += 1 2438 time_idx += skip 2439 need_loop = skip == 0 2440 2441 # this rarely happens, but we manually increment the `skip` number 2442 # if blank is emitted and duration=0 is predicted. This prevents possible 2443 # infinite loops. 2444 if skip == 0: 2445 skip = 1 2446 2447 if self.preserve_alignments: 2448 # convert Ti-th logits into a torch array 2449 hypothesis.alignments.append([]) # blank buffer for next timestep 2450 2451 if self.preserve_frame_confidence: 2452 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 2453 2454 if symbols_added == self.max_symbols: 2455 time_idx += 1 2456 2457 # Remove trailing empty list of Alignments 2458 if self.preserve_alignments: 2459 if len(hypothesis.alignments[-1]) == 0: 2460 del hypothesis.alignments[-1] 2461 2462 # Remove trailing empty list of per-frame confidence 2463 if self.preserve_frame_confidence: 2464 if len(hypothesis.frame_confidence[-1]) == 0: 2465 del hypothesis.frame_confidence[-1] 2466 2467 # Unpack the hidden states 2468 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 2469 2470 return hypothesis 2471 2472 2473 class GreedyBatchedTDTInfer(_GreedyRNNTInfer): 2474 """A batch level greedy TDT decoder. 2475 Batch level greedy decoding, performed auto-regressively. 2476 Args: 2477 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2478 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2479 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2480 durations: a list containing durations. 2481 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2482 to a sequence in a single time step; if set to None then there is 2483 no limit. 2484 preserve_alignments: Bool flag which preserves the history of alignments generated during 2485 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2486 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2487 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2488 The length of the list corresponds to the Acoustic Length (T). 2489 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2490 U is the number of target tokens for the current timestep Ti. 2491 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2492 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2493 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2494 The length of the list corresponds to the Acoustic Length (T). 2495 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2496 U is the number of target tokens for the current timestep Ti. 2497 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 2498 confidence scores. 2499 2500 name: The method name (str). 2501 Supported values: 2502 - 'max_prob' for using the maximum token probability as a confidence. 2503 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2504 2505 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 2506 Supported values: 2507 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2508 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2509 Note that for this entropy, the alpha should comply the following inequality: 2510 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2511 where V is the model vocabulary size. 2512 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2513 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2514 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2515 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2516 - 'renyi' for the Rรฉnyi entropy. 2517 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2518 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2519 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2520 2521 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2522 When the alpha equals one, scaling is not applied to 'max_prob', 2523 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2524 2525 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2526 Supported values: 2527 - 'lin' for using the linear mapping. 2528 - 'exp' for using exponential mapping with linear shift. 2529 """ 2530 2531 def __init__( 2532 self, 2533 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2534 joint_model: rnnt_abstract.AbstractRNNTJoint, 2535 blank_index: int, 2536 durations: List[int], 2537 max_symbols_per_step: Optional[int] = None, 2538 preserve_alignments: bool = False, 2539 preserve_frame_confidence: bool = False, 2540 confidence_method_cfg: Optional[DictConfig] = None, 2541 ): 2542 super().__init__( 2543 decoder_model=decoder_model, 2544 joint_model=joint_model, 2545 blank_index=blank_index, 2546 max_symbols_per_step=max_symbols_per_step, 2547 preserve_alignments=preserve_alignments, 2548 preserve_frame_confidence=preserve_frame_confidence, 2549 confidence_method_cfg=confidence_method_cfg, 2550 ) 2551 self.durations = durations 2552 2553 # Depending on availability of `blank_as_pad` support 2554 # switch between more efficient batch decoding technique 2555 if self.decoder.blank_as_pad: 2556 self._greedy_decode = self._greedy_decode_blank_as_pad 2557 else: 2558 self._greedy_decode = self._greedy_decode_masked 2559 2560 @typecheck() 2561 def forward( 2562 self, 2563 encoder_output: torch.Tensor, 2564 encoded_lengths: torch.Tensor, 2565 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2566 ): 2567 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2568 Output token is generated auto-regressively. 2569 Args: 2570 encoder_output: A tensor of size (batch, features, timesteps). 2571 encoded_lengths: list of int representing the length of each sequence 2572 output sequence. 2573 Returns: 2574 packed list containing batch number of sentences (Hypotheses). 2575 """ 2576 # Preserve decoder and joint training state 2577 decoder_training_state = self.decoder.training 2578 joint_training_state = self.joint.training 2579 2580 with torch.inference_mode(): 2581 # Apply optional preprocessing 2582 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2583 logitlen = encoded_lengths 2584 2585 self.decoder.eval() 2586 self.joint.eval() 2587 2588 with self.decoder.as_frozen(), self.joint.as_frozen(): 2589 inseq = encoder_output # [B, T, D] 2590 hypotheses = self._greedy_decode( 2591 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 2592 ) 2593 2594 # Pack the hypotheses results 2595 packed_result = pack_hypotheses(hypotheses, logitlen) 2596 2597 self.decoder.train(decoder_training_state) 2598 self.joint.train(joint_training_state) 2599 2600 return (packed_result,) 2601 2602 def _greedy_decode_blank_as_pad( 2603 self, 2604 x: torch.Tensor, 2605 out_len: torch.Tensor, 2606 device: torch.device, 2607 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2608 ): 2609 if partial_hypotheses is not None: 2610 raise NotImplementedError("`partial_hypotheses` support is not supported") 2611 2612 with torch.inference_mode(): 2613 # x: [B, T, D] 2614 # out_len: [B] 2615 # device: torch.device 2616 2617 # Initialize list of Hypothesis 2618 batchsize = x.shape[0] 2619 hypotheses = [ 2620 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 2621 ] 2622 2623 # Initialize Hidden state matrix (shared by entire batch) 2624 hidden = None 2625 2626 # If alignments need to be preserved, register a danling list to hold the values 2627 if self.preserve_alignments: 2628 # alignments is a 3-dimensional dangling list representing B x T x U 2629 for hyp in hypotheses: 2630 hyp.alignments = [[]] 2631 2632 # If confidence scores need to be preserved, register a danling list to hold the values 2633 if self.preserve_frame_confidence: 2634 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2635 for hyp in hypotheses: 2636 hyp.frame_confidence = [[]] 2637 2638 # Last Label buffer + Last Label without blank buffer 2639 # batch level equivalent of the last_label 2640 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2641 2642 # Mask buffers 2643 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2644 2645 # Get max sequence length 2646 max_out_len = out_len.max() 2647 2648 # skip means the number of frames the next decoding step should "jump" to. When skip == 1 2649 # it means the next decoding step will just use the next input frame. 2650 skip = 1 2651 for time_idx in range(max_out_len): 2652 if skip > 1: # if skip > 1 at the current step, we decrement it and skip the current frame. 2653 skip -= 1 2654 continue 2655 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2656 2657 # need_to_stay is a boolean indicates whether the next decoding step should remain in the same frame. 2658 need_to_stay = True 2659 symbols_added = 0 2660 2661 # Reset blank mask 2662 blank_mask.mul_(False) 2663 2664 # Update blank mask with time mask 2665 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2666 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2667 blank_mask = time_idx >= out_len 2668 2669 # Start inner loop 2670 while need_to_stay and (self.max_symbols is None or symbols_added < self.max_symbols): 2671 # Batch prediction and joint network steps 2672 # If very first prediction step, submit SOS tag (blank) to pred_step. 2673 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2674 if time_idx == 0 and symbols_added == 0 and hidden is None: 2675 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2676 else: 2677 # Perform batch step prediction of decoder, getting new states and scores ("g") 2678 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 2679 2680 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2681 # Note: log_normalize must not be True here since the joiner output is contanetation of both token logits and duration logits, 2682 # and they need to be normalized independently. 2683 joined = self._joint_step(f, g, log_normalize=None) 2684 logp = joined[:, 0, 0, : -len(self.durations)] 2685 duration_logp = joined[:, 0, 0, -len(self.durations) :] 2686 2687 if logp.dtype != torch.float32: 2688 logp = logp.float() 2689 duration_logp = duration_logp.float() 2690 2691 # get the max for both token and duration predictions. 2692 v, k = logp.max(1) 2693 dv, dk = duration_logp.max(1) 2694 2695 # here we set the skip value to be the minimum of all predicted durations, hense the "torch.min(dk)" call there. 2696 # Please refer to Section 5.2 of our paper https://arxiv.org/pdf/2304.06795.pdf for explanation of this. 2697 skip = self.durations[int(torch.min(dk))] 2698 2699 # this is a special case: if all batches emit blanks, we require that skip be at least 1 2700 # so we don't loop forever at the current frame. 2701 if blank_mask.all(): 2702 if skip == 0: 2703 skip = 1 2704 2705 need_to_stay = skip == 0 2706 del g 2707 2708 # Update blank mask with current predicted blanks 2709 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2710 k_is_blank = k == self._blank_index 2711 blank_mask.bitwise_or_(k_is_blank) 2712 2713 del k_is_blank 2714 del logp, duration_logp 2715 2716 # If all samples predict / have predicted prior blanks, exit loop early 2717 # This is equivalent to if single sample predicted k 2718 if not blank_mask.all(): 2719 # Collect batch indices where blanks occurred now/past 2720 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2721 2722 # Recover prior state for all samples which predicted blank now/past 2723 if hidden is not None: 2724 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2725 2726 elif len(blank_indices) > 0 and hidden is None: 2727 # Reset state if there were some blank and other non-blank predictions in batch 2728 # Original state is filled with zeros so we just multiply 2729 # LSTM has 2 states 2730 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2731 2732 # Recover prior predicted label for all samples which predicted blank now/past 2733 k[blank_indices] = last_label[blank_indices, 0] 2734 2735 # Update new label and hidden state for next iteration 2736 last_label = k.clone().view(-1, 1) 2737 hidden = hidden_prime 2738 2739 # Update predicted labels, accounting for time mask 2740 # If blank was predicted even once, now or in the past, 2741 # Force the current predicted label to also be blank 2742 # This ensures that blanks propogate across all timesteps 2743 # once they have occured (normally stopping condition of sample level loop). 2744 for kidx, ki in enumerate(k): 2745 if blank_mask[kidx] == 0: 2746 hypotheses[kidx].y_sequence.append(ki) 2747 hypotheses[kidx].timestep.append(time_idx) 2748 hypotheses[kidx].score += float(v[kidx]) 2749 2750 symbols_added += 1 2751 2752 # Remove trailing empty list of alignments at T_{am-len} x Uj 2753 if self.preserve_alignments: 2754 for batch_idx in range(batchsize): 2755 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2756 del hypotheses[batch_idx].alignments[-1] 2757 2758 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2759 if self.preserve_frame_confidence: 2760 for batch_idx in range(batchsize): 2761 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2762 del hypotheses[batch_idx].frame_confidence[-1] 2763 2764 # Preserve states 2765 for batch_idx in range(batchsize): 2766 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2767 2768 return hypotheses 2769 2770 def _greedy_decode_masked( 2771 self, 2772 x: torch.Tensor, 2773 out_len: torch.Tensor, 2774 device: torch.device, 2775 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2776 ): 2777 raise NotImplementedError("masked greedy-batched decode is not supported for TDT models.") 2778 [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from abc import ABC, abstractmethod 17 from dataclasses import dataclass 18 from functools import partial 19 from typing import List, Optional 20 21 import torch 22 from omegaconf import DictConfig, OmegaConf 23 24 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 25 from nemo.utils import logging 26 27 28 class ConfidenceMethodConstants: 29 NAMES = ("max_prob", "entropy") 30 ENTROPY_TYPES = ("gibbs", "tsallis", "renyi") 31 ENTROPY_NORMS = ("lin", "exp") 32 33 @classmethod 34 def print(cls): 35 return ( 36 cls.__name__ 37 + ": " 38 + str({"NAMES": cls.NAMES, "ENTROPY_TYPES": cls.ENTROPY_TYPES, "ENTROPY_NORMS": cls.ENTROPY_NORMS}) 39 ) 40 41 42 class ConfidenceConstants: 43 AGGREGATIONS = ("mean", "min", "max", "prod") 44 45 @classmethod 46 def print(cls): 47 return cls.__name__ + ": " + str({"AGGREGATIONS": cls.AGGREGATIONS}) 48 49 50 @dataclass 51 class ConfidenceMethodConfig: 52 """A Config which contains the method name and settings to compute per-frame confidence scores. 53 54 Args: 55 name: The method name (str). 56 Supported values: 57 - 'max_prob' for using the maximum token probability as a confidence. 58 - 'entropy' for using a normalized entropy of a log-likelihood vector. 59 60 entropy_type: Which type of entropy to use (str). 61 Used if confidence_method_cfg.name is set to `entropy`. 62 Supported values: 63 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 64 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 65 Note that for this entropy, the alpha should comply the following inequality: 66 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 67 where V is the model vocabulary size. 68 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 69 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 70 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 71 More: https://en.wikipedia.org/wiki/Tsallis_entropy 72 - 'renyi' for the Rรฉnyi entropy. 73 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 74 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 75 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 76 77 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 78 When the alpha equals one, scaling is not applied to 'max_prob', 79 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 80 81 entropy_norm: A mapping of the entropy value to the interval [0,1]. 82 Supported values: 83 - 'lin' for using the linear mapping. 84 - 'exp' for using exponential mapping with linear shift. 85 """ 86 87 name: str = "entropy" 88 entropy_type: str = "tsallis" 89 alpha: float = 0.33 90 entropy_norm: str = "exp" 91 temperature: str = "DEPRECATED" 92 93 def __post_init__(self): 94 if self.temperature != "DEPRECATED": 95 # self.temperature has type str 96 self.alpha = float(self.temperature) 97 self.temperature = "DEPRECATED" 98 if self.name not in ConfidenceMethodConstants.NAMES: 99 raise ValueError( 100 f"`name` must be one of the following: " 101 f"{'`' + '`, `'.join(ConfidenceMethodConstants.NAMES) + '`'}. Provided: `{self.name}`" 102 ) 103 if self.entropy_type not in ConfidenceMethodConstants.ENTROPY_TYPES: 104 raise ValueError( 105 f"`entropy_type` must be one of the following: " 106 f"{'`' + '`, `'.join(ConfidenceMethodConstants.ENTROPY_TYPES) + '`'}. Provided: `{self.entropy_type}`" 107 ) 108 if self.alpha <= 0.0: 109 raise ValueError(f"`alpha` must be > 0. Provided: {self.alpha}") 110 if self.entropy_norm not in ConfidenceMethodConstants.ENTROPY_NORMS: 111 raise ValueError( 112 f"`entropy_norm` must be one of the following: " 113 f"{'`' + '`, `'.join(ConfidenceMethodConstants.ENTROPY_NORMS) + '`'}. Provided: `{self.entropy_norm}`" 114 ) 115 116 117 @dataclass 118 class ConfidenceConfig: 119 """A config which contains the following key-value pairs related to confidence scores. 120 121 Args: 122 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 123 generated during decoding. When set to true, the Hypothesis will contain 124 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 125 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 126 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 127 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 128 129 The length of the list corresponds to the number of recognized tokens. 130 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 131 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 132 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 133 134 The length of the list corresponds to the number of recognized words. 135 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 136 from the `token_confidence`. 137 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 138 Valid options are `mean`, `min`, `max`, `prod`. 139 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 140 confidence scores. 141 142 name: The method name (str). 143 Supported values: 144 - 'max_prob' for using the maximum token probability as a confidence. 145 - 'entropy' for using a normalized entropy of a log-likelihood vector. 146 147 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 148 Supported values: 149 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 150 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 151 Note that for this entropy, the alpha should comply the following inequality: 152 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 153 where V is the model vocabulary size. 154 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 155 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 156 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 157 More: https://en.wikipedia.org/wiki/Tsallis_entropy 158 - 'renyi' for the Rรฉnyi entropy. 159 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 160 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 161 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 162 163 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 164 When the alpha equals one, scaling is not applied to 'max_prob', 165 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 166 167 entropy_norm: A mapping of the entropy value to the interval [0,1]. 168 Supported values: 169 - 'lin' for using the linear mapping. 170 - 'exp' for using exponential mapping with linear shift. 171 """ 172 173 preserve_frame_confidence: bool = False 174 preserve_token_confidence: bool = False 175 preserve_word_confidence: bool = False 176 exclude_blank: bool = True 177 aggregation: str = "min" 178 method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() 179 180 def __post_init__(self): 181 # OmegaConf.structured ensures that post_init check is always executed 182 self.method_cfg = OmegaConf.structured( 183 self.method_cfg 184 if isinstance(self.method_cfg, ConfidenceMethodConfig) 185 else ConfidenceMethodConfig(**self.method_cfg) 186 ) 187 if self.aggregation not in ConfidenceConstants.AGGREGATIONS: 188 raise ValueError( 189 f"`aggregation` has to be one of the following: " 190 f"{'`' + '`, `'.join(ConfidenceConstants.AGGREGATIONS) + '`'}. Provided: `{self.aggregation}`" 191 ) 192 193 194 def get_confidence_measure_bank(): 195 """Generate a dictionary with confidence measure functionals. 196 197 Supported confidence measures: 198 max_prob: normalized maximum probability 199 entropy_gibbs_lin: Gibbs entropy with linear normalization 200 entropy_gibbs_exp: Gibbs entropy with exponential normalization 201 entropy_tsallis_lin: Tsallis entropy with linear normalization 202 entropy_tsallis_exp: Tsallis entropy with exponential normalization 203 entropy_renyi_lin: Rรฉnyi entropy with linear normalization 204 entropy_renyi_exp: Rรฉnyi entropy with exponential normalization 205 206 Returns: 207 dictionary with lambda functions. 208 """ 209 # helper functions 210 # Gibbs entropy is implemented without alpha 211 neg_entropy_gibbs = lambda x: (x.exp() * x).sum(-1) 212 neg_entropy_alpha = lambda x, t: (x * t).exp().sum(-1) 213 neg_entropy_alpha_gibbs = lambda x, t: ((x * t).exp() * x).sum(-1) 214 # too big for a lambda 215 def entropy_tsallis_exp(x, v, t): 216 exp_neg_max_ent = math.exp((1 - math.pow(v, 1 - t)) / (1 - t)) 217 return (((1 - neg_entropy_alpha(x, t)) / (1 - t)).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 218 219 def entropy_gibbs_exp(x, v, t): 220 exp_neg_max_ent = math.pow(v, -t * math.pow(v, 1 - t)) 221 return ((neg_entropy_alpha_gibbs(x, t) * t).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 222 223 # use Gibbs entropies for Tsallis and Rรฉnyi with t == 1.0 224 entropy_gibbs_lin_baseline = lambda x, v: 1 + neg_entropy_gibbs(x) / math.log(v) 225 entropy_gibbs_exp_baseline = lambda x, v: (neg_entropy_gibbs(x).exp() * v - 1) / (v - 1) 226 # fill the measure bank 227 confidence_measure_bank = {} 228 # Maximum probability measure is implemented without alpha 229 confidence_measure_bank["max_prob"] = ( 230 lambda x, v, t: (x.max(dim=-1)[0].exp() * v - 1) / (v - 1) 231 if t == 1.0 232 else ((x.max(dim=-1)[0] * t).exp() * math.pow(v, t) - 1) / (math.pow(v, t) - 1) 233 ) 234 confidence_measure_bank["entropy_gibbs_lin"] = ( 235 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 236 if t == 1.0 237 else 1 + neg_entropy_alpha_gibbs(x, t) / math.log(v) / math.pow(v, 1 - t) 238 ) 239 confidence_measure_bank["entropy_gibbs_exp"] = ( 240 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_gibbs_exp(x, v, t) 241 ) 242 confidence_measure_bank["entropy_tsallis_lin"] = ( 243 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 244 if t == 1.0 245 else 1 + (1 - neg_entropy_alpha(x, t)) / (math.pow(v, 1 - t) - 1) 246 ) 247 confidence_measure_bank["entropy_tsallis_exp"] = ( 248 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_tsallis_exp(x, v, t) 249 ) 250 confidence_measure_bank["entropy_renyi_lin"] = ( 251 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 252 if t == 1.0 253 else 1 + neg_entropy_alpha(x, t).log2() / (t - 1) / math.log(v, 2) 254 ) 255 confidence_measure_bank["entropy_renyi_exp"] = ( 256 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) 257 if t == 1.0 258 else (neg_entropy_alpha(x, t).pow(1 / (t - 1)) * v - 1) / (v - 1) 259 ) 260 return confidence_measure_bank 261 262 263 def get_confidence_aggregation_bank(): 264 """Generate a dictionary with confidence aggregation functions. 265 266 Supported confidence aggregation functions: 267 min: minimum 268 max: maximum 269 mean: arithmetic mean 270 prod: product 271 272 Returns: 273 dictionary with functions. 274 """ 275 confidence_aggregation_bank = {"mean": lambda x: sum(x) / len(x), "min": min, "max": max} 276 # python 3.7 and earlier do not have math.prod 277 if hasattr(math, "prod"): 278 confidence_aggregation_bank["prod"] = math.prod 279 else: 280 import operator 281 from functools import reduce 282 283 confidence_aggregation_bank["prod"] = lambda x: reduce(operator.mul, x, 1) 284 return confidence_aggregation_bank 285 286 287 class ConfidenceMethodMixin(ABC): 288 """Confidence Method Mixin class. 289 290 It initializes per-frame confidence method. 291 """ 292 293 def _init_confidence_method(self, confidence_method_cfg: Optional[DictConfig] = None): 294 """Initialize per-frame confidence method from config. 295 """ 296 # OmegaConf.structured ensures that post_init check is always executed 297 confidence_method_cfg = OmegaConf.structured( 298 ConfidenceMethodConfig() 299 if confidence_method_cfg is None 300 else ConfidenceMethodConfig(**confidence_method_cfg) 301 ) 302 303 # set confidence calculation method 304 # we suppose that self.blank_id == len(vocabulary) 305 self.num_tokens = (self.blank_id if hasattr(self, "blank_id") else self._blank_index) + 1 306 self.alpha = confidence_method_cfg.alpha 307 308 # init confidence measure bank 309 self.confidence_measure_bank = get_confidence_measure_bank() 310 311 measure = None 312 # construct measure_name 313 measure_name = "" 314 if confidence_method_cfg.name == "max_prob": 315 measure_name = "max_prob" 316 elif confidence_method_cfg.name == "entropy": 317 measure_name = '_'.join( 318 [confidence_method_cfg.name, confidence_method_cfg.entropy_type, confidence_method_cfg.entropy_norm] 319 ) 320 else: 321 raise ValueError(f"Unsupported `confidence_method_cfg.name`: `{confidence_method_cfg.name}`") 322 if measure_name not in self.confidence_measure_bank: 323 raise ValueError(f"Unsupported measure setup: `{measure_name}`") 324 measure = partial(self.confidence_measure_bank[measure_name], v=self.num_tokens, t=self.alpha) 325 self._get_confidence = lambda x: measure(torch.nan_to_num(x)).tolist() 326 327 328 class ConfidenceMixin(ABC): 329 """Confidence Mixin class. 330 331 It is responsible for confidence estimation method initialization and high-level confidence score calculation. 332 """ 333 334 def _init_confidence(self, confidence_cfg: Optional[DictConfig] = None): 335 """Initialize confidence-related fields and confidence aggregation function from config. 336 """ 337 # OmegaConf.structured ensures that post_init check is always executed 338 confidence_cfg = OmegaConf.structured( 339 ConfidenceConfig() if confidence_cfg is None else ConfidenceConfig(**confidence_cfg) 340 ) 341 self.confidence_method_cfg = confidence_cfg.method_cfg 342 343 # extract the config 344 self.preserve_word_confidence = confidence_cfg.get('preserve_word_confidence', False) 345 # set preserve_frame_confidence and preserve_token_confidence to True 346 # if preserve_word_confidence is True 347 self.preserve_token_confidence = ( 348 confidence_cfg.get('preserve_token_confidence', False) | self.preserve_word_confidence 349 ) 350 # set preserve_frame_confidence to True if preserve_token_confidence is True 351 self.preserve_frame_confidence = ( 352 confidence_cfg.get('preserve_frame_confidence', False) | self.preserve_token_confidence 353 ) 354 self.exclude_blank_from_confidence = confidence_cfg.get('exclude_blank', True) 355 self.word_confidence_aggregation = confidence_cfg.get('aggregation', "min") 356 357 # define aggregation functions 358 self.confidence_aggregation_bank = get_confidence_aggregation_bank() 359 self._aggregate_confidence = self.confidence_aggregation_bank[self.word_confidence_aggregation] 360 361 # Update preserve frame confidence 362 if self.preserve_frame_confidence is False: 363 if self.cfg.strategy in ['greedy', 'greedy_batch']: 364 self.preserve_frame_confidence = self.cfg.greedy.get('preserve_frame_confidence', False) 365 # OmegaConf.structured ensures that post_init check is always executed 366 confidence_method_cfg = OmegaConf.structured(self.cfg.greedy).get('confidence_method_cfg', None) 367 self.confidence_method_cfg = ( 368 OmegaConf.structured(ConfidenceMethodConfig()) 369 if confidence_method_cfg is None 370 else OmegaConf.structured(ConfidenceMethodConfig(**confidence_method_cfg)) 371 ) 372 373 @abstractmethod 374 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 375 """Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 376 Assumes that `frame_confidence` is present in the hypotheses. 377 378 Args: 379 hypotheses_list: List of Hypothesis. 380 381 Returns: 382 A list of hypotheses with high-level confidence scores. 383 """ 384 raise NotImplementedError() 385 386 @abstractmethod 387 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 388 """Implemented by subclass in order to aggregate token confidence to a word-level confidence. 389 390 Args: 391 hypothesis: Hypothesis 392 393 Returns: 394 A list of word-level confidence scores. 395 """ 396 raise NotImplementedError() 397 398 def _aggregate_token_confidence_chars(self, words: List[str], token_confidence: List[float]) -> List[float]: 399 """Implementation of token confidence aggregation for character-based models. 400 401 Args: 402 words: List of words of a hypothesis. 403 token_confidence: List of token-level confidence scores of a hypothesis. 404 405 Returns: 406 A list of word-level confidence scores. 407 """ 408 word_confidence = [] 409 i = 0 410 for word in words: 411 word_len = len(word) 412 word_confidence.append(self._aggregate_confidence(token_confidence[i : i + word_len])) 413 # we assume that there is exactly one space token between words and exclude it from word confidence 414 i += word_len + 1 415 return word_confidence 416 417 def _aggregate_token_confidence_subwords_sentencepiece( 418 self, words: List[str], token_confidence: List[float], token_ids: List[int] 419 ) -> List[float]: 420 """Implementation of token confidence aggregation for subword-based models. 421 422 **Note**: Only supports Sentencepiece based tokenizers ! 423 424 Args: 425 words: List of words of a hypothesis. 426 token_confidence: List of token-level confidence scores of a hypothesis. 427 token_ids: List of token ids of a hypothesis. 428 429 Returns: 430 A list of word-level confidence scores. 431 """ 432 word_confidence = [] 433 # run only if there are final words 434 if len(words) > 0: 435 j = 0 436 prev_unk = False 437 prev_underline = False 438 for i, token_id in enumerate(token_ids): 439 token = self.decode_ids_to_tokens([int(token_id)])[0] 440 token_text = self.decode_tokens_to_str([int(token_id)]) 441 # treat `<unk>` as a separate word regardless of the next token 442 # to match the result of `tokenizer.ids_to_text` 443 if (token != token_text or prev_unk) and i > j: 444 # do not add confidence for `โ–` if the current token starts with `โ–` 445 # to match the result of `tokenizer.ids_to_text` 446 if not prev_underline: 447 word_confidence.append(self._aggregate_confidence(token_confidence[j:i])) 448 j = i 449 prev_unk = token == '<unk>' 450 prev_underline = token == 'โ–' 451 if not prev_underline: 452 word_confidence.append(self._aggregate_confidence(token_confidence[j : len(token_ids)])) 453 if len(words) != len(word_confidence): 454 raise RuntimeError( 455 f"""Something went wrong with word-level confidence aggregation.\n 456 Please check these values for debugging:\n 457 len(words): {len(words)},\n 458 len(word_confidence): {len(word_confidence)},\n 459 recognized text: `{' '.join(words)}`""" 460 ) 461 return word_confidence 462 [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of nemo/collections/common/parts/adapter_modules.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, is_dataclass 16 from typing import Any, Optional 17 18 from hydra.utils import instantiate 19 from omegaconf import OmegaConf 20 from torch import nn as nn 21 22 from nemo.collections.common.parts.utils import activation_registry 23 from nemo.core.classes.mixins import access_mixins, adapter_mixin_strategies 24 25 26 class AdapterModuleUtil(access_mixins.AccessMixin): 27 """ 28 Base class of Adapter Modules, providing common functionality to all Adapter Modules. 29 """ 30 31 def setup_adapter_strategy(self, adapter_strategy: Optional[adapter_mixin_strategies.AbstractAdapterStrategy]): 32 """ 33 Setup adapter strategy of this class, enabling dynamic change in the way the adapter output is 34 merged with the input. 35 36 When called successfully, will assign the variable `adapter_strategy` to the module. 37 38 Args: 39 adapter_strategy: Can be a None or an implementation of AbstractAdapterStrategy. 40 """ 41 # set default adapter strategy 42 if adapter_strategy is None: 43 adapter_strategy = self.get_default_strategy_config() 44 45 if is_dataclass(adapter_strategy): 46 adapter_strategy = OmegaConf.structured(adapter_strategy) 47 OmegaConf.set_struct(adapter_strategy, False) 48 49 # The config must have the `_target_` field pointing to the actual adapter strategy class 50 # which will load that strategy dynamically to this module. 51 if isinstance(adapter_strategy, dict) or OmegaConf.is_config(adapter_strategy): 52 self.adapter_strategy = instantiate(adapter_strategy) 53 elif isinstance(adapter_strategy, adapter_mixin_strategies.AbstractAdapterStrategy): 54 self.adapter_strategy = adapter_strategy 55 else: 56 raise AttributeError(f'`adapter_strategy` provided is invalid : {adapter_strategy}') 57 58 def get_default_strategy_config(self) -> 'dataclass': 59 """ 60 Returns a default adapter module strategy. 61 """ 62 return adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 63 64 def adapter_unfreeze(self,): 65 """ 66 Sets the requires grad for all parameters in the adapter to True. 67 This method should be overridden for any custom unfreeze behavior that is required. 68 For example, if not all params of the adapter should be unfrozen. 69 """ 70 for param in self.parameters(): 71 param.requires_grad_(True) 72 73 74 class LinearAdapter(nn.Module, AdapterModuleUtil): 75 76 """ 77 Simple Linear Feedforward Adapter module with LayerNorm and singe hidden layer with activation function. 78 Note: The adapter explicitly initializes its final layer with all zeros in order to avoid affecting the 79 original model when all adapters are disabled. 80 81 Args: 82 in_features: Input dimension of the module. Note that for adapters, input_dim == output_dim. 83 dim: Hidden dimension of the feed forward network. 84 activation: Str name for an activation function. 85 norm_position: Str, can be `pre` or `post`. Defaults to `pre`. Determines whether the normalization 86 will occur in the first layer or the last layer. Certain architectures may prefer one over the other. 87 dropout: float value, whether to perform dropout on the output of the last layer of the adapter. 88 adapter_strategy: By default, ResidualAddAdapterStrategyConfig. An adapter composition function object. 89 """ 90 91 def __init__( 92 self, 93 in_features: int, 94 dim: int, 95 activation: str = 'swish', 96 norm_position: str = 'pre', 97 dropout: float = 0.0, 98 adapter_strategy: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig = None, 99 ): 100 super().__init__() 101 102 activation = activation_registry[activation]() 103 # If the activation can be executed in place, do so. 104 if hasattr(activation, 'inplace'): 105 activation.inplace = True 106 107 assert norm_position in ['pre', 'post'] 108 self.norm_position = norm_position 109 110 if norm_position == 'pre': 111 self.module = nn.Sequential( 112 nn.LayerNorm(in_features), 113 nn.Linear(in_features, dim, bias=False), 114 activation, 115 nn.Linear(dim, in_features, bias=False), 116 ) 117 118 elif norm_position == 'post': 119 self.module = nn.Sequential( 120 nn.Linear(in_features, dim, bias=False), 121 activation, 122 nn.Linear(dim, in_features, bias=False), 123 nn.LayerNorm(in_features), 124 ) 125 126 if dropout > 0.0: 127 self.dropout = nn.Dropout(dropout) 128 else: 129 self.dropout = None 130 131 # Setup adapter strategy 132 self.setup_adapter_strategy(adapter_strategy) 133 134 # reset parameters 135 self.reset_parameters() 136 137 def reset_parameters(self): 138 # Final layer initializations must be 0 139 if self.norm_position == 'pre': 140 self.module[-1].weight.data *= 0 141 142 elif self.norm_position == 'post': 143 self.module[-1].weight.data *= 0 144 self.module[-1].bias.data *= 0 145 146 def forward(self, x): 147 x = self.module(x) 148 149 # Add dropout if available 150 if self.dropout is not None: 151 x = self.dropout(x) 152 153 return x 154 155 156 @dataclass 157 class LinearAdapterConfig: 158 in_features: int 159 dim: int 160 activation: str = 'swish' 161 norm_position: str = 'pre' 162 dropout: float = 0.0 163 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 164 _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) 165 [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import re 15 from typing import List 16 17 import ipadic 18 import MeCab 19 from pangu import spacing 20 from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer 21 22 23 class EnJaProcessor: 24 """ 25 Tokenizer, Detokenizer and Normalizer utilities for Japanese & English 26 Args: 27 lang_id: One of ['en', 'ja']. 28 """ 29 30 def __init__(self, lang_id: str): 31 self.lang_id = lang_id 32 self.moses_tokenizer = MosesTokenizer(lang=lang_id) 33 self.moses_detokenizer = MosesDetokenizer(lang=lang_id) 34 self.normalizer = MosesPunctNormalizer( 35 lang=lang_id, pre_replace_unicode_punct=True, post_remove_control_chars=True 36 ) 37 38 def detokenize(self, tokens: List[str]) -> str: 39 """ 40 Detokenizes a list of tokens 41 Args: 42 tokens: list of strings as tokens 43 Returns: 44 detokenized Japanese or English string 45 """ 46 return self.moses_detokenizer.detokenize(tokens) 47 48 def tokenize(self, text) -> str: 49 """ 50 Tokenizes text using Moses. Returns a string of tokens. 51 """ 52 tokens = self.moses_tokenizer.tokenize(text) 53 return ' '.join(tokens) 54 55 def normalize(self, text) -> str: 56 # Normalization doesn't handle Japanese periods correctly; 57 # 'ใ€‚'becomes '.'. 58 if self.lang_id == 'en': 59 return self.normalizer.normalize(text) 60 else: 61 return text 62 63 64 class JaMecabProcessor: 65 """ 66 Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English 67 """ 68 69 def __init__(self): 70 self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") 71 72 def detokenize(self, text: List[str]) -> str: 73 RE_WS_IN_FW = re.compile( 74 r'([\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])\s+(?=[\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])' 75 ) 76 77 detokenize = lambda s: spacing(RE_WS_IN_FW.sub(r'\1', s)).strip() 78 return detokenize(' '.join(text)) 79 80 def tokenize(self, text) -> str: 81 """ 82 Tokenizes text using Moses. Returns a string of tokens. 83 """ 84 return self.mecab_tokenizer.parse(text).strip() 85 86 def normalize(self, text) -> str: 87 return text 88 [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Optional, Tuple 17 18 from omegaconf.omegaconf import MISSING 19 20 from nemo.collections.nlp.data.machine_translation.machine_translation_dataset import TranslationDataConfig 21 from nemo.collections.nlp.models.enc_dec_nlp_model import EncDecNLPModelConfig 22 from nemo.collections.nlp.modules.common.token_classifier import TokenClassifierConfig 23 from nemo.collections.nlp.modules.common.tokenizer_utils import TokenizerConfig 24 from nemo.collections.nlp.modules.common.transformer.transformer import ( 25 NeMoTransformerConfig, 26 NeMoTransformerEncoderConfig, 27 ) 28 from nemo.collections.nlp.modules.common.transformer.transformer_bottleneck import ( 29 NeMoTransformerBottleneckDecoderConfig, 30 NeMoTransformerBottleneckEncoderConfig, 31 ) 32 from nemo.core.config.modelPT import OptimConfig, SchedConfig 33 34 35 @dataclass 36 class MTSchedConfig(SchedConfig): 37 name: str = 'InverseSquareRootAnnealing' 38 warmup_ratio: Optional[float] = None 39 last_epoch: int = -1 40 41 42 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). 43 @dataclass 44 class MTOptimConfig(OptimConfig): 45 name: str = 'adam' 46 lr: float = 1e-3 47 betas: Tuple[float, float] = (0.9, 0.98) 48 weight_decay: float = 0.0 49 sched: Optional[MTSchedConfig] = MTSchedConfig() 50 51 52 @dataclass 53 class MTEncDecModelConfig(EncDecNLPModelConfig): 54 # machine translation configurations 55 num_val_examples: int = 3 56 num_test_examples: int = 3 57 max_generation_delta: int = 10 58 label_smoothing: Optional[float] = 0.0 59 beam_size: int = 4 60 len_pen: float = 0.0 61 src_language: Any = 'en' # Any = str or List[str] 62 tgt_language: Any = 'en' # Any = str or List[str] 63 find_unused_parameters: Optional[bool] = True 64 shared_tokenizer: Optional[bool] = True 65 multilingual: Optional[bool] = False 66 preproc_out_dir: Optional[str] = None 67 validate_input_ids: Optional[bool] = True 68 shared_embeddings: bool = False 69 70 # network architecture configuration 71 encoder_tokenizer: Any = MISSING 72 encoder: Any = MISSING 73 74 decoder_tokenizer: Any = MISSING 75 decoder: Any = MISSING 76 77 head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) 78 79 # dataset configurations 80 train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 81 src_file_name=MISSING, 82 tgt_file_name=MISSING, 83 tokens_in_batch=512, 84 clean=True, 85 shuffle=True, 86 cache_ids=False, 87 use_cache=False, 88 ) 89 validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 90 src_file_name=MISSING, 91 tgt_file_name=MISSING, 92 tokens_in_batch=512, 93 clean=False, 94 shuffle=False, 95 cache_ids=False, 96 use_cache=False, 97 ) 98 test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 99 src_file_name=MISSING, 100 tgt_file_name=MISSING, 101 tokens_in_batch=512, 102 clean=False, 103 shuffle=False, 104 cache_ids=False, 105 use_cache=False, 106 ) 107 optim: Optional[OptimConfig] = MTOptimConfig() 108 109 110 @dataclass 111 class AAYNBaseConfig(MTEncDecModelConfig): 112 113 # Attention is All You Need Base Configuration 114 encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 115 decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 116 117 encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( 118 library='nemo', 119 model_name=None, 120 pretrained=False, 121 hidden_size=512, 122 inner_size=2048, 123 num_layers=6, 124 num_attention_heads=8, 125 ffn_dropout=0.1, 126 attn_score_dropout=0.1, 127 attn_layer_dropout=0.1, 128 ) 129 130 decoder: NeMoTransformerConfig = NeMoTransformerConfig( 131 library='nemo', 132 model_name=None, 133 pretrained=False, 134 hidden_size=512, 135 inner_size=2048, 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass 145 class MTBottleneckModelConfig(AAYNBaseConfig): 146 model_type: str = 'nll' 147 min_logv: float = -6 148 latent_size: int = -1 # -1 will take value of encoder hidden 149 non_recon_warmup_batches: int = 200000 150 recon_per_token: bool = True 151 log_timing: bool = True 152 153 encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( 154 library='nemo', 155 model_name=None, 156 pretrained=False, 157 hidden_size=512, 158 inner_size=2048, 159 num_layers=6, 160 num_attention_heads=8, 161 ffn_dropout=0.1, 162 attn_score_dropout=0.1, 163 attn_layer_dropout=0.1, 164 arch='seq2seq', 165 hidden_steps=32, 166 hidden_blocks=1, 167 hidden_init_method='params', 168 ) 169 170 decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( 171 library='nemo', 172 model_name=None, 173 pretrained=False, 174 inner_size=2048, 175 num_layers=6, 176 num_attention_heads=8, 177 ffn_dropout=0.1, 178 attn_score_dropout=0.1, 179 attn_layer_dropout=0.1, 180 arch='seq2seq', 181 ) 182 [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, Optional 17 18 from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict 19 20 from nemo.collections.common.parts.adapter_modules import LinearAdapterConfig 21 from nemo.collections.nlp.data.token_classification.punctuation_capitalization_dataset import ( 22 PunctuationCapitalizationEvalDataConfig, 23 PunctuationCapitalizationTrainDataConfig, 24 legacy_data_config_to_new_data_config, 25 ) 26 from nemo.core.config import TrainerConfig 27 from nemo.core.config.modelPT import NemoConfig 28 from nemo.utils.exp_manager import ExpManagerConfig 29 30 31 @dataclass 32 class FreezeConfig: 33 is_enabled: bool = False 34 """Freeze audio encoder weight and add Conformer Layers on top of it""" 35 d_model: Optional[int] = 256 36 """`d_model` parameter of ``ConformerLayer``""" 37 d_ff: Optional[int] = 1024 38 """``d_ff`` parameter of ``ConformerLayer``""" 39 num_layers: Optional[int] = 8 40 """``num_layers`` number of ``ConformerLayer`` modules to add on top of audio encoder""" 41 42 43 @dataclass 44 class AdapterConfig: 45 config: Optional[LinearAdapterConfig] = None 46 """Linear adapter config see ``collections.common.parts.LinearAdapterConfig``""" 47 enable: bool = False 48 """Use adapters for audio encoder""" 49 50 51 @dataclass 52 class FusionConfig: 53 num_layers: Optional[int] = 4 54 """"Number of layers to use in fusion""" 55 num_attention_heads: Optional[int] = 4 56 """Number of attention heads to use in fusion""" 57 inner_size: Optional[int] = 2048 58 """Fusion inner size""" 59 60 61 @dataclass 62 class AudioEncoderConfig: 63 pretrained_model: str = MISSING 64 """A configuration for restoring pretrained audio encoder""" 65 freeze: Optional[FreezeConfig] = None 66 adapter: Optional[AdapterConfig] = None 67 fusion: Optional[FusionConfig] = None 68 69 70 @dataclass 71 class TokenizerConfig: 72 """A structure and default values of source text tokenizer.""" 73 74 vocab_file: Optional[str] = None 75 """A path to vocabulary file which is used in ``'word'``, ``'char'``, and HuggingFace tokenizers""" 76 77 tokenizer_name: str = MISSING 78 """A name of the tokenizer used for tokenization of source sequences. Possible options are ``'sentencepiece'``, 79 ``'word'``, ``'char'``, HuggingFace tokenizers (e.g. ``'bert-base-uncased'``). For more options see function 80 ``nemo.collections.nlp.modules.common.get_tokenizer``. The tokenizer must have properties ``cls_id``, ``pad_id``, 81 ``sep_id``, ``unk_id``.""" 82 83 special_tokens: Optional[Dict[str, str]] = None 84 """A dictionary with special tokens passed to constructors of ``'char'``, ``'word'``, ``'sentencepiece'``, and 85 various HuggingFace tokenizers.""" 86 87 tokenizer_model: Optional[str] = None 88 """A path to a tokenizer model required for ``'sentencepiece'`` tokenizer.""" 89 90 91 @dataclass 92 class LanguageModelConfig: 93 """ 94 A structure and default values of language model configuration of punctuation and capitalization model. BERT like 95 HuggingFace models are supported. Provide a valid ``pretrained_model_name`` and, optionally, you may 96 reinitialize model via ``config_file`` or ``config``. 97 98 Alternatively you can initialize the language model using ``lm_checkpoint``. 99 100 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 101 """ 102 103 pretrained_model_name: str = MISSING 104 """A mandatory parameter containing name of HuggingFace pretrained model. For example, ``'bert-base-uncased'``.""" 105 106 config_file: Optional[str] = None 107 """A path to a file with HuggingFace model config which is used to reinitialize language model.""" 108 109 config: Optional[Dict] = None 110 """A HuggingFace config which is used to reinitialize language model.""" 111 112 lm_checkpoint: Optional[str] = None 113 """A path to a ``torch`` checkpoint of a language model.""" 114 115 116 @dataclass 117 class HeadConfig: 118 """ 119 A structure and default values of configuration of capitalization or punctuation model head. This config defines a 120 multilayer perceptron which is applied to output of a language model. Number of units in the hidden layer is equal 121 to the dimension of the language model. 122 123 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 124 """ 125 126 num_fc_layers: int = 1 127 """A number of hidden layers in a multilayer perceptron.""" 128 129 fc_dropout: float = 0.1 130 """A dropout used in an MLP.""" 131 132 activation: str = 'relu' 133 """An activation used in hidden layers.""" 134 135 use_transformer_init: bool = True 136 """Whether to initialize the weights of the classifier head with the approach that was used for language model 137 initialization.""" 138 139 140 @dataclass 141 class ClassLabelsConfig: 142 """ 143 A structure and default values of a mandatory part of config which contains names of files which are saved in .nemo 144 checkpoint. These files can also be used for passing label vocabulary to the model. For using them as label 145 vocabularies you will need to provide path these files in parameter 146 ``model.common_dataset_parameters.label_vocab_dir``. Each line in labels files 147 contains 1 label. The values are sorted, ``<line number>==<label id>``, starting from ``0``. A label with ``0`` id 148 must contain neutral label which must be equal to ``model.common_dataset_parameters.pad_label``. 149 150 This config is a part of :class:`~CommonDatasetParametersConfig`. 151 """ 152 153 punct_labels_file: str = MISSING 154 """A name of punctuation labels file.""" 155 156 capit_labels_file: str = MISSING 157 """A name of capitalization labels file.""" 158 159 160 @dataclass 161 class CommonDatasetParametersConfig: 162 """ 163 A structure and default values of common dataset parameters config which includes label and loss mask information. 164 If you omit parameters ``punct_label_ids``, ``capit_label_ids``, ``label_vocab_dir``, then labels will be inferred 165 from a training dataset or loaded from a checkpoint. 166 167 Parameters ``ignore_extra_tokens`` and ``ignore_start_end`` are responsible for forming loss mask. A loss mask 168 defines on which tokens loss is computed. 169 170 This parameter is a part of config :class:`~PunctuationCapitalizationModelConfig`. 171 """ 172 173 pad_label: str = MISSING 174 """A mandatory parameter which should contain label used for punctuation and capitalization label padding. It 175 also serves as a neutral label for both punctuation and capitalization. If any of ``punct_label_ids``, 176 ``capit_label_ids`` parameters is provided, then ``pad_label`` must have ``0`` id in them. In addition, if ``label_vocab_dir`` 177 is provided, then ``pad_label`` must be on the first lines in files ``class_labels.punct_labels_file`` and 178 ``class_labels.capit_labels_file``.""" 179 180 ignore_extra_tokens: bool = False 181 """Whether to compute loss on not first tokens in words. If this parameter is ``True``, then loss mask is ``False`` 182 for all tokens in a word except the first.""" 183 184 ignore_start_end: bool = True 185 """If ``False``, then loss is computed on [CLS] and [SEP] tokens.""" 186 187 punct_label_ids: Optional[Dict[str, int]] = None 188 """A dictionary with punctuation label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this 189 parameter and pass label ids through ``class_labels.punct_labels_file`` or let the model to infer label ids from 190 dataset or load them from checkpoint.""" 191 192 capit_label_ids: Optional[Dict[str, int]] = None 193 """A dictionary with capitalization label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit 194 this parameter and pass label ids through ``class_labels.capit_labels_file`` or let model to infer label ids from 195 dataset or load them from checkpoint.""" 196 197 label_vocab_dir: Optional[str] = None 198 """A path to directory which contains class labels files. See :class:`ClassLabelsConfig`. If this parameter is 199 provided, then labels will be loaded from files which are located in ``label_vocab_dir`` and have names specified 200 in ``model.class_labels`` configuration section. A label specified in ``pad_label`` has to be on the first lines 201 of ``model.class_labels`` files.""" 202 203 204 @dataclass 205 class PunctuationCapitalizationModelConfig: 206 """ 207 A configuration of 208 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 209 model. 210 211 See an example of model config in 212 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 213 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 214 215 This config is a part of :class:`~PunctuationCapitalizationConfig`. 216 """ 217 218 class_labels: ClassLabelsConfig = ClassLabelsConfig() 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" 229 230 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 231 """A configuration for creating validation datasets and data loaders.""" 232 233 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 234 """A configuration for creating test datasets and data loaders.""" 235 236 punct_head: HeadConfig = HeadConfig() 237 """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" 238 239 capit_head: HeadConfig = HeadConfig() 240 """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" 241 242 tokenizer: Any = TokenizerConfig() 243 """A configuration for source text tokenizer.""" 244 245 language_model: LanguageModelConfig = LanguageModelConfig() 246 """A configuration of a BERT-like language model which serves as a model body.""" 247 248 optim: Optional[Any] = None 249 """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For 250 description see `Optimizers 251 <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/core/core.html#optimizers>`_ section in 252 documentation and `primer <https://github.com/NVIDIA/NeMo/blob/main/tutorials/00_NeMo_Primer.ipynb>_ tutorial.""" 253 254 255 @dataclass 256 class PunctuationCapitalizationLexicalAudioModelConfig(PunctuationCapitalizationModelConfig): 257 """ 258 A configuration of 259 :class:`~nemo.collections.nlp.models.token_classification.punctuation_lexical_audio_capitalization_model.PunctuationCapitalizationLexicalAudioModel` 260 model. 261 262 See an example of model config in 263 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 264 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_lexical_audio_config.yaml>`_ 265 266 Audio encoder can be frozen during training with ``freeze_audio_encoder`` parameter. 267 Adapter can be added to audio encoder with ``use_adapters`` and ``adapter_config`` parameters. 268 More conformer layers can be added on top of pretrained audio encoder with ``frozen_conf_d_model``, ``frozen_conf_d_ff`` and ``frozen_conf_num_layers`` parameters. 269 """ 270 271 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 272 """A configuration for creating training dataset and data loader.""" 273 274 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 275 """A configuration for creating validation datasets and data loaders.""" 276 277 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 278 """A configuration for creating test datasets and data loaders.""" 279 280 audio_encoder: Optional[AudioEncoderConfig] = None 281 282 restore_lexical_encoder_from: Optional[str] = None 283 """"Path to .nemo checkpoint to load weights from""" # add more comments 284 285 use_weighted_loss: Optional[bool] = False 286 """If set to ``True`` CrossEntropyLoss will be weighted""" 287 288 289 @dataclass 290 class PunctuationCapitalizationConfig(NemoConfig): 291 """ 292 A config for punctuation model training and testing. 293 294 See an example of full config in 295 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 296 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 297 """ 298 299 pretrained_model: Optional[str] = None 300 """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options 301 by calling method 302 :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. 303 """ 304 305 name: Optional[str] = 'Punctuation_and_Capitalization' 306 """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" 307 308 do_training: bool = True 309 """Whether to perform training of the model.""" 310 311 do_testing: bool = False 312 """Whether ot perform testing of the model.""" 313 314 model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() 315 """A configuration for the 316 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 317 model.""" 318 319 trainer: Optional[TrainerConfig] = TrainerConfig() 320 """Contains ``Trainer`` Lightning class constructor parameters.""" 321 322 exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) 323 """A configuration with various NeMo training options such as output directories, resuming from checkpoint, 324 tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" 325 326 327 @dataclass 328 class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): 329 model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() 330 331 332 def is_legacy_model_config(model_cfg: DictConfig) -> bool: 333 """ 334 Test if model config is old style config. Old style configs are configs which were used before 335 ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of 336 ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support 337 tarred datasets. 338 339 Args: 340 model_cfg: model configuration 341 342 Returns: 343 whether ``model_config`` is legacy 344 """ 345 return 'common_dataset_parameters' not in model_cfg 346 347 348 def legacy_model_config_to_new_model_config(model_cfg: DictConfig) -> DictConfig: 349 """ 350 Transform old style config into 351 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig`. 352 Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style 353 datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. 354 Old style configs do not support tarred datasets. 355 356 Args: 357 model_cfg: old style config 358 359 Returns: 360 model config which follows dataclass 361 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig` 362 """ 363 train_ds = model_cfg.get('train_ds') 364 validation_ds = model_cfg.get('validation_ds') 365 test_ds = model_cfg.get('test_ds') 366 dataset = model_cfg.dataset 367 punct_head_config = model_cfg.get('punct_head', {}) 368 capit_head_config = model_cfg.get('capit_head', {}) 369 omega_conf = OmegaConf.structured( 370 PunctuationCapitalizationModelConfig( 371 class_labels=model_cfg.class_labels, 372 common_dataset_parameters=CommonDatasetParametersConfig( 373 pad_label=dataset.pad_label, 374 ignore_extra_tokens=dataset.get( 375 'ignore_extra_tokens', CommonDatasetParametersConfig.ignore_extra_tokens 376 ), 377 ignore_start_end=dataset.get('ignore_start_end', CommonDatasetParametersConfig.ignore_start_end), 378 punct_label_ids=model_cfg.punct_label_ids, 379 capit_label_ids=model_cfg.capit_label_ids, 380 ), 381 train_ds=None 382 if train_ds is None 383 else legacy_data_config_to_new_data_config(train_ds, dataset, train=True), 384 validation_ds=None 385 if validation_ds is None 386 else legacy_data_config_to_new_data_config(validation_ds, dataset, train=False), 387 test_ds=None if test_ds is None else legacy_data_config_to_new_data_config(test_ds, dataset, train=False), 388 punct_head=HeadConfig( 389 num_fc_layers=punct_head_config.get('punct_num_fc_layers', HeadConfig.num_fc_layers), 390 fc_dropout=punct_head_config.get('fc_dropout', HeadConfig.fc_dropout), 391 activation=punct_head_config.get('activation', HeadConfig.activation), 392 use_transformer_init=punct_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 393 ), 394 capit_head=HeadConfig( 395 num_fc_layers=capit_head_config.get('capit_num_fc_layers', HeadConfig.num_fc_layers), 396 fc_dropout=capit_head_config.get('fc_dropout', HeadConfig.fc_dropout), 397 activation=capit_head_config.get('activation', HeadConfig.activation), 398 use_transformer_init=capit_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 399 ), 400 tokenizer=model_cfg.tokenizer, 401 language_model=model_cfg.language_model, 402 optim=model_cfg.optim, 403 ) 404 ) 405 with open_dict(omega_conf): 406 retain_during_legacy_conversion = model_cfg.get('retain_during_legacy_conversion', {}) 407 for key in retain_during_legacy_conversion.keys(): 408 omega_conf[key] = retain_during_legacy_conversion[key] 409 return omega_conf 410 [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """Transformer based language model.""" 16 from MeCab import Model 17 from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule 18 from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule 19 from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( 20 MegatronRetrievalTransformerEncoderModule, 21 ) 22 from nemo.collections.nlp.modules.common.megatron.utils import ( 23 ApexGuardDefaults, 24 init_method_normal, 25 scaled_init_method_normal, 26 ) 27 28 try: 29 from apex.transformer.enums import AttnMaskType, ModelType 30 31 HAVE_APEX = True 32 except (ImportError, ModuleNotFoundError): 33 HAVE_APEX = False 34 # fake missing classes with None attributes 35 AttnMaskType = ApexGuardDefaults() 36 ModelType = ApexGuardDefaults() 37 38 try: 39 from megatron.core import ModelParallelConfig 40 41 HAVE_MEGATRON_CORE = True 42 43 except (ImportError, ModuleNotFoundError): 44 45 ModelParallelConfig = ApexGuardDefaults 46 47 HAVE_MEGATRON_CORE = False 48 49 __all__ = [] 50 51 AVAILABLE_ENCODERS = ["transformer", "perceiver", "retro"] 52 53 54 def get_encoder_model( 55 config: ModelParallelConfig, 56 arch, 57 hidden_size, 58 ffn_hidden_size, 59 num_layers, 60 num_attention_heads, 61 apply_query_key_layer_scaling=False, 62 kv_channels=None, 63 init_method=None, 64 scaled_init_method=None, 65 encoder_attn_mask_type=AttnMaskType.padding, 66 pre_process=True, 67 post_process=True, 68 init_method_std=0.02, 69 megatron_amp_O2=False, 70 hidden_dropout=0.1, 71 attention_dropout=0.1, 72 ffn_dropout=0.0, 73 precision=16, 74 fp32_residual_connection=False, 75 activations_checkpoint_method=None, 76 activations_checkpoint_num_layers=1, 77 activations_checkpoint_granularity=None, 78 layernorm_epsilon=1e-5, 79 bias_activation_fusion=True, 80 bias_dropout_add_fusion=True, 81 masked_softmax_fusion=True, 82 persist_layer_norm=False, 83 openai_gelu=False, 84 activation="gelu", 85 onnx_safe=False, 86 bias=True, 87 normalization="layernorm", 88 headscale=False, 89 transformer_block_type="pre_ln", 90 hidden_steps=32, 91 parent_model_type=ModelType.encoder_or_decoder, 92 layer_type=None, 93 chunk_size=64, 94 num_self_attention_per_cross_attention=1, 95 layer_number_offset=0, # this is use only for attention norm_factor scaling 96 megatron_legacy=False, 97 normalize_attention_scores=True, 98 sequence_parallel=False, 99 num_moe_experts=1, 100 moe_frequency=1, 101 moe_dropout=0.0, 102 turn_off_rop=False, # turn off the RoP positional embedding 103 version=1, # model version 104 position_embedding_type='learned_absolute', 105 use_flash_attention=False, 106 ): 107 """Build language model and return along with the key to save.""" 108 109 if kv_channels is None: 110 assert ( 111 hidden_size % num_attention_heads == 0 112 ), 'hidden_size must be divisible by num_attention_heads if kv_channels is None' 113 kv_channels = hidden_size // num_attention_heads 114 115 if init_method is None: 116 init_method = init_method_normal(init_method_std) 117 118 if scaled_init_method is None: 119 scaled_init_method = scaled_init_method_normal(init_method_std, num_layers) 120 121 if arch == "transformer": 122 # Language encoder. 123 encoder = MegatronTransformerEncoderModule( 124 config=config, 125 init_method=init_method, 126 output_layer_init_method=scaled_init_method, 127 hidden_size=hidden_size, 128 num_layers=num_layers, 129 num_attention_heads=num_attention_heads, 130 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 131 kv_channels=kv_channels, 132 ffn_hidden_size=ffn_hidden_size, 133 encoder_attn_mask_type=encoder_attn_mask_type, 134 pre_process=pre_process, 135 post_process=post_process, 136 megatron_amp_O2=megatron_amp_O2, 137 hidden_dropout=hidden_dropout, 138 attention_dropout=attention_dropout, 139 ffn_dropout=ffn_dropout, 140 precision=precision, 141 fp32_residual_connection=fp32_residual_connection, 142 activations_checkpoint_method=activations_checkpoint_method, 143 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 144 activations_checkpoint_granularity=activations_checkpoint_granularity, 145 layernorm_epsilon=layernorm_epsilon, 146 bias_activation_fusion=bias_activation_fusion, 147 bias_dropout_add_fusion=bias_dropout_add_fusion, 148 masked_softmax_fusion=masked_softmax_fusion, 149 persist_layer_norm=persist_layer_norm, 150 openai_gelu=openai_gelu, 151 onnx_safe=onnx_safe, 152 activation=activation, 153 bias=bias, 154 normalization=normalization, 155 transformer_block_type=transformer_block_type, 156 headscale=headscale, 157 parent_model_type=parent_model_type, 158 megatron_legacy=megatron_legacy, 159 normalize_attention_scores=normalize_attention_scores, 160 num_moe_experts=num_moe_experts, 161 moe_frequency=moe_frequency, 162 moe_dropout=moe_dropout, 163 position_embedding_type=position_embedding_type, 164 use_flash_attention=use_flash_attention, 165 ) 166 elif arch == "retro": 167 encoder = MegatronRetrievalTransformerEncoderModule( 168 config=config, 169 init_method=init_method, 170 output_layer_init_method=scaled_init_method, 171 hidden_size=hidden_size, 172 num_layers=num_layers, 173 num_attention_heads=num_attention_heads, 174 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 175 kv_channels=kv_channels, 176 layer_type=layer_type, 177 ffn_hidden_size=ffn_hidden_size, 178 pre_process=pre_process, 179 post_process=post_process, 180 megatron_amp_O2=megatron_amp_O2, 181 hidden_dropout=hidden_dropout, 182 attention_dropout=attention_dropout, 183 precision=precision, 184 fp32_residual_connection=fp32_residual_connection, 185 activations_checkpoint_method=activations_checkpoint_method, 186 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 187 activations_checkpoint_granularity=activations_checkpoint_granularity, 188 layernorm_epsilon=layernorm_epsilon, 189 bias_activation_fusion=bias_activation_fusion, 190 bias_dropout_add_fusion=bias_dropout_add_fusion, 191 masked_softmax_fusion=masked_softmax_fusion, 192 persist_layer_norm=persist_layer_norm, 193 openai_gelu=openai_gelu, 194 onnx_safe=onnx_safe, 195 activation=activation, 196 bias=bias, 197 normalization=normalization, 198 transformer_block_type=transformer_block_type, 199 parent_model_type=parent_model_type, 200 chunk_size=chunk_size, 201 layer_number_offset=layer_number_offset, 202 megatron_legacy=megatron_legacy, 203 normalize_attention_scores=normalize_attention_scores, 204 turn_off_rop=turn_off_rop, 205 version=version, 206 ) 207 elif arch == "perceiver": 208 encoder = MegatronPerceiverEncoderModule( 209 config=config, 210 init_method=init_method, 211 output_layer_init_method=scaled_init_method, 212 hidden_size=hidden_size, 213 num_layers=num_layers, 214 num_attention_heads=num_attention_heads, 215 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 216 kv_channels=kv_channels, 217 ffn_hidden_size=ffn_hidden_size, 218 encoder_attn_mask_type=encoder_attn_mask_type, 219 pre_process=pre_process, 220 post_process=post_process, 221 megatron_amp_O2=megatron_amp_O2, 222 hidden_dropout=hidden_dropout, 223 attention_dropout=attention_dropout, 224 ffn_dropout=ffn_dropout, 225 precision=precision, 226 fp32_residual_connection=fp32_residual_connection, 227 activations_checkpoint_method=activations_checkpoint_method, 228 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 229 activations_checkpoint_granularity=activations_checkpoint_granularity, 230 layernorm_epsilon=layernorm_epsilon, 231 bias_activation_fusion=bias_activation_fusion, 232 bias_dropout_add_fusion=bias_dropout_add_fusion, 233 masked_softmax_fusion=masked_softmax_fusion, 234 persist_layer_norm=persist_layer_norm, 235 openai_gelu=openai_gelu, 236 onnx_safe=onnx_safe, 237 activation=activation, 238 bias=bias, 239 normalization=normalization, 240 transformer_block_type=transformer_block_type, 241 headscale=headscale, 242 parent_model_type=parent_model_type, 243 hidden_steps=hidden_steps, 244 num_self_attention_per_cross_attention=num_self_attention_per_cross_attention, 245 megatron_legacy=megatron_legacy, 246 normalize_attention_scores=normalize_attention_scores, 247 ) 248 else: 249 raise ValueError(f"Unknown encoder arch = {arch}. Available encoder arch = {AVAILABLE_ENCODERS}") 250 251 return encoder 252 [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/tts/models/fastpitch.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import contextlib 15 from dataclasses import dataclass 16 from pathlib import Path 17 from typing import List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import DictConfig, OmegaConf, open_dict 22 from pytorch_lightning import Trainer 23 from pytorch_lightning.loggers import TensorBoardLogger 24 25 from nemo.collections.common.parts.preprocessing import parsers 26 from nemo.collections.tts.losses.aligner_loss import BinLoss, ForwardSumLoss 27 from nemo.collections.tts.losses.fastpitchloss import DurationLoss, EnergyLoss, MelLoss, PitchLoss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.modules.fastpitch import FastPitchModule 30 from nemo.collections.tts.parts.mixins import FastPitchAdapterModelMixin 31 from nemo.collections.tts.parts.utils.callbacks import LoggingCallback 32 from nemo.collections.tts.parts.utils.helpers import ( 33 batch_from_ragged, 34 g2p_backward_compatible_support, 35 plot_alignment_to_numpy, 36 plot_spectrogram_to_numpy, 37 process_batch, 38 sample_tts_input, 39 ) 40 from nemo.core.classes import Exportable 41 from nemo.core.classes.common import PretrainedModelInfo, typecheck 42 from nemo.core.neural_types.elements import ( 43 Index, 44 LengthsType, 45 MelSpectrogramType, 46 ProbsType, 47 RegressionValuesType, 48 TokenDurationType, 49 TokenIndex, 50 TokenLogDurationType, 51 ) 52 from nemo.core.neural_types.neural_type import NeuralType 53 from nemo.utils import logging, model_utils 54 55 56 @dataclass 57 class G2PConfig: 58 _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" 59 phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" 60 heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" 61 phoneme_probability: float = 0.5 62 63 64 @dataclass 65 class TextTokenizer: 66 _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" 67 punct: bool = True 68 stresses: bool = True 69 chars: bool = True 70 apostrophe: bool = True 71 pad_with_space: bool = True 72 add_blank_at: bool = True 73 g2p: G2PConfig = G2PConfig() 74 75 76 @dataclass 77 class TextTokenizerConfig: 78 text_tokenizer: TextTokenizer = TextTokenizer() 79 80 81 class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): 82 """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" 83 84 def __init__(self, cfg: DictConfig, trainer: Trainer = None): 85 # Convert to Hydra 1.0 compatible DictConfig 86 cfg = model_utils.convert_model_config_to_dict_config(cfg) 87 cfg = model_utils.maybe_update_config_version(cfg) 88 89 # Setup normalizer 90 self.normalizer = None 91 self.text_normalizer_call = None 92 self.text_normalizer_call_kwargs = {} 93 self._setup_normalizer(cfg) 94 95 self.learn_alignment = cfg.get("learn_alignment", False) 96 97 # Setup vocabulary (=tokenizer) and input_fft_kwargs (supported only with self.learn_alignment=True) 98 input_fft_kwargs = {} 99 if self.learn_alignment: 100 self.vocab = None 101 102 self.ds_class = cfg.train_ds.dataset._target_ 103 self.ds_class_name = self.ds_class.split(".")[-1] 104 if not self.ds_class in [ 105 "nemo.collections.tts.data.dataset.TTSDataset", 106 "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset", 107 "nemo.collections.tts.torch.data.TTSDataset", 108 ]: 109 raise ValueError(f"Unknown dataset class: {self.ds_class}.") 110 111 self._setup_tokenizer(cfg) 112 assert self.vocab is not None 113 input_fft_kwargs["n_embed"] = len(self.vocab.tokens) 114 input_fft_kwargs["padding_idx"] = self.vocab.pad 115 116 self._parser = None 117 self._tb_logger = None 118 super().__init__(cfg=cfg, trainer=trainer) 119 120 self.bin_loss_warmup_epochs = cfg.get("bin_loss_warmup_epochs", 100) 121 self.log_images = cfg.get("log_images", False) 122 self.log_train_images = False 123 124 default_prosody_loss_scale = 0.1 if self.learn_alignment else 1.0 125 dur_loss_scale = cfg.get("dur_loss_scale", default_prosody_loss_scale) 126 pitch_loss_scale = cfg.get("pitch_loss_scale", default_prosody_loss_scale) 127 energy_loss_scale = cfg.get("energy_loss_scale", default_prosody_loss_scale) 128 129 self.mel_loss_fn = MelLoss() 130 self.pitch_loss_fn = PitchLoss(loss_scale=pitch_loss_scale) 131 self.duration_loss_fn = DurationLoss(loss_scale=dur_loss_scale) 132 self.energy_loss_fn = EnergyLoss(loss_scale=energy_loss_scale) 133 134 self.aligner = None 135 if self.learn_alignment: 136 aligner_loss_scale = cfg.get("aligner_loss_scale", 1.0) 137 self.aligner = instantiate(self._cfg.alignment_module) 138 self.forward_sum_loss_fn = ForwardSumLoss(loss_scale=aligner_loss_scale) 139 self.bin_loss_fn = BinLoss(loss_scale=aligner_loss_scale) 140 141 self.preprocessor = instantiate(self._cfg.preprocessor) 142 input_fft = instantiate(self._cfg.input_fft, **input_fft_kwargs) 143 output_fft = instantiate(self._cfg.output_fft) 144 duration_predictor = instantiate(self._cfg.duration_predictor) 145 pitch_predictor = instantiate(self._cfg.pitch_predictor) 146 speaker_encoder = instantiate(self._cfg.get("speaker_encoder", None)) 147 energy_embedding_kernel_size = cfg.get("energy_embedding_kernel_size", 0) 148 energy_predictor = instantiate(self._cfg.get("energy_predictor", None)) 149 150 # [TODO] may remove if we change the pre-trained config 151 # cfg: condition_types = [ "add" ] 152 n_speakers = cfg.get("n_speakers", 0) 153 speaker_emb_condition_prosody = cfg.get("speaker_emb_condition_prosody", False) 154 speaker_emb_condition_decoder = cfg.get("speaker_emb_condition_decoder", False) 155 speaker_emb_condition_aligner = cfg.get("speaker_emb_condition_aligner", False) 156 min_token_duration = cfg.get("min_token_duration", 0) 157 use_log_energy = cfg.get("use_log_energy", True) 158 if n_speakers > 1 and "add" not in input_fft.cond_input.condition_types: 159 input_fft.cond_input.condition_types.append("add") 160 if speaker_emb_condition_prosody: 161 duration_predictor.cond_input.condition_types.append("add") 162 pitch_predictor.cond_input.condition_types.append("add") 163 if speaker_emb_condition_decoder: 164 output_fft.cond_input.condition_types.append("add") 165 if speaker_emb_condition_aligner and self.aligner is not None: 166 self.aligner.cond_input.condition_types.append("add") 167 168 self.fastpitch = FastPitchModule( 169 input_fft, 170 output_fft, 171 duration_predictor, 172 pitch_predictor, 173 energy_predictor, 174 self.aligner, 175 speaker_encoder, 176 n_speakers, 177 cfg.symbols_embedding_dim, 178 cfg.pitch_embedding_kernel_size, 179 energy_embedding_kernel_size, 180 cfg.n_mel_channels, 181 min_token_duration, 182 cfg.max_token_duration, 183 use_log_energy, 184 ) 185 self._input_types = self._output_types = None 186 self.export_config = { 187 "emb_range": (0, self.fastpitch.encoder.word_emb.num_embeddings), 188 "enable_volume": False, 189 "enable_ragged_batches": False, 190 } 191 if self.fastpitch.speaker_emb is not None: 192 self.export_config["num_speakers"] = cfg.n_speakers 193 194 self.log_config = cfg.get("log_config", None) 195 196 # Adapter modules setup (from FastPitchAdapterModelMixin) 197 self.setup_adapters() 198 199 def _get_default_text_tokenizer_conf(self): 200 text_tokenizer: TextTokenizerConfig = TextTokenizerConfig() 201 return OmegaConf.create(OmegaConf.to_yaml(text_tokenizer)) 202 203 def _setup_normalizer(self, cfg): 204 if "text_normalizer" in cfg: 205 normalizer_kwargs = {} 206 207 if "whitelist" in cfg.text_normalizer: 208 normalizer_kwargs["whitelist"] = self.register_artifact( 209 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 210 ) 211 try: 212 import nemo_text_processing 213 214 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 215 except Exception as e: 216 logging.error(e) 217 raise ImportError( 218 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 219 ) 220 221 self.text_normalizer_call = self.normalizer.normalize 222 if "text_normalizer_call_kwargs" in cfg: 223 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 224 225 def _setup_tokenizer(self, cfg): 226 text_tokenizer_kwargs = {} 227 228 if "g2p" in cfg.text_tokenizer: 229 # for backward compatibility 230 if ( 231 self._is_model_being_restored() 232 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 233 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 234 ): 235 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 236 cfg.text_tokenizer.g2p["_target_"] 237 ) 238 239 g2p_kwargs = {} 240 241 if "phoneme_dict" in cfg.text_tokenizer.g2p: 242 g2p_kwargs["phoneme_dict"] = self.register_artifact( 243 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 244 ) 245 246 if "heteronyms" in cfg.text_tokenizer.g2p: 247 g2p_kwargs["heteronyms"] = self.register_artifact( 248 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 249 ) 250 251 # for backward compatability 252 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 253 254 # TODO @xueyang: rename the instance of tokenizer because vocab is misleading. 255 self.vocab = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 256 257 @property 258 def tb_logger(self): 259 if self._tb_logger is None: 260 if self.logger is None and self.logger.experiment is None: 261 return None 262 tb_logger = self.logger.experiment 263 for logger in self.trainer.loggers: 264 if isinstance(logger, TensorBoardLogger): 265 tb_logger = logger.experiment 266 break 267 self._tb_logger = tb_logger 268 return self._tb_logger 269 270 @property 271 def parser(self): 272 if self._parser is not None: 273 return self._parser 274 275 if self.learn_alignment: 276 self._parser = self.vocab.encode 277 else: 278 self._parser = parsers.make_parser( 279 labels=self._cfg.labels, 280 name='en', 281 unk_id=-1, 282 blank_id=-1, 283 do_normalize=True, 284 abbreviation_version="fastpitch", 285 make_table=False, 286 ) 287 return self._parser 288 289 def parse(self, str_input: str, normalize=True) -> torch.tensor: 290 if self.training: 291 logging.warning("parse() is meant to be called in eval mode.") 292 293 if normalize and self.text_normalizer_call is not None: 294 str_input = self.text_normalizer_call(str_input, **self.text_normalizer_call_kwargs) 295 296 if self.learn_alignment: 297 eval_phon_mode = contextlib.nullcontext() 298 if hasattr(self.vocab, "set_phone_prob"): 299 eval_phon_mode = self.vocab.set_phone_prob(prob=1.0) 300 301 # Disable mixed g2p representation if necessary 302 with eval_phon_mode: 303 tokens = self.parser(str_input) 304 else: 305 tokens = self.parser(str_input) 306 307 x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device) 308 return x 309 310 @typecheck( 311 input_types={ 312 "text": NeuralType(('B', 'T_text'), TokenIndex()), 313 "durs": NeuralType(('B', 'T_text'), TokenDurationType()), 314 "pitch": NeuralType(('B', 'T_audio'), RegressionValuesType()), 315 "energy": NeuralType(('B', 'T_audio'), RegressionValuesType(), optional=True), 316 "speaker": NeuralType(('B'), Index(), optional=True), 317 "pace": NeuralType(optional=True), 318 "spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 319 "attn_prior": NeuralType(('B', 'T_spec', 'T_text'), ProbsType(), optional=True), 320 "mel_lens": NeuralType(('B'), LengthsType(), optional=True), 321 "input_lens": NeuralType(('B'), LengthsType(), optional=True), 322 # reference_* data is used for multi-speaker FastPitch training 323 "reference_spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 324 "reference_spec_lens": NeuralType(('B'), LengthsType(), optional=True), 325 } 326 ) 327 def forward( 328 self, 329 *, 330 text, 331 durs=None, 332 pitch=None, 333 energy=None, 334 speaker=None, 335 pace=1.0, 336 spec=None, 337 attn_prior=None, 338 mel_lens=None, 339 input_lens=None, 340 reference_spec=None, 341 reference_spec_lens=None, 342 ): 343 return self.fastpitch( 344 text=text, 345 durs=durs, 346 pitch=pitch, 347 energy=energy, 348 speaker=speaker, 349 pace=pace, 350 spec=spec, 351 attn_prior=attn_prior, 352 mel_lens=mel_lens, 353 input_lens=input_lens, 354 reference_spec=reference_spec, 355 reference_spec_lens=reference_spec_lens, 356 ) 357 358 @typecheck(output_types={"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType())}) 359 def generate_spectrogram( 360 self, 361 tokens: 'torch.tensor', 362 speaker: Optional[int] = None, 363 pace: float = 1.0, 364 reference_spec: Optional['torch.tensor'] = None, 365 reference_spec_lens: Optional['torch.tensor'] = None, 366 ) -> torch.tensor: 367 if self.training: 368 logging.warning("generate_spectrogram() is meant to be called in eval mode.") 369 if isinstance(speaker, int): 370 speaker = torch.tensor([speaker]).to(self.device) 371 spect, *_ = self( 372 text=tokens, 373 durs=None, 374 pitch=None, 375 speaker=speaker, 376 pace=pace, 377 reference_spec=reference_spec, 378 reference_spec_lens=reference_spec_lens, 379 ) 380 return spect 381 382 def training_step(self, batch, batch_idx): 383 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 384 None, 385 None, 386 None, 387 None, 388 None, 389 None, 390 ) 391 if self.learn_alignment: 392 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 393 batch_dict = batch 394 else: 395 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 396 audio = batch_dict.get("audio") 397 audio_lens = batch_dict.get("audio_lens") 398 text = batch_dict.get("text") 399 text_lens = batch_dict.get("text_lens") 400 attn_prior = batch_dict.get("align_prior_matrix", None) 401 pitch = batch_dict.get("pitch", None) 402 energy = batch_dict.get("energy", None) 403 speaker = batch_dict.get("speaker_id", None) 404 reference_audio = batch_dict.get("reference_audio", None) 405 reference_audio_len = batch_dict.get("reference_audio_lens", None) 406 else: 407 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 408 409 mels, spec_len = self.preprocessor(input_signal=audio, length=audio_lens) 410 reference_spec, reference_spec_len = None, None 411 if reference_audio is not None: 412 reference_spec, reference_spec_len = self.preprocessor( 413 input_signal=reference_audio, length=reference_audio_len 414 ) 415 416 ( 417 mels_pred, 418 _, 419 _, 420 log_durs_pred, 421 pitch_pred, 422 attn_soft, 423 attn_logprob, 424 attn_hard, 425 attn_hard_dur, 426 pitch, 427 energy_pred, 428 energy_tgt, 429 ) = self( 430 text=text, 431 durs=durs, 432 pitch=pitch, 433 energy=energy, 434 speaker=speaker, 435 pace=1.0, 436 spec=mels if self.learn_alignment else None, 437 reference_spec=reference_spec, 438 reference_spec_lens=reference_spec_len, 439 attn_prior=attn_prior, 440 mel_lens=spec_len, 441 input_lens=text_lens, 442 ) 443 if durs is None: 444 durs = attn_hard_dur 445 446 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 447 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 448 loss = mel_loss + dur_loss 449 if self.learn_alignment: 450 ctc_loss = self.forward_sum_loss_fn(attn_logprob=attn_logprob, in_lens=text_lens, out_lens=spec_len) 451 bin_loss_weight = min(self.current_epoch / self.bin_loss_warmup_epochs, 1.0) * 1.0 452 bin_loss = self.bin_loss_fn(hard_attention=attn_hard, soft_attention=attn_soft) * bin_loss_weight 453 loss += ctc_loss + bin_loss 454 455 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 456 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 457 loss += pitch_loss + energy_loss 458 459 self.log("t_loss", loss) 460 self.log("t_mel_loss", mel_loss) 461 self.log("t_dur_loss", dur_loss) 462 self.log("t_pitch_loss", pitch_loss) 463 if energy_tgt is not None: 464 self.log("t_energy_loss", energy_loss) 465 if self.learn_alignment: 466 self.log("t_ctc_loss", ctc_loss) 467 self.log("t_bin_loss", bin_loss) 468 469 # Log images to tensorboard 470 if self.log_images and self.log_train_images and isinstance(self.logger, TensorBoardLogger): 471 self.log_train_images = False 472 473 self.tb_logger.add_image( 474 "train_mel_target", 475 plot_spectrogram_to_numpy(mels[0].data.cpu().float().numpy()), 476 self.global_step, 477 dataformats="HWC", 478 ) 479 spec_predict = mels_pred[0].data.cpu().float().numpy() 480 self.tb_logger.add_image( 481 "train_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 482 ) 483 if self.learn_alignment: 484 attn = attn_hard[0].data.cpu().float().numpy().squeeze() 485 self.tb_logger.add_image( 486 "train_attn", plot_alignment_to_numpy(attn.T), self.global_step, dataformats="HWC", 487 ) 488 soft_attn = attn_soft[0].data.cpu().float().numpy().squeeze() 489 self.tb_logger.add_image( 490 "train_soft_attn", plot_alignment_to_numpy(soft_attn.T), self.global_step, dataformats="HWC", 491 ) 492 493 return loss 494 495 def validation_step(self, batch, batch_idx): 496 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 497 None, 498 None, 499 None, 500 None, 501 None, 502 None, 503 ) 504 if self.learn_alignment: 505 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 506 batch_dict = batch 507 else: 508 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 509 audio = batch_dict.get("audio") 510 audio_lens = batch_dict.get("audio_lens") 511 text = batch_dict.get("text") 512 text_lens = batch_dict.get("text_lens") 513 attn_prior = batch_dict.get("align_prior_matrix", None) 514 pitch = batch_dict.get("pitch", None) 515 energy = batch_dict.get("energy", None) 516 speaker = batch_dict.get("speaker_id", None) 517 reference_audio = batch_dict.get("reference_audio", None) 518 reference_audio_len = batch_dict.get("reference_audio_lens", None) 519 else: 520 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 521 522 mels, mel_lens = self.preprocessor(input_signal=audio, length=audio_lens) 523 reference_spec, reference_spec_len = None, None 524 if reference_audio is not None: 525 reference_spec, reference_spec_len = self.preprocessor( 526 input_signal=reference_audio, length=reference_audio_len 527 ) 528 529 # Calculate val loss on ground truth durations to better align L2 loss in time 530 (mels_pred, _, _, log_durs_pred, pitch_pred, _, _, _, attn_hard_dur, pitch, energy_pred, energy_tgt,) = self( 531 text=text, 532 durs=durs, 533 pitch=pitch, 534 energy=energy, 535 speaker=speaker, 536 pace=1.0, 537 spec=mels if self.learn_alignment else None, 538 reference_spec=reference_spec, 539 reference_spec_lens=reference_spec_len, 540 attn_prior=attn_prior, 541 mel_lens=mel_lens, 542 input_lens=text_lens, 543 ) 544 if durs is None: 545 durs = attn_hard_dur 546 547 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 548 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 549 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 550 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 551 loss = mel_loss + dur_loss + pitch_loss + energy_loss 552 553 val_outputs = { 554 "val_loss": loss, 555 "mel_loss": mel_loss, 556 "dur_loss": dur_loss, 557 "pitch_loss": pitch_loss, 558 "energy_loss": energy_loss if energy_tgt is not None else None, 559 "mel_target": mels if batch_idx == 0 else None, 560 "mel_pred": mels_pred if batch_idx == 0 else None, 561 } 562 self.validation_step_outputs.append(val_outputs) 563 return val_outputs 564 565 def on_validation_epoch_end(self): 566 collect = lambda key: torch.stack([x[key] for x in self.validation_step_outputs]).mean() 567 val_loss = collect("val_loss") 568 mel_loss = collect("mel_loss") 569 dur_loss = collect("dur_loss") 570 pitch_loss = collect("pitch_loss") 571 self.log("val_loss", val_loss, sync_dist=True) 572 self.log("val_mel_loss", mel_loss, sync_dist=True) 573 self.log("val_dur_loss", dur_loss, sync_dist=True) 574 self.log("val_pitch_loss", pitch_loss, sync_dist=True) 575 if self.validation_step_outputs[0]["energy_loss"] is not None: 576 energy_loss = collect("energy_loss") 577 self.log("val_energy_loss", energy_loss, sync_dist=True) 578 579 _, _, _, _, _, spec_target, spec_predict = self.validation_step_outputs[0].values() 580 581 if self.log_images and isinstance(self.logger, TensorBoardLogger): 582 self.tb_logger.add_image( 583 "val_mel_target", 584 plot_spectrogram_to_numpy(spec_target[0].data.cpu().float().numpy()), 585 self.global_step, 586 dataformats="HWC", 587 ) 588 spec_predict = spec_predict[0].data.cpu().float().numpy() 589 self.tb_logger.add_image( 590 "val_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 591 ) 592 self.log_train_images = True 593 self.validation_step_outputs.clear() # free memory) 594 595 def _setup_train_dataloader(self, cfg): 596 phon_mode = contextlib.nullcontext() 597 if hasattr(self.vocab, "set_phone_prob"): 598 phon_mode = self.vocab.set_phone_prob(self.vocab.phoneme_probability) 599 600 with phon_mode: 601 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 602 603 sampler = dataset.get_sampler(cfg.dataloader_params.batch_size) 604 return torch.utils.data.DataLoader( 605 dataset, collate_fn=dataset.collate_fn, sampler=sampler, **cfg.dataloader_params 606 ) 607 608 def _setup_test_dataloader(self, cfg): 609 phon_mode = contextlib.nullcontext() 610 if hasattr(self.vocab, "set_phone_prob"): 611 phon_mode = self.vocab.set_phone_prob(0.0) 612 613 with phon_mode: 614 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 615 616 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 617 618 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 619 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 620 raise ValueError(f"No dataset for {name}") 621 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 622 raise ValueError(f"No dataloader_params for {name}") 623 if shuffle_should_be: 624 if 'shuffle' not in cfg.dataloader_params: 625 logging.warning( 626 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 627 "config. Manually setting to True" 628 ) 629 with open_dict(cfg.dataloader_params): 630 cfg.dataloader_params.shuffle = True 631 elif not cfg.dataloader_params.shuffle: 632 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 633 elif cfg.dataloader_params.shuffle: 634 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 635 636 if self.ds_class == "nemo.collections.tts.data.dataset.TTSDataset": 637 phon_mode = contextlib.nullcontext() 638 if hasattr(self.vocab, "set_phone_prob"): 639 phon_mode = self.vocab.set_phone_prob(prob=None if name == "val" else self.vocab.phoneme_probability) 640 641 with phon_mode: 642 dataset = instantiate( 643 cfg.dataset, 644 text_normalizer=self.normalizer, 645 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 646 text_tokenizer=self.vocab, 647 ) 648 else: 649 dataset = instantiate(cfg.dataset) 650 651 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 652 653 def setup_training_data(self, cfg): 654 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 655 self._train_dl = self._setup_train_dataloader(cfg) 656 else: 657 self._train_dl = self.__setup_dataloader_from_config(cfg) 658 659 def setup_validation_data(self, cfg): 660 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 661 self._validation_dl = self._setup_test_dataloader(cfg) 662 else: 663 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="val") 664 665 def setup_test_data(self, cfg): 666 """Omitted.""" 667 pass 668 669 def configure_callbacks(self): 670 if not self.log_config: 671 return [] 672 673 sample_ds_class = self.log_config.dataset._target_ 674 if sample_ds_class != "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 675 raise ValueError(f"Logging callback only supported for TextToSpeechDataset, got {sample_ds_class}") 676 677 data_loader = self._setup_test_dataloader(self.log_config) 678 679 generators = instantiate(self.log_config.generators) 680 log_dir = Path(self.log_config.log_dir) if self.log_config.log_dir else None 681 log_callback = LoggingCallback( 682 generators=generators, 683 data_loader=data_loader, 684 log_epochs=self.log_config.log_epochs, 685 epoch_frequency=self.log_config.epoch_frequency, 686 output_dir=log_dir, 687 loggers=self.trainer.loggers, 688 log_tensorboard=self.log_config.log_tensorboard, 689 log_wandb=self.log_config.log_wandb, 690 ) 691 692 return [log_callback] 693 694 @classmethod 695 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 696 """ 697 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 698 Returns: 699 List of available pre-trained models. 700 """ 701 list_of_models = [] 702 703 # en-US, single speaker, 22050Hz, LJSpeech (ARPABET). 704 model = PretrainedModelInfo( 705 pretrained_model_name="tts_en_fastpitch", 706 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/1.8.1/files/tts_en_fastpitch_align.nemo", 707 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is ARPABET-based.", 708 class_=cls, 709 ) 710 list_of_models.append(model) 711 712 # en-US, single speaker, 22050Hz, LJSpeech (IPA). 713 model = PretrainedModelInfo( 714 pretrained_model_name="tts_en_fastpitch_ipa", 715 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/IPA_1.13.0/files/tts_en_fastpitch_align_ipa.nemo", 716 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is IPA-based.", 717 class_=cls, 718 ) 719 list_of_models.append(model) 720 721 # en-US, multi-speaker, 44100Hz, HiFiTTS. 722 model = PretrainedModelInfo( 723 pretrained_model_name="tts_en_fastpitch_multispeaker", 724 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_multispeaker_fastpitchhifigan/versions/1.10.0/files/tts_en_fastpitch_multispeaker.nemo", 725 description="This model is trained on HiFITTS sampled at 44100Hz with and can be used to generate male and female English voices with an American accent.", 726 class_=cls, 727 ) 728 list_of_models.append(model) 729 730 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 21.02 731 model = PretrainedModelInfo( 732 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2102", 733 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2102.nemo", 734 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 21.02 Dataset sampled at 22050Hz and can be used to generate male German voices.", 735 class_=cls, 736 ) 737 list_of_models.append(model) 738 739 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 22.10 740 model = PretrainedModelInfo( 741 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2210", 742 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2210.nemo", 743 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 22.10 Dataset sampled at 22050Hz and can be used to generate male German voices.", 744 class_=cls, 745 ) 746 list_of_models.append(model) 747 748 # de-DE, multi-speaker, 5 speakers, 44100 Hz, HUI-Audio-Corpus-German Clean. 749 model = PretrainedModelInfo( 750 pretrained_model_name="tts_de_fastpitch_multispeaker_5", 751 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitch_multispeaker_5/versions/1.11.0/files/tts_de_fastpitch_multispeaker_5.nemo", 752 description="This model is trained on 5 speakers in HUI-Audio-Corpus-German clean subset sampled at 44100Hz with and can be used to generate male and female German voices.", 753 class_=cls, 754 ) 755 list_of_models.append(model) 756 757 # es, 174 speakers, 44100Hz, OpenSLR (IPA) 758 model = PretrainedModelInfo( 759 pretrained_model_name="tts_es_fastpitch_multispeaker", 760 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_es_multispeaker_fastpitchhifigan/versions/1.15.0/files/tts_es_fastpitch_multispeaker.nemo", 761 description="This model is trained on 174 speakers in 6 crowdsourced Latin American Spanish OpenSLR datasets sampled at 44100Hz and can be used to generate male and female Spanish voices with Latin American accents.", 762 class_=cls, 763 ) 764 list_of_models.append(model) 765 766 # zh, single female speaker, 22050Hz, SFSpeech Bilingual Chinese/English dataset, improved model using richer 767 # dict and jieba word segmenter for polyphone disambiguation. 768 model = PretrainedModelInfo( 769 pretrained_model_name="tts_zh_fastpitch_sfspeech", 770 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_zh_fastpitch_hifigan_sfspeech/versions/1.15.0/files/tts_zh_fastpitch_sfspeech.nemo", 771 description="This model is trained on a single female speaker in SFSpeech Bilingual Chinese/English dataset" 772 " sampled at 22050Hz and can be used to generate female Mandarin Chinese voices. It is improved" 773 " using richer dict and jieba word segmenter for polyphone disambiguation.", 774 class_=cls, 775 ) 776 list_of_models.append(model) 777 778 # en, multi speaker, LibriTTS, 16000 Hz 779 # stft 25ms 10ms matching ASR params 780 # for use during Enhlish ASR training/adaptation 781 model = PretrainedModelInfo( 782 pretrained_model_name="tts_en_fastpitch_for_asr_finetuning", 783 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch_spectrogram_enhancer_for_asr_finetuning/versions/1.20.0/files/tts_en_fastpitch_for_asr_finetuning.nemo", 784 description="This model is trained on LibriSpeech, train-960 subset." 785 " STFT parameters follow those commonly used in ASR: 25 ms window, 10 ms hop." 786 " This model is supposed to be used with its companion SpetrogramEnhancer for " 787 " ASR fine-tuning. Usage for regular TTS tasks is not advised.", 788 class_=cls, 789 ) 790 list_of_models.append(model) 791 792 return list_of_models 793 794 # Methods for model exportability 795 def _prepare_for_export(self, **kwargs): 796 super()._prepare_for_export(**kwargs) 797 798 tensor_shape = ('T') if self.export_config["enable_ragged_batches"] else ('B', 'T') 799 800 # Define input_types and output_types as required by export() 801 self._input_types = { 802 "text": NeuralType(tensor_shape, TokenIndex()), 803 "pitch": NeuralType(tensor_shape, RegressionValuesType()), 804 "pace": NeuralType(tensor_shape), 805 "volume": NeuralType(tensor_shape, optional=True), 806 "batch_lengths": NeuralType(('B'), optional=True), 807 "speaker": NeuralType(('B'), Index(), optional=True), 808 } 809 self._output_types = { 810 "spect": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 811 "num_frames": NeuralType(('B'), TokenDurationType()), 812 "durs_predicted": NeuralType(('B', 'T'), TokenDurationType()), 813 "log_durs_predicted": NeuralType(('B', 'T'), TokenLogDurationType()), 814 "pitch_predicted": NeuralType(('B', 'T'), RegressionValuesType()), 815 } 816 if self.export_config["enable_volume"]: 817 self._output_types["volume_aligned"] = NeuralType(('B', 'T'), RegressionValuesType()) 818 819 def _export_teardown(self): 820 self._input_types = self._output_types = None 821 822 @property 823 def disabled_deployment_input_names(self): 824 """Implement this method to return a set of input names disabled for export""" 825 disabled_inputs = set() 826 if self.fastpitch.speaker_emb is None: 827 disabled_inputs.add("speaker") 828 if not self.export_config["enable_ragged_batches"]: 829 disabled_inputs.add("batch_lengths") 830 if not self.export_config["enable_volume"]: 831 disabled_inputs.add("volume") 832 return disabled_inputs 833 834 @property 835 def input_types(self): 836 return self._input_types 837 838 @property 839 def output_types(self): 840 return self._output_types 841 842 def input_example(self, max_batch=1, max_dim=44): 843 """ 844 Generates input examples for tracing etc. 845 Returns: 846 A tuple of input examples. 847 """ 848 par = next(self.fastpitch.parameters()) 849 inputs = sample_tts_input(self.export_config, par.device, max_batch=max_batch, max_dim=max_dim) 850 if 'enable_ragged_batches' not in self.export_config: 851 inputs.pop('batch_lengths', None) 852 return (inputs,) 853 854 def forward_for_export(self, text, pitch, pace, volume=None, batch_lengths=None, speaker=None): 855 if self.export_config["enable_ragged_batches"]: 856 text, pitch, pace, volume_tensor, lens = batch_from_ragged( 857 text, pitch, pace, batch_lengths, padding_idx=self.fastpitch.encoder.padding_idx, volume=volume 858 ) 859 if volume is not None: 860 volume = volume_tensor 861 return self.fastpitch.infer(text=text, pitch=pitch, pace=pace, volume=volume, speaker=speaker) 862 863 def interpolate_speaker( 864 self, original_speaker_1, original_speaker_2, weight_speaker_1, weight_speaker_2, new_speaker_id 865 ): 866 """ 867 This method performs speaker interpolation between two original speakers the model is trained on. 868 869 Inputs: 870 original_speaker_1: Integer speaker ID of first existing speaker in the model 871 original_speaker_2: Integer speaker ID of second existing speaker in the model 872 weight_speaker_1: Floating point weight associated in to first speaker during weight combination 873 weight_speaker_2: Floating point weight associated in to second speaker during weight combination 874 new_speaker_id: Integer speaker ID of new interpolated speaker in the model 875 """ 876 if self.fastpitch.speaker_emb is None: 877 raise Exception( 878 "Current FastPitch model is not a multi-speaker FastPitch model. Speaker interpolation can only \ 879 be performed with a multi-speaker model" 880 ) 881 n_speakers = self.fastpitch.speaker_emb.weight.data.size()[0] 882 if original_speaker_1 >= n_speakers or original_speaker_2 >= n_speakers or new_speaker_id >= n_speakers: 883 raise Exception( 884 f"Parameters original_speaker_1, original_speaker_2, new_speaker_id should be less than the total \ 885 total number of speakers FastPitch was trained on (n_speakers = {n_speakers})." 886 ) 887 speaker_emb_1 = ( 888 self.fastpitch.speaker_emb(torch.tensor(original_speaker_1, dtype=torch.int32).cuda()).clone().detach() 889 ) 890 speaker_emb_2 = ( 891 self.fastpitch.speaker_emb(torch.tensor(original_speaker_2, dtype=torch.int32).cuda()).clone().detach() 892 ) 893 new_speaker_emb = weight_speaker_1 * speaker_emb_1 + weight_speaker_2 * speaker_emb_2 894 self.fastpitch.speaker_emb.weight.data[new_speaker_id] = new_speaker_emb 895 [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/tts/models/tacotron2.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import contextlib 16 from dataclasses import dataclass 17 from typing import Any, Dict, List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import MISSING, DictConfig, OmegaConf, open_dict 22 from omegaconf.errors import ConfigAttributeError 23 from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger 24 from torch import nn 25 26 from nemo.collections.common.parts.preprocessing import parsers 27 from nemo.collections.tts.losses.tacotron2loss import Tacotron2Loss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.parts.utils.helpers import ( 30 g2p_backward_compatible_support, 31 get_mask_from_lengths, 32 tacotron2_log_to_tb_func, 33 tacotron2_log_to_wandb_func, 34 ) 35 from nemo.core.classes.common import PretrainedModelInfo, typecheck 36 from nemo.core.neural_types.elements import ( 37 AudioSignal, 38 EmbeddedTextType, 39 LengthsType, 40 LogitsType, 41 MelSpectrogramType, 42 SequenceToSequenceAlignmentType, 43 ) 44 from nemo.core.neural_types.neural_type import NeuralType 45 from nemo.utils import logging, model_utils 46 47 48 @dataclass 49 class Preprocessor: 50 _target_: str = MISSING 51 pad_value: float = MISSING 52 53 54 @dataclass 55 class Tacotron2Config: 56 preprocessor: Preprocessor = Preprocessor() 57 encoder: Dict[Any, Any] = MISSING 58 decoder: Dict[Any, Any] = MISSING 59 postnet: Dict[Any, Any] = MISSING 60 labels: List = MISSING 61 train_ds: Optional[Dict[Any, Any]] = None 62 validation_ds: Optional[Dict[Any, Any]] = None 63 64 65 class Tacotron2Model(SpectrogramGenerator): 66 """Tacotron 2 Model that is used to generate mel spectrograms from text""" 67 68 def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None): 69 # Convert to Hydra 1.0 compatible DictConfig 70 cfg = model_utils.convert_model_config_to_dict_config(cfg) 71 cfg = model_utils.maybe_update_config_version(cfg) 72 73 # setup normalizer 74 self.normalizer = None 75 self.text_normalizer_call = None 76 self.text_normalizer_call_kwargs = {} 77 self._setup_normalizer(cfg) 78 79 # setup tokenizer 80 self.tokenizer = None 81 if hasattr(cfg, 'text_tokenizer'): 82 self._setup_tokenizer(cfg) 83 84 self.num_tokens = len(self.tokenizer.tokens) 85 self.tokenizer_pad = self.tokenizer.pad 86 self.tokenizer_unk = self.tokenizer.oov 87 # assert self.tokenizer is not None 88 else: 89 self.num_tokens = len(cfg.labels) + 3 90 91 super().__init__(cfg=cfg, trainer=trainer) 92 93 schema = OmegaConf.structured(Tacotron2Config) 94 # ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes 95 if isinstance(cfg, dict): 96 cfg = OmegaConf.create(cfg) 97 elif not isinstance(cfg, DictConfig): 98 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 99 # Ensure passed cfg is compliant with schema 100 try: 101 OmegaConf.merge(cfg, schema) 102 self.pad_value = cfg.preprocessor.pad_value 103 except ConfigAttributeError: 104 self.pad_value = cfg.preprocessor.params.pad_value 105 logging.warning( 106 "Your config is using an old NeMo yaml configuration. Please ensure that the yaml matches the " 107 "current version in the main branch for future compatibility." 108 ) 109 110 self._parser = None 111 self.audio_to_melspec_precessor = instantiate(cfg.preprocessor) 112 self.text_embedding = nn.Embedding(self.num_tokens, 512) 113 self.encoder = instantiate(self._cfg.encoder) 114 self.decoder = instantiate(self._cfg.decoder) 115 self.postnet = instantiate(self._cfg.postnet) 116 self.loss = Tacotron2Loss() 117 self.calculate_loss = True 118 119 @property 120 def parser(self): 121 if self._parser is not None: 122 return self._parser 123 124 ds_class_name = self._cfg.train_ds.dataset._target_.split(".")[-1] 125 if ds_class_name == "TTSDataset": 126 self._parser = None 127 elif hasattr(self._cfg, "labels"): 128 self._parser = parsers.make_parser( 129 labels=self._cfg.labels, 130 name='en', 131 unk_id=-1, 132 blank_id=-1, 133 do_normalize=True, 134 abbreviation_version="fastpitch", 135 make_table=False, 136 ) 137 else: 138 raise ValueError("Wanted to setup parser, but model does not have necessary paramaters") 139 140 return self._parser 141 142 def parse(self, text: str, normalize=True) -> torch.Tensor: 143 if self.training: 144 logging.warning("parse() is meant to be called in eval mode.") 145 if normalize and self.text_normalizer_call is not None: 146 text = self.text_normalizer_call(text, **self.text_normalizer_call_kwargs) 147 148 eval_phon_mode = contextlib.nullcontext() 149 if hasattr(self.tokenizer, "set_phone_prob"): 150 eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0) 151 152 with eval_phon_mode: 153 if self.tokenizer is not None: 154 tokens = self.tokenizer.encode(text) 155 else: 156 tokens = self.parser(text) 157 # Old parser doesn't add bos and eos ids, so maunally add it 158 tokens = [len(self._cfg.labels)] + tokens + [len(self._cfg.labels) + 1] 159 tokens_tensor = torch.tensor(tokens).unsqueeze_(0).to(self.device) 160 return tokens_tensor 161 162 @property 163 def input_types(self): 164 if self.training: 165 return { 166 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 167 "token_len": NeuralType(('B'), LengthsType()), 168 "audio": NeuralType(('B', 'T'), AudioSignal()), 169 "audio_len": NeuralType(('B'), LengthsType()), 170 } 171 else: 172 return { 173 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 174 "token_len": NeuralType(('B'), LengthsType()), 175 "audio": NeuralType(('B', 'T'), AudioSignal(), optional=True), 176 "audio_len": NeuralType(('B'), LengthsType(), optional=True), 177 } 178 179 @property 180 def output_types(self): 181 if not self.calculate_loss and not self.training: 182 return { 183 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 184 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 185 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 186 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 187 "pred_length": NeuralType(('B'), LengthsType()), 188 } 189 return { 190 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 191 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 192 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 193 "spec_target": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 194 "spec_target_len": NeuralType(('B'), LengthsType()), 195 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 196 } 197 198 @typecheck() 199 def forward(self, *, tokens, token_len, audio=None, audio_len=None): 200 if audio is not None and audio_len is not None: 201 spec_target, spec_target_len = self.audio_to_melspec_precessor(audio, audio_len) 202 else: 203 if self.training or self.calculate_loss: 204 raise ValueError( 205 f"'audio' and 'audio_len' can not be None when either 'self.training' or 'self.calculate_loss' is True." 206 ) 207 208 token_embedding = self.text_embedding(tokens).transpose(1, 2) 209 encoder_embedding = self.encoder(token_embedding=token_embedding, token_len=token_len) 210 211 if self.training: 212 spec_pred_dec, gate_pred, alignments = self.decoder( 213 memory=encoder_embedding, decoder_inputs=spec_target, memory_lengths=token_len 214 ) 215 else: 216 spec_pred_dec, gate_pred, alignments, pred_length = self.decoder( 217 memory=encoder_embedding, memory_lengths=token_len 218 ) 219 220 spec_pred_postnet = self.postnet(mel_spec=spec_pred_dec) 221 222 if not self.calculate_loss and not self.training: 223 return spec_pred_dec, spec_pred_postnet, gate_pred, alignments, pred_length 224 225 return spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments 226 227 @typecheck( 228 input_types={"tokens": NeuralType(('B', 'T'), EmbeddedTextType())}, 229 output_types={"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType())}, 230 ) 231 def generate_spectrogram(self, *, tokens): 232 self.eval() 233 self.calculate_loss = False 234 token_len = torch.tensor([len(i) for i in tokens]).to(self.device) 235 tensors = self(tokens=tokens, token_len=token_len) 236 spectrogram_pred = tensors[1] 237 238 if spectrogram_pred.shape[0] > 1: 239 # Silence all frames past the predicted end 240 mask = ~get_mask_from_lengths(tensors[-1]) 241 mask = mask.expand(spectrogram_pred.shape[1], mask.size(0), mask.size(1)) 242 mask = mask.permute(1, 0, 2) 243 spectrogram_pred.data.masked_fill_(mask, self.pad_value) 244 245 return spectrogram_pred 246 247 def training_step(self, batch, batch_idx): 248 audio, audio_len, tokens, token_len = batch 249 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, _ = self.forward( 250 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 251 ) 252 253 loss, _ = self.loss( 254 spec_pred_dec=spec_pred_dec, 255 spec_pred_postnet=spec_pred_postnet, 256 gate_pred=gate_pred, 257 spec_target=spec_target, 258 spec_target_len=spec_target_len, 259 pad_value=self.pad_value, 260 ) 261 262 output = { 263 'loss': loss, 264 'progress_bar': {'training_loss': loss}, 265 'log': {'loss': loss}, 266 } 267 return output 268 269 def validation_step(self, batch, batch_idx): 270 audio, audio_len, tokens, token_len = batch 271 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments = self.forward( 272 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 273 ) 274 275 loss, gate_target = self.loss( 276 spec_pred_dec=spec_pred_dec, 277 spec_pred_postnet=spec_pred_postnet, 278 gate_pred=gate_pred, 279 spec_target=spec_target, 280 spec_target_len=spec_target_len, 281 pad_value=self.pad_value, 282 ) 283 loss = { 284 "val_loss": loss, 285 "mel_target": spec_target, 286 "mel_postnet": spec_pred_postnet, 287 "gate": gate_pred, 288 "gate_target": gate_target, 289 "alignments": alignments, 290 } 291 self.validation_step_outputs.append(loss) 292 return loss 293 294 def on_validation_epoch_end(self): 295 if self.logger is not None and self.logger.experiment is not None: 296 logger = self.logger.experiment 297 for logger in self.trainer.loggers: 298 if isinstance(logger, TensorBoardLogger): 299 logger = logger.experiment 300 break 301 if isinstance(logger, TensorBoardLogger): 302 tacotron2_log_to_tb_func( 303 logger, 304 self.validation_step_outputs[0].values(), 305 self.global_step, 306 tag="val", 307 log_images=True, 308 add_audio=False, 309 ) 310 elif isinstance(logger, WandbLogger): 311 tacotron2_log_to_wandb_func( 312 logger, 313 self.validation_step_outputs[0].values(), 314 self.global_step, 315 tag="val", 316 log_images=True, 317 add_audio=False, 318 ) 319 avg_loss = torch.stack( 320 [x['val_loss'] for x in self.validation_step_outputs] 321 ).mean() # This reduces across batches, not workers! 322 self.log('val_loss', avg_loss) 323 self.validation_step_outputs.clear() # free memory 324 325 def _setup_normalizer(self, cfg): 326 if "text_normalizer" in cfg: 327 normalizer_kwargs = {} 328 329 if "whitelist" in cfg.text_normalizer: 330 normalizer_kwargs["whitelist"] = self.register_artifact( 331 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 332 ) 333 334 try: 335 import nemo_text_processing 336 337 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 338 except Exception as e: 339 logging.error(e) 340 raise ImportError( 341 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 342 ) 343 344 self.text_normalizer_call = self.normalizer.normalize 345 if "text_normalizer_call_kwargs" in cfg: 346 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 347 348 def _setup_tokenizer(self, cfg): 349 text_tokenizer_kwargs = {} 350 if "g2p" in cfg.text_tokenizer and cfg.text_tokenizer.g2p is not None: 351 # for backward compatibility 352 if ( 353 self._is_model_being_restored() 354 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 355 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 356 ): 357 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 358 cfg.text_tokenizer.g2p["_target_"] 359 ) 360 361 g2p_kwargs = {} 362 363 if "phoneme_dict" in cfg.text_tokenizer.g2p: 364 g2p_kwargs["phoneme_dict"] = self.register_artifact( 365 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 366 ) 367 368 if "heteronyms" in cfg.text_tokenizer.g2p: 369 g2p_kwargs["heteronyms"] = self.register_artifact( 370 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 371 ) 372 373 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 374 375 self.tokenizer = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 376 377 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 378 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 379 raise ValueError(f"No dataset for {name}") 380 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 381 raise ValueError(f"No dataloder_params for {name}") 382 if shuffle_should_be: 383 if 'shuffle' not in cfg.dataloader_params: 384 logging.warning( 385 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 386 "config. Manually setting to True" 387 ) 388 with open_dict(cfg.dataloader_params): 389 cfg.dataloader_params.shuffle = True 390 elif not cfg.dataloader_params.shuffle: 391 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 392 elif not shuffle_should_be and cfg.dataloader_params.shuffle: 393 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 394 395 dataset = instantiate( 396 cfg.dataset, 397 text_normalizer=self.normalizer, 398 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 399 text_tokenizer=self.tokenizer, 400 ) 401 402 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 403 404 def setup_training_data(self, cfg): 405 self._train_dl = self.__setup_dataloader_from_config(cfg) 406 407 def setup_validation_data(self, cfg): 408 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="validation") 409 410 @classmethod 411 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 412 """ 413 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 414 Returns: 415 List of available pre-trained models. 416 """ 417 list_of_models = [] 418 model = PretrainedModelInfo( 419 pretrained_model_name="tts_en_tacotron2", 420 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_tacotron2/versions/1.10.0/files/tts_en_tacotron2.nemo", 421 description="This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.", 422 class_=cls, 423 aliases=["Tacotron2-22050Hz"], 424 ) 425 list_of_models.append(model) 426 return list_of_models 427 [end of nemo/collections/tts/models/tacotron2.py] [start of nemo/core/config/modelPT.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Dict, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.core import config 21 from nemo.core.classes.dataset import DatasetConfig 22 from nemo.utils import exp_manager 23 24 25 @dataclass 26 class SchedConfig: 27 name: str = MISSING 28 min_lr: float = 0.0 29 last_epoch: int = -1 30 31 32 @dataclass 33 class OptimConfig: 34 name: str = MISSING 35 sched: Optional[SchedConfig] = None 36 37 38 @dataclass 39 class ModelConfig: 40 """ 41 Model component inside ModelPT 42 """ 43 44 # ... 45 train_ds: Optional[DatasetConfig] = None 46 validation_ds: Optional[DatasetConfig] = None 47 test_ds: Optional[DatasetConfig] = None 48 optim: Optional[OptimConfig] = None 49 50 51 @dataclass 52 class HydraConfig: 53 run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) 54 job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) 55 56 57 @dataclass 58 class NemoConfig: 59 name: str = MISSING 60 model: ModelConfig = MISSING 61 trainer: config.TrainerConfig = config.TrainerConfig( 62 strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' 63 ) 64 exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() 65 hydra: HydraConfig = HydraConfig() 66 67 68 class ModelConfigBuilder: 69 def __init__(self, model_cfg: ModelConfig): 70 """ 71 Base class for any Model Config Builder. 72 73 A Model Config Builder is a utility class that accepts a ModelConfig dataclass, 74 and via a set of utility methods (that are implemented by the subclassed ModelConfigBuilder), 75 builds a finalized ModelConfig that can be supplied to a NemoModel dataclass as 76 the `model` component. 77 78 Subclasses *must* implement the private method `_finalize_cfg`. 79 Inside this method, they must update `self.model_cfg` with all interdependent config 80 options that need to be set (either updated by user explicitly or with their default value). 81 82 The updated model config must then be preserved in `self.model_cfg`. 83 84 Example: 85 # Create the config builder 86 config_builder = <subclass>ModelConfigBuilder() 87 88 # Update the components of the config that are modifiable 89 config_builder.set_X(X) 90 config_builder.set_Y(Y) 91 92 # Create a "finalized" config dataclass that will contain all the updates 93 # that were specified by the builder 94 model_config = config_builder.build() 95 96 # Use model config as is (or further update values), then create a new Model 97 model = nemo.<domain>.models.<ModelName>Model(cfg=model_config, trainer=Trainer()) 98 99 Supported build methods: 100 - set_train_ds: All model configs can accept a subclass of `DatasetConfig` as their 101 training config. Subclasses can override this method to enable auto-complete 102 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 103 104 - set_validation_ds: All model configs can accept a subclass of `DatasetConfig` as their 105 validation config. Subclasses can override this method to enable auto-complete 106 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 107 108 - set_test_ds: All model configs can accept a subclass of `DatasetConfig` as their 109 test config. Subclasses can override this method to enable auto-complete 110 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 111 112 - set_optim: A build method that supports changes to the Optimizer (and optionally, 113 the Scheduler) used for training the model. The function accepts two inputs - 114 115 `cfg`: A subclass of `OptimizerParams` - any OptimizerParams subclass can be used, 116 in order to select an appropriate Optimizer. Examples: AdamParams. 117 118 `sched_cfg`: A subclass of `SchedulerParams` - any SchedulerParams subclass can be used, 119 in order to select an appropriate Scheduler. Examples: CosineAnnealingParams. 120 Note that this argument is optional. 121 122 - build(): The method which should return a "finalized" ModelConfig dataclass. 123 Subclasses *should* always override this method, and update the signature 124 of this method with the return type of the Dataclass, so that it enables 125 autocomplete for the user. 126 127 Example: 128 def build(self) -> EncDecCTCConfig: 129 return super().build() 130 131 Any additional build methods must be added by subclasses of ModelConfigBuilder. 132 133 Args: 134 model_cfg: 135 """ 136 self.model_cfg = model_cfg 137 self.train_ds_cfg = None 138 self.validation_ds_cfg = None 139 self.test_ds_cfg = None 140 self.optim_cfg = None 141 142 def set_train_ds(self, cfg: Optional[DatasetConfig] = None): 143 self.model_cfg.train_ds = cfg 144 145 def set_validation_ds(self, cfg: Optional[DatasetConfig] = None): 146 self.model_cfg.validation_ds = cfg 147 148 def set_test_ds(self, cfg: Optional[DatasetConfig] = None): 149 self.model_cfg.test_ds = cfg 150 151 def set_optim(self, cfg: config.OptimizerParams, sched_cfg: Optional[config.SchedulerParams] = None): 152 @dataclass 153 class WrappedOptimConfig(OptimConfig, cfg.__class__): 154 pass 155 156 # Setup optim 157 optim_name = cfg.__class__.__name__.replace("Params", "").lower() 158 wrapped_cfg = WrappedOptimConfig(name=optim_name, sched=None, **vars(cfg)) 159 160 if sched_cfg is not None: 161 162 @dataclass 163 class WrappedSchedConfig(SchedConfig, sched_cfg.__class__): 164 pass 165 166 # Setup scheduler 167 sched_name = sched_cfg.__class__.__name__.replace("Params", "") 168 wrapped_sched_cfg = WrappedSchedConfig(name=sched_name, **vars(sched_cfg)) 169 170 wrapped_cfg.sched = wrapped_sched_cfg 171 172 self.model_cfg.optim = wrapped_cfg 173 174 def _finalize_cfg(self): 175 raise NotImplementedError() 176 177 def build(self) -> ModelConfig: 178 # validate config 179 self._finalize_cfg() 180 181 return self.model_cfg 182 [end of nemo/core/config/modelPT.py] [start of nemo/utils/exp_manager.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import glob 16 import os 17 import subprocess 18 import sys 19 import time 20 import warnings 21 from dataclasses import dataclass 22 from datetime import timedelta 23 from pathlib import Path 24 from shutil import copy, move 25 from typing import Any, Dict, List, Optional, Tuple, Union 26 27 import pytorch_lightning 28 import torch 29 from hydra.core.hydra_config import HydraConfig 30 from hydra.utils import get_original_cwd 31 from omegaconf import DictConfig, OmegaConf, open_dict 32 from pytorch_lightning.callbacks import Callback, ModelCheckpoint 33 from pytorch_lightning.callbacks.early_stopping import EarlyStopping 34 from pytorch_lightning.callbacks.timer import Interval, Timer 35 from pytorch_lightning.loggers import MLFlowLogger, TensorBoardLogger, WandbLogger 36 from pytorch_lightning.loops import _TrainingEpochLoop 37 from pytorch_lightning.strategies.ddp import DDPStrategy 38 39 from nemo.collections.common.callbacks import EMA 40 from nemo.constants import NEMO_ENV_VARNAME_TESTING, NEMO_ENV_VARNAME_VERSION 41 from nemo.utils import logging, timers 42 from nemo.utils.app_state import AppState 43 from nemo.utils.callbacks import NeMoModelCheckpoint, PreemptionCallback 44 from nemo.utils.env_var_parsing import get_envbool 45 from nemo.utils.exceptions import NeMoBaseException 46 from nemo.utils.get_rank import is_global_rank_zero 47 from nemo.utils.lightning_logger_patch import add_filehandlers_to_pl_logger 48 from nemo.utils.loggers import ClearMLLogger, ClearMLParams, DLLogger, DLLoggerParams, MLFlowParams 49 from nemo.utils.model_utils import uninject_model_parallel_rank 50 51 52 class NotFoundError(NeMoBaseException): 53 """ Raised when a file or folder is not found""" 54 55 56 class LoggerMisconfigurationError(NeMoBaseException): 57 """ Raised when a mismatch between trainer.logger and exp_manager occurs""" 58 59 def __init__(self, message): 60 message = ( 61 message 62 + " You can disable lighning's trainer from creating a logger by passing logger=False to its constructor." 63 ) 64 super().__init__(message) 65 66 67 class CheckpointMisconfigurationError(NeMoBaseException): 68 """ Raised when a mismatch between trainer.callbacks and exp_manager occurs""" 69 70 71 @dataclass 72 class EarlyStoppingParams: 73 monitor: str = "val_loss" # The metric that early stopping should consider. 74 mode: str = "min" # inform early stopping whether to look for increase or decrease in monitor. 75 min_delta: float = 0.001 # smallest change to consider as improvement. 76 patience: int = 10 # how many (continuous) validation cycles to wait with no improvement and stopping training. 77 verbose: bool = True 78 strict: bool = True 79 check_finite: bool = True 80 stopping_threshold: Optional[float] = None 81 divergence_threshold: Optional[float] = None 82 check_on_train_epoch_end: Optional[bool] = None 83 log_rank_zero_only: bool = False 84 85 86 @dataclass 87 class CallbackParams: 88 filepath: Optional[str] = None # Deprecated 89 dirpath: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 90 filename: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 91 monitor: Optional[str] = "val_loss" 92 verbose: Optional[bool] = True 93 save_last: Optional[bool] = True 94 save_top_k: Optional[int] = 3 95 save_weights_only: Optional[bool] = False 96 mode: Optional[str] = "min" 97 auto_insert_metric_name: bool = True 98 every_n_epochs: Optional[int] = 1 99 every_n_train_steps: Optional[int] = None 100 train_time_interval: Optional[str] = None 101 prefix: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 102 postfix: str = ".nemo" 103 save_best_model: bool = False 104 always_save_nemo: bool = False 105 save_nemo_on_train_end: Optional[bool] = True # Whether to automatically save .nemo file durin on_train_end hook 106 model_parallel_size: Optional[int] = None # tensor parallel size * pipeline parallel size 107 save_on_train_epoch_end: Optional[bool] = False # Save after training, not after validation 108 109 110 @dataclass 111 class StepTimingParams: 112 reduction: Optional[str] = "mean" 113 # if True torch.cuda.synchronize() is called on start/stop 114 sync_cuda: Optional[bool] = False 115 # if positive, defines the size of a sliding window for computing mean 116 buffer_size: Optional[int] = 1 117 118 119 @dataclass 120 class EMAParams: 121 enable: Optional[bool] = False 122 decay: Optional[float] = 0.999 123 cpu_offload: Optional[bool] = False 124 validate_original_weights: Optional[bool] = False 125 every_n_steps: int = 1 126 127 128 @dataclass 129 class ExpManagerConfig: 130 """Experiment Manager config for validation of passed arguments. 131 """ 132 133 # Log dir creation parameters 134 explicit_log_dir: Optional[str] = None 135 exp_dir: Optional[str] = None 136 name: Optional[str] = None 137 version: Optional[str] = None 138 use_datetime_version: Optional[bool] = True 139 resume_if_exists: Optional[bool] = False 140 resume_past_end: Optional[bool] = False 141 resume_ignore_no_checkpoint: Optional[bool] = False 142 resume_from_checkpoint: Optional[str] = None 143 # Logging parameters 144 create_tensorboard_logger: Optional[bool] = True 145 summary_writer_kwargs: Optional[Dict[Any, Any]] = None 146 create_wandb_logger: Optional[bool] = False 147 wandb_logger_kwargs: Optional[Dict[Any, Any]] = None 148 create_mlflow_logger: Optional[bool] = False 149 mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() 150 create_dllogger_logger: Optional[bool] = False 151 dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() 152 create_clearml_logger: Optional[bool] = False 153 clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() 154 # Checkpointing parameters 155 create_checkpoint_callback: Optional[bool] = True 156 checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() 157 create_early_stopping_callback: Optional[bool] = False 158 early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() 159 create_preemption_callback: Optional[bool] = True 160 # Additional exp_manager arguments 161 files_to_copy: Optional[List[str]] = None 162 # logs timing of train/val/test steps 163 log_step_timing: Optional[bool] = True 164 step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() 165 # Configures creation of log files for different ranks 166 log_local_rank_0_only: Optional[bool] = False 167 log_global_rank_0_only: Optional[bool] = False 168 # disable initial validation when resuming from a checkpoint saved during validation 169 disable_validation_on_resume: Optional[bool] = True 170 ema: Optional[EMAParams] = EMAParams() 171 # Wall clock time limit 172 max_time_per_run: Optional[str] = None 173 # time to sleep non 0 ranks during initialization 174 seconds_to_sleep: float = 5 175 176 177 class TimingCallback(Callback): 178 """ 179 Logs execution time of train/val/test steps 180 """ 181 182 def __init__(self, timer_kwargs={}): 183 self.timer = timers.NamedTimer(**timer_kwargs) 184 185 def _on_batch_start(self, name): 186 # reset only if we do not return mean of a sliding window 187 if self.timer.buffer_size <= 0: 188 self.timer.reset(name) 189 190 self.timer.start(name) 191 192 def _on_batch_end(self, name, pl_module): 193 self.timer.stop(name) 194 # Set the `batch_size=1` as WAR for `dataloader_iter`, which is not used for any metric 195 pl_module.log( 196 name + ' in s', 197 self.timer[name], 198 on_step=True, 199 on_epoch=False, 200 batch_size=1, 201 prog_bar=(name == "train_step_timing"), 202 ) 203 204 def on_train_batch_start(self, trainer, pl_module, batch, batch_idx): 205 self._on_batch_start("train_step_timing") 206 207 def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): 208 self._on_batch_end("train_step_timing", pl_module) 209 210 def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 211 self._on_batch_start("validation_step_timing") 212 213 def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 214 self._on_batch_end("validation_step_timing", pl_module) 215 216 def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 217 self._on_batch_start("test_step_timing") 218 219 def on_test_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 220 self._on_batch_end("test_step_timing", pl_module) 221 222 def on_before_backward(self, trainer, pl_module, loss): 223 self._on_batch_start("train_backward_timing") 224 225 def on_after_backward(self, trainer, pl_module): 226 self._on_batch_end("train_backward_timing", pl_module) 227 228 229 def exp_manager(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None) -> Optional[Path]: 230 """ 231 exp_manager is a helper function used to manage folders for experiments. It follows the pytorch lightning paradigm 232 of exp_dir/model_or_experiment_name/version. If the lightning trainer has a logger, exp_manager will get exp_dir, 233 name, and version from the logger. Otherwise it will use the exp_dir and name arguments to create the logging 234 directory. exp_manager also allows for explicit folder creation via explicit_log_dir. 235 236 The version can be a datetime string or an integer. Datestime version can be disabled if use_datetime_version is set 237 to False. It optionally creates TensorBoardLogger, WandBLogger, DLLogger, MLFlowLogger, ClearMLLogger, 238 ModelCheckpoint objects from pytorch lightning. 239 It copies sys.argv, and git information if available to the logging directory. It creates a log file for each 240 process to log their output into. 241 242 exp_manager additionally has a resume feature (resume_if_exists) which can be used to continuing training from 243 the constructed log_dir. When you need to continue the training repeatedly (like on a cluster which you need 244 multiple consecutive jobs), you need to avoid creating the version folders. Therefore from v1.0.0, when 245 resume_if_exists is set to True, creating the version folders is ignored. 246 247 Args: 248 trainer (pytorch_lightning.Trainer): The lightning trainer. 249 cfg (DictConfig, dict): Can have the following keys: 250 251 - explicit_log_dir (str, Path): Can be used to override exp_dir/name/version folder creation. Defaults to 252 None, which will use exp_dir, name, and version to construct the logging directory. 253 - exp_dir (str, Path): The base directory to create the logging directory. Defaults to None, which logs to 254 ./nemo_experiments. 255 - name (str): The name of the experiment. Defaults to None which turns into "default" via name = name or 256 "default". 257 - version (str): The version of the experiment. Defaults to None which uses either a datetime string or 258 lightning's TensorboardLogger system of using version_{int}. 259 - use_datetime_version (bool): Whether to use a datetime string for version. Defaults to True. 260 - resume_if_exists (bool): Whether this experiment is resuming from a previous run. If True, it sets 261 trainer._checkpoint_connector._ckpt_path so that the trainer should auto-resume. exp_manager will move files 262 under log_dir to log_dir/run_{int}. Defaults to False. From v1.0.0, when resume_if_exists is True, 263 we would not create version folders to make it easier to find the log folder for next runs. 264 - resume_past_end (bool): exp_manager errors out if resume_if_exists is True and a checkpoint matching 265 ``*end.ckpt`` indicating a previous training run fully completed. This behaviour can be disabled, in which 266 case the ``*end.ckpt`` will be loaded by setting resume_past_end to True. Defaults to False. 267 - resume_ignore_no_checkpoint (bool): exp_manager errors out if resume_if_exists is True and no checkpoint 268 could be found. This behaviour can be disabled, in which case exp_manager will print a message and 269 continue without restoring, by setting resume_ignore_no_checkpoint to True. Defaults to False. 270 - resume_from_checkpoint (str): Can be used to specify a path to a specific checkpoint file to load from. This will 271 override any checkpoint found when resume_if_exists is True. Defaults to None. 272 - create_tensorboard_logger (bool): Whether to create a tensorboard logger and attach it to the pytorch 273 lightning trainer. Defaults to True. 274 - summary_writer_kwargs (dict): A dictionary of kwargs that can be passed to lightning's TensorboardLogger 275 class. Note that log_dir is passed by exp_manager and cannot exist in this dict. Defaults to None. 276 - create_wandb_logger (bool): Whether to create a Weights and Baises logger and attach it to the pytorch 277 lightning trainer. Defaults to False. 278 - wandb_logger_kwargs (dict): A dictionary of kwargs that can be passed to lightning's WandBLogger 279 class. Note that name and project are required parameters if create_wandb_logger is True. 280 Defaults to None. 281 - create_mlflow_logger (bool): Whether to create an MLFlow logger and attach it to the pytorch lightning 282 training. Defaults to False 283 - mlflow_logger_kwargs (dict): optional parameters for the MLFlow logger 284 - create_dllogger_logger (bool): Whether to create an DLLogger logger and attach it to the pytorch lightning 285 training. Defaults to False 286 - dllogger_logger_kwargs (dict): optional parameters for the DLLogger logger 287 - create_clearml_logger (bool): Whether to create an ClearML logger and attach it to the pytorch lightning 288 training. Defaults to False 289 - clearml_logger_kwargs (dict): optional parameters for the ClearML logger 290 - create_checkpoint_callback (bool): Whether to create a ModelCheckpoint callback and attach it to the 291 pytorch lightning trainer. The ModelCheckpoint saves the top 3 models with the best "val_loss", the most 292 recent checkpoint under ``*last.ckpt``, and the final checkpoint after training completes under ``*end.ckpt``. 293 Defaults to True. 294 - create_early_stopping_callback (bool): Flag to decide if early stopping should be used to stop training. Default is False. 295 See EarlyStoppingParams dataclass above. 296 - create_preemption_callback (bool): Flag to decide whether to enable preemption callback to save checkpoints and exit training 297 immediately upon preemption. Default is True. 298 - files_to_copy (list): A list of files to copy to the experiment logging directory. Defaults to None which 299 copies no files. 300 - log_local_rank_0_only (bool): Whether to only create log files for local rank 0. Defaults to False. 301 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 302 - log_global_rank_0_only (bool): Whether to only create log files for global rank 0. Defaults to False. 303 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 304 - max_time (str): The maximum wall clock time *per run*. This is intended to be used on clusters where you want 305 a checkpoint to be saved after this specified time and be able to resume from that checkpoint. Defaults to None. 306 - seconds_to_sleep (float): seconds to sleep non rank 0 processes for. Used to give enough time for rank 0 to initialize 307 308 returns: 309 log_dir (Path): The final logging directory where logging files are saved. Usually the concatenation of 310 exp_dir, name, and version. 311 """ 312 # Add rank information to logger 313 # Note: trainer.global_rank and trainer.is_global_zero are not set until trainer.fit, so have to hack around it 314 local_rank = int(os.environ.get("LOCAL_RANK", 0)) 315 global_rank = trainer.node_rank * trainer.num_devices + local_rank 316 logging.rank = global_rank 317 318 if cfg is None: 319 logging.error("exp_manager did not receive a cfg argument. It will be disabled.") 320 return 321 if trainer.fast_dev_run: 322 logging.info("Trainer was called with fast_dev_run. exp_manager will return without any functionality.") 323 return 324 325 # Ensure passed cfg is compliant with ExpManagerConfig 326 schema = OmegaConf.structured(ExpManagerConfig) 327 if isinstance(cfg, dict): 328 cfg = OmegaConf.create(cfg) 329 elif not isinstance(cfg, DictConfig): 330 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 331 cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True)) 332 cfg = OmegaConf.merge(schema, cfg) 333 334 error_checks(trainer, cfg) # Ensures that trainer options are compliant with NeMo and exp_manager arguments 335 336 log_dir, exp_dir, name, version = get_log_dir( 337 trainer=trainer, 338 exp_dir=cfg.exp_dir, 339 name=cfg.name, 340 version=cfg.version, 341 explicit_log_dir=cfg.explicit_log_dir, 342 use_datetime_version=cfg.use_datetime_version, 343 resume_if_exists=cfg.resume_if_exists, 344 ) 345 346 check_resume( 347 trainer, 348 log_dir, 349 cfg.resume_if_exists, 350 cfg.resume_past_end, 351 cfg.resume_ignore_no_checkpoint, 352 cfg.checkpoint_callback_params.dirpath, 353 cfg.resume_from_checkpoint, 354 ) 355 356 checkpoint_name = name 357 # If name returned from get_log_dir is "", use cfg.name for checkpointing 358 if checkpoint_name is None or checkpoint_name == '': 359 checkpoint_name = cfg.name or "default" 360 361 # Set mlflow name if it's not set, before the main name is erased 362 if cfg.create_mlflow_logger and (not cfg.mlflow_logger_kwargs.get("experiment_name", None)): 363 cfg.mlflow_logger_kwargs.experiment_name = cfg.name 364 logging.warning( 365 'mlflow logger specified but no experiment name set. Using the same as Tensorboard: %s', 366 cfg.mlflow_logger_kwargs.experiment_name, 367 ) 368 369 cfg.name = name # Used for configure_loggers so that the log_dir is properly set even if name is "" 370 cfg.version = version 371 372 # update app_state with log_dir, exp_dir, etc 373 app_state = AppState() 374 app_state.log_dir = log_dir 375 app_state.exp_dir = exp_dir 376 app_state.name = name 377 app_state.version = version 378 app_state.checkpoint_name = checkpoint_name 379 app_state.create_checkpoint_callback = cfg.create_checkpoint_callback 380 app_state.checkpoint_callback_params = cfg.checkpoint_callback_params 381 382 # Create the logging directory if it does not exist 383 os.makedirs(log_dir, exist_ok=True) # Cannot limit creation to global zero as all ranks write to own log file 384 logging.info(f'Experiments will be logged at {log_dir}') 385 trainer._default_root_dir = log_dir 386 387 if cfg.log_local_rank_0_only is True and cfg.log_global_rank_0_only is True: 388 raise ValueError( 389 f"Cannot set both log_local_rank_0_only and log_global_rank_0_only to True. Please set either one or neither." 390 ) 391 392 # This is set if the env var NEMO_TESTING is set to True. 393 nemo_testing = get_envbool(NEMO_ENV_VARNAME_TESTING, False) 394 395 # Handle logging to file 396 log_file = log_dir / f'nemo_log_globalrank-{global_rank}_localrank-{local_rank}.txt' 397 if cfg.log_local_rank_0_only is True and not nemo_testing: 398 if local_rank == 0: 399 logging.add_file_handler(log_file) 400 elif cfg.log_global_rank_0_only is True and not nemo_testing: 401 if global_rank == 0: 402 logging.add_file_handler(log_file) 403 else: 404 # Logs on all ranks. 405 logging.add_file_handler(log_file) 406 407 # For some reason, LearningRateLogger requires trainer to have a logger. Safer to create logger on all ranks 408 # not just global rank 0. 409 if ( 410 cfg.create_tensorboard_logger 411 or cfg.create_wandb_logger 412 or cfg.create_mlflow_logger 413 or cfg.create_dllogger_logger 414 or cfg.create_clearml_logger 415 ): 416 configure_loggers( 417 trainer, 418 exp_dir, 419 log_dir, 420 cfg.name, 421 cfg.version, 422 cfg.checkpoint_callback_params, 423 cfg.create_tensorboard_logger, 424 cfg.summary_writer_kwargs, 425 cfg.create_wandb_logger, 426 cfg.wandb_logger_kwargs, 427 cfg.create_mlflow_logger, 428 cfg.mlflow_logger_kwargs, 429 cfg.create_dllogger_logger, 430 cfg.dllogger_logger_kwargs, 431 cfg.create_clearml_logger, 432 cfg.clearml_logger_kwargs, 433 ) 434 435 # add loggers timing callbacks 436 if cfg.log_step_timing: 437 timing_callback = TimingCallback(timer_kwargs=cfg.step_timing_kwargs or {}) 438 trainer.callbacks.insert(0, timing_callback) 439 440 if cfg.ema.enable: 441 ema_callback = EMA( 442 decay=cfg.ema.decay, 443 validate_original_weights=cfg.ema.validate_original_weights, 444 cpu_offload=cfg.ema.cpu_offload, 445 every_n_steps=cfg.ema.every_n_steps, 446 ) 447 trainer.callbacks.append(ema_callback) 448 449 if cfg.create_early_stopping_callback: 450 early_stop_callback = EarlyStopping(**cfg.early_stopping_callback_params) 451 trainer.callbacks.append(early_stop_callback) 452 453 if cfg.create_checkpoint_callback: 454 configure_checkpointing( 455 trainer, 456 log_dir, 457 checkpoint_name, 458 cfg.resume_if_exists, 459 cfg.checkpoint_callback_params, 460 cfg.create_preemption_callback, 461 ) 462 463 if cfg.disable_validation_on_resume: 464 # extend training loop to skip initial validation when resuming from checkpoint 465 configure_no_restart_validation_training_loop(trainer) 466 # Setup a stateless timer for use on clusters. 467 if cfg.max_time_per_run is not None: 468 found_ptl_timer = False 469 for idx, callback in enumerate(trainer.callbacks): 470 if isinstance(callback, Timer): 471 # NOTE: PTL does not expose a `trainer.max_time`. By the time we are in this function, PTL has already setup a timer if the user specifies `trainer.max_time` so best we can do is replace that. 472 # Working: If only `trainer.max_time` is set - it behaves as a normal PTL timer. If only `exp_manager.max_time_per_run` is set - it behaves as a StateLessTimer. If both are set, it also behaves as a StateLessTimer. 473 logging.warning( 474 f'Found a PTL Timer callback, replacing with a StatelessTimer callback. This will happen if you set trainer.max_time as well as exp_manager.max_time_per_run.' 475 ) 476 trainer.callbacks[idx] = StatelessTimer(cfg.max_time_per_run) 477 found_ptl_timer = True 478 break 479 480 if not found_ptl_timer: 481 trainer.max_time = cfg.max_time_per_run 482 trainer.callbacks.append(StatelessTimer(cfg.max_time_per_run)) 483 484 if is_global_rank_zero(): 485 # Move files_to_copy to folder and add git information if present 486 if cfg.files_to_copy: 487 for _file in cfg.files_to_copy: 488 copy(Path(_file), log_dir) 489 490 # Create files for cmd args and git info 491 with open(log_dir / 'cmd-args.log', 'w', encoding='utf-8') as _file: 492 _file.write(" ".join(sys.argv)) 493 494 # Try to get git hash 495 git_repo, git_hash = get_git_hash() 496 if git_repo: 497 with open(log_dir / 'git-info.log', 'w', encoding='utf-8') as _file: 498 _file.write(f'commit hash: {git_hash}') 499 _file.write(get_git_diff()) 500 501 # Add err_file logging to global_rank zero 502 logging.add_err_file_handler(log_dir / 'nemo_error_log.txt') 503 504 # Add lightning file logging to global_rank zero 505 add_filehandlers_to_pl_logger(log_dir / 'lightning_logs.txt', log_dir / 'nemo_error_log.txt') 506 507 elif trainer.num_nodes * trainer.num_devices > 1: 508 # sleep other ranks so rank 0 can finish 509 # doing the initialization such as moving files 510 time.sleep(cfg.seconds_to_sleep) 511 512 return log_dir 513 514 515 def error_checks(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None): 516 """ 517 Checks that the passed trainer is compliant with NeMo and exp_manager's passed configuration. Checks that: 518 - Throws error when hydra has changed the working directory. This causes issues with lightning's DDP 519 - Throws error when trainer has loggers defined but create_tensorboard_logger or create_wandB_logger 520 or create_mlflow_logger or create_dllogger_logger is True 521 - Prints error messages when 1) run on multi-node and not Slurm, and 2) run on multi-gpu without DDP 522 """ 523 if HydraConfig.initialized() and get_original_cwd() != os.getcwd(): 524 raise ValueError( 525 "Hydra changed the working directory. This interferes with ExpManger's functionality. Please pass " 526 "hydra.run.dir=. to your python script." 527 ) 528 if trainer.logger is not None and ( 529 cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger 530 ): 531 raise LoggerMisconfigurationError( 532 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and either " 533 f"create_tensorboard_logger: {cfg.create_tensorboard_logger} or create_wandb_logger: " 534 f"{cfg.create_wandb_logger} or create_mlflow_logger: {cfg.create_mlflow_logger}" 535 f"or create_dllogger_logger: {cfg.create_mlflow_logger} was set to True. " 536 "These can only be used if trainer does not already have a logger." 537 ) 538 if trainer.num_nodes > 1 and not check_slurm(trainer): 539 logging.error( 540 "You are running multi-node training without SLURM handling the processes." 541 " Please note that this is not tested in NeMo and could result in errors." 542 ) 543 if trainer.num_devices > 1 and not isinstance(trainer.strategy, DDPStrategy): 544 logging.error( 545 "You are running multi-gpu without ddp.Please note that this is not tested in NeMo and could result in " 546 "errors." 547 ) 548 549 550 def check_resume( 551 trainer: 'pytorch_lightning.Trainer', 552 log_dir: str, 553 resume_if_exists: bool = False, 554 resume_past_end: bool = False, 555 resume_ignore_no_checkpoint: bool = False, 556 dirpath: str = None, 557 resume_from_checkpoint: str = None, 558 ): 559 """Checks that resume=True was used correctly with the arguments pass to exp_manager. Sets 560 trainer._checkpoint_connector._ckpt_path as necessary. 561 562 Returns: 563 log_dir (Path): The log_dir 564 exp_dir (str): The base exp_dir without name nor version 565 name (str): The name of the experiment 566 version (str): The version of the experiment 567 568 Raises: 569 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 570 ValueError: If resume is True, and there were more than 1 checkpoint could found. 571 """ 572 573 if not log_dir: 574 raise ValueError(f"Resuming requires the log_dir {log_dir} to be passed to exp_manager") 575 576 checkpoint = None 577 if resume_from_checkpoint: 578 checkpoint = resume_from_checkpoint 579 if resume_if_exists: 580 # Use <log_dir>/checkpoints/ unless `dirpath` is set 581 checkpoint_dir = Path(dirpath) if dirpath else Path(Path(log_dir) / "checkpoints") 582 583 # when using distributed checkpointing, checkpoint_dir is a directory of directories 584 # we check for this here 585 dist_checkpoints = [d for d in list(checkpoint_dir.glob("*")) if d.is_dir()] 586 end_dist_checkpoints = [d for d in dist_checkpoints if d.match("*end")] 587 last_dist_checkpoints = [d for d in dist_checkpoints if d.match("*last")] 588 589 end_checkpoints = end_dist_checkpoints if end_dist_checkpoints else list(checkpoint_dir.rglob("*end.ckpt")) 590 last_checkpoints = last_dist_checkpoints if last_dist_checkpoints else list(checkpoint_dir.rglob("*last.ckpt")) 591 592 if not checkpoint_dir.exists() or (not len(end_checkpoints) > 0 and not len(last_checkpoints) > 0): 593 if resume_ignore_no_checkpoint: 594 warn = f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. " 595 if checkpoint is None: 596 warn += "Training from scratch." 597 elif checkpoint == resume_from_checkpoint: 598 warn += f"Training from {resume_from_checkpoint}." 599 logging.warning(warn) 600 else: 601 raise NotFoundError( 602 f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. Cannot resume." 603 ) 604 elif len(end_checkpoints) > 0: 605 if resume_past_end: 606 if len(end_checkpoints) > 1: 607 if 'mp_rank' in str(end_checkpoints[0]): 608 checkpoint = end_checkpoints[0] 609 else: 610 raise ValueError(f"Multiple checkpoints {end_checkpoints} that matches *end.ckpt.") 611 else: 612 raise ValueError( 613 f"Found {end_checkpoints[0]} indicating that the last training run has already completed." 614 ) 615 elif len(last_checkpoints) > 1: 616 if 'mp_rank' in str(last_checkpoints[0]) or 'tp_rank' in str(last_checkpoints[0]): 617 checkpoint = last_checkpoints[0] 618 checkpoint = uninject_model_parallel_rank(checkpoint) 619 else: 620 raise ValueError(f"Multiple checkpoints {last_checkpoints} that matches *last.ckpt.") 621 else: 622 checkpoint = last_checkpoints[0] 623 624 # PTL 2.0 supports ckpt_path instead of resume_from_checkpoint as the trainer flag 625 if checkpoint is not None: 626 trainer.ckpt_path = str(checkpoint) 627 logging.info(f'Resuming training from checkpoint: {trainer.ckpt_path}') 628 629 if is_global_rank_zero(): 630 # Check to see if any files exist that need to be moved 631 files_to_move = [] 632 if Path(log_dir).exists(): 633 for child in Path(log_dir).iterdir(): 634 if child.is_file(): 635 files_to_move.append(child) 636 637 if len(files_to_move) > 0: 638 # Move old files to a new folder 639 other_run_dirs = Path(log_dir).glob("run_*") 640 run_count = 0 641 for fold in other_run_dirs: 642 if fold.is_dir(): 643 run_count += 1 644 new_run_dir = Path(Path(log_dir) / f"run_{run_count}") 645 new_run_dir.mkdir() 646 for _file in files_to_move: 647 move(str(_file), str(new_run_dir)) 648 649 650 def check_explicit_log_dir( 651 trainer: 'pytorch_lightning.Trainer', explicit_log_dir: Union[Path, str], exp_dir: str, name: str, version: str 652 ) -> Tuple[Path, str, str, str]: 653 """ Checks that the passed arguments are compatible with explicit_log_dir. 654 655 Returns: 656 log_dir (Path): the log_dir 657 exp_dir (str): the base exp_dir without name nor version 658 name (str): The name of the experiment 659 version (str): The version of the experiment 660 661 Raise: 662 LoggerMisconfigurationError 663 """ 664 if trainer.logger is not None: 665 raise LoggerMisconfigurationError( 666 "The pytorch lightning trainer that was passed to exp_manager contained a logger and explicit_log_dir: " 667 f"{explicit_log_dir} was pass to exp_manager. Please remove the logger from the lightning trainer." 668 ) 669 # Checking only (explicit_log_dir) vs (exp_dir and version). 670 # The `name` will be used as the actual name of checkpoint/archive. 671 if exp_dir or version: 672 logging.error( 673 f"exp_manager received explicit_log_dir: {explicit_log_dir} and at least one of exp_dir: {exp_dir}, " 674 f"or version: {version}. Please note that exp_dir, name, and version will be ignored." 675 ) 676 if is_global_rank_zero() and Path(explicit_log_dir).exists(): 677 logging.warning(f"Exp_manager is logging to {explicit_log_dir}, but it already exists.") 678 return Path(explicit_log_dir), str(explicit_log_dir), "", "" 679 680 681 def get_log_dir( 682 trainer: 'pytorch_lightning.Trainer', 683 exp_dir: str = None, 684 name: str = None, 685 version: str = None, 686 explicit_log_dir: str = None, 687 use_datetime_version: bool = True, 688 resume_if_exists: bool = False, 689 ) -> Tuple[Path, str, str, str]: 690 """ 691 Obtains the log_dir used for exp_manager. 692 693 Returns: 694 log_dir (Path): the log_dir 695 exp_dir (str): the base exp_dir without name nor version 696 name (str): The name of the experiment 697 version (str): The version of the experiment 698 explicit_log_dir (str): The explicit path to the log folder. Defaults to False. 699 use_datetime_version (bool): Uses date and time as the version of the log folder. Defaults to True. 700 resume_if_exists (bool): if resume_if_exists of the exp_manager's config is enabled or not. When enabled, the 701 version folders would not get created. 702 703 Raise: 704 LoggerMisconfigurationError: If trainer is incompatible with arguments 705 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 706 ValueError: If resume is True, and there were more than 1 checkpoint could found. 707 """ 708 if explicit_log_dir: # If explicit log_dir was passed, short circuit 709 return check_explicit_log_dir(trainer, explicit_log_dir, exp_dir, name, version) 710 711 # Default exp_dir to ./nemo_experiments if None was passed 712 _exp_dir = exp_dir 713 if exp_dir is None: 714 _exp_dir = str(Path.cwd() / 'nemo_experiments') 715 716 # If the user has already defined a logger for the trainer, use the logger defaults for logging directory 717 if trainer.logger is not None: 718 if trainer.logger.save_dir: 719 if exp_dir: 720 raise LoggerMisconfigurationError( 721 "The pytorch lightning trainer that was passed to exp_manager contained a logger, the logger's " 722 f"save_dir was not None, and exp_dir ({exp_dir}) was not None. If trainer.logger.save_dir " 723 "exists, exp_manager will use trainer.logger.save_dir as the logging directory and exp_dir " 724 "must be None." 725 ) 726 _exp_dir = trainer.logger.save_dir 727 if name: 728 raise LoggerMisconfigurationError( 729 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and name: " 730 f"{name} was also passed to exp_manager. If the trainer contains a " 731 "logger, exp_manager will use trainer.logger.name, and name passed to exp_manager must be None." 732 ) 733 name = trainer.logger.name 734 version = f"version_{trainer.logger.version}" 735 # Use user-defined exp_dir, project_name, exp_name, and versioning options 736 else: 737 name = name or "default" 738 version = version or os.environ.get(NEMO_ENV_VARNAME_VERSION, None) 739 740 if not version: 741 if resume_if_exists: 742 logging.warning( 743 "No version folders would be created under the log folder as 'resume_if_exists' is enabled." 744 ) 745 version = None 746 elif is_global_rank_zero(): 747 if use_datetime_version: 748 version = time.strftime('%Y-%m-%d_%H-%M-%S') 749 else: 750 tensorboard_logger = TensorBoardLogger(save_dir=Path(_exp_dir), name=name, version=version) 751 version = f"version_{tensorboard_logger.version}" 752 os.environ[NEMO_ENV_VARNAME_VERSION] = "" if version is None else version 753 754 log_dir = Path(_exp_dir) / Path(str(name)) / Path("" if version is None else str(version)) 755 return log_dir, str(_exp_dir), name, version 756 757 758 def get_git_hash(): 759 """ 760 Helper function that tries to get the commit hash if running inside a git folder 761 762 returns: 763 Bool: Whether the git subprocess ran without error 764 str: git subprocess output or error message 765 """ 766 try: 767 return ( 768 True, 769 subprocess.check_output(['git', 'rev-parse', 'HEAD'], stderr=subprocess.STDOUT).decode(), 770 ) 771 except subprocess.CalledProcessError as err: 772 return False, "{}\n".format(err.output.decode("utf-8")) 773 774 775 def get_git_diff(): 776 """ 777 Helper function that tries to get the git diff if running inside a git folder 778 779 returns: 780 Bool: Whether the git subprocess ran without error 781 str: git subprocess output or error message 782 """ 783 try: 784 return subprocess.check_output(['git', 'diff'], stderr=subprocess.STDOUT).decode() 785 except subprocess.CalledProcessError as err: 786 return "{}\n".format(err.output.decode("utf-8")) 787 788 789 def configure_loggers( 790 trainer: 'pytorch_lightning.Trainer', 791 exp_dir: [Path, str], 792 log_dir: [Path, str], 793 name: str, 794 version: str, 795 checkpoint_callback_params: dict, 796 create_tensorboard_logger: bool, 797 summary_writer_kwargs: dict, 798 create_wandb_logger: bool, 799 wandb_kwargs: dict, 800 create_mlflow_logger: bool, 801 mlflow_kwargs: dict, 802 create_dllogger_logger: bool, 803 dllogger_kwargs: dict, 804 create_clearml_logger: bool, 805 clearml_kwargs: dict, 806 ): 807 """ 808 Creates TensorboardLogger and/or WandBLogger / MLFlowLogger / DLlogger / ClearMLLogger and attach them to trainer. 809 Raises ValueError if summary_writer_kwargs or wandb_kwargs are misconfigured. 810 """ 811 # Potentially create tensorboard logger and/or WandBLogger / MLFlowLogger / DLLogger 812 logger_list = [] 813 if create_tensorboard_logger: 814 if summary_writer_kwargs is None: 815 summary_writer_kwargs = {} 816 elif "log_dir" in summary_writer_kwargs: 817 raise ValueError( 818 "You cannot pass `log_dir` as part of `summary_writer_kwargs`. `log_dir` is handled by lightning's " 819 "TensorBoardLogger logger." 820 ) 821 tensorboard_logger = TensorBoardLogger(save_dir=exp_dir, name=name, version=version, **summary_writer_kwargs) 822 logger_list.append(tensorboard_logger) 823 logging.info("TensorboardLogger has been set up") 824 825 if create_wandb_logger: 826 if wandb_kwargs is None: 827 wandb_kwargs = {} 828 if "name" not in wandb_kwargs and "project" not in wandb_kwargs: 829 raise ValueError("name and project are required for wandb_logger") 830 831 # Update the wandb save_dir 832 if wandb_kwargs.get('save_dir', None) is None: 833 wandb_kwargs['save_dir'] = exp_dir 834 os.makedirs(wandb_kwargs['save_dir'], exist_ok=True) 835 wandb_logger = WandbLogger(version=version, **wandb_kwargs) 836 837 logger_list.append(wandb_logger) 838 logging.info("WandBLogger has been set up") 839 840 if create_mlflow_logger: 841 mlflow_logger = MLFlowLogger(run_name=version, **mlflow_kwargs) 842 843 logger_list.append(mlflow_logger) 844 logging.info("MLFlowLogger has been set up") 845 846 if create_dllogger_logger: 847 dllogger_logger = DLLogger(**dllogger_kwargs) 848 849 logger_list.append(dllogger_logger) 850 logging.info("DLLogger has been set up") 851 852 if create_clearml_logger: 853 clearml_logger = ClearMLLogger( 854 clearml_cfg=clearml_kwargs, 855 log_dir=log_dir, 856 prefix=name, 857 save_best_model=checkpoint_callback_params.save_best_model, 858 ) 859 860 logger_list.append(clearml_logger) 861 logging.info("ClearMLLogger has been set up") 862 863 trainer._logger_connector.configure_logger(logger_list) 864 865 866 def configure_checkpointing( 867 trainer: 'pytorch_lightning.Trainer', 868 log_dir: Path, 869 name: str, 870 resume: bool, 871 params: 'DictConfig', 872 create_preemption_callback: bool, 873 ): 874 """ Adds ModelCheckpoint to trainer. Raises CheckpointMisconfigurationError if trainer already has a ModelCheckpoint 875 callback 876 """ 877 for callback in trainer.callbacks: 878 if isinstance(callback, ModelCheckpoint): 879 raise CheckpointMisconfigurationError( 880 "The pytorch lightning trainer that was passed to exp_manager contained a ModelCheckpoint " 881 "and create_checkpoint_callback was set to True. Please either set create_checkpoint_callback " 882 "to False, or remove ModelCheckpoint from the lightning trainer" 883 ) 884 # Create the callback and attach it to trainer 885 if "filepath" in params: 886 if params.filepath is not None: 887 logging.warning("filepath is deprecated. Please switch to dirpath and filename instead") 888 if params.dirpath is None: 889 params.dirpath = Path(params.filepath).parent 890 if params.filename is None: 891 params.filename = Path(params.filepath).name 892 with open_dict(params): 893 del params["filepath"] 894 if params.dirpath is None: 895 params.dirpath = Path(log_dir / 'checkpoints') 896 if params.filename is None: 897 params.filename = f'{name}--{{{params.monitor}:.4f}}-{{epoch}}' 898 if params.prefix is None: 899 params.prefix = name 900 NeMoModelCheckpoint.CHECKPOINT_NAME_LAST = params.filename + '-last' 901 902 logging.debug(params.dirpath) 903 logging.debug(params.filename) 904 logging.debug(params.prefix) 905 906 if "val" in params.monitor: 907 if ( 908 trainer.max_epochs is not None 909 and trainer.max_epochs != -1 910 and trainer.max_epochs < trainer.check_val_every_n_epoch 911 ): 912 logging.error( 913 "The checkpoint callback was told to monitor a validation value but trainer.max_epochs(" 914 f"{trainer.max_epochs}) was less than trainer.check_val_every_n_epoch({trainer.check_val_every_n_epoch}" 915 f"). It is very likely this run will fail with ModelCheckpoint(monitor='{params.monitor}') not found " 916 "in the returned metrics. Please ensure that validation is run within trainer.max_epochs." 917 ) 918 elif trainer.max_steps is not None and trainer.max_steps != -1: 919 logging.warning( 920 "The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to " 921 f"{trainer.max_steps}. Please ensure that max_steps will run for at least " 922 f"{trainer.check_val_every_n_epoch} epochs to ensure that checkpointing will not error out." 923 ) 924 925 checkpoint_callback = NeMoModelCheckpoint(n_resume=resume, **params) 926 checkpoint_callback.last_model_path = trainer.ckpt_path or "" 927 if 'mp_rank' in checkpoint_callback.last_model_path or 'tp_rank' in checkpoint_callback.last_model_path: 928 checkpoint_callback.last_model_path = uninject_model_parallel_rank(checkpoint_callback.last_model_path) 929 trainer.callbacks.append(checkpoint_callback) 930 if create_preemption_callback: 931 # Check if cuda is avialable as preemption is supported only on GPUs 932 if torch.cuda.is_available(): 933 ## By default PreemptionCallback handles SIGTERM. To handle other signals pass the signal in the call as below: 934 ## PreemptionCallback(checkpoint_callback, signal.SIGCHLD) 935 preemption_callback = PreemptionCallback(checkpoint_callback) 936 trainer.callbacks.append(preemption_callback) 937 else: 938 logging.info("Preemption is supported only on GPUs, disabling preemption") 939 940 941 def check_slurm(trainer): 942 try: 943 return trainer.accelerator_connector.is_slurm_managing_tasks 944 except AttributeError: 945 return False 946 947 948 class StatelessTimer(Timer): 949 """Extension of PTL timers to be per run.""" 950 951 def __init__(self, duration: timedelta = None, interval: str = Interval.step, verbose: bool = True,) -> None: 952 super().__init__(duration, interval, verbose) 953 954 # Override PTL Timer's state dict to not store elapsed time information so that we can restore and continue training. 955 def state_dict(self) -> Dict[str, Any]: 956 return {} 957 958 def load_state_dict(self, state_dict: Dict[str, Any]) -> None: 959 return 960 961 962 def configure_no_restart_validation_training_loop(trainer: pytorch_lightning.Trainer) -> None: 963 if type(trainer.fit_loop.epoch_loop) != _TrainingEpochLoop: 964 warnings.warn("Detected custom epoch loop. Skipping no validation on restart support.", UserWarning) 965 return 966 ## Pass trainer object to avoid trainer getting overwritten as None 967 loop = SkipResumeTrainingValidationLoop(trainer, trainer.min_steps, trainer.max_steps) 968 trainer.fit_loop.epoch_loop = loop 969 970 971 class SkipResumeTrainingValidationLoop(_TrainingEpochLoop): 972 """ 973 Extend the PTL Epoch loop to skip validating when resuming. 974 This happens when resuming a checkpoint that has already run validation, but loading restores 975 the training state before validation has run. 976 """ 977 978 def _should_check_val_fx(self) -> bool: 979 if self.restarting and self.global_step % self.trainer.val_check_batch == 0: 980 return False 981 return super()._should_check_val_fx() 982 983 984 def clean_exp_ckpt(exp_log_dir: Union[str, Path], remove_ckpt: bool = True, remove_nemo: bool = False): 985 """ 986 Helper method that removes Pytorch Lightning .ckpt files or NeMo .nemo files from the checkpoint directory 987 988 Args: 989 exp_log_dir: str path to the root directory of the current experiment. 990 remove_ckpt: bool, whether to remove all *.ckpt files in the checkpoints directory. 991 remove_nemo: bool, whether to remove all *.nemo files in the checkpoints directory. 992 """ 993 exp_log_dir = str(exp_log_dir) 994 995 if remove_ckpt: 996 logging.info("Deleting *.ckpt files ...") 997 ckpt_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.ckpt")) 998 for filepath in ckpt_files: 999 os.remove(filepath) 1000 logging.info(f"Deleted file : {filepath}") 1001 1002 if remove_nemo: 1003 logging.info("Deleting *.nemo files ...") 1004 nemo_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.nemo")) 1005 for filepath in nemo_files: 1006 os.remove(filepath) 1007 logging.info(f"Deleted file : {filepath}") 1008 [end of nemo/utils/exp_manager.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR model with CTC decoder. To evaluate a model with 19 # Transducer (RNN-T) decoder use another script 'scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py'. 20 # NeMo's beam search decoders are capable of using the KenLM's N-gram models 21 # to find the best candidates. This script supports both character level and BPE level 22 # encodings and models which is detected automatically from the type of the model. 23 # You may train the LM model with 'scripts/asr_language_modeling/ngram_lm/train_kenlm.py'. 24 25 # Config Help 26 27 To discover all arguments of the script, please run : 28 python eval_beamsearch_ngram.py --help 29 python eval_beamsearch_ngram.py --cfg job 30 31 # USAGE 32 33 python eval_beamsearch_ngram.py nemo_model_file=<path to the .nemo file of the model> \ 34 input_manifest=<path to the evaluation JSON manifest file> \ 35 kenlm_model_file=<path to the binary KenLM model> \ 36 beam_width=[<list of the beam widths, separated with commas>] \ 37 beam_alpha=[<list of the beam alphas, separated with commas>] \ 38 beam_beta=[<list of the beam betas, separated with commas>] \ 39 preds_output_folder=<optional folder to store the predictions> \ 40 probs_cache_file=null \ 41 decoding_mode=beamsearch_ngram 42 ... 43 44 45 # Grid Search for Hyper parameters 46 47 For grid search, you can provide a list of arguments as follows - 48 49 beam_width=[4,8,16,....] \ 50 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 51 beam_beta=[-1.0,-0.5,0.0,...,1.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 from dataclasses import dataclass, field, is_dataclass 64 from pathlib import Path 65 from typing import List, Optional 66 67 import editdistance 68 import numpy as np 69 import torch 70 from omegaconf import MISSING, OmegaConf 71 from sklearn.model_selection import ParameterGrid 72 from tqdm.auto import tqdm 73 74 import nemo.collections.asr as nemo_asr 75 from nemo.collections.asr.models import EncDecHybridRNNTCTCModel 76 from nemo.collections.asr.parts.submodules import ctc_beam_decoding 77 from nemo.collections.asr.parts.utils.transcribe_utils import PunctuationCapitalization, TextProcessingConfig 78 from nemo.core.config import hydra_runner 79 from nemo.utils import logging 80 81 # fmt: off 82 83 84 @dataclass 85 class EvalBeamSearchNGramConfig: 86 """ 87 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 88 """ 89 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 90 nemo_model_file: str = MISSING 91 92 # File paths 93 input_manifest: str = MISSING # The manifest file of the evaluation set 94 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 95 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 96 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 97 98 # Parameters for inference 99 acoustic_batch_size: int = 16 # The batch size to calculate log probabilities 100 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 101 device: str = "cuda" # The device to load the model onto to calculate log probabilities 102 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 103 104 # Beam Search hyperparameters 105 106 # The decoding scheme to be used for evaluation. 107 # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] 108 decoding_mode: str = "beamsearch_ngram" 109 110 beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding 111 beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding 112 beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding 113 114 decoding_strategy: str = "beam" 115 decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) 116 117 text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( 118 punctuation_marks = ".,?", 119 separate_punctuation = False, 120 do_lowercase = False, 121 rm_punctuation = False, 122 ) 123 # fmt: on 124 125 126 def beam_search_eval( 127 model: nemo_asr.models.ASRModel, 128 cfg: EvalBeamSearchNGramConfig, 129 all_probs: List[torch.Tensor], 130 target_transcripts: List[str], 131 preds_output_file: str = None, 132 lm_path: str = None, 133 beam_alpha: float = 1.0, 134 beam_beta: float = 0.0, 135 beam_width: int = 128, 136 beam_batch_size: int = 128, 137 progress_bar: bool = True, 138 punctuation_capitalization: PunctuationCapitalization = None, 139 ): 140 level = logging.getEffectiveLevel() 141 logging.setLevel(logging.CRITICAL) 142 # Reset config 143 model.change_decoding_strategy(None) 144 145 # Override the beam search config with current search candidate configuration 146 cfg.decoding.beam_size = beam_width 147 cfg.decoding.beam_alpha = beam_alpha 148 cfg.decoding.beam_beta = beam_beta 149 cfg.decoding.return_best_hypothesis = False 150 cfg.decoding.kenlm_path = cfg.kenlm_model_file 151 152 # Update model's decoding strategy config 153 model.cfg.decoding.strategy = cfg.decoding_strategy 154 model.cfg.decoding.beam = cfg.decoding 155 156 # Update model's decoding strategy 157 if isinstance(model, EncDecHybridRNNTCTCModel): 158 model.change_decoding_strategy(model.cfg.decoding, decoder_type='ctc') 159 decoding = model.ctc_decoding 160 else: 161 model.change_decoding_strategy(model.cfg.decoding) 162 decoding = model.decoding 163 logging.setLevel(level) 164 165 wer_dist_first = cer_dist_first = 0 166 wer_dist_best = cer_dist_best = 0 167 words_count = 0 168 chars_count = 0 169 sample_idx = 0 170 if preds_output_file: 171 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 172 173 if progress_bar: 174 it = tqdm( 175 range(int(np.ceil(len(all_probs) / beam_batch_size))), 176 desc=f"Beam search decoding with width={beam_width}, alpha={beam_alpha}, beta={beam_beta}", 177 ncols=120, 178 ) 179 else: 180 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 181 for batch_idx in it: 182 # disabling type checking 183 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 184 probs_lens = torch.tensor([prob.shape[0] for prob in probs_batch]) 185 with torch.no_grad(): 186 packed_batch = torch.zeros(len(probs_batch), max(probs_lens), probs_batch[0].shape[-1], device='cpu') 187 188 for prob_index in range(len(probs_batch)): 189 packed_batch[prob_index, : probs_lens[prob_index], :] = torch.tensor( 190 probs_batch[prob_index], device=packed_batch.device, dtype=packed_batch.dtype 191 ) 192 193 _, beams_batch = decoding.ctc_decoder_predictions_tensor( 194 packed_batch, decoder_lengths=probs_lens, return_hypotheses=True, 195 ) 196 197 for beams_idx, beams in enumerate(beams_batch): 198 target = target_transcripts[sample_idx + beams_idx] 199 target_split_w = target.split() 200 target_split_c = list(target) 201 words_count += len(target_split_w) 202 chars_count += len(target_split_c) 203 wer_dist_min = cer_dist_min = 10000 204 for candidate_idx, candidate in enumerate(beams): # type: (int, ctc_beam_decoding.rnnt_utils.Hypothesis) 205 pred_text = candidate.text 206 if cfg.text_processing.do_lowercase: 207 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 208 if cfg.text_processing.rm_punctuation: 209 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 210 if cfg.text_processing.separate_punctuation: 211 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 212 pred_split_w = pred_text.split() 213 wer_dist = editdistance.eval(target_split_w, pred_split_w) 214 pred_split_c = list(pred_text) 215 cer_dist = editdistance.eval(target_split_c, pred_split_c) 216 217 wer_dist_min = min(wer_dist_min, wer_dist) 218 cer_dist_min = min(cer_dist_min, cer_dist) 219 220 if candidate_idx == 0: 221 # first candidate 222 wer_dist_first += wer_dist 223 cer_dist_first += cer_dist 224 225 score = candidate.score 226 if preds_output_file: 227 out_file.write('{}\t{}\n'.format(pred_text, score)) 228 wer_dist_best += wer_dist_min 229 cer_dist_best += cer_dist_min 230 sample_idx += len(probs_batch) 231 232 if preds_output_file: 233 out_file.close() 234 logging.info(f"Stored the predictions of beam search decoding at '{preds_output_file}'.") 235 236 if lm_path: 237 logging.info( 238 'WER/CER with beam search decoding and N-gram model = {:.2%}/{:.2%}'.format( 239 wer_dist_first / words_count, cer_dist_first / chars_count 240 ) 241 ) 242 else: 243 logging.info( 244 'WER/CER with beam search decoding = {:.2%}/{:.2%}'.format( 245 wer_dist_first / words_count, cer_dist_first / chars_count 246 ) 247 ) 248 logging.info( 249 'Oracle WER/CER in candidates with perfect LM= {:.2%}/{:.2%}'.format( 250 wer_dist_best / words_count, cer_dist_best / chars_count 251 ) 252 ) 253 logging.info(f"=================================================================================") 254 255 return wer_dist_first / words_count, cer_dist_first / chars_count 256 257 258 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 259 def main(cfg: EvalBeamSearchNGramConfig): 260 logging.warning("This file will be renamed to eval_beamsearch_ngram_ctc.py in the future NeMo (1.21) release.") 261 if is_dataclass(cfg): 262 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 263 264 valid_decoding_modes = ["greedy", "beamsearch", "beamsearch_ngram"] 265 if cfg.decoding_mode not in valid_decoding_modes: 266 raise ValueError( 267 f"Given decoding_mode={cfg.decoding_mode} is invalid. Available options are :\n" f"{valid_decoding_modes}" 268 ) 269 270 if cfg.nemo_model_file.endswith('.nemo'): 271 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 272 else: 273 logging.warning( 274 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 275 ) 276 asr_model = nemo_asr.models.ASRModel.from_pretrained( 277 cfg.nemo_model_file, map_location=torch.device(cfg.device) 278 ) 279 280 target_transcripts = [] 281 manifest_dir = Path(cfg.input_manifest).parent 282 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 283 audio_file_paths = [] 284 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 285 data = json.loads(line) 286 audio_file = Path(data['audio_filepath']) 287 if not audio_file.is_file() and not audio_file.is_absolute(): 288 audio_file = manifest_dir / audio_file 289 target_transcripts.append(data['text']) 290 audio_file_paths.append(str(audio_file.absolute())) 291 292 punctuation_capitalization = PunctuationCapitalization(cfg.text_processing.punctuation_marks) 293 if cfg.text_processing.do_lowercase: 294 target_transcripts = punctuation_capitalization.do_lowercase(target_transcripts) 295 if cfg.text_processing.rm_punctuation: 296 target_transcripts = punctuation_capitalization.rm_punctuation(target_transcripts) 297 if cfg.text_processing.separate_punctuation: 298 target_transcripts = punctuation_capitalization.separate_punctuation(target_transcripts) 299 300 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 301 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 302 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 303 with open(cfg.probs_cache_file, 'rb') as probs_file: 304 all_probs = pickle.load(probs_file) 305 306 if len(all_probs) != len(audio_file_paths): 307 raise ValueError( 308 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 309 f"match the manifest file. You may need to delete the probabilities cached file." 310 ) 311 else: 312 313 @contextlib.contextmanager 314 def default_autocast(): 315 yield 316 317 if cfg.use_amp: 318 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 319 logging.info("AMP is enabled!\n") 320 autocast = torch.cuda.amp.autocast 321 322 else: 323 autocast = default_autocast 324 else: 325 326 autocast = default_autocast 327 328 with autocast(): 329 with torch.no_grad(): 330 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 331 asr_model.cur_decoder = 'ctc' 332 all_logits = asr_model.transcribe(audio_file_paths, batch_size=cfg.acoustic_batch_size, logprobs=True) 333 334 all_probs = all_logits 335 if cfg.probs_cache_file: 336 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 337 with open(cfg.probs_cache_file, 'wb') as f_dump: 338 pickle.dump(all_probs, f_dump) 339 340 wer_dist_greedy = 0 341 cer_dist_greedy = 0 342 words_count = 0 343 chars_count = 0 344 for batch_idx, probs in enumerate(all_probs): 345 preds = np.argmax(probs, axis=1) 346 preds_tensor = torch.tensor(preds, device='cpu').unsqueeze(0) 347 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 348 pred_text = asr_model.ctc_decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 349 else: 350 pred_text = asr_model._wer.decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 351 352 if cfg.text_processing.do_lowercase: 353 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 354 if cfg.text_processing.rm_punctuation: 355 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 356 if cfg.text_processing.separate_punctuation: 357 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 358 359 pred_split_w = pred_text.split() 360 target_split_w = target_transcripts[batch_idx].split() 361 pred_split_c = list(pred_text) 362 target_split_c = list(target_transcripts[batch_idx]) 363 364 wer_dist = editdistance.eval(target_split_w, pred_split_w) 365 cer_dist = editdistance.eval(target_split_c, pred_split_c) 366 367 wer_dist_greedy += wer_dist 368 cer_dist_greedy += cer_dist 369 words_count += len(target_split_w) 370 chars_count += len(target_split_c) 371 372 logging.info('Greedy WER/CER = {:.2%}/{:.2%}'.format(wer_dist_greedy / words_count, cer_dist_greedy / chars_count)) 373 374 asr_model = asr_model.to('cpu') 375 376 if cfg.decoding_mode == "beamsearch_ngram": 377 if not os.path.exists(cfg.kenlm_model_file): 378 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 379 lm_path = cfg.kenlm_model_file 380 else: 381 lm_path = None 382 383 # 'greedy' decoding_mode would skip the beam search decoding 384 if cfg.decoding_mode in ["beamsearch_ngram", "beamsearch"]: 385 if cfg.beam_width is None or cfg.beam_alpha is None or cfg.beam_beta is None: 386 raise ValueError("beam_width, beam_alpha and beam_beta are needed to perform beam search decoding.") 387 params = {'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'beam_beta': cfg.beam_beta} 388 hp_grid = ParameterGrid(params) 389 hp_grid = list(hp_grid) 390 391 best_wer_beam_size, best_cer_beam_size = None, None 392 best_wer_alpha, best_cer_alpha = None, None 393 best_wer_beta, best_cer_beta = None, None 394 best_wer, best_cer = 1e6, 1e6 395 396 logging.info(f"==============================Starting the beam search decoding===============================") 397 logging.info(f"Grid search size: {len(hp_grid)}") 398 logging.info(f"It may take some time...") 399 logging.info(f"==============================================================================================") 400 401 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 402 os.mkdir(cfg.preds_output_folder) 403 for hp in hp_grid: 404 if cfg.preds_output_folder: 405 preds_output_file = os.path.join( 406 cfg.preds_output_folder, 407 f"preds_out_width{hp['beam_width']}_alpha{hp['beam_alpha']}_beta{hp['beam_beta']}.tsv", 408 ) 409 else: 410 preds_output_file = None 411 412 candidate_wer, candidate_cer = beam_search_eval( 413 asr_model, 414 cfg, 415 all_probs=all_probs, 416 target_transcripts=target_transcripts, 417 preds_output_file=preds_output_file, 418 lm_path=lm_path, 419 beam_width=hp["beam_width"], 420 beam_alpha=hp["beam_alpha"], 421 beam_beta=hp["beam_beta"], 422 beam_batch_size=cfg.beam_batch_size, 423 progress_bar=True, 424 punctuation_capitalization=punctuation_capitalization, 425 ) 426 427 if candidate_cer < best_cer: 428 best_cer_beam_size = hp["beam_width"] 429 best_cer_alpha = hp["beam_alpha"] 430 best_cer_beta = hp["beam_beta"] 431 best_cer = candidate_cer 432 433 if candidate_wer < best_wer: 434 best_wer_beam_size = hp["beam_width"] 435 best_wer_alpha = hp["beam_alpha"] 436 best_wer_beta = hp["beam_beta"] 437 best_wer = candidate_wer 438 439 logging.info( 440 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 441 f'Beam alpha = {best_wer_alpha}, Beam beta = {best_wer_beta}' 442 ) 443 444 logging.info( 445 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 446 f'Beam alpha = {best_cer_alpha}, Beam beta = {best_cer_beta}' 447 ) 448 logging.info(f"=================================================================================") 449 450 451 if __name__ == '__main__': 452 main() 453 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR Transducer model. NeMo's beam search decoders are capable of using the 19 # KenLM's N-gram models to find the best candidates. This script supports both character level and BPE level 20 # encodings and models which is detected automatically from the type of the model. 21 # You may train the LM model with 'scripts/ngram_lm/train_kenlm.py'. 22 23 # Config Help 24 25 To discover all arguments of the script, please run : 26 python eval_beamsearch_ngram.py --help 27 python eval_beamsearch_ngram.py --cfg job 28 29 # USAGE 30 31 python eval_beamsearch_ngram_transducer.py nemo_model_file=<path to the .nemo file of the model> \ 32 input_manifest=<path to the evaluation JSON manifest file \ 33 kenlm_model_file=<path to the binary KenLM model> \ 34 beam_width=[<list of the beam widths, separated with commas>] \ 35 beam_alpha=[<list of the beam alphas, separated with commas>] \ 36 preds_output_folder=<optional folder to store the predictions> \ 37 probs_cache_file=null \ 38 decoding_strategy=<greedy_batch or maes decoding> 39 maes_prefix_alpha=[<list of the maes prefix alphas, separated with commas>] \ 40 maes_expansion_gamma=[<list of the maes expansion gammas, separated with commas>] \ 41 hat_subtract_ilm=<in case of HAT model: subtract internal LM or not> \ 42 hat_ilm_weight=[<in case of HAT model: list of the HAT internal LM weights, separated with commas>] \ 43 ... 44 45 46 # Grid Search for Hyper parameters 47 48 For grid search, you can provide a list of arguments as follows - 49 50 beam_width=[4,8,16,....] \ 51 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 import tempfile 64 from dataclasses import dataclass, field, is_dataclass 65 from pathlib import Path 66 from typing import List, Optional 67 68 import editdistance 69 import numpy as np 70 import torch 71 from omegaconf import MISSING, OmegaConf 72 from sklearn.model_selection import ParameterGrid 73 from tqdm.auto import tqdm 74 75 import nemo.collections.asr as nemo_asr 76 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding 77 from nemo.core.config import hydra_runner 78 from nemo.utils import logging 79 80 # fmt: off 81 82 83 @dataclass 84 class EvalBeamSearchNGramConfig: 85 """ 86 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 87 """ 88 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 89 nemo_model_file: str = MISSING 90 91 # File paths 92 input_manifest: str = MISSING # The manifest file of the evaluation set 93 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 94 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 95 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 96 97 # Parameters for inference 98 acoustic_batch_size: int = 128 # The batch size to calculate log probabilities 99 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 100 device: str = "cuda" # The device to load the model onto to calculate log probabilities 101 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 102 num_workers: int = 1 # Number of workers for DataLoader 103 104 # The decoding scheme to be used for evaluation 105 decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] 106 107 # Beam Search hyperparameters 108 beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding 109 beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding 110 111 maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding 112 maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding 113 114 # HAT related parameters (only for internal lm subtraction) 115 hat_subtract_ilm: bool = False 116 hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) 117 118 decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) 119 120 121 # fmt: on 122 123 124 def decoding_step( 125 model: nemo_asr.models.ASRModel, 126 cfg: EvalBeamSearchNGramConfig, 127 all_probs: List[torch.Tensor], 128 target_transcripts: List[str], 129 preds_output_file: str = None, 130 beam_batch_size: int = 128, 131 progress_bar: bool = True, 132 ): 133 level = logging.getEffectiveLevel() 134 logging.setLevel(logging.CRITICAL) 135 # Reset config 136 model.change_decoding_strategy(None) 137 138 cfg.decoding.hat_ilm_weight = cfg.decoding.hat_ilm_weight * cfg.hat_subtract_ilm 139 # Override the beam search config with current search candidate configuration 140 cfg.decoding.return_best_hypothesis = False 141 cfg.decoding.ngram_lm_model = cfg.kenlm_model_file 142 cfg.decoding.hat_subtract_ilm = cfg.hat_subtract_ilm 143 144 # Update model's decoding strategy config 145 model.cfg.decoding.strategy = cfg.decoding_strategy 146 model.cfg.decoding.beam = cfg.decoding 147 148 # Update model's decoding strategy 149 model.change_decoding_strategy(model.cfg.decoding) 150 logging.setLevel(level) 151 152 wer_dist_first = cer_dist_first = 0 153 wer_dist_best = cer_dist_best = 0 154 words_count = 0 155 chars_count = 0 156 sample_idx = 0 157 if preds_output_file: 158 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 159 160 if progress_bar: 161 if cfg.decoding_strategy == "greedy_batch": 162 description = "Greedy_batch decoding.." 163 else: 164 description = f"{cfg.decoding_strategy} decoding with bw={cfg.decoding.beam_size}, ba={cfg.decoding.ngram_lm_alpha}, ma={cfg.decoding.maes_prefix_alpha}, mg={cfg.decoding.maes_expansion_gamma}, hat_ilmw={cfg.decoding.hat_ilm_weight}" 165 it = tqdm(range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=description, ncols=120) 166 else: 167 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 168 for batch_idx in it: 169 # disabling type checking 170 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 171 probs_lens = torch.tensor([prob.shape[-1] for prob in probs_batch]) 172 with torch.no_grad(): 173 packed_batch = torch.zeros(len(probs_batch), probs_batch[0].shape[0], max(probs_lens), device='cpu') 174 175 for prob_index in range(len(probs_batch)): 176 packed_batch[prob_index, :, : probs_lens[prob_index]] = torch.tensor( 177 probs_batch[prob_index].unsqueeze(0), device=packed_batch.device, dtype=packed_batch.dtype 178 ) 179 best_hyp_batch, beams_batch = model.decoding.rnnt_decoder_predictions_tensor( 180 packed_batch, probs_lens, return_hypotheses=True, 181 ) 182 if cfg.decoding_strategy == "greedy_batch": 183 beams_batch = [[x] for x in best_hyp_batch] 184 185 for beams_idx, beams in enumerate(beams_batch): 186 target = target_transcripts[sample_idx + beams_idx] 187 target_split_w = target.split() 188 target_split_c = list(target) 189 words_count += len(target_split_w) 190 chars_count += len(target_split_c) 191 wer_dist_min = cer_dist_min = 10000 192 for candidate_idx, candidate in enumerate(beams): # type: (int, rnnt_beam_decoding.rnnt_utils.Hypothesis) 193 pred_text = candidate.text 194 pred_split_w = pred_text.split() 195 wer_dist = editdistance.eval(target_split_w, pred_split_w) 196 pred_split_c = list(pred_text) 197 cer_dist = editdistance.eval(target_split_c, pred_split_c) 198 199 wer_dist_min = min(wer_dist_min, wer_dist) 200 cer_dist_min = min(cer_dist_min, cer_dist) 201 202 if candidate_idx == 0: 203 # first candidate 204 wer_dist_first += wer_dist 205 cer_dist_first += cer_dist 206 207 score = candidate.score 208 if preds_output_file: 209 out_file.write('{}\t{}\n'.format(pred_text, score)) 210 wer_dist_best += wer_dist_min 211 cer_dist_best += cer_dist_min 212 sample_idx += len(probs_batch) 213 214 if cfg.decoding_strategy == "greedy_batch": 215 return wer_dist_first / words_count, cer_dist_first / chars_count 216 217 if preds_output_file: 218 out_file.close() 219 logging.info(f"Stored the predictions of {cfg.decoding_strategy} decoding at '{preds_output_file}'.") 220 221 if cfg.decoding.ngram_lm_model: 222 logging.info( 223 f"WER/CER with {cfg.decoding_strategy} decoding and N-gram model = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 224 ) 225 else: 226 logging.info( 227 f"WER/CER with {cfg.decoding_strategy} decoding = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 228 ) 229 logging.info( 230 f"Oracle WER/CER in candidates with perfect LM= {wer_dist_best / words_count:.2%}/{cer_dist_best / chars_count:.2%}" 231 ) 232 logging.info(f"=================================================================================") 233 234 return wer_dist_first / words_count, cer_dist_first / chars_count 235 236 237 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 238 def main(cfg: EvalBeamSearchNGramConfig): 239 if is_dataclass(cfg): 240 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 241 242 valid_decoding_strategis = ["greedy_batch", "beam", "tsd", "alsd", "maes"] 243 if cfg.decoding_strategy not in valid_decoding_strategis: 244 raise ValueError( 245 f"Given decoding_strategy={cfg.decoding_strategy} is invalid. Available options are :\n" 246 f"{valid_decoding_strategis}" 247 ) 248 249 if cfg.nemo_model_file.endswith('.nemo'): 250 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 251 else: 252 logging.warning( 253 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 254 ) 255 asr_model = nemo_asr.models.ASRModel.from_pretrained( 256 cfg.nemo_model_file, map_location=torch.device(cfg.device) 257 ) 258 259 if cfg.kenlm_model_file: 260 if not os.path.exists(cfg.kenlm_model_file): 261 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 262 if cfg.decoding_strategy != "maes": 263 raise ValueError(f"Decoding with kenlm model is supported only for maes decoding algorithm.") 264 lm_path = cfg.kenlm_model_file 265 else: 266 lm_path = None 267 cfg.beam_alpha = [0.0] 268 if cfg.hat_subtract_ilm: 269 assert lm_path, "kenlm must be set for hat internal lm subtraction" 270 271 if cfg.decoding_strategy != "maes": 272 cfg.maes_prefix_alpha, cfg.maes_expansion_gamma, cfg.hat_ilm_weight = [0], [0], [0] 273 274 target_transcripts = [] 275 manifest_dir = Path(cfg.input_manifest).parent 276 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 277 audio_file_paths = [] 278 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 279 data = json.loads(line) 280 audio_file = Path(data['audio_filepath']) 281 if not audio_file.is_file() and not audio_file.is_absolute(): 282 audio_file = manifest_dir / audio_file 283 target_transcripts.append(data['text']) 284 audio_file_paths.append(str(audio_file.absolute())) 285 286 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 287 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 288 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 289 with open(cfg.probs_cache_file, 'rb') as probs_file: 290 all_probs = pickle.load(probs_file) 291 292 if len(all_probs) != len(audio_file_paths): 293 raise ValueError( 294 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 295 f"match the manifest file. You may need to delete the probabilities cached file." 296 ) 297 else: 298 299 @contextlib.contextmanager 300 def default_autocast(): 301 yield 302 303 if cfg.use_amp: 304 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 305 logging.info("AMP is enabled!\n") 306 autocast = torch.cuda.amp.autocast 307 308 else: 309 autocast = default_autocast 310 else: 311 312 autocast = default_autocast 313 314 # manual calculation of encoder_embeddings 315 with autocast(): 316 with torch.no_grad(): 317 asr_model.eval() 318 asr_model.encoder.freeze() 319 device = next(asr_model.parameters()).device 320 all_probs = [] 321 with tempfile.TemporaryDirectory() as tmpdir: 322 with open(os.path.join(tmpdir, 'manifest.json'), 'w', encoding='utf-8') as fp: 323 for audio_file in audio_file_paths: 324 entry = {'audio_filepath': audio_file, 'duration': 100000, 'text': ''} 325 fp.write(json.dumps(entry) + '\n') 326 config = { 327 'paths2audio_files': audio_file_paths, 328 'batch_size': cfg.acoustic_batch_size, 329 'temp_dir': tmpdir, 330 'num_workers': cfg.num_workers, 331 'channel_selector': None, 332 'augmentor': None, 333 } 334 temporary_datalayer = asr_model._setup_transcribe_dataloader(config) 335 for test_batch in tqdm(temporary_datalayer, desc="Transcribing", disable=True): 336 encoded, encoded_len = asr_model.forward( 337 input_signal=test_batch[0].to(device), input_signal_length=test_batch[1].to(device) 338 ) 339 # dump encoder embeddings per file 340 for idx in range(encoded.shape[0]): 341 encoded_no_pad = encoded[idx, :, : encoded_len[idx]] 342 all_probs.append(encoded_no_pad) 343 344 if cfg.probs_cache_file: 345 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 346 with open(cfg.probs_cache_file, 'wb') as f_dump: 347 pickle.dump(all_probs, f_dump) 348 349 if cfg.decoding_strategy == "greedy_batch": 350 asr_model = asr_model.to('cpu') 351 candidate_wer, candidate_cer = decoding_step( 352 asr_model, 353 cfg, 354 all_probs=all_probs, 355 target_transcripts=target_transcripts, 356 beam_batch_size=cfg.beam_batch_size, 357 progress_bar=True, 358 ) 359 logging.info(f"Greedy batch WER/CER = {candidate_wer:.2%}/{candidate_cer:.2%}") 360 361 asr_model = asr_model.to('cpu') 362 363 # 'greedy_batch' decoding_strategy would skip the beam search decoding 364 if cfg.decoding_strategy in ["beam", "tsd", "alsd", "maes"]: 365 if cfg.beam_width is None or cfg.beam_alpha is None: 366 raise ValueError("beam_width and beam_alpha are needed to perform beam search decoding.") 367 params = { 368 'beam_width': cfg.beam_width, 369 'beam_alpha': cfg.beam_alpha, 370 'maes_prefix_alpha': cfg.maes_prefix_alpha, 371 'maes_expansion_gamma': cfg.maes_expansion_gamma, 372 'hat_ilm_weight': cfg.hat_ilm_weight, 373 } 374 hp_grid = ParameterGrid(params) 375 hp_grid = list(hp_grid) 376 377 best_wer_beam_size, best_cer_beam_size = None, None 378 best_wer_alpha, best_cer_alpha = None, None 379 best_wer, best_cer = 1e6, 1e6 380 381 logging.info( 382 f"==============================Starting the {cfg.decoding_strategy} decoding===============================" 383 ) 384 logging.info(f"Grid search size: {len(hp_grid)}") 385 logging.info(f"It may take some time...") 386 logging.info(f"==============================================================================================") 387 388 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 389 os.mkdir(cfg.preds_output_folder) 390 for hp in hp_grid: 391 if cfg.preds_output_folder: 392 results_file = f"preds_out_{cfg.decoding_strategy}_bw{hp['beam_width']}" 393 if cfg.decoding_strategy == "maes": 394 results_file = f"{results_file}_ma{hp['maes_prefix_alpha']}_mg{hp['maes_expansion_gamma']}" 395 if cfg.kenlm_model_file: 396 results_file = f"{results_file}_ba{hp['beam_alpha']}" 397 if cfg.hat_subtract_ilm: 398 results_file = f"{results_file}_hat_ilmw{hp['hat_ilm_weight']}" 399 preds_output_file = os.path.join(cfg.preds_output_folder, f"{results_file}.tsv") 400 else: 401 preds_output_file = None 402 403 cfg.decoding.beam_size = hp["beam_width"] 404 cfg.decoding.ngram_lm_alpha = hp["beam_alpha"] 405 cfg.decoding.maes_prefix_alpha = hp["maes_prefix_alpha"] 406 cfg.decoding.maes_expansion_gamma = hp["maes_expansion_gamma"] 407 cfg.decoding.hat_ilm_weight = hp["hat_ilm_weight"] 408 409 candidate_wer, candidate_cer = decoding_step( 410 asr_model, 411 cfg, 412 all_probs=all_probs, 413 target_transcripts=target_transcripts, 414 preds_output_file=preds_output_file, 415 beam_batch_size=cfg.beam_batch_size, 416 progress_bar=True, 417 ) 418 419 if candidate_cer < best_cer: 420 best_cer_beam_size = hp["beam_width"] 421 best_cer_alpha = hp["beam_alpha"] 422 best_cer_ma = hp["maes_prefix_alpha"] 423 best_cer_mg = hp["maes_expansion_gamma"] 424 best_cer_hat_ilm_weight = hp["hat_ilm_weight"] 425 best_cer = candidate_cer 426 427 if candidate_wer < best_wer: 428 best_wer_beam_size = hp["beam_width"] 429 best_wer_alpha = hp["beam_alpha"] 430 best_wer_ma = hp["maes_prefix_alpha"] 431 best_wer_ga = hp["maes_expansion_gamma"] 432 best_wer_hat_ilm_weight = hp["hat_ilm_weight"] 433 best_wer = candidate_wer 434 435 wer_hat_parameter = "" 436 if cfg.hat_subtract_ilm: 437 wer_hat_parameter = f"HAT ilm weight = {best_wer_hat_ilm_weight}, " 438 logging.info( 439 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 440 f'Beam alpha = {best_wer_alpha}, {wer_hat_parameter}' 441 f'maes_prefix_alpha = {best_wer_ma}, maes_expansion_gamma = {best_wer_ga} ' 442 ) 443 444 cer_hat_parameter = "" 445 if cfg.hat_subtract_ilm: 446 cer_hat_parameter = f"HAT ilm weight = {best_cer_hat_ilm_weight}" 447 logging.info( 448 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 449 f'Beam alpha = {best_cer_alpha}, {cer_hat_parameter} ' 450 f'maes_prefix_alpha = {best_cer_ma}, maes_expansion_gamma = {best_cer_mg}' 451 ) 452 logging.info(f"=================================================================================") 453 454 455 if __name__ == '__main__': 456 main() 457 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of scripts/confidence_ensembles/build_ensemble.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 This script provides a functionality to create confidence-based ensembles 17 from a collection of pretrained models. 18 19 For more details see the paper https://arxiv.org/abs/2306.15824 20 or tutorial in tutorials/asr/Confidence_Ensembles.ipynb 21 22 You would typically use this script by providing a yaml config file or overriding 23 default options from command line. 24 25 Usage examples: 26 27 1. Building an ensemble of two monolingual models with default settings (no confidence tuning). 28 29 python build_ensemble.py --config-path=. --config-name=ensemble_config.yaml 30 ensemble.0.model=stt_it_conformer_ctc_large 31 ensemble.0.training_manifest=<path to the Italian data of 100+ utterances (no transcription required)> 32 ensemble.1.model=stt_es_conformer_ctc_large 33 ensemble.1.training_manifest=<path to the Spanish data of 100+ utterances (no transcription required)> 34 output_path=<path to the desired location of the .nemo checkpoint> 35 36 You can have more than 2 models and can control transcription settings (e.g., batch size) 37 with ``transcription.<any argument of examples/asr/transcribe_speech.py>`` parameters. 38 39 2. If you want to get improved results, you can enable tuning of the confidence and logistic regression (LR) parameters. 40 E.g. 41 42 python build_ensemble.py 43 <all arguments like in the previous example> 44 ensemble.0.dev_manifest=<path to the dev data that's required for tuning> 45 ... 46 # IMPORTANT: see the note below if you use > 2 models! 47 ensemble.N.dev_manifest=<path to the dev data that's required for tuning> 48 tune_confidence=True # to allow confidence tuning. LR is tuned by default 49 50 As with any tuning, it is recommended to have reasonably large validation set for each model, 51 otherwise you might overfit to the validation data. 52 53 Note that if you add additional models (> 2) you will need to modify ensemble_config.yaml 54 or create a new one with added models in there. While it's theoretically possible to 55 fully override such parameters from commandline, hydra is very unfriendly for such 56 use-cases, so it's strongly recommended to be creating new configs. 57 58 3. If you want to precisely control tuning grid search, you can do that with 59 60 python build_ensemble.py 61 <all arguments as in the previous examples> 62 tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set 63 tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set 64 65 You can check the dataclasses in this file for the full list of supported 66 arguments and their default values. 67 """ 68 69 import atexit 70 71 # using default logging to be able to silence unnecessary messages from nemo 72 import logging 73 import os 74 import random 75 import sys 76 import tempfile 77 from copy import deepcopy 78 from dataclasses import dataclass 79 from pathlib import Path 80 from typing import Dict, List, Optional, Tuple 81 82 import joblib 83 import numpy as np 84 import pytorch_lightning as pl 85 from omegaconf import MISSING, DictConfig, OmegaConf 86 from sklearn.linear_model import LogisticRegression 87 from sklearn.metrics import confusion_matrix 88 from sklearn.pipeline import Pipeline, make_pipeline 89 from sklearn.preprocessing import StandardScaler 90 from tqdm import tqdm 91 92 from nemo.collections.asr.models.confidence_ensemble import ( 93 ConfidenceEnsembleModel, 94 ConfidenceSpec, 95 compute_confidence, 96 get_filtered_logprobs, 97 ) 98 from nemo.collections.asr.parts.utils.asr_confidence_utils import ( 99 ConfidenceConfig, 100 ConfidenceMethodConfig, 101 get_confidence_aggregation_bank, 102 get_confidence_measure_bank, 103 ) 104 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 105 from nemo.core.config import hydra_runner 106 107 LOG = logging.getLogger(__file__) 108 109 # adding Python path. If not found, asking user to get the file 110 try: 111 sys.path.append(str(Path(__file__).parents[2] / "examples" / "asr")) 112 import transcribe_speech 113 except ImportError: 114 # if users run script normally from nemo repo, this shouldn't be triggered as 115 # we modify the path above. But if they downloaded the build_ensemble.py as 116 # an isolated script, we'd ask them to also download corresponding version 117 # of the transcribe_speech.py 118 print( 119 "Current script depends on 'examples/asr/transcribe_speech.py', but can't find it. " 120 "If it's not present, download it from the NeMo github manually and put inside this folder." 121 ) 122 123 124 @dataclass 125 class EnsembleConfig: 126 # .nemo path or pretrained name 127 model: str = MISSING 128 # path to the training data manifest (non-tarred) 129 training_manifest: str = MISSING 130 # specify to limit the number of training samples 131 # 100 is most likely enough, but setting higher default just in case 132 max_training_samples: int = 1000 133 # specify to provide dev data manifest for HP tuning 134 dev_manifest: Optional[str] = None 135 136 137 @dataclass 138 class TuneConfidenceConfig: 139 # important parameter, so should always be tuned 140 exclude_blank: Tuple[bool] = (True, False) 141 # prod is pretty much always worse, so not including by default 142 aggregation: Tuple[str] = ("mean", "min", "max") 143 # not including max prob, as there is always an entropy-based metric 144 # that's better but otherwise including everything 145 confidence_type: Tuple[str] = ( 146 "entropy_renyi_exp", 147 "entropy_renyi_lin", 148 "entropy_tsallis_exp", 149 "entropy_tsallis_lin", 150 "entropy_gibbs_lin", 151 "entropy_gibbs_exp", 152 ) 153 154 # TODO: currently it's not possible to efficiently tune temperature, as we always 155 # apply log-softmax in the decoder, so to try different values it will be required 156 # to rerun the decoding, which is very slow. To support this for one-off experiments 157 # it's possible to modify the code of CTC decoder / Transducer joint to 158 # remove log-softmax and then apply it directly in this script with the temperature 159 # 160 # Alternatively, one can run this script multiple times with different values of 161 # temperature and pick the best performing ensemble. Note that this will increase 162 # tuning time by the number of temperature values tried. On the other hand, 163 # the above approach is a lot more efficient and will only slightly increase 164 # the total tuning runtime. 165 166 # very important to tune for max prob, but for entropy metrics 1.0 is almost always best 167 # temperature: Tuple[float] = (1.0,) 168 169 # not that important, but can sometimes make a small difference 170 alpha: Tuple[float] = (0.25, 0.33, 0.5, 1.0) 171 172 def get_grid_size(self) -> int: 173 """Returns the total number of points in the search space.""" 174 if "max_prob" in self.confidence_type: 175 return ( 176 len(self.exclude_blank) 177 * len(self.aggregation) 178 * ((len(self.confidence_type) - 1) * len(self.alpha) + 1) 179 ) 180 return len(self.exclude_blank) * len(self.aggregation) * len(self.confidence_type) * len(self.alpha) 181 182 183 @dataclass 184 class TuneLogisticRegressionConfig: 185 # will have log-uniform grid over this range with that many points 186 # note that a value of 10000.0 (not regularization) is always added 187 C_num_points: int = 10 188 C_min: float = 0.0001 189 C_max: float = 10.0 190 191 # not too important 192 multi_class: Tuple[str] = ("ovr", "multinomial") 193 194 # should try to include weights directly if the data is too imbalanced 195 class_weight: Tuple = (None, "balanced") 196 197 # increase if getting many warnings that algorithm didn't converge 198 max_iter: int = 1000 199 200 201 @dataclass 202 class BuildEnsembleConfig: 203 # where to save the resulting ensemble model 204 output_path: str = MISSING 205 206 # each model specification 207 ensemble: List[EnsembleConfig] = MISSING 208 209 random_seed: int = 0 # for reproducibility 210 211 # default confidence, can override 212 confidence: ConfidenceConfig = ConfidenceConfig( 213 # we keep frame confidences and apply aggregation manually to get full-utterance confidence 214 preserve_frame_confidence=True, 215 exclude_blank=True, 216 aggregation="mean", 217 method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False 229 # used to specify what to tune over. By default runs tuning over some 230 # reasonalbe grid, so that it does not take forever. 231 # Can be changed as needed 232 tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() 233 234 # very fast to tune and can be important in case of imbalanced datasets 235 # will automatically set to False if dev data is not available 236 tune_logistic_regression: bool = True 237 tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() 238 239 def __post_init__(self): 240 """Checking that if any dev data is provided, all are provided. 241 242 Will also auto-set tune_logistic_regression to False if no dev data 243 is available. 244 245 If tune_confidence is set to True (user choice) and no dev data is 246 provided, will raise an error. 247 """ 248 num_dev_data = 0 249 for ensemble_cfg in self.ensemble: 250 num_dev_data += ensemble_cfg.dev_manifest is not None 251 if num_dev_data == 0: 252 if self.tune_confidence: 253 raise ValueError("tune_confidence is set to True, but no dev data is provided") 254 LOG.info("Setting tune_logistic_regression = False since no dev data is provided") 255 self.tune_logistic_regression = False 256 return 257 258 if num_dev_data < len(self.ensemble): 259 raise ValueError( 260 "Some ensemble configs specify dev data, but some don't. Either all have to specify it or none!" 261 ) 262 263 264 def calculate_score(features: np.ndarray, labels: np.ndarray, pipe: Pipeline) -> Tuple[float, np.ndarray]: 265 """Score is always calculated as mean of the per-class scores. 266 267 This is done to account for possible class imbalances. 268 269 Args: 270 features: numpy array of features of shape [N x D], where N is the 271 number of objects (typically a total number of utterances in 272 all datasets) and D is the total number of confidence scores 273 used to train the model (typically = number of models). 274 labels: numpy array of shape [N] contatining ground-truth model indices. 275 pipe: classification pipeline (currently, standardization + logistic 276 regression). 277 278 Returns: 279 tuple: score value in [0, 1] and full classification confusion matrix. 280 """ 281 predictions = pipe.predict(features) 282 conf_m = confusion_matrix(labels, predictions) 283 score = np.diag(conf_m).sum() / conf_m.sum() 284 return score, conf_m 285 286 287 def train_model_selection( 288 training_features: np.ndarray, 289 training_labels: np.ndarray, 290 dev_features: Optional[np.ndarray] = None, 291 dev_labels: Optional[np.ndarray] = None, 292 tune_lr: bool = False, 293 tune_lr_cfg: Optional[TuneLogisticRegressionConfig] = None, 294 verbose: bool = False, 295 ) -> Tuple[Pipeline, float]: 296 """Trains model selection block with an (optional) tuning of the parameters. 297 298 Returns a pipeline consisting of feature standardization and logistic 299 regression. If tune_lr is set to True, dev features/labels will be used 300 to tune the hyperparameters of the logistic regression with the grid 301 search that's defined via ``tune_lr_cfg``. 302 303 If no tuning is requested, uses the following parameters:: 304 305 best_pipe = make_pipeline( 306 StandardScaler(), 307 LogisticRegression( 308 multi_class="multinomial", 309 C=10000.0, 310 max_iter=1000, 311 class_weight="balanced", 312 ), 313 ) 314 315 Args: 316 training_features: numpy array of features of shape [N x D], where N is 317 the number of objects (typically a total number of utterances in 318 all training datasets) and D is the total number of confidence 319 scores used to train the model (typically = number of models). 320 training_labels: numpy array of shape [N] contatining ground-truth 321 model indices. 322 dev_features: same as training, but for the validation subset. 323 dev_labels: same as training, but for the validation subset. 324 tune_lr: controls whether tuning of LR hyperparameters is performed. 325 If set to True, it's required to also provide dev features/labels. 326 tune_lr_cfg: specifies what values of LR hyperparameters to try. 327 verbose: if True, will output final training/dev scores. 328 329 Returns: 330 tuple: trained model selection pipeline, best score (or -1 if no tuning 331 was done). 332 """ 333 if not tune_lr: 334 # default parameters: C=10000.0 disables regularization 335 best_pipe = make_pipeline( 336 StandardScaler(), 337 LogisticRegression(multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced"), 338 ) 339 max_score = -1 340 else: 341 C_pms = np.append( 342 np.exp(np.linspace(np.log(tune_lr_cfg.C_min), np.log(tune_lr_cfg.C_max), tune_lr_cfg.C_num_points)), 343 10000.0, 344 ) 345 max_score = 0 346 best_pipe = None 347 for class_weight in tune_lr_cfg.class_weight: 348 for multi_class in tune_lr_cfg.multi_class: 349 for C in C_pms: 350 pipe = make_pipeline( 351 StandardScaler(), 352 LogisticRegression( 353 multi_class=multi_class, C=C, max_iter=tune_lr_cfg.max_iter, class_weight=class_weight 354 ), 355 ) 356 pipe.fit(training_features, training_labels) 357 score, confusion = calculate_score(dev_features, dev_labels, pipe) 358 if score > max_score: 359 max_score = score 360 best_pipe = pipe 361 362 best_pipe.fit(training_features, training_labels) 363 if verbose: 364 accuracy, confusion = calculate_score(training_features, training_labels, best_pipe) 365 LOG.info("Training fit accuracy: %.4f", accuracy * 100.0) 366 LOG.info("Training confusion matrix:\n%s", str(confusion)) 367 if dev_features is not None and verbose: 368 accuracy, confusion = calculate_score(dev_features, dev_labels, best_pipe) 369 LOG.info("Dev fit accuracy: %.4f", accuracy * 100.0) 370 LOG.info("Dev confusion matrix:\n%s", str(confusion)) 371 372 return best_pipe, max_score 373 374 375 def subsample_manifest(manifest_file: str, max_samples: int) -> str: 376 """Will save a subsampled version of the manifest to the same folder. 377 378 Have to save to the same folder to support relative paths. 379 380 Args: 381 manifest_file: path to the manifest file that needs subsampling. 382 max_samples: how many samples to retain. Will randomly select that 383 many lines from the manifest. 384 385 Returns: 386 str: the path to the subsampled manifest file. 387 """ 388 with open(manifest_file, "rt", encoding="utf-8") as fin: 389 lines = fin.readlines() 390 if max_samples < len(lines): 391 lines = random.sample(lines, max_samples) 392 output_file = manifest_file + "-subsampled" 393 with open(output_file, "wt", encoding="utf-8") as fout: 394 fout.write("".join(lines)) 395 return output_file 396 397 398 def cleanup_subsampled_manifests(subsampled_manifests: List[str]): 399 """Removes all generated subsamples manifests.""" 400 for manifest in subsampled_manifests: 401 os.remove(manifest) 402 403 404 def compute_all_confidences( 405 hypothesis: Hypothesis, tune_confidence_cfg: TuneConfidenceConfig 406 ) -> Dict[ConfidenceSpec, float]: 407 """Computes a set of confidence scores from a given hypothesis. 408 409 Works with the output of both CTC and Transducer decoding. 410 411 Args: 412 hypothesis: generated hypothesis as returned from the transcribe 413 method of the ASR model. 414 tune_confidence_cfg: config specifying what confidence scores to 415 compute. 416 417 Returns: 418 dict: dictionary with confidenct spec -> confidence score mapping. 419 """ 420 conf_values = {} 421 422 for exclude_blank in tune_confidence_cfg.exclude_blank: 423 filtered_logprobs = get_filtered_logprobs(hypothesis, exclude_blank) 424 vocab_size = filtered_logprobs.shape[1] 425 for aggregation in tune_confidence_cfg.aggregation: 426 aggr_func = get_confidence_aggregation_bank()[aggregation] 427 for conf_type in tune_confidence_cfg.confidence_type: 428 conf_func = get_confidence_measure_bank()[conf_type] 429 if conf_type == "max_prob": # skipping alpha in this case 430 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=1.0)).cpu().item() 431 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, 1.0)] = conf_value 432 else: 433 for alpha in tune_confidence_cfg.alpha: 434 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=alpha)).cpu().item() 435 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, alpha)] = conf_value 436 437 return conf_values 438 439 440 def find_best_confidence( 441 train_confidences: List[List[Dict[ConfidenceSpec, float]]], 442 train_labels: List[int], 443 dev_confidences: List[List[Dict[ConfidenceSpec, float]]], 444 dev_labels: List[int], 445 tune_lr: bool, 446 tune_lr_config: TuneConfidenceConfig, 447 ) -> Tuple[ConfidenceConfig, Pipeline]: 448 """Finds the best confidence configuration for model selection. 449 450 Will loop over all values in the confidence dictionary and fit the LR 451 model (optionally tuning its HPs). The best performing confidence (on the 452 dev set) will be used for the final LR model. 453 454 Args: 455 train_confidences: this is an object of type 456 ``List[List[Dict[ConfidenceSpec, float]]]``. The shape of this 457 object is [M, N, S], where 458 M: number of models 459 N: number of utterances in all training sets 460 S: number of confidence scores to try 461 462 This argument will be used to construct np.array objects for each 463 of the confidence scores with the shape [M, N] 464 465 train_labels: ground-truth labels of the correct model for each data 466 points. This is a list of size [N] 467 dev_confidences: same as training, but for the validation subset. 468 dev_labels: same as training, but for the validation subset. 469 tune_lr: controls whether tuning of LR hyperparameters is performed. 470 tune_lr_cfg: specifies what values of LR hyperparameters to try. 471 472 Returns: 473 tuple: best confidence config, best model selection pipeline 474 """ 475 max_score = 0 476 best_pipe = None 477 best_conf_spec = None 478 LOG.info("Evaluation all confidences. Total grid size: %d", len(train_confidences[0][0].keys())) 479 for conf_spec in tqdm(train_confidences[0][0].keys()): 480 cur_train_confidences = [] 481 for model_confs in train_confidences: 482 cur_train_confidences.append([]) 483 for model_conf in model_confs: 484 cur_train_confidences[-1].append(model_conf[conf_spec]) 485 cur_dev_confidences = [] 486 for model_confs in dev_confidences: 487 cur_dev_confidences.append([]) 488 for model_conf in model_confs: 489 cur_dev_confidences[-1].append(model_conf[conf_spec]) 490 # transposing with zip(*list) 491 training_features = np.array(list(zip(*cur_train_confidences))) 492 training_labels = np.array(train_labels) 493 dev_features = np.array(list(zip(*cur_dev_confidences))) 494 dev_labels = np.array(dev_labels) 495 pipe, score = train_model_selection( 496 training_features, training_labels, dev_features, dev_labels, tune_lr, tune_lr_config, 497 ) 498 if max_score < score: 499 max_score = score 500 best_pipe = pipe 501 best_conf_spec = conf_spec 502 LOG.info("Found better parameters: %s. New score: %.4f", str(conf_spec), max_score) 503 504 return best_conf_spec.to_confidence_config(), best_pipe 505 506 507 @hydra_runner(config_name="BuildEnsembleConfig", schema=BuildEnsembleConfig) 508 def main(cfg: BuildEnsembleConfig): 509 # silencing all messages from nemo/ptl to avoid dumping tons of configs to the stdout 510 logging.getLogger('pytorch_lightning').setLevel(logging.CRITICAL) 511 logging.getLogger('nemo_logger').setLevel(logging.CRITICAL) 512 LOG.info(f'Build ensemble config:\n{OmegaConf.to_yaml(cfg)}') 513 514 # to ensure post init is called 515 cfg = BuildEnsembleConfig(**cfg) 516 517 pl.seed_everything(cfg.random_seed) 518 cfg.transcription.random_seed = None # seed is already applied 519 cfg.transcription.return_transcriptions = True 520 cfg.transcription.preserve_alignment = True 521 cfg.transcription.ctc_decoding.temperature = cfg.temperature 522 cfg.transcription.rnnt_decoding.temperature = cfg.temperature 523 # this ensures that generated output is after log-softmax for consistency with CTC 524 525 train_confidences = [] 526 dev_confidences = [] 527 train_labels = [] 528 dev_labels = [] 529 530 # registering clean-up function that will hold on to this list and 531 # should clean up even if there is partial error in some of the transcribe 532 # calls 533 subsampled_manifests = [] 534 atexit.register(cleanup_subsampled_manifests, subsampled_manifests) 535 536 # note that we loop over the same config. 537 # This is intentional, as we need to run all models on all datasets 538 # this loop will do the following things: 539 # 1. Goes through each model X each training dataset 540 # 2. Computes predictions by directly calling transcribe_speech.main 541 # 3. Converts transcription to the confidence score(s) as specified in the config 542 # 4. If dev sets are provided, computes the same for them 543 # 5. Creates a list of ground-truth model indices by mapping each model 544 # to its own training dataset as specified in the config. 545 # 6. After the loop, we either run tuning over all confidence scores or 546 # directly use a single score to fit logistic regression and save the 547 # final ensemble model. 548 for model_idx, model_cfg in enumerate(cfg.ensemble): 549 train_model_confidences = [] 550 dev_model_confidences = [] 551 for data_idx, data_cfg in enumerate(cfg.ensemble): 552 if model_idx == 0: # generating subsampled manifests only one time 553 subsampled_manifests.append( 554 subsample_manifest(data_cfg.training_manifest, data_cfg.max_training_samples) 555 ) 556 subsampled_manifest = subsampled_manifests[data_idx] 557 558 if model_cfg.model.endswith(".nemo"): 559 cfg.transcription.model_path = model_cfg.model 560 else: # assuming pretrained model 561 cfg.transcription.pretrained_name = model_cfg.model 562 563 cfg.transcription.dataset_manifest = subsampled_manifest 564 565 # training 566 with tempfile.NamedTemporaryFile() as output_file: 567 cfg.transcription.output_filename = output_file.name 568 LOG.info("Transcribing training dataset %d with model %d", data_idx, model_idx) 569 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 570 LOG.info("Generating confidence scores") 571 # TODO: parallelize this loop? 572 for transcription in tqdm(transcriptions): 573 if cfg.tune_confidence: 574 train_model_confidences.append( 575 compute_all_confidences(transcription, cfg.tune_confidence_config) 576 ) 577 else: 578 train_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 579 if model_idx == 0: # labels are the same for all models 580 train_labels.append(data_idx) 581 582 # optional dev 583 if data_cfg.dev_manifest is not None: 584 cfg.transcription.dataset_manifest = data_cfg.dev_manifest 585 with tempfile.NamedTemporaryFile() as output_file: 586 cfg.transcription.output_filename = output_file.name 587 LOG.info("Transcribing dev dataset %d with model %d", data_idx, model_idx) 588 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 589 LOG.info("Generating confidence scores") 590 for transcription in tqdm(transcriptions): 591 if cfg.tune_confidence: 592 dev_model_confidences.append( 593 compute_all_confidences(transcription, cfg.tune_confidence_config) 594 ) 595 else: 596 dev_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 597 if model_idx == 0: # labels are the same for all models 598 dev_labels.append(data_idx) 599 600 train_confidences.append(train_model_confidences) 601 if dev_model_confidences: 602 dev_confidences.append(dev_model_confidences) 603 604 if cfg.tune_confidence: 605 best_confidence, model_selection_block = find_best_confidence( 606 train_confidences, 607 train_labels, 608 dev_confidences, 609 dev_labels, 610 cfg.tune_logistic_regression, 611 cfg.tune_logistic_regression_config, 612 ) 613 else: 614 best_confidence = cfg.confidence 615 # transposing with zip(*list) 616 training_features = np.array(list(zip(*train_confidences))) 617 training_labels = np.array(train_labels) 618 if dev_confidences: 619 dev_features = np.array(list(zip(*dev_confidences))) 620 dev_labels = np.array(dev_labels) 621 else: 622 dev_features = None 623 dev_labels = None 624 model_selection_block, _ = train_model_selection( 625 training_features, 626 training_labels, 627 dev_features, 628 dev_labels, 629 cfg.tune_logistic_regression, 630 cfg.tune_logistic_regression_config, 631 verbose=True, 632 ) 633 634 with tempfile.TemporaryDirectory() as tmpdir: 635 model_selection_block_path = os.path.join(tmpdir, 'model_selection_block.pkl') 636 joblib.dump(model_selection_block, model_selection_block_path) 637 638 # creating ensemble checkpoint 639 ensemble_model = ConfidenceEnsembleModel( 640 cfg=DictConfig( 641 { 642 'model_selection_block': model_selection_block_path, 643 'confidence': best_confidence, 644 'temperature': cfg.temperature, 645 'load_models': [model_cfg.model for model_cfg in cfg.ensemble], 646 } 647 ), 648 trainer=None, 649 ) 650 ensemble_model.save_to(cfg.output_path) 651 652 653 if __name__ == '__main__': 654 main() 655 [end of scripts/confidence_ensembles/build_ensemble.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import json 16 import os 17 from dataclasses import dataclass, is_dataclass 18 from pathlib import Path 19 from typing import Optional 20 21 import pytorch_lightning as pl 22 import torch 23 from omegaconf import MISSING, OmegaConf 24 from sklearn.model_selection import ParameterGrid 25 26 from nemo.collections.asr.metrics.rnnt_wer import RNNTDecodingConfig 27 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 28 from nemo.collections.asr.models import ASRModel, EncDecRNNTModel 29 from nemo.collections.asr.parts.utils.asr_confidence_benchmarking_utils import ( 30 apply_confidence_parameters, 31 run_confidence_benchmark, 32 ) 33 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig 34 from nemo.core.config import hydra_runner 35 from nemo.utils import logging, model_utils 36 37 """ 38 Get confidence metrics and curve plots for a given model, dataset, and confidence parameters. 39 40 # Arguments 41 model_path: Path to .nemo ASR checkpoint 42 pretrained_name: Name of pretrained ASR model (from NGC registry) 43 dataset_manifest: Path to dataset JSON manifest file (in NeMo format) 44 output_dir: Output directory to store a report and curve plot directories 45 46 batch_size: batch size during inference 47 num_workers: number of workers during inference 48 49 cuda: Optional int to enable or disable execution of model on certain CUDA device 50 amp: Bool to decide if Automatic Mixed Precision should be used during inference 51 audio_type: Str filetype of the audio. Supported = wav, flac, mp3 52 53 target_level: Word- or token-level confidence. Supported = word, token, auto (for computing both word and token) 54 confidence_cfg: Config with confidence parameters 55 grid_params: Dictionary with lists of parameters to iteratively benchmark on 56 57 # Usage 58 ASR model can be specified by either "model_path" or "pretrained_name". 59 Data for transcription are defined with "dataset_manifest". 60 Results are returned as a benchmark report and curve plots. 61 62 python benchmark_asr_confidence.py \ 63 model_path=null \ 64 pretrained_name=null \ 65 dataset_manifest="" \ 66 output_dir="" \ 67 batch_size=64 \ 68 num_workers=8 \ 69 cuda=0 \ 70 amp=True \ 71 target_level="word" \ 72 confidence_cfg.exclude_blank=False \ 73 'grid_params="{\"aggregation\": [\"min\", \"prod\"], \"alpha\": [0.33, 0.5]}"' 74 """ 75 76 77 def get_experiment_params(cfg): 78 """Get experiment parameters from a confidence config and generate the experiment name. 79 80 Returns: 81 List of experiment parameters. 82 String with the experiment name. 83 """ 84 blank = "no_blank" if cfg.exclude_blank else "blank" 85 aggregation = cfg.aggregation 86 method_name = cfg.method_cfg.name 87 alpha = cfg.method_cfg.alpha 88 if method_name == "entropy": 89 entropy_type = cfg.method_cfg.entropy_type 90 entropy_norm = cfg.method_cfg.entropy_norm 91 experiment_param_list = [ 92 aggregation, 93 str(cfg.exclude_blank), 94 method_name, 95 entropy_type, 96 entropy_norm, 97 str(alpha), 98 ] 99 experiment_str = "-".join([aggregation, blank, method_name, entropy_type, entropy_norm, str(alpha)]) 100 else: 101 experiment_param_list = [aggregation, str(cfg.exclude_blank), method_name, "-", "-", str(alpha)] 102 experiment_str = "-".join([aggregation, blank, method_name, str(alpha)]) 103 return experiment_param_list, experiment_str 104 105 106 @dataclass 107 class ConfidenceBenchmarkingConfig: 108 # Required configs 109 model_path: Optional[str] = None # Path to a .nemo file 110 pretrained_name: Optional[str] = None # Name of a pretrained model 111 dataset_manifest: str = MISSING 112 output_dir: str = MISSING 113 114 # General configs 115 batch_size: int = 32 116 num_workers: int = 4 117 118 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA 119 # device anyway, and do inference on CPU only if CUDA device is not found. 120 # If `cuda` is a negative number, inference will be on CPU only. 121 cuda: Optional[int] = None 122 amp: bool = False 123 audio_type: str = "wav" 124 125 # Confidence configs 126 target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) 127 confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) 128 grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on 129 130 131 @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) 132 def main(cfg: ConfidenceBenchmarkingConfig): 133 torch.set_grad_enabled(False) 134 135 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 136 137 if is_dataclass(cfg): 138 cfg = OmegaConf.structured(cfg) 139 140 if cfg.model_path is None and cfg.pretrained_name is None: 141 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None!") 142 143 # setup GPU 144 if cfg.cuda is None: 145 if torch.cuda.is_available(): 146 device = [0] # use 0th CUDA device 147 accelerator = 'gpu' 148 else: 149 device = 1 150 accelerator = 'cpu' 151 else: 152 device = [cfg.cuda] 153 accelerator = 'gpu' 154 155 map_location = torch.device('cuda:{}'.format(device[0]) if accelerator == 'gpu' else 'cpu') 156 157 # setup model 158 if cfg.model_path is not None: 159 # restore model from .nemo file path 160 model_cfg = ASRModel.restore_from(restore_path=cfg.model_path, return_config=True) 161 classpath = model_cfg.target # original class path 162 imported_class = model_utils.import_class_by_path(classpath) # type: ASRModel 163 logging.info(f"Restoring model : {imported_class.__name__}") 164 asr_model = imported_class.restore_from( 165 restore_path=cfg.model_path, map_location=map_location 166 ) # type: ASRModel 167 else: 168 # restore model by name 169 asr_model = ASRModel.from_pretrained( 170 model_name=cfg.pretrained_name, map_location=map_location 171 ) # type: ASRModel 172 173 trainer = pl.Trainer(devices=device, accelerator=accelerator) 174 asr_model.set_trainer(trainer) 175 asr_model = asr_model.eval() 176 177 # Check if ctc or rnnt model 178 is_rnnt = isinstance(asr_model, EncDecRNNTModel) 179 180 # Check that the model has the `change_decoding_strategy` method 181 if not hasattr(asr_model, 'change_decoding_strategy'): 182 raise RuntimeError("The asr_model you are using must have the `change_decoding_strategy` method.") 183 184 # get filenames and reference texts from manifest 185 filepaths = [] 186 reference_texts = [] 187 if os.stat(cfg.dataset_manifest).st_size == 0: 188 logging.error(f"The input dataset_manifest {cfg.dataset_manifest} is empty. Exiting!") 189 return None 190 manifest_dir = Path(cfg.dataset_manifest).parent 191 with open(cfg.dataset_manifest, 'r') as f: 192 for line in f: 193 item = json.loads(line) 194 audio_file = Path(item['audio_filepath']) 195 if not audio_file.is_file() and not audio_file.is_absolute(): 196 audio_file = manifest_dir / audio_file 197 filepaths.append(str(audio_file.absolute())) 198 reference_texts.append(item['text']) 199 200 # setup AMP (optional) 201 autocast = None 202 if cfg.amp and torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 203 logging.info("AMP enabled!\n") 204 autocast = torch.cuda.amp.autocast 205 206 # do grid-based benchmarking if grid_params is provided, otherwise a regular one 207 work_dir = Path(cfg.output_dir) 208 os.makedirs(work_dir, exist_ok=True) 209 report_legend = ( 210 ",".join( 211 [ 212 "model_type", 213 "aggregation", 214 "blank", 215 "method_name", 216 "entropy_type", 217 "entropy_norm", 218 "alpha", 219 "target_level", 220 "auc_roc", 221 "auc_pr", 222 "auc_nt", 223 "nce", 224 "ece", 225 "auc_yc", 226 "std_yc", 227 "max_yc", 228 ] 229 ) 230 + "\n" 231 ) 232 model_typename = "RNNT" if is_rnnt else "CTC" 233 report_file = work_dir / Path("report.csv") 234 if cfg.grid_params: 235 asr_model.change_decoding_strategy( 236 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 237 if is_rnnt 238 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 239 ) 240 params = json.loads(cfg.grid_params) 241 hp_grid = ParameterGrid(params) 242 hp_grid = list(hp_grid) 243 244 logging.info(f"==============================Running a benchmarking with grid search=========================") 245 logging.info(f"Grid search size: {len(hp_grid)}") 246 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directories near the file") 247 logging.info(f"==============================================================================================") 248 249 with open(report_file, "tw", encoding="utf-8") as f: 250 f.write(report_legend) 251 f.flush() 252 for i, hp in enumerate(hp_grid): 253 logging.info(f"Run # {i + 1}, grid: `{hp}`") 254 asr_model.change_decoding_strategy(apply_confidence_parameters(asr_model.cfg.decoding, hp)) 255 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 256 plot_dir = work_dir / Path(experiment_name) 257 results = run_confidence_benchmark( 258 asr_model, 259 cfg.target_level, 260 filepaths, 261 reference_texts, 262 cfg.batch_size, 263 cfg.num_workers, 264 plot_dir, 265 autocast, 266 ) 267 for level, result in results.items(): 268 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 269 f.flush() 270 else: 271 asr_model.change_decoding_strategy( 272 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 273 if is_rnnt 274 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 275 ) 276 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 277 plot_dir = work_dir / Path(experiment_name) 278 279 logging.info(f"==============================Running a single benchmarking===================================") 280 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directory `{plot_dir}`") 281 282 with open(report_file, "tw", encoding="utf-8") as f: 283 f.write(report_legend) 284 f.flush() 285 results = run_confidence_benchmark( 286 asr_model, 287 cfg.batch_size, 288 cfg.num_workers, 289 cfg.target_level, 290 filepaths, 291 reference_texts, 292 plot_dir, 293 autocast, 294 ) 295 for level, result in results.items(): 296 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 297 logging.info(f"===========================================Done===============================================") 298 299 300 if __name__ == '__main__': 301 main() 302 [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 """ 15 # This script converts an existing audio dataset with a manifest to 16 # a tarred and sharded audio dataset that can be read by the 17 # TarredAudioToTextDataLayer. 18 19 # Please make sure your audio_filepath DOES NOT CONTAIN '-sub'! 20 # Because we will use it to handle files which have duplicate filenames but with different offsets 21 # (see function create_shard for details) 22 23 24 # Bucketing can help to improve the training speed. You may use --buckets_num to specify the number of buckets. 25 # It creates multiple tarred datasets, one per bucket, based on the audio durations. 26 # The range of [min_duration, max_duration) is split into equal sized buckets. 27 # Recommend to use --sort_in_shards to speedup the training by reducing the paddings in the batches 28 # More info on how to use bucketing feature: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/datasets.html 29 30 # If valid NVIDIA DALI version is installed, will also generate the corresponding DALI index files that need to be 31 # supplied to the config in order to utilize webdataset for efficient large dataset handling. 32 # NOTE: DALI + Webdataset is NOT compatible with Bucketing support ! 33 34 # Usage: 35 1) Creating a new tarfile dataset 36 37 python convert_to_tarred_audio_dataset.py \ 38 --manifest_path=<path to the manifest file> \ 39 --target_dir=<path to output directory> \ 40 --num_shards=<number of tarfiles that will contain the audio> \ 41 --max_duration=<float representing maximum duration of audio samples> \ 42 --min_duration=<float representing minimum duration of audio samples> \ 43 --shuffle --shuffle_seed=1 \ 44 --sort_in_shards \ 45 --workers=-1 46 47 48 2) Concatenating more tarfiles to a pre-existing tarred dataset 49 50 python convert_to_tarred_audio_dataset.py \ 51 --manifest_path=<path to the tarred manifest file> \ 52 --metadata_path=<path to the metadata.yaml (or metadata_version_{X}.yaml) file> \ 53 --target_dir=<path to output directory where the original tarfiles are contained> \ 54 --max_duration=<float representing maximum duration of audio samples> \ 55 --min_duration=<float representing minimum duration of audio samples> \ 56 --shuffle --shuffle_seed=1 \ 57 --sort_in_shards \ 58 --workers=-1 \ 59 --concat_manifest_paths \ 60 <space separated paths to 1 or more manifest files to concatenate into the original tarred dataset> 61 62 3) Writing an empty metadata file 63 64 python convert_to_tarred_audio_dataset.py \ 65 --target_dir=<path to output directory> \ 66 # any other optional argument 67 --num_shards=8 \ 68 --max_duration=16.7 \ 69 --min_duration=0.01 \ 70 --shuffle \ 71 --workers=-1 \ 72 --sort_in_shards \ 73 --shuffle_seed=1 \ 74 --write_metadata 75 76 """ 77 import argparse 78 import copy 79 import json 80 import os 81 import random 82 import tarfile 83 from collections import defaultdict 84 from dataclasses import dataclass, field 85 from datetime import datetime 86 from typing import Any, List, Optional 87 88 from joblib import Parallel, delayed 89 from omegaconf import DictConfig, OmegaConf, open_dict 90 91 try: 92 import create_dali_tarred_dataset_index as dali_index 93 94 DALI_INDEX_SCRIPT_AVAILABLE = True 95 except (ImportError, ModuleNotFoundError, FileNotFoundError): 96 DALI_INDEX_SCRIPT_AVAILABLE = False 97 98 parser = argparse.ArgumentParser( 99 description="Convert an existing ASR dataset to tarballs compatible with TarredAudioToTextDataLayer." 100 ) 101 parser.add_argument( 102 "--manifest_path", default=None, type=str, required=False, help="Path to the existing dataset's manifest." 103 ) 104 105 parser.add_argument( 106 '--concat_manifest_paths', 107 nargs='+', 108 default=None, 109 type=str, 110 required=False, 111 help="Path to the additional dataset's manifests that will be concatenated with base dataset.", 112 ) 113 114 # Optional arguments 115 parser.add_argument( 116 "--target_dir", 117 default='./tarred', 118 type=str, 119 help="Target directory for resulting tarballs and manifest. Defaults to `./tarred`. Creates the path if necessary.", 120 ) 121 122 parser.add_argument( 123 "--metadata_path", required=False, default=None, type=str, help="Path to metadata file for the dataset.", 124 ) 125 126 parser.add_argument( 127 "--num_shards", 128 default=-1, 129 type=int, 130 help="Number of shards (tarballs) to create. Used for partitioning data among workers.", 131 ) 132 parser.add_argument( 133 '--max_duration', 134 default=None, 135 required=True, 136 type=float, 137 help='Maximum duration of audio clip in the dataset. By default, it is None and is required to be set.', 138 ) 139 parser.add_argument( 140 '--min_duration', 141 default=None, 142 type=float, 143 help='Minimum duration of audio clip in the dataset. By default, it is None and will not filter files.', 144 ) 145 parser.add_argument( 146 "--shuffle", 147 action='store_true', 148 help="Whether or not to randomly shuffle the samples in the manifest before tarring/sharding.", 149 ) 150 151 parser.add_argument( 152 "--keep_files_together", 153 action='store_true', 154 help="Whether or not to keep entries from the same file (but different offsets) together when sorting before tarring/sharding.", 155 ) 156 157 parser.add_argument( 158 "--sort_in_shards", 159 action='store_true', 160 help="Whether or not to sort samples inside the shards based on their duration.", 161 ) 162 163 parser.add_argument( 164 "--buckets_num", type=int, default=1, help="Number of buckets to create based on duration.", 165 ) 166 167 parser.add_argument("--shuffle_seed", type=int, default=None, help="Random seed for use if shuffling is enabled.") 168 parser.add_argument( 169 '--write_metadata', 170 action='store_true', 171 help=( 172 "Flag to write a blank metadata with the current call config. " 173 "Note that the metadata will not contain the number of shards, " 174 "and it must be filled out by the user." 175 ), 176 ) 177 parser.add_argument( 178 "--no_shard_manifests", 179 action='store_true', 180 help="Do not write sharded manifests along with the aggregated manifest.", 181 ) 182 parser.add_argument('--workers', type=int, default=1, help='Number of worker processes') 183 args = parser.parse_args() 184 185 186 @dataclass 187 class ASRTarredDatasetConfig: 188 num_shards: int = -1 189 shuffle: bool = False 190 max_duration: Optional[float] = None 191 min_duration: Optional[float] = None 192 shuffle_seed: Optional[int] = None 193 sort_in_shards: bool = True 194 shard_manifests: bool = True 195 keep_files_together: bool = False 196 197 198 @dataclass 199 class ASRTarredDatasetMetadata: 200 created_datetime: Optional[str] = None 201 version: int = 0 202 num_samples_per_shard: Optional[int] = None 203 is_concatenated_manifest: bool = False 204 205 dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() 206 history: Optional[List[Any]] = field(default_factory=lambda: []) 207 208 def __post_init__(self): 209 self.created_datetime = self.get_current_datetime() 210 211 def get_current_datetime(self): 212 return datetime.now().strftime("%m-%d-%Y %H-%M-%S") 213 214 @classmethod 215 def from_config(cls, config: DictConfig): 216 obj = cls() 217 obj.__dict__.update(**config) 218 return obj 219 220 @classmethod 221 def from_file(cls, filepath: str): 222 config = OmegaConf.load(filepath) 223 return ASRTarredDatasetMetadata.from_config(config=config) 224 225 226 class ASRTarredDatasetBuilder: 227 """ 228 Helper class that constructs a tarred dataset from scratch, or concatenates tarred datasets 229 together and constructs manifests for them. 230 """ 231 232 def __init__(self): 233 self.config = None 234 235 def configure(self, config: ASRTarredDatasetConfig): 236 """ 237 Sets the config generated from command line overrides. 238 239 Args: 240 config: ASRTarredDatasetConfig dataclass object. 241 """ 242 self.config = config # type: ASRTarredDatasetConfig 243 244 if self.config.num_shards < 0: 245 raise ValueError("`num_shards` must be > 0. Please fill in the metadata information correctly.") 246 247 def create_new_dataset(self, manifest_path: str, target_dir: str = "./tarred/", num_workers: int = 0): 248 """ 249 Creates a new tarred dataset from a given manifest file. 250 251 Args: 252 manifest_path: Path to the original ASR manifest. 253 target_dir: Output directory. 254 num_workers: Integer denoting number of parallel worker processes which will write tarfiles. 255 Defaults to 1 - which denotes sequential worker process. 256 257 Output: 258 Writes tarfiles, along with the tarred dataset compatible manifest file. 259 Also preserves a record of the metadata used to construct this tarred dataset. 260 """ 261 if self.config is None: 262 raise ValueError("Config has not been set. Please call `configure(config: ASRTarredDatasetConfig)`") 263 264 if manifest_path is None: 265 raise FileNotFoundError("Manifest filepath cannot be None !") 266 267 config = self.config # type: ASRTarredDatasetConfig 268 269 if not os.path.exists(target_dir): 270 os.makedirs(target_dir) 271 272 # Read the existing manifest 273 entries, total_duration, filtered_entries, filtered_duration = self._read_manifest(manifest_path, config) 274 275 if len(filtered_entries) > 0: 276 print(f"Filtered {len(filtered_entries)} files which amounts to {filtered_duration} seconds of audio.") 277 print( 278 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 279 ) 280 281 if len(entries) == 0: 282 print("No tarred dataset was created as there were 0 valid samples after filtering!") 283 return 284 if config.shuffle: 285 random.seed(config.shuffle_seed) 286 print("Shuffling...") 287 if config.keep_files_together: 288 filename_entries = defaultdict(list) 289 for ent in entries: 290 filename_entries[ent["audio_filepath"]].append(ent) 291 filenames = list(filename_entries.keys()) 292 random.shuffle(filenames) 293 shuffled_entries = [] 294 for filename in filenames: 295 shuffled_entries += filename_entries[filename] 296 entries = shuffled_entries 297 else: 298 random.shuffle(entries) 299 300 # Create shards and updated manifest entries 301 print(f"Number of samples added : {len(entries)}") 302 print(f"Remainder: {len(entries) % config.num_shards}") 303 304 start_indices = [] 305 end_indices = [] 306 # Build indices 307 for i in range(config.num_shards): 308 start_idx = (len(entries) // config.num_shards) * i 309 end_idx = start_idx + (len(entries) // config.num_shards) 310 print(f"Shard {i} has entries {start_idx} ~ {end_idx}") 311 files = set() 312 for ent_id in range(start_idx, end_idx): 313 files.add(entries[ent_id]["audio_filepath"]) 314 print(f"Shard {i} contains {len(files)} files") 315 if i == config.num_shards - 1: 316 # We discard in order to have the same number of entries per shard. 317 print(f"Have {len(entries) - end_idx} entries left over that will be discarded.") 318 319 start_indices.append(start_idx) 320 end_indices.append(end_idx) 321 322 manifest_folder, _ = os.path.split(manifest_path) 323 324 with Parallel(n_jobs=num_workers, verbose=config.num_shards) as parallel: 325 # Call parallel tarfile construction 326 new_entries_list = parallel( 327 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, i, manifest_folder) 328 for i, (start_idx, end_idx) in enumerate(zip(start_indices, end_indices)) 329 ) 330 331 if config.shard_manifests: 332 sharded_manifests_dir = target_dir + '/sharded_manifests' 333 if not os.path.exists(sharded_manifests_dir): 334 os.makedirs(sharded_manifests_dir) 335 336 for manifest in new_entries_list: 337 shard_id = manifest[0]['shard_id'] 338 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 339 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 340 for entry in manifest: 341 json.dump(entry, m2) 342 m2.write('\n') 343 344 # Flatten the list of list of entries to a list of entries 345 new_entries = [sample for manifest in new_entries_list for sample in manifest] 346 del new_entries_list 347 348 print("Total number of entries in manifest :", len(new_entries)) 349 350 # Write manifest 351 new_manifest_path = os.path.join(target_dir, 'tarred_audio_manifest.json') 352 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 353 for entry in new_entries: 354 json.dump(entry, m2) 355 m2.write('\n') 356 357 # Write metadata (default metadata for new datasets) 358 new_metadata_path = os.path.join(target_dir, 'metadata.yaml') 359 metadata = ASRTarredDatasetMetadata() 360 361 # Update metadata 362 metadata.dataset_config = config 363 metadata.num_samples_per_shard = len(new_entries) // config.num_shards 364 365 # Write metadata 366 metadata_yaml = OmegaConf.structured(metadata) 367 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 368 369 def create_concatenated_dataset( 370 self, 371 base_manifest_path: str, 372 manifest_paths: List[str], 373 metadata: ASRTarredDatasetMetadata, 374 target_dir: str = "./tarred_concatenated/", 375 num_workers: int = 1, 376 ): 377 """ 378 Creates new tarfiles in order to create a concatenated dataset, whose manifest contains the data for 379 both the original dataset as well as the new data submitted in manifest paths. 380 381 Args: 382 base_manifest_path: Path to the manifest file which contains the information for the original 383 tarred dataset (with flattened paths). 384 manifest_paths: List of one or more paths to manifest files that will be concatenated with above 385 base tarred dataset. 386 metadata: ASRTarredDatasetMetadata dataclass instance with overrides from command line. 387 target_dir: Output directory 388 389 Output: 390 Writes tarfiles which with indices mapping to a "concatenated" tarred dataset, 391 along with the tarred dataset compatible manifest file which includes information 392 about all the datasets that comprise the concatenated dataset. 393 394 Also preserves a record of the metadata used to construct this tarred dataset. 395 """ 396 if not os.path.exists(target_dir): 397 os.makedirs(target_dir) 398 399 if base_manifest_path is None: 400 raise FileNotFoundError("Base manifest filepath cannot be None !") 401 402 if manifest_paths is None or len(manifest_paths) == 0: 403 raise FileNotFoundError("List of additional manifest filepaths cannot be None !") 404 405 config = ASRTarredDatasetConfig(**(metadata.dataset_config)) 406 407 # Read the existing manifest (no filtering here) 408 base_entries, _, _, _ = self._read_manifest(base_manifest_path, config) 409 print(f"Read base manifest containing {len(base_entries)} samples.") 410 411 # Precompute number of samples per shard 412 if metadata.num_samples_per_shard is None: 413 num_samples_per_shard = len(base_entries) // config.num_shards 414 else: 415 num_samples_per_shard = metadata.num_samples_per_shard 416 417 print("Number of samples per shard :", num_samples_per_shard) 418 419 # Compute min and max duration and update config (if no metadata passed) 420 print(f"Selected max duration : {config.max_duration}") 421 print(f"Selected min duration : {config.min_duration}") 422 423 entries = [] 424 for new_manifest_idx in range(len(manifest_paths)): 425 new_entries, total_duration, filtered_new_entries, filtered_duration = self._read_manifest( 426 manifest_paths[new_manifest_idx], config 427 ) 428 429 if len(filtered_new_entries) > 0: 430 print( 431 f"Filtered {len(filtered_new_entries)} files which amounts to {filtered_duration:0.2f}" 432 f" seconds of audio from manifest {manifest_paths[new_manifest_idx]}." 433 ) 434 print( 435 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 436 ) 437 438 entries.extend(new_entries) 439 440 if len(entries) == 0: 441 print("No tarred dataset was created as there were 0 valid samples after filtering!") 442 return 443 444 if config.shuffle: 445 random.seed(config.shuffle_seed) 446 print("Shuffling...") 447 random.shuffle(entries) 448 449 # Drop last section of samples that cannot be added onto a chunk 450 drop_count = len(entries) % num_samples_per_shard 451 total_new_entries = len(entries) 452 entries = entries[:-drop_count] 453 454 print( 455 f"Dropping {drop_count} samples from total new samples {total_new_entries} since they cannot " 456 f"be added into a uniformly sized chunk." 457 ) 458 459 # Create shards and updated manifest entries 460 num_added_shards = len(entries) // num_samples_per_shard 461 462 print(f"Number of samples in base dataset : {len(base_entries)}") 463 print(f"Number of samples in additional datasets : {len(entries)}") 464 print(f"Number of added shards : {num_added_shards}") 465 print(f"Remainder: {len(entries) % num_samples_per_shard}") 466 467 start_indices = [] 468 end_indices = [] 469 shard_indices = [] 470 for i in range(num_added_shards): 471 start_idx = (len(entries) // num_added_shards) * i 472 end_idx = start_idx + (len(entries) // num_added_shards) 473 shard_idx = i + config.num_shards 474 print(f"Shard {shard_idx} has entries {start_idx + len(base_entries)} ~ {end_idx + len(base_entries)}") 475 476 start_indices.append(start_idx) 477 end_indices.append(end_idx) 478 shard_indices.append(shard_idx) 479 480 manifest_folder, _ = os.path.split(base_manifest_path) 481 482 with Parallel(n_jobs=num_workers, verbose=num_added_shards) as parallel: 483 # Call parallel tarfile construction 484 new_entries_list = parallel( 485 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, shard_idx, manifest_folder) 486 for i, (start_idx, end_idx, shard_idx) in enumerate(zip(start_indices, end_indices, shard_indices)) 487 ) 488 489 if config.shard_manifests: 490 sharded_manifests_dir = target_dir + '/sharded_manifests' 491 if not os.path.exists(sharded_manifests_dir): 492 os.makedirs(sharded_manifests_dir) 493 494 for manifest in new_entries_list: 495 shard_id = manifest[0]['shard_id'] 496 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 497 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 498 for entry in manifest: 499 json.dump(entry, m2) 500 m2.write('\n') 501 502 # Flatten the list of list of entries to a list of entries 503 new_entries = [sample for manifest in new_entries_list for sample in manifest] 504 del new_entries_list 505 506 # Write manifest 507 if metadata is None: 508 new_version = 1 # start with `1`, where `0` indicates the base manifest + dataset 509 else: 510 new_version = metadata.version + 1 511 512 print("Total number of entries in manifest :", len(base_entries) + len(new_entries)) 513 514 new_manifest_path = os.path.join(target_dir, f'tarred_audio_manifest_version_{new_version}.json') 515 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 516 # First write all the entries of base manifest 517 for entry in base_entries: 518 json.dump(entry, m2) 519 m2.write('\n') 520 521 # Finally write the new entries 522 for entry in new_entries: 523 json.dump(entry, m2) 524 m2.write('\n') 525 526 # Preserve historical metadata 527 base_metadata = metadata 528 529 # Write metadata (updated metadata for concatenated datasets) 530 new_metadata_path = os.path.join(target_dir, f'metadata_version_{new_version}.yaml') 531 metadata = ASRTarredDatasetMetadata() 532 533 # Update config 534 config.num_shards = config.num_shards + num_added_shards 535 536 # Update metadata 537 metadata.version = new_version 538 metadata.dataset_config = config 539 metadata.num_samples_per_shard = num_samples_per_shard 540 metadata.is_concatenated_manifest = True 541 metadata.created_datetime = metadata.get_current_datetime() 542 543 # Attach history 544 current_metadata = OmegaConf.structured(base_metadata.history) 545 metadata.history = current_metadata 546 547 # Write metadata 548 metadata_yaml = OmegaConf.structured(metadata) 549 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 550 551 def _read_manifest(self, manifest_path: str, config: ASRTarredDatasetConfig): 552 """Read and filters data from the manifest""" 553 # Read the existing manifest 554 entries = [] 555 total_duration = 0.0 556 filtered_entries = [] 557 filtered_duration = 0.0 558 with open(manifest_path, 'r', encoding='utf-8') as m: 559 for line in m: 560 entry = json.loads(line) 561 if (config.max_duration is None or entry['duration'] < config.max_duration) and ( 562 config.min_duration is None or entry['duration'] >= config.min_duration 563 ): 564 entries.append(entry) 565 total_duration += entry["duration"] 566 else: 567 filtered_entries.append(entry) 568 filtered_duration += entry['duration'] 569 570 return entries, total_duration, filtered_entries, filtered_duration 571 572 def _create_shard(self, entries, target_dir, shard_id, manifest_folder): 573 """Creates a tarball containing the audio files from `entries`. 574 """ 575 if self.config.sort_in_shards: 576 entries.sort(key=lambda x: x["duration"], reverse=False) 577 578 new_entries = [] 579 tar = tarfile.open(os.path.join(target_dir, f'audio_{shard_id}.tar'), mode='w', dereference=True) 580 581 count = dict() 582 for entry in entries: 583 # We squash the filename since we do not preserve directory structure of audio files in the tarball. 584 if os.path.exists(entry["audio_filepath"]): 585 audio_filepath = entry["audio_filepath"] 586 else: 587 audio_filepath = os.path.join(manifest_folder, entry["audio_filepath"]) 588 if not os.path.exists(audio_filepath): 589 raise FileNotFoundError(f"Could not find {entry['audio_filepath']}!") 590 591 base, ext = os.path.splitext(audio_filepath) 592 base = base.replace('/', '_') 593 # Need the following replacement as long as WebDataset splits on first period 594 base = base.replace('.', '_') 595 squashed_filename = f'{base}{ext}' 596 if squashed_filename not in count: 597 tar.add(audio_filepath, arcname=squashed_filename) 598 to_write = squashed_filename 599 count[squashed_filename] = 1 600 else: 601 to_write = base + "-sub" + str(count[squashed_filename]) + ext 602 count[squashed_filename] += 1 603 604 new_entry = { 605 'audio_filepath': to_write, 606 'duration': entry['duration'], 607 'shard_id': shard_id, # Keep shard ID for recordkeeping 608 } 609 610 if 'label' in entry: 611 new_entry['label'] = entry['label'] 612 613 if 'text' in entry: 614 new_entry['text'] = entry['text'] 615 616 if 'offset' in entry: 617 new_entry['offset'] = entry['offset'] 618 619 if 'lang' in entry: 620 new_entry['lang'] = entry['lang'] 621 622 new_entries.append(new_entry) 623 624 tar.close() 625 return new_entries 626 627 @classmethod 628 def setup_history(cls, base_metadata: ASRTarredDatasetMetadata, history: List[Any]): 629 if 'history' in base_metadata.keys(): 630 for history_val in base_metadata.history: 631 cls.setup_history(history_val, history) 632 633 if base_metadata is not None: 634 metadata_copy = copy.deepcopy(base_metadata) 635 with open_dict(metadata_copy): 636 metadata_copy.pop('history', None) 637 history.append(metadata_copy) 638 639 640 def main(): 641 if args.buckets_num > 1: 642 bucket_length = (args.max_duration - args.min_duration) / float(args.buckets_num) 643 for i in range(args.buckets_num): 644 min_duration = args.min_duration + i * bucket_length 645 max_duration = min_duration + bucket_length 646 if i == args.buckets_num - 1: 647 # add a small number to cover the samples with exactly duration of max_duration in the last bucket. 648 max_duration += 1e-5 649 target_dir = os.path.join(args.target_dir, f"bucket{i+1}") 650 print(f"Creating bucket {i+1} with min_duration={min_duration} and max_duration={max_duration} ...") 651 print(f"Results are being saved at: {target_dir}.") 652 create_tar_datasets(min_duration=min_duration, max_duration=max_duration, target_dir=target_dir) 653 print(f"Bucket {i+1} is created.") 654 else: 655 create_tar_datasets(min_duration=args.min_duration, max_duration=args.max_duration, target_dir=args.target_dir) 656 657 658 def create_tar_datasets(min_duration: float, max_duration: float, target_dir: str): 659 builder = ASRTarredDatasetBuilder() 660 661 shard_manifests = False if args.no_shard_manifests else True 662 663 if args.write_metadata: 664 metadata = ASRTarredDatasetMetadata() 665 dataset_cfg = ASRTarredDatasetConfig( 666 num_shards=args.num_shards, 667 shuffle=args.shuffle, 668 max_duration=max_duration, 669 min_duration=min_duration, 670 shuffle_seed=args.shuffle_seed, 671 sort_in_shards=args.sort_in_shards, 672 shard_manifests=shard_manifests, 673 keep_files_together=args.keep_files_together, 674 ) 675 metadata.dataset_config = dataset_cfg 676 677 output_path = os.path.join(target_dir, 'default_metadata.yaml') 678 OmegaConf.save(metadata, output_path, resolve=True) 679 print(f"Default metadata written to {output_path}") 680 exit(0) 681 682 if args.concat_manifest_paths is None or len(args.concat_manifest_paths) == 0: 683 print("Creating new tarred dataset ...") 684 685 # Create a tarred dataset from scratch 686 config = ASRTarredDatasetConfig( 687 num_shards=args.num_shards, 688 shuffle=args.shuffle, 689 max_duration=max_duration, 690 min_duration=min_duration, 691 shuffle_seed=args.shuffle_seed, 692 sort_in_shards=args.sort_in_shards, 693 shard_manifests=shard_manifests, 694 keep_files_together=args.keep_files_together, 695 ) 696 builder.configure(config) 697 builder.create_new_dataset(manifest_path=args.manifest_path, target_dir=target_dir, num_workers=args.workers) 698 699 else: 700 if args.buckets_num > 1: 701 raise ValueError("Concatenation feature does not support buckets_num > 1.") 702 print("Concatenating multiple tarred datasets ...") 703 704 # Implicitly update config from base details 705 if args.metadata_path is not None: 706 metadata = ASRTarredDatasetMetadata.from_file(args.metadata_path) 707 else: 708 raise ValueError("`metadata` yaml file path must be provided!") 709 710 # Preserve history 711 history = [] 712 builder.setup_history(OmegaConf.structured(metadata), history) 713 metadata.history = history 714 715 # Add command line overrides (everything other than num_shards) 716 metadata.dataset_config.max_duration = max_duration 717 metadata.dataset_config.min_duration = min_duration 718 metadata.dataset_config.shuffle = args.shuffle 719 metadata.dataset_config.shuffle_seed = args.shuffle_seed 720 metadata.dataset_config.sort_in_shards = args.sort_in_shards 721 metadata.dataset_config.shard_manifests = shard_manifests 722 723 builder.configure(metadata.dataset_config) 724 725 # Concatenate a tarred dataset onto a previous one 726 builder.create_concatenated_dataset( 727 base_manifest_path=args.manifest_path, 728 manifest_paths=args.concat_manifest_paths, 729 metadata=metadata, 730 target_dir=target_dir, 731 num_workers=args.workers, 732 ) 733 734 if DALI_INDEX_SCRIPT_AVAILABLE and dali_index.INDEX_CREATOR_AVAILABLE: 735 print("Constructing DALI Tarfile Index - ", target_dir) 736 index_config = dali_index.DALITarredIndexConfig(tar_dir=target_dir, workers=args.workers) 737 dali_index.main(index_config) 738 739 740 if __name__ == "__main__": 741 main() 742 [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of tools/nemo_forced_aligner/align.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import math 17 import os 18 from dataclasses import dataclass, field, is_dataclass 19 from pathlib import Path 20 from typing import List, Optional 21 22 import torch 23 from omegaconf import OmegaConf 24 from utils.data_prep import ( 25 add_t_start_end_to_utt_obj, 26 get_batch_starts_ends, 27 get_batch_variables, 28 get_manifest_lines_batch, 29 is_entry_in_all_lines, 30 is_entry_in_any_lines, 31 ) 32 from utils.make_ass_files import make_ass_files 33 from utils.make_ctm_files import make_ctm_files 34 from utils.make_output_manifest import write_manifest_out_line 35 from utils.viterbi_decoding import viterbi_decoding 36 37 from nemo.collections.asr.models.ctc_models import EncDecCTCModel 38 from nemo.collections.asr.models.hybrid_rnnt_ctc_models import EncDecHybridRNNTCTCModel 39 from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchASR 40 from nemo.collections.asr.parts.utils.transcribe_utils import setup_model 41 from nemo.core.config import hydra_runner 42 from nemo.utils import logging 43 44 """ 45 Align the utterances in manifest_filepath. 46 Results are saved in ctm files in output_dir. 47 48 Arguments: 49 pretrained_name: string specifying the name of a CTC NeMo ASR model which will be automatically downloaded 50 from NGC and used for generating the log-probs which we will use to do alignment. 51 Note: NFA can only use CTC models (not Transducer models) at the moment. 52 model_path: string specifying the local filepath to a CTC NeMo ASR model which will be used to generate the 53 log-probs which we will use to do alignment. 54 Note: NFA can only use CTC models (not Transducer models) at the moment. 55 Note: if a model_path is provided, it will override the pretrained_name. 56 manifest_filepath: filepath to the manifest of the data you want to align, 57 containing 'audio_filepath' and 'text' fields. 58 output_dir: the folder where output CTM files and new JSON manifest will be saved. 59 align_using_pred_text: if True, will transcribe the audio using the specified model and then use that transcription 60 as the reference text for the forced alignment. 61 transcribe_device: None, or a string specifying the device that will be used for generating log-probs (i.e. "transcribing"). 62 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 63 (otherwise will set it to 'cpu'). 64 viterbi_device: None, or string specifying the device that will be used for doing Viterbi decoding. 65 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 66 (otherwise will set it to 'cpu'). 67 batch_size: int specifying batch size that will be used for generating log-probs and doing Viterbi decoding. 68 use_local_attention: boolean flag specifying whether to try to use local attention for the ASR Model (will only 69 work if the ASR Model is a Conformer model). If local attention is used, we will set the local attention context 70 size to [64,64]. 71 additional_segment_grouping_separator: an optional string used to separate the text into smaller segments. 72 If this is not specified, then the whole text will be treated as a single segment. 73 remove_blank_tokens_from_ctm: a boolean denoting whether to remove <blank> tokens from token-level output CTMs. 74 audio_filepath_parts_in_utt_id: int specifying how many of the 'parts' of the audio_filepath 75 we will use (starting from the final part of the audio_filepath) to determine the 76 utt_id that will be used in the CTM files. Note also that any spaces that are present in the audio_filepath 77 will be replaced with dashes, so as not to change the number of space-separated elements in the 78 CTM files. 79 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 1 => utt_id will be "e1" 80 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 2 => utt_id will be "d_e1" 81 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 3 => utt_id will be "c_d_e1" 82 use_buffered_infer: False, if set True, using streaming to do get the logits for alignment 83 This flag is useful when aligning large audio file. 84 However, currently the chunk streaming inference does not support batch inference, 85 which means even you set batch_size > 1, it will only infer one by one instead of doing 86 the whole batch inference together. 87 chunk_len_in_secs: float chunk length in seconds 88 total_buffer_in_secs: float Length of buffer (chunk + left and right padding) in seconds 89 chunk_batch_size: int batch size for buffered chunk inference, 90 which will cut one audio into segments and do inference on chunk_batch_size segments at a time 91 92 simulate_cache_aware_streaming: False, if set True, using cache aware streaming to do get the logits for alignment 93 94 save_output_file_formats: List of strings specifying what type of output files to save (default: ["ctm", "ass"]) 95 ctm_file_config: CTMFileConfig to specify the configuration of the output CTM files 96 ass_file_config: ASSFileConfig to specify the configuration of the output ASS files 97 """ 98 99 100 @dataclass 101 class CTMFileConfig: 102 remove_blank_tokens: bool = False 103 # minimum duration (in seconds) for timestamps in the CTM.If any line in the CTM has a 104 # duration lower than this, it will be enlarged from the middle outwards until it 105 # meets the minimum_timestamp_duration, or reaches the beginning or end of the audio file. 106 # Note that this may cause timestamps to overlap. 107 minimum_timestamp_duration: float = 0 108 109 110 @dataclass 111 class ASSFileConfig: 112 fontsize: int = 20 113 vertical_alignment: str = "center" 114 # if resegment_text_to_fill_space is True, the ASS files will use new segments 115 # such that each segment will not take up more than (approximately) max_lines_per_segment 116 # when the ASS file is applied to a video 117 resegment_text_to_fill_space: bool = False 118 max_lines_per_segment: int = 2 119 text_already_spoken_rgb: List[int] = field(default_factory=lambda: [49, 46, 61]) # dark gray 120 text_being_spoken_rgb: List[int] = field(default_factory=lambda: [57, 171, 9]) # dark green 121 text_not_yet_spoken_rgb: List[int] = field(default_factory=lambda: [194, 193, 199]) # light gray 122 123 124 @dataclass 125 class AlignmentConfig: 126 # Required configs 127 pretrained_name: Optional[str] = None 128 model_path: Optional[str] = None 129 manifest_filepath: Optional[str] = None 130 output_dir: Optional[str] = None 131 132 # General configs 133 align_using_pred_text: bool = False 134 transcribe_device: Optional[str] = None 135 viterbi_device: Optional[str] = None 136 batch_size: int = 1 137 use_local_attention: bool = True 138 additional_segment_grouping_separator: Optional[str] = None 139 audio_filepath_parts_in_utt_id: int = 1 140 141 # Buffered chunked streaming configs 142 use_buffered_chunked_streaming: bool = False 143 chunk_len_in_secs: float = 1.6 144 total_buffer_in_secs: float = 4.0 145 chunk_batch_size: int = 32 146 147 # Cache aware streaming configs 148 simulate_cache_aware_streaming: Optional[bool] = False 149 150 # Output file configs 151 save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) 152 ctm_file_config: CTMFileConfig = CTMFileConfig() 153 ass_file_config: ASSFileConfig = ASSFileConfig() 154 155 156 @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) 157 def main(cfg: AlignmentConfig): 158 159 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 160 161 if is_dataclass(cfg): 162 cfg = OmegaConf.structured(cfg) 163 164 # Validate config 165 if cfg.model_path is None and cfg.pretrained_name is None: 166 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None") 167 168 if cfg.model_path is not None and cfg.pretrained_name is not None: 169 raise ValueError("One of cfg.model_path and cfg.pretrained_name must be None") 170 171 if cfg.manifest_filepath is None: 172 raise ValueError("cfg.manifest_filepath must be specified") 173 174 if cfg.output_dir is None: 175 raise ValueError("cfg.output_dir must be specified") 176 177 if cfg.batch_size < 1: 178 raise ValueError("cfg.batch_size cannot be zero or a negative number") 179 180 if cfg.additional_segment_grouping_separator == "" or cfg.additional_segment_grouping_separator == " ": 181 raise ValueError("cfg.additional_grouping_separator cannot be empty string or space character") 182 183 if cfg.ctm_file_config.minimum_timestamp_duration < 0: 184 raise ValueError("cfg.minimum_timestamp_duration cannot be a negative number") 185 186 if cfg.ass_file_config.vertical_alignment not in ["top", "center", "bottom"]: 187 raise ValueError("cfg.ass_file_config.vertical_alignment must be one of 'top', 'center' or 'bottom'") 188 189 for rgb_list in [ 190 cfg.ass_file_config.text_already_spoken_rgb, 191 cfg.ass_file_config.text_already_spoken_rgb, 192 cfg.ass_file_config.text_already_spoken_rgb, 193 ]: 194 if len(rgb_list) != 3: 195 raise ValueError( 196 "cfg.ass_file_config.text_already_spoken_rgb," 197 " cfg.ass_file_config.text_being_spoken_rgb," 198 " and cfg.ass_file_config.text_already_spoken_rgb all need to contain" 199 " exactly 3 elements." 200 ) 201 202 # Validate manifest contents 203 if not is_entry_in_all_lines(cfg.manifest_filepath, "audio_filepath"): 204 raise RuntimeError( 205 "At least one line in cfg.manifest_filepath does not contain an 'audio_filepath' entry. " 206 "All lines must contain an 'audio_filepath' entry." 207 ) 208 209 if cfg.align_using_pred_text: 210 if is_entry_in_any_lines(cfg.manifest_filepath, "pred_text"): 211 raise RuntimeError( 212 "Cannot specify cfg.align_using_pred_text=True when the manifest at cfg.manifest_filepath " 213 "contains 'pred_text' entries. This is because the audio will be transcribed and may produce " 214 "a different 'pred_text'. This may cause confusion." 215 ) 216 else: 217 if not is_entry_in_all_lines(cfg.manifest_filepath, "text"): 218 raise RuntimeError( 219 "At least one line in cfg.manifest_filepath does not contain a 'text' entry. " 220 "NFA requires all lines to contain a 'text' entry when cfg.align_using_pred_text=False." 221 ) 222 223 # init devices 224 if cfg.transcribe_device is None: 225 transcribe_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 226 else: 227 transcribe_device = torch.device(cfg.transcribe_device) 228 logging.info(f"Device to be used for transcription step (`transcribe_device`) is {transcribe_device}") 229 230 if cfg.viterbi_device is None: 231 viterbi_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 232 else: 233 viterbi_device = torch.device(cfg.viterbi_device) 234 logging.info(f"Device to be used for viterbi step (`viterbi_device`) is {viterbi_device}") 235 236 if transcribe_device.type == 'cuda' or viterbi_device.type == 'cuda': 237 logging.warning( 238 'One or both of transcribe_device and viterbi_device are GPUs. If you run into OOM errors ' 239 'it may help to change both devices to be the CPU.' 240 ) 241 242 # load model 243 model, _ = setup_model(cfg, transcribe_device) 244 model.eval() 245 246 if isinstance(model, EncDecHybridRNNTCTCModel): 247 model.change_decoding_strategy(decoder_type="ctc") 248 249 if cfg.use_local_attention: 250 logging.info( 251 "Flag use_local_attention is set to True => will try to use local attention for model if it allows it" 252 ) 253 model.change_attention_model(self_attention_model="rel_pos_local_attn", att_context_size=[64, 64]) 254 255 if not (isinstance(model, EncDecCTCModel) or isinstance(model, EncDecHybridRNNTCTCModel)): 256 raise NotImplementedError( 257 f"Model is not an instance of NeMo EncDecCTCModel or ENCDecHybridRNNTCTCModel." 258 " Currently only instances of these models are supported" 259 ) 260 261 if cfg.ctm_file_config.minimum_timestamp_duration > 0: 262 logging.warning( 263 f"cfg.ctm_file_config.minimum_timestamp_duration has been set to {cfg.ctm_file_config.minimum_timestamp_duration} seconds. " 264 "This may cause the alignments for some tokens/words/additional segments to be overlapping." 265 ) 266 267 buffered_chunk_params = {} 268 if cfg.use_buffered_chunked_streaming: 269 model_cfg = copy.deepcopy(model._cfg) 270 271 OmegaConf.set_struct(model_cfg.preprocessor, False) 272 # some changes for streaming scenario 273 model_cfg.preprocessor.dither = 0.0 274 model_cfg.preprocessor.pad_to = 0 275 276 if model_cfg.preprocessor.normalize != "per_feature": 277 logging.error( 278 "Only EncDecCTCModelBPE models trained with per_feature normalization are supported currently" 279 ) 280 # Disable config overwriting 281 OmegaConf.set_struct(model_cfg.preprocessor, True) 282 283 feature_stride = model_cfg.preprocessor['window_stride'] 284 model_stride_in_secs = feature_stride * cfg.model_downsample_factor 285 total_buffer = cfg.total_buffer_in_secs 286 chunk_len = float(cfg.chunk_len_in_secs) 287 tokens_per_chunk = math.ceil(chunk_len / model_stride_in_secs) 288 mid_delay = math.ceil((chunk_len + (total_buffer - chunk_len) / 2) / model_stride_in_secs) 289 logging.info(f"tokens_per_chunk is {tokens_per_chunk}, mid_delay is {mid_delay}") 290 291 model = FrameBatchASR( 292 asr_model=model, 293 frame_len=chunk_len, 294 total_buffer=cfg.total_buffer_in_secs, 295 batch_size=cfg.chunk_batch_size, 296 ) 297 buffered_chunk_params = { 298 "delay": mid_delay, 299 "model_stride_in_secs": model_stride_in_secs, 300 "tokens_per_chunk": tokens_per_chunk, 301 } 302 # get start and end line IDs of batches 303 starts, ends = get_batch_starts_ends(cfg.manifest_filepath, cfg.batch_size) 304 305 # init output_timestep_duration = None and we will calculate and update it during the first batch 306 output_timestep_duration = None 307 308 # init f_manifest_out 309 os.makedirs(cfg.output_dir, exist_ok=True) 310 tgt_manifest_name = str(Path(cfg.manifest_filepath).stem) + "_with_output_file_paths.json" 311 tgt_manifest_filepath = str(Path(cfg.output_dir) / tgt_manifest_name) 312 f_manifest_out = open(tgt_manifest_filepath, 'w') 313 314 # get alignment and save in CTM batch-by-batch 315 for start, end in zip(starts, ends): 316 manifest_lines_batch = get_manifest_lines_batch(cfg.manifest_filepath, start, end) 317 318 (log_probs_batch, y_batch, T_batch, U_batch, utt_obj_batch, output_timestep_duration,) = get_batch_variables( 319 manifest_lines_batch, 320 model, 321 cfg.additional_segment_grouping_separator, 322 cfg.align_using_pred_text, 323 cfg.audio_filepath_parts_in_utt_id, 324 output_timestep_duration, 325 cfg.simulate_cache_aware_streaming, 326 cfg.use_buffered_chunked_streaming, 327 buffered_chunk_params, 328 ) 329 330 alignments_batch = viterbi_decoding(log_probs_batch, y_batch, T_batch, U_batch, viterbi_device) 331 332 for utt_obj, alignment_utt in zip(utt_obj_batch, alignments_batch): 333 334 utt_obj = add_t_start_end_to_utt_obj(utt_obj, alignment_utt, output_timestep_duration) 335 336 if "ctm" in cfg.save_output_file_formats: 337 utt_obj = make_ctm_files(utt_obj, cfg.output_dir, cfg.ctm_file_config,) 338 339 if "ass" in cfg.save_output_file_formats: 340 utt_obj = make_ass_files(utt_obj, cfg.output_dir, cfg.ass_file_config) 341 342 write_manifest_out_line( 343 f_manifest_out, utt_obj, 344 ) 345 346 f_manifest_out.close() 347 348 return None 349 350 351 if __name__ == "__main__": 352 main() 353 [end of tools/nemo_forced_aligner/align.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
NVIDIA/NeMo
15db83ec4a65e649d83b61d7a4a58d911586e853
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-10-03T19:14:38Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,7 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,7 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed @@ -2201,7 +2201,7 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -175,7 +175,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() + method_cfg: ConfidenceMethodConfig = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -118,7 +118,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
slackapi__python-slack-events-api-71
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): </issue> <code> [start of README.rst] 1 Slack Events API adapter for Python 2 =================================== 3 4 .. image:: https://badge.fury.io/py/slackeventsapi.svg 5 :target: https://pypi.org/project/slackeventsapi/ 6 .. image:: https://travis-ci.org/slackapi/python-slack-events-api.svg?branch=master 7 :target: https://travis-ci.org/slackapi/python-slack-events-api 8 .. image:: https://codecov.io/gh/slackapi/python-slack-events-api/branch/master/graph/badge.svg 9 :target: https://codecov.io/gh/slackapi/python-slack-events-api 10 11 12 The Slack Events Adapter is a Python-based solution to receive and parse events 13 from Slackโ€™s Events API. This library uses an event emitter framework to allow 14 you to easily process Slack events by simply attaching functions 15 to event listeners. 16 17 This adapter enhances and simplifies Slack's Events API by incorporating useful best practices, patterns, and opportunities to abstract out common tasks. 18 19 ๐Ÿ’ก We wrote a `blog post which explains how`_ the Events API can help you, why we built these tools, and how you can use them to build production-ready Slack apps. 20 21 .. _blog post which explains how: https://medium.com/@SlackAPI/enhancing-slacks-events-api-7535827829ab 22 23 24 ๐Ÿค– Installation 25 ---------------- 26 27 .. code:: shell 28 29 pip install slackeventsapi 30 31 ๐Ÿค– App Setup 32 -------------------- 33 34 Before you can use the `Events API`_ you must 35 `create a Slack App`_, and turn on 36 `Event Subscriptions`_. 37 38 ๐Ÿ’ก When you add the Request URL to your app's Event Subscription settings, 39 Slack will send a request containing a `challenge` code to verify that your 40 server is alive. This package handles that URL Verification event for you, so 41 all you need to do is start the example app, start ngrok and configure your 42 URL accordingly. 43 44 โœ… Once you have your `Request URL` verified, your app is ready to start 45 receiving Team Events. 46 47 ๐Ÿ”‘ Your server will begin receiving Events from Slack's Events API as soon as a 48 user has authorized your app. 49 50 ๐Ÿค– Development workflow: 51 =========================== 52 53 (1) Create a Slack app on https://api.slack.com/apps 54 (2) Add a `bot user` for your app 55 (3) Start the example app on your **Request URL** endpoint 56 (4) Start ngrok and copy the **HTTPS** URL 57 (5) Add your **Request URL** and subscribe your app to events 58 (6) Go to your ngrok URL (e.g. https://myapp12.ngrok.com/) and auth your app 59 60 **๐ŸŽ‰ Once your app has been authorized, you will begin receiving Slack Events** 61 62 โš ๏ธ Ngrok is a great tool for developing Slack apps, but we don't recommend using ngrok 63 for production apps. 64 65 ๐Ÿค– Usage 66 ---------- 67 **โš ๏ธ Keep your app's credentials safe!** 68 69 - For development, keep them in virtualenv variables. 70 71 - For production, use a secure data store. 72 73 - Never post your app's credentials to github. 74 75 .. code:: python 76 77 SLACK_SIGNING_SECRET = os.environ["SLACK_SIGNING_SECRET"] 78 79 Create a Slack Event Adapter for receiving actions via the Events API 80 ----------------------------------------------------------------------- 81 **Using the built-in Flask server:** 82 83 .. code:: python 84 85 from slackeventsapi import SlackEventAdapter 86 87 88 slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, endpoint="/slack/events") 89 90 91 # Create an event listener for "reaction_added" events and print the emoji name 92 @slack_events_adapter.on("reaction_added") 93 def reaction_added(event_data): 94 emoji = event_data["event"]["reaction"] 95 print(emoji) 96 97 98 # Start the server on port 3000 99 slack_events_adapter.start(port=3000) 100 101 102 **Using your existing Flask instance:** 103 104 105 .. code:: python 106 107 from flask import Flask 108 from slackeventsapi import SlackEventAdapter 109 110 111 # This `app` represents your existing Flask app 112 app = Flask(__name__) 113 114 115 # An example of one of your Flask app's routes 116 @app.route("/") 117 def hello(): 118 return "Hello there!" 119 120 121 # Bind the Events API route to your existing Flask app by passing the server 122 # instance as the last param, or with `server=app`. 123 slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, "/slack/events", app) 124 125 126 # Create an event listener for "reaction_added" events and print the emoji name 127 @slack_events_adapter.on("reaction_added") 128 def reaction_added(event_data): 129 emoji = event_data["event"]["reaction"] 130 print(emoji) 131 132 133 # Start the server on port 3000 134 if __name__ == "__main__": 135 app.run(port=3000) 136 137 For a comprehensive list of available Slack `Events` and more information on 138 `Scopes`, see https://api.slack.com/events-api 139 140 ๐Ÿค– Example event listeners 141 ----------------------------- 142 143 See `example.py`_ for usage examples. This example also utilizes the 144 SlackClient Web API client. 145 146 .. _example.py: /example/ 147 148 ๐Ÿค” Support 149 ----------- 150 151 Need help? Join `Slack Community`_ and talk to us in `#slack-api`_. 152 153 You can also `create an Issue`_ right here on GitHub. 154 155 .. _Events API: https://api.slack.com/events-api 156 .. _create a Slack App: https://api.slack.com/apps/new 157 .. _Event Subscriptions: https://api.slack.com/events-api#subscriptions 158 .. _Slack Community: http://slackcommunity.com/ 159 .. _#slack-api: https://dev4slack.slack.com/messages/slack-api/ 160 .. _create an Issue: https://github.com/slackapi/python-slack-events-api/issues/new 161 [end of README.rst] [start of /dev/null] 1 [end of /dev/null] [start of slackeventsapi/server.py] 1 from flask import Flask, request, make_response, Blueprint 2 import json 3 import platform 4 import sys 5 import hmac 6 import hashlib 7 from time import time 8 from .version import __version__ 9 10 11 class SlackServer(Flask): 12 def __init__(self, signing_secret, endpoint, emitter, server): 13 self.signing_secret = signing_secret 14 self.emitter = emitter 15 self.endpoint = endpoint 16 self.package_info = self.get_package_info() 17 18 # If a server is passed in, bind the event handler routes to it, 19 # otherwise create a new Flask instance. 20 if server: 21 if isinstance(server, Flask) or isinstance(server, Blueprint): 22 self.bind_route(server) 23 else: 24 raise TypeError("Server must be an instance of Flask or Blueprint") 25 else: 26 Flask.__init__(self, __name__) 27 self.bind_route(self) 28 29 def get_package_info(self): 30 client_name = __name__.split('.')[0] 31 client_version = __version__ # Version is returned from version.py 32 33 # Collect the package info, Python version and OS version. 34 package_info = { 35 "client": "{0}/{1}".format(client_name, client_version), 36 "python": "Python/{v.major}.{v.minor}.{v.micro}".format(v=sys.version_info), 37 "system": "{0}/{1}".format(platform.system(), platform.release()) 38 } 39 40 # Concatenate and format the user-agent string to be passed into request headers 41 ua_string = [] 42 for key, val in package_info.items(): 43 ua_string.append(val) 44 45 return " ".join(ua_string) 46 47 def verify_signature(self, timestamp, signature): 48 # Verify the request signature of the request sent from Slack 49 # Generate a new hash using the app's signing secret and request data 50 51 # Compare the generated hash and incoming request signature 52 # Python 2.7.6 doesn't support compare_digest 53 # It's recommended to use Python 2.7.7+ 54 # noqa See https://docs.python.org/2/whatsnew/2.7.html#pep-466-network-security-enhancements-for-python-2-7 55 req = str.encode('v0:' + str(timestamp) + ':') + request.get_data() 56 request_hash = 'v0=' + hmac.new( 57 str.encode(self.signing_secret), 58 req, hashlib.sha256 59 ).hexdigest() 60 61 if hasattr(hmac, "compare_digest"): 62 # Compare byte strings for Python 2 63 if (sys.version_info[0] == 2): 64 return hmac.compare_digest(bytes(request_hash), bytes(signature)) 65 else: 66 return hmac.compare_digest(request_hash, signature) 67 else: 68 if len(request_hash) != len(signature): 69 return False 70 result = 0 71 if isinstance(request_hash, bytes) and isinstance(signature, bytes): 72 for x, y in zip(request_hash, signature): 73 result |= x ^ y 74 else: 75 for x, y in zip(request_hash, signature): 76 result |= ord(x) ^ ord(y) 77 return result == 0 78 79 def bind_route(self, server): 80 @server.route(self.endpoint, methods=['GET', 'POST']) 81 def event(): 82 # If a GET request is made, return 404. 83 if request.method == 'GET': 84 return make_response("These are not the slackbots you're looking for.", 404) 85 86 # Each request comes with request timestamp and request signature 87 # emit an error if the timestamp is out of range 88 req_timestamp = request.headers.get('X-Slack-Request-Timestamp') 89 if abs(time() - int(req_timestamp)) > 60 * 5: 90 slack_exception = SlackEventAdapterException('Invalid request timestamp') 91 self.emitter.emit('error', slack_exception) 92 return make_response("", 403) 93 94 # Verify the request signature using the app's signing secret 95 # emit an error if the signature can't be verified 96 req_signature = request.headers.get('X-Slack-Signature') 97 if not self.verify_signature(req_timestamp, req_signature): 98 slack_exception = SlackEventAdapterException('Invalid request signature') 99 self.emitter.emit('error', slack_exception) 100 return make_response("", 403) 101 102 # Parse the request payload into JSON 103 event_data = json.loads(request.data.decode('utf-8')) 104 105 # Echo the URL verification challenge code back to Slack 106 if "challenge" in event_data: 107 return make_response( 108 event_data.get("challenge"), 200, {"content_type": "application/json"} 109 ) 110 111 # Parse the Event payload and emit the event to the event listener 112 if "event" in event_data: 113 event_type = event_data["event"]["type"] 114 self.emitter.emit(event_type, event_data) 115 response = make_response("", 200) 116 response.headers['X-Slack-Powered-By'] = self.package_info 117 return response 118 119 120 class SlackEventAdapterException(Exception): 121 """ 122 Base exception for all errors raised by the SlackClient library 123 """ 124 125 def __init__(self, msg=None): 126 if msg is None: 127 # default error message 128 msg = "An error occurred in the SlackEventsApiAdapter library" 129 super(SlackEventAdapterException, self).__init__(msg) 130 [end of slackeventsapi/server.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
slackapi/python-slack-events-api
0c0ce604b502508622fb14c278a0d64841fa32e3
Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s):
2020-06-12T06:58:10Z
<patch> <patch> diff --git a/example/current_app/main.py b/example/current_app/main.py new file mode 100644 --- /dev/null +++ b/example/current_app/main.py @@ -0,0 +1,49 @@ +# ------------------ +# Only for running this script here +import sys +from os.path import dirname +sys.path.insert(1, f"{dirname(__file__)}/../..") +# ------------------ + +import os +from slack import WebClient +import logging +logging.basicConfig(level=logging.DEBUG) + +from flask import Flask + +app = Flask(__name__) + +with app.app_context(): + from test_module.slack_app import slack_events_adapter + + slack_bot_token = os.environ["SLACK_BOT_TOKEN"] + slack_client = WebClient(slack_bot_token) + + + @slack_events_adapter.on("message") + def handle_message(event_data): + message = event_data["event"] + if message.get("subtype") is None and "hi" in message.get('text'): + channel = message["channel"] + message = "Hi <@%s>! :tada:" % message["user"] + slack_client.chat_postMessage(channel=channel, text=message) + + + @slack_events_adapter.on("error") + def error_handler(err): + print("ERROR: " + str(err)) + +# (Terminal A) +# source env/bin/activate +# (env) $ export SLACK_BOT_TOKEN=xoxb-*** +# (env) $ export SLACK_SIGNING_SECRET=** +# (env) $ cd example/current_app +# (env) $ FLASK_APP=main.py FLASK_ENV=development flask run --port 3000 + +# (Terminal B) +# ngrok http 3000 + +# in Slack +# /invite @{your app's bot user} +# post a message "hi" in the channel diff --git a/slackeventsapi/server.py b/slackeventsapi/server.py --- a/slackeventsapi/server.py +++ b/slackeventsapi/server.py @@ -1,10 +1,13 @@ -from flask import Flask, request, make_response, Blueprint +import hashlib +import hmac import json import platform import sys -import hmac -import hashlib from time import time + +from flask import Flask, request, make_response, Blueprint +from werkzeug.local import LocalProxy + from .version import __version__ @@ -18,10 +21,10 @@ def __init__(self, signing_secret, endpoint, emitter, server): # If a server is passed in, bind the event handler routes to it, # otherwise create a new Flask instance. if server: - if isinstance(server, Flask) or isinstance(server, Blueprint): + if isinstance(server, (Flask, Blueprint, LocalProxy)): self.bind_route(server) else: - raise TypeError("Server must be an instance of Flask or Blueprint") + raise TypeError("Server must be an instance of Flask, Blueprint, or LocalProxy") else: Flask.__init__(self, __name__) self.bind_route(self) </patch> </s> </patch>
diff --git a/example/current_app/test_module/__init__.py b/example/current_app/test_module/__init__.py new file mode 100644 diff --git a/example/current_app/test_module/slack_app.py b/example/current_app/test_module/slack_app.py new file mode 100644 --- /dev/null +++ b/example/current_app/test_module/slack_app.py @@ -0,0 +1,16 @@ +# ------------------ +# Only for running this script here +import logging +import sys +from os.path import dirname + +sys.path.insert(1, f"{dirname(__file__)}/../../..") +logging.basicConfig(level=logging.DEBUG) +# ------------------ + +from flask import current_app as app +from slackeventsapi import SlackEventAdapter +import os + +slack_signing_secret = os.environ["SLACK_SIGNING_SECRET"] +slack_events_adapter = SlackEventAdapter(slack_signing_secret, "/slack/events", app) diff --git a/tests/test_server.py b/tests/test_server.py --- a/tests/test_server.py +++ b/tests/test_server.py @@ -18,7 +18,7 @@ def test_server_not_flask(): with pytest.raises(TypeError) as e: invalid_flask = "I am not a Flask" SlackEventAdapter("SIGNING_SECRET", "/slack/events", invalid_flask) - assert e.value.args[0] == 'Server must be an instance of Flask or Blueprint' + assert e.value.args[0] == 'Server must be an instance of Flask, Blueprint, or LocalProxy' def test_blueprint_server():
1.0
celery__celery-2598
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ``` </issue> <code> [start of README.rst] 1 ================================= 2 celery - Distributed Task Queue 3 ================================= 4 5 .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png 6 7 |build-status| |coverage-status| 8 9 :Version: 3.2.0a1 (Cipater) 10 :Web: http://celeryproject.org/ 11 :Download: http://pypi.python.org/pypi/celery/ 12 :Source: http://github.com/celery/celery/ 13 :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, 14 python, webhooks, queue, distributed 15 16 -- 17 18 What is a Task Queue? 19 ===================== 20 21 Task queues are used as a mechanism to distribute work across threads or 22 machines. 23 24 A task queue's input is a unit of work, called a task, dedicated worker 25 processes then constantly monitor the queue for new work to perform. 26 27 Celery communicates via messages, usually using a broker 28 to mediate between clients and workers. To initiate a task a client puts a 29 message on the queue, the broker then delivers the message to a worker. 30 31 A Celery system can consist of multiple workers and brokers, giving way 32 to high availability and horizontal scaling. 33 34 Celery is a library written in Python, but the protocol can be implemented in 35 any language. So far there's RCelery_ for the Ruby programming language, and a 36 `PHP client`, but language interoperability can also be achieved 37 by using webhooks. 38 39 .. _RCelery: https://github.com/leapfrogonline/rcelery 40 .. _`PHP client`: https://github.com/gjedeer/celery-php 41 .. _`using webhooks`: 42 http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html 43 44 What do I need? 45 =============== 46 47 Celery version 3.0 runs on, 48 49 - Python (2.6, 2.7, 3.3, 3.4) 50 - PyPy (1.8, 1.9) 51 - Jython (2.5, 2.7). 52 53 This is the last version to support Python 2.5, 54 and from Celery 3.1, Python 2.6 or later is required. 55 The last version to support Python 2.4 was Celery series 2.2. 56 57 *Celery* is usually used with a message broker to send and receive messages. 58 The RabbitMQ, Redis transports are feature complete, 59 but there's also experimental support for a myriad of other solutions, including 60 using SQLite for local development. 61 62 *Celery* can run on a single machine, on multiple machines, or even 63 across datacenters. 64 65 Get Started 66 =========== 67 68 If this is the first time you're trying to use Celery, or you are 69 new to Celery 3.0 coming from previous versions then you should read our 70 getting started tutorials: 71 72 - `First steps with Celery`_ 73 74 Tutorial teaching you the bare minimum needed to get started with Celery. 75 76 - `Next steps`_ 77 78 A more complete overview, showing more features. 79 80 .. _`First steps with Celery`: 81 http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html 82 83 .. _`Next steps`: 84 http://docs.celeryproject.org/en/latest/getting-started/next-steps.html 85 86 Celery is... 87 ========== 88 89 - **Simple** 90 91 Celery is easy to use and maintain, and does *not need configuration files*. 92 93 It has an active, friendly community you can talk to for support, 94 including a `mailing-list`_ and and an IRC channel. 95 96 Here's one of the simplest applications you can make:: 97 98 from celery import Celery 99 100 app = Celery('hello', broker='amqp://guest@localhost//') 101 102 @app.task 103 def hello(): 104 return 'hello world' 105 106 - **Highly Available** 107 108 Workers and clients will automatically retry in the event 109 of connection loss or failure, and some brokers support 110 HA in way of *Master/Master* or *Master/Slave* replication. 111 112 - **Fast** 113 114 A single Celery process can process millions of tasks a minute, 115 with sub-millisecond round-trip latency (using RabbitMQ, 116 py-librabbitmq, and optimized settings). 117 118 - **Flexible** 119 120 Almost every part of *Celery* can be extended or used on its own, 121 Custom pool implementations, serializers, compression schemes, logging, 122 schedulers, consumers, producers, autoscalers, broker transports and much more. 123 124 It supports... 125 ============ 126 127 - **Message Transports** 128 129 - RabbitMQ_, Redis_, 130 - MongoDB_ (experimental), Amazon SQS (experimental), 131 - CouchDB_ (experimental), SQLAlchemy_ (experimental), 132 - Django ORM (experimental), `IronMQ`_ 133 - and more... 134 135 - **Concurrency** 136 137 - Prefork, Eventlet_, gevent_, threads/single threaded 138 139 - **Result Stores** 140 141 - AMQP, Redis 142 - memcached, MongoDB 143 - SQLAlchemy, Django ORM 144 - Apache Cassandra, IronCache 145 146 - **Serialization** 147 148 - *pickle*, *json*, *yaml*, *msgpack*. 149 - *zlib*, *bzip2* compression. 150 - Cryptographic message signing. 151 152 .. _`Eventlet`: http://eventlet.net/ 153 .. _`gevent`: http://gevent.org/ 154 155 .. _RabbitMQ: http://rabbitmq.com 156 .. _Redis: http://redis.io 157 .. _MongoDB: http://mongodb.org 158 .. _Beanstalk: http://kr.github.com/beanstalkd 159 .. _CouchDB: http://couchdb.apache.org 160 .. _SQLAlchemy: http://sqlalchemy.org 161 .. _`IronMQ`: http://iron.io 162 163 Framework Integration 164 ===================== 165 166 Celery is easy to integrate with web frameworks, some of which even have 167 integration packages: 168 169 +--------------------+------------------------+ 170 | `Django`_ | not needed | 171 +--------------------+------------------------+ 172 | `Pyramid`_ | `pyramid_celery`_ | 173 +--------------------+------------------------+ 174 | `Pylons`_ | `celery-pylons`_ | 175 +--------------------+------------------------+ 176 | `Flask`_ | not needed | 177 +--------------------+------------------------+ 178 | `web2py`_ | `web2py-celery`_ | 179 +--------------------+------------------------+ 180 | `Tornado`_ | `tornado-celery`_ | 181 +--------------------+------------------------+ 182 183 The integration packages are not strictly necessary, but they can make 184 development easier, and sometimes they add important hooks like closing 185 database connections at ``fork``. 186 187 .. _`Django`: http://djangoproject.com/ 188 .. _`Pylons`: http://www.pylonsproject.org/ 189 .. _`Flask`: http://flask.pocoo.org/ 190 .. _`web2py`: http://web2py.com/ 191 .. _`Bottle`: http://bottlepy.org/ 192 .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html 193 .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ 194 .. _`django-celery`: http://pypi.python.org/pypi/django-celery 195 .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons 196 .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ 197 .. _`Tornado`: http://www.tornadoweb.org/ 198 .. _`tornado-celery`: http://github.com/mher/tornado-celery/ 199 200 .. _celery-documentation: 201 202 Documentation 203 ============= 204 205 The `latest documentation`_ with user guides, tutorials and API reference 206 is hosted at Read The Docs. 207 208 .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ 209 210 .. _celery-installation: 211 212 Installation 213 ============ 214 215 You can install Celery either via the Python Package Index (PyPI) 216 or from source. 217 218 To install using `pip`,:: 219 220 $ pip install -U Celery 221 222 To install using `easy_install`,:: 223 224 $ easy_install -U Celery 225 226 .. _bundles: 227 228 Bundles 229 ------- 230 231 Celery also defines a group of bundles that can be used 232 to install Celery and the dependencies for a given feature. 233 234 You can specify these in your requirements or on the ``pip`` comand-line 235 by using brackets. Multiple bundles can be specified by separating them by 236 commas. 237 :: 238 239 $ pip install "celery[librabbitmq]" 240 241 $ pip install "celery[librabbitmq,redis,auth,msgpack]" 242 243 The following bundles are available: 244 245 Serializers 246 ~~~~~~~~~~~ 247 248 :celery[auth]: 249 for using the auth serializer. 250 251 :celery[msgpack]: 252 for using the msgpack serializer. 253 254 :celery[yaml]: 255 for using the yaml serializer. 256 257 Concurrency 258 ~~~~~~~~~~~ 259 260 :celery[eventlet]: 261 for using the eventlet pool. 262 263 :celery[gevent]: 264 for using the gevent pool. 265 266 :celery[threads]: 267 for using the thread pool. 268 269 Transports and Backends 270 ~~~~~~~~~~~~~~~~~~~~~~~ 271 272 :celery[librabbitmq]: 273 for using the librabbitmq C library. 274 275 :celery[redis]: 276 for using Redis as a message transport or as a result backend. 277 278 :celery[mongodb]: 279 for using MongoDB as a message transport (*experimental*), 280 or as a result backend (*supported*). 281 282 :celery[sqs]: 283 for using Amazon SQS as a message transport (*experimental*). 284 285 :celery[memcache]: 286 for using memcached as a result backend. 287 288 :celery[cassandra]: 289 for using Apache Cassandra as a result backend. 290 291 :celery[couchdb]: 292 for using CouchDB as a message transport (*experimental*). 293 294 :celery[couchbase]: 295 for using CouchBase as a result backend. 296 297 :celery[beanstalk]: 298 for using Beanstalk as a message transport (*experimental*). 299 300 :celery[zookeeper]: 301 for using Zookeeper as a message transport. 302 303 :celery[zeromq]: 304 for using ZeroMQ as a message transport (*experimental*). 305 306 :celery[sqlalchemy]: 307 for using SQLAlchemy as a message transport (*experimental*), 308 or as a result backend (*supported*). 309 310 :celery[pyro]: 311 for using the Pyro4 message transport (*experimental*). 312 313 :celery[slmq]: 314 for using the SoftLayer Message Queue transport (*experimental*). 315 316 .. _celery-installing-from-source: 317 318 Downloading and installing from source 319 -------------------------------------- 320 321 Download the latest version of Celery from 322 http://pypi.python.org/pypi/celery/ 323 324 You can install it by doing the following,:: 325 326 $ tar xvfz celery-0.0.0.tar.gz 327 $ cd celery-0.0.0 328 $ python setup.py build 329 # python setup.py install 330 331 The last command must be executed as a privileged user if 332 you are not currently using a virtualenv. 333 334 .. _celery-installing-from-git: 335 336 Using the development version 337 ----------------------------- 338 339 With pip 340 ~~~~~~~~ 341 342 The Celery development version also requires the development 343 versions of ``kombu``, ``amqp`` and ``billiard``. 344 345 You can install the latest snapshot of these using the following 346 pip commands:: 347 348 $ pip install https://github.com/celery/celery/zipball/master#egg=celery 349 $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard 350 $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp 351 $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu 352 353 With git 354 ~~~~~~~~ 355 356 Please the Contributing section. 357 358 .. _getting-help: 359 360 Getting Help 361 ============ 362 363 .. _mailing-list: 364 365 Mailing list 366 ------------ 367 368 For discussions about the usage, development, and future of celery, 369 please join the `celery-users`_ mailing list. 370 371 .. _`celery-users`: http://groups.google.com/group/celery-users/ 372 373 .. _irc-channel: 374 375 IRC 376 --- 377 378 Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ 379 network. 380 381 .. _`Freenode`: http://freenode.net 382 383 .. _bug-tracker: 384 385 Bug tracker 386 =========== 387 388 If you have any suggestions, bug reports or annoyances please report them 389 to our issue tracker at http://github.com/celery/celery/issues/ 390 391 .. _wiki: 392 393 Wiki 394 ==== 395 396 http://wiki.github.com/celery/celery/ 397 398 .. _contributing-short: 399 400 Contributing 401 ============ 402 403 Development of `celery` happens at Github: http://github.com/celery/celery 404 405 You are highly encouraged to participate in the development 406 of `celery`. If you don't like Github (for some reason) you're welcome 407 to send regular patches. 408 409 Be sure to also read the `Contributing to Celery`_ section in the 410 documentation. 411 412 .. _`Contributing to Celery`: 413 http://docs.celeryproject.org/en/master/contributing.html 414 415 .. _license: 416 417 License 418 ======= 419 420 This software is licensed under the `New BSD License`. See the ``LICENSE`` 421 file in the top distribution directory for the full license text. 422 423 .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround 424 425 426 .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png 427 :alt: Bitdeli badge 428 :target: https://bitdeli.com/free 429 430 .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master 431 :target: https://travis-ci.org/celery/celery 432 .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg 433 :target: https://coveralls.io/r/celery/celery 434 [end of README.rst] [start of celery/backends/amqp.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.backends.amqp 4 ~~~~~~~~~~~~~~~~~~~~ 5 6 The AMQP result backend. 7 8 This backend publishes results as messages. 9 10 """ 11 from __future__ import absolute_import 12 13 import socket 14 15 from collections import deque 16 from operator import itemgetter 17 18 from kombu import Exchange, Queue, Producer, Consumer 19 20 from celery import states 21 from celery.exceptions import TimeoutError 22 from celery.five import range, monotonic 23 from celery.utils.functional import dictfilter 24 from celery.utils.log import get_logger 25 from celery.utils.timeutils import maybe_s_to_ms 26 27 from .base import BaseBackend 28 29 __all__ = ['BacklogLimitExceeded', 'AMQPBackend'] 30 31 logger = get_logger(__name__) 32 33 34 class BacklogLimitExceeded(Exception): 35 """Too much state history to fast-forward.""" 36 37 38 def repair_uuid(s): 39 # Historically the dashes in UUIDS are removed from AMQ entity names, 40 # but there is no known reason to. Hopefully we'll be able to fix 41 # this in v4.0. 42 return '%s-%s-%s-%s-%s' % (s[:8], s[8:12], s[12:16], s[16:20], s[20:]) 43 44 45 class NoCacheQueue(Queue): 46 can_cache_declaration = False 47 48 49 class AMQPBackend(BaseBackend): 50 """Publishes results by sending messages.""" 51 Exchange = Exchange 52 Queue = NoCacheQueue 53 Consumer = Consumer 54 Producer = Producer 55 56 BacklogLimitExceeded = BacklogLimitExceeded 57 58 persistent = True 59 supports_autoexpire = True 60 supports_native_join = True 61 62 retry_policy = { 63 'max_retries': 20, 64 'interval_start': 0, 65 'interval_step': 1, 66 'interval_max': 1, 67 } 68 69 def __init__(self, app, connection=None, exchange=None, exchange_type=None, 70 persistent=None, serializer=None, auto_delete=True, **kwargs): 71 super(AMQPBackend, self).__init__(app, **kwargs) 72 conf = self.app.conf 73 self._connection = connection 74 self.persistent = self.prepare_persistent(persistent) 75 self.delivery_mode = 2 if self.persistent else 1 76 exchange = exchange or conf.CELERY_RESULT_EXCHANGE 77 exchange_type = exchange_type or conf.CELERY_RESULT_EXCHANGE_TYPE 78 self.exchange = self._create_exchange( 79 exchange, exchange_type, self.delivery_mode, 80 ) 81 self.serializer = serializer or conf.CELERY_RESULT_SERIALIZER 82 self.auto_delete = auto_delete 83 self.queue_arguments = dictfilter({ 84 'x-expires': maybe_s_to_ms(self.expires), 85 }) 86 87 def _create_exchange(self, name, type='direct', delivery_mode=2): 88 return self.Exchange(name=name, 89 type=type, 90 delivery_mode=delivery_mode, 91 durable=self.persistent, 92 auto_delete=False) 93 94 def _create_binding(self, task_id): 95 name = self.rkey(task_id) 96 return self.Queue(name=name, 97 exchange=self.exchange, 98 routing_key=name, 99 durable=self.persistent, 100 auto_delete=self.auto_delete, 101 queue_arguments=self.queue_arguments) 102 103 def revive(self, channel): 104 pass 105 106 def rkey(self, task_id): 107 return task_id.replace('-', '') 108 109 def destination_for(self, task_id, request): 110 if request: 111 return self.rkey(task_id), request.correlation_id or task_id 112 return self.rkey(task_id), task_id 113 114 def store_result(self, task_id, result, status, 115 traceback=None, request=None, **kwargs): 116 """Send task return value and status.""" 117 routing_key, correlation_id = self.destination_for(task_id, request) 118 if not routing_key: 119 return 120 with self.app.amqp.producer_pool.acquire(block=True) as producer: 121 producer.publish( 122 {'task_id': task_id, 'status': status, 123 'result': self.encode_result(result, status), 124 'traceback': traceback, 125 'children': self.current_task_children(request)}, 126 exchange=self.exchange, 127 routing_key=routing_key, 128 correlation_id=correlation_id, 129 serializer=self.serializer, 130 retry=True, retry_policy=self.retry_policy, 131 declare=self.on_reply_declare(task_id), 132 delivery_mode=self.delivery_mode, 133 ) 134 return result 135 136 def on_reply_declare(self, task_id): 137 return [self._create_binding(task_id)] 138 139 def wait_for(self, task_id, timeout=None, cache=True, 140 no_ack=True, on_interval=None, 141 READY_STATES=states.READY_STATES, 142 PROPAGATE_STATES=states.PROPAGATE_STATES, 143 **kwargs): 144 cached_meta = self._cache.get(task_id) 145 if cache and cached_meta and \ 146 cached_meta['status'] in READY_STATES: 147 return cached_meta 148 else: 149 try: 150 return self.consume(task_id, timeout=timeout, no_ack=no_ack, 151 on_interval=on_interval) 152 except socket.timeout: 153 raise TimeoutError('The operation timed out.') 154 155 def get_task_meta(self, task_id, backlog_limit=1000): 156 # Polling and using basic_get 157 with self.app.pool.acquire_channel(block=True) as (_, channel): 158 binding = self._create_binding(task_id)(channel) 159 binding.declare() 160 161 prev = latest = acc = None 162 for i in range(backlog_limit): # spool ffwd 163 acc = binding.get( 164 accept=self.accept, no_ack=False, 165 ) 166 if not acc: # no more messages 167 break 168 if acc.payload['task_id'] == task_id: 169 prev, latest = latest, acc 170 if prev: 171 # backends are not expected to keep history, 172 # so we delete everything except the most recent state. 173 prev.ack() 174 prev = None 175 else: 176 raise self.BacklogLimitExceeded(task_id) 177 178 if latest: 179 payload = self._cache[task_id] = latest.payload 180 latest.requeue() 181 return payload 182 else: 183 # no new state, use previous 184 try: 185 return self._cache[task_id] 186 except KeyError: 187 # result probably pending. 188 return {'status': states.PENDING, 'result': None} 189 poll = get_task_meta # XXX compat 190 191 def drain_events(self, connection, consumer, 192 timeout=None, on_interval=None, now=monotonic, wait=None): 193 wait = wait or connection.drain_events 194 results = {} 195 196 def callback(meta, message): 197 if meta['status'] in states.READY_STATES: 198 results[meta['task_id']] = meta 199 200 consumer.callbacks[:] = [callback] 201 time_start = now() 202 203 while 1: 204 # Total time spent may exceed a single call to wait() 205 if timeout and now() - time_start >= timeout: 206 raise socket.timeout() 207 try: 208 wait(timeout=1) 209 except socket.timeout: 210 pass 211 if on_interval: 212 on_interval() 213 if results: # got event on the wanted channel. 214 break 215 self._cache.update(results) 216 return results 217 218 def consume(self, task_id, timeout=None, no_ack=True, on_interval=None): 219 wait = self.drain_events 220 with self.app.pool.acquire_channel(block=True) as (conn, channel): 221 binding = self._create_binding(task_id) 222 with self.Consumer(channel, binding, 223 no_ack=no_ack, accept=self.accept) as consumer: 224 while 1: 225 try: 226 return wait( 227 conn, consumer, timeout, on_interval)[task_id] 228 except KeyError: 229 continue 230 231 def _many_bindings(self, ids): 232 return [self._create_binding(task_id) for task_id in ids] 233 234 def get_many(self, task_ids, timeout=None, no_ack=True, on_message=None, 235 now=monotonic, getfields=itemgetter('status', 'task_id'), 236 READY_STATES=states.READY_STATES, 237 PROPAGATE_STATES=states.PROPAGATE_STATES, **kwargs): 238 with self.app.pool.acquire_channel(block=True) as (conn, channel): 239 ids = set(task_ids) 240 cached_ids = set() 241 mark_cached = cached_ids.add 242 for task_id in ids: 243 try: 244 cached = self._cache[task_id] 245 except KeyError: 246 pass 247 else: 248 if cached['status'] in READY_STATES: 249 yield task_id, cached 250 mark_cached(task_id) 251 ids.difference_update(cached_ids) 252 results = deque() 253 push_result = results.append 254 push_cache = self._cache.__setitem__ 255 decode_result = self.meta_from_decoded 256 257 def _on_message(message): 258 body = decode_result(message.decode()) 259 if on_message is not None: 260 on_message(body) 261 state, uid = getfields(body) 262 if state in READY_STATES: 263 push_result(body) \ 264 if uid in task_ids else push_cache(uid, body) 265 266 bindings = self._many_bindings(task_ids) 267 with self.Consumer(channel, bindings, on_message=_on_message, 268 accept=self.accept, no_ack=no_ack): 269 wait = conn.drain_events 270 popleft = results.popleft 271 while ids: 272 wait(timeout=timeout) 273 while results: 274 state = popleft() 275 task_id = state['task_id'] 276 ids.discard(task_id) 277 push_cache(task_id, state) 278 yield task_id, state 279 280 def reload_task_result(self, task_id): 281 raise NotImplementedError( 282 'reload_task_result is not supported by this backend.') 283 284 def reload_group_result(self, task_id): 285 """Reload group result, even if it has been previously fetched.""" 286 raise NotImplementedError( 287 'reload_group_result is not supported by this backend.') 288 289 def save_group(self, group_id, result): 290 raise NotImplementedError( 291 'save_group is not supported by this backend.') 292 293 def restore_group(self, group_id, cache=True): 294 raise NotImplementedError( 295 'restore_group is not supported by this backend.') 296 297 def delete_group(self, group_id): 298 raise NotImplementedError( 299 'delete_group is not supported by this backend.') 300 301 def __reduce__(self, args=(), kwargs={}): 302 kwargs.update( 303 connection=self._connection, 304 exchange=self.exchange.name, 305 exchange_type=self.exchange.type, 306 persistent=self.persistent, 307 serializer=self.serializer, 308 auto_delete=self.auto_delete, 309 expires=self.expires, 310 ) 311 return super(AMQPBackend, self).__reduce__(args, kwargs) 312 [end of celery/backends/amqp.py] </code> I need you to solve this issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the following format. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch>
celery/celery
6592ff64b6b024a4b68abcc53b151888fdf0dee3
CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ```
This is biting me as well. Any news?
2015-04-29T14:52:17Z
<patch> <patch> diff --git a/celery/backends/amqp.py b/celery/backends/amqp.py --- a/celery/backends/amqp.py +++ b/celery/backends/amqp.py @@ -195,7 +195,7 @@ def drain_events(self, connection, consumer, def callback(meta, message): if meta['status'] in states.READY_STATES: - results[meta['task_id']] = meta + results[meta['task_id']] = self.meta_from_decoded(meta) consumer.callbacks[:] = [callback] time_start = now() </patch> </s> </patch>
diff --git a/celery/tests/backends/test_amqp.py b/celery/tests/backends/test_amqp.py --- a/celery/tests/backends/test_amqp.py +++ b/celery/tests/backends/test_amqp.py @@ -13,6 +13,7 @@ from celery.backends.amqp import AMQPBackend from celery.exceptions import TimeoutError from celery.five import Empty, Queue, range +from celery.result import AsyncResult from celery.utils import uuid from celery.tests.case import ( @@ -246,10 +247,20 @@ def test_wait_for(self): with self.assertRaises(TimeoutError): b.wait_for(tid, timeout=0.01, cache=False) - def test_drain_events_remaining_timeouts(self): + def test_drain_events_decodes_exceptions_in_meta(self): + tid = uuid() + b = self.create_backend(serializer="json") + b.store_result(tid, RuntimeError("aap"), states.FAILURE) + result = AsyncResult(tid, backend=b) - class Connection(object): + with self.assertRaises(Exception) as cm: + result.get() + self.assertEqual(cm.exception.__class__.__name__, "RuntimeError") + self.assertEqual(str(cm.exception), "aap") + + def test_drain_events_remaining_timeouts(self): + class Connection(object): def drain_events(self, timeout=None): pass
1.0
celery__celery-2840
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost. </issue> <code> [start of README.rst] 1 ================================= 2 celery - Distributed Task Queue 3 ================================= 4 5 .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png 6 7 |build-status| |coverage-status| 8 9 :Version: 3.2.0a1 (Cipater) 10 :Web: http://celeryproject.org/ 11 :Download: http://pypi.python.org/pypi/celery/ 12 :Source: http://github.com/celery/celery/ 13 :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, 14 python, webhooks, queue, distributed 15 16 -- 17 18 What is a Task Queue? 19 ===================== 20 21 Task queues are used as a mechanism to distribute work across threads or 22 machines. 23 24 A task queue's input is a unit of work, called a task, dedicated worker 25 processes then constantly monitor the queue for new work to perform. 26 27 Celery communicates via messages, usually using a broker 28 to mediate between clients and workers. To initiate a task a client puts a 29 message on the queue, the broker then delivers the message to a worker. 30 31 A Celery system can consist of multiple workers and brokers, giving way 32 to high availability and horizontal scaling. 33 34 Celery is a library written in Python, but the protocol can be implemented in 35 any language. So far there's RCelery_ for the Ruby programming language, and a 36 `PHP client`, but language interoperability can also be achieved 37 by using webhooks. 38 39 .. _RCelery: https://github.com/leapfrogonline/rcelery 40 .. _`PHP client`: https://github.com/gjedeer/celery-php 41 .. _`using webhooks`: 42 http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html 43 44 What do I need? 45 =============== 46 47 Celery version 3.0 runs on, 48 49 - Python (2.6, 2.7, 3.3, 3.4) 50 - PyPy (1.8, 1.9) 51 - Jython (2.5, 2.7). 52 53 This is the last version to support Python 2.5, 54 and from Celery 3.1, Python 2.6 or later is required. 55 The last version to support Python 2.4 was Celery series 2.2. 56 57 *Celery* is usually used with a message broker to send and receive messages. 58 The RabbitMQ, Redis transports are feature complete, 59 but there's also experimental support for a myriad of other solutions, including 60 using SQLite for local development. 61 62 *Celery* can run on a single machine, on multiple machines, or even 63 across datacenters. 64 65 Get Started 66 =========== 67 68 If this is the first time you're trying to use Celery, or you are 69 new to Celery 3.0 coming from previous versions then you should read our 70 getting started tutorials: 71 72 - `First steps with Celery`_ 73 74 Tutorial teaching you the bare minimum needed to get started with Celery. 75 76 - `Next steps`_ 77 78 A more complete overview, showing more features. 79 80 .. _`First steps with Celery`: 81 http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html 82 83 .. _`Next steps`: 84 http://docs.celeryproject.org/en/latest/getting-started/next-steps.html 85 86 Celery is... 87 ========== 88 89 - **Simple** 90 91 Celery is easy to use and maintain, and does *not need configuration files*. 92 93 It has an active, friendly community you can talk to for support, 94 including a `mailing-list`_ and and an IRC channel. 95 96 Here's one of the simplest applications you can make:: 97 98 from celery import Celery 99 100 app = Celery('hello', broker='amqp://guest@localhost//') 101 102 @app.task 103 def hello(): 104 return 'hello world' 105 106 - **Highly Available** 107 108 Workers and clients will automatically retry in the event 109 of connection loss or failure, and some brokers support 110 HA in way of *Master/Master* or *Master/Slave* replication. 111 112 - **Fast** 113 114 A single Celery process can process millions of tasks a minute, 115 with sub-millisecond round-trip latency (using RabbitMQ, 116 py-librabbitmq, and optimized settings). 117 118 - **Flexible** 119 120 Almost every part of *Celery* can be extended or used on its own, 121 Custom pool implementations, serializers, compression schemes, logging, 122 schedulers, consumers, producers, autoscalers, broker transports and much more. 123 124 It supports... 125 ============ 126 127 - **Message Transports** 128 129 - RabbitMQ_, Redis_, 130 - MongoDB_ (experimental), Amazon SQS (experimental), 131 - CouchDB_ (experimental), SQLAlchemy_ (experimental), 132 - Django ORM (experimental), `IronMQ`_ 133 - and more... 134 135 - **Concurrency** 136 137 - Prefork, Eventlet_, gevent_, threads/single threaded 138 139 - **Result Stores** 140 141 - AMQP, Redis 142 - memcached, MongoDB 143 - SQLAlchemy, Django ORM 144 - Apache Cassandra, IronCache 145 146 - **Serialization** 147 148 - *pickle*, *json*, *yaml*, *msgpack*. 149 - *zlib*, *bzip2* compression. 150 - Cryptographic message signing. 151 152 .. _`Eventlet`: http://eventlet.net/ 153 .. _`gevent`: http://gevent.org/ 154 155 .. _RabbitMQ: http://rabbitmq.com 156 .. _Redis: http://redis.io 157 .. _MongoDB: http://mongodb.org 158 .. _Beanstalk: http://kr.github.com/beanstalkd 159 .. _CouchDB: http://couchdb.apache.org 160 .. _SQLAlchemy: http://sqlalchemy.org 161 .. _`IronMQ`: http://iron.io 162 163 Framework Integration 164 ===================== 165 166 Celery is easy to integrate with web frameworks, some of which even have 167 integration packages: 168 169 +--------------------+----------------------------------------------------+ 170 | `Django`_ | not needed | 171 +--------------------+----------------------------------------------------+ 172 | `Pyramid`_ | `pyramid_celery`_ | 173 +--------------------+----------------------------------------------------+ 174 | `Pylons`_ | `celery-pylons`_ | 175 +--------------------+----------------------------------------------------+ 176 | `Flask`_ | not needed | 177 +--------------------+----------------------------------------------------+ 178 | `web2py`_ | `web2py-celery`_ | 179 +--------------------+----------------------------------------------------+ 180 | `Tornado`_ | `tornado-celery`_ | `another tornado-celery`_ | 181 +--------------------+----------------------------------------------------+ 182 183 The integration packages are not strictly necessary, but they can make 184 development easier, and sometimes they add important hooks like closing 185 database connections at ``fork``. 186 187 .. _`Django`: http://djangoproject.com/ 188 .. _`Pylons`: http://www.pylonsproject.org/ 189 .. _`Flask`: http://flask.pocoo.org/ 190 .. _`web2py`: http://web2py.com/ 191 .. _`Bottle`: http://bottlepy.org/ 192 .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html 193 .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ 194 .. _`django-celery`: http://pypi.python.org/pypi/django-celery 195 .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons 196 .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ 197 .. _`Tornado`: http://www.tornadoweb.org/ 198 .. _`tornado-celery`: http://github.com/mher/tornado-celery/ 199 .. _`another tornado-celery`: https://github.com/mayflaver/tornado-celery 200 201 .. _celery-documentation: 202 203 Documentation 204 ============= 205 206 The `latest documentation`_ with user guides, tutorials and API reference 207 is hosted at Read The Docs. 208 209 .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ 210 211 .. _celery-installation: 212 213 Installation 214 ============ 215 216 You can install Celery either via the Python Package Index (PyPI) 217 or from source. 218 219 To install using `pip`,:: 220 221 $ pip install -U Celery 222 223 To install using `easy_install`,:: 224 225 $ easy_install -U Celery 226 227 .. _bundles: 228 229 Bundles 230 ------- 231 232 Celery also defines a group of bundles that can be used 233 to install Celery and the dependencies for a given feature. 234 235 You can specify these in your requirements or on the ``pip`` comand-line 236 by using brackets. Multiple bundles can be specified by separating them by 237 commas. 238 :: 239 240 $ pip install "celery[librabbitmq]" 241 242 $ pip install "celery[librabbitmq,redis,auth,msgpack]" 243 244 The following bundles are available: 245 246 Serializers 247 ~~~~~~~~~~~ 248 249 :celery[auth]: 250 for using the auth serializer. 251 252 :celery[msgpack]: 253 for using the msgpack serializer. 254 255 :celery[yaml]: 256 for using the yaml serializer. 257 258 Concurrency 259 ~~~~~~~~~~~ 260 261 :celery[eventlet]: 262 for using the eventlet pool. 263 264 :celery[gevent]: 265 for using the gevent pool. 266 267 :celery[threads]: 268 for using the thread pool. 269 270 Transports and Backends 271 ~~~~~~~~~~~~~~~~~~~~~~~ 272 273 :celery[librabbitmq]: 274 for using the librabbitmq C library. 275 276 :celery[redis]: 277 for using Redis as a message transport or as a result backend. 278 279 :celery[mongodb]: 280 for using MongoDB as a message transport (*experimental*), 281 or as a result backend (*supported*). 282 283 :celery[sqs]: 284 for using Amazon SQS as a message transport (*experimental*). 285 286 :celery[memcache]: 287 for using memcached as a result backend. 288 289 :celery[cassandra]: 290 for using Apache Cassandra as a result backend. 291 292 :celery[couchdb]: 293 for using CouchDB as a message transport (*experimental*). 294 295 :celery[couchbase]: 296 for using CouchBase as a result backend. 297 298 :celery[beanstalk]: 299 for using Beanstalk as a message transport (*experimental*). 300 301 :celery[zookeeper]: 302 for using Zookeeper as a message transport. 303 304 :celery[zeromq]: 305 for using ZeroMQ as a message transport (*experimental*). 306 307 :celery[sqlalchemy]: 308 for using SQLAlchemy as a message transport (*experimental*), 309 or as a result backend (*supported*). 310 311 :celery[pyro]: 312 for using the Pyro4 message transport (*experimental*). 313 314 :celery[slmq]: 315 for using the SoftLayer Message Queue transport (*experimental*). 316 317 .. _celery-installing-from-source: 318 319 Downloading and installing from source 320 -------------------------------------- 321 322 Download the latest version of Celery from 323 http://pypi.python.org/pypi/celery/ 324 325 You can install it by doing the following,:: 326 327 $ tar xvfz celery-0.0.0.tar.gz 328 $ cd celery-0.0.0 329 $ python setup.py build 330 # python setup.py install 331 332 The last command must be executed as a privileged user if 333 you are not currently using a virtualenv. 334 335 .. _celery-installing-from-git: 336 337 Using the development version 338 ----------------------------- 339 340 With pip 341 ~~~~~~~~ 342 343 The Celery development version also requires the development 344 versions of ``kombu``, ``amqp`` and ``billiard``. 345 346 You can install the latest snapshot of these using the following 347 pip commands:: 348 349 $ pip install https://github.com/celery/celery/zipball/master#egg=celery 350 $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard 351 $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp 352 $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu 353 354 With git 355 ~~~~~~~~ 356 357 Please the Contributing section. 358 359 .. _getting-help: 360 361 Getting Help 362 ============ 363 364 .. _mailing-list: 365 366 Mailing list 367 ------------ 368 369 For discussions about the usage, development, and future of celery, 370 please join the `celery-users`_ mailing list. 371 372 .. _`celery-users`: http://groups.google.com/group/celery-users/ 373 374 .. _irc-channel: 375 376 IRC 377 --- 378 379 Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ 380 network. 381 382 .. _`Freenode`: http://freenode.net 383 384 .. _bug-tracker: 385 386 Bug tracker 387 =========== 388 389 If you have any suggestions, bug reports or annoyances please report them 390 to our issue tracker at http://github.com/celery/celery/issues/ 391 392 .. _wiki: 393 394 Wiki 395 ==== 396 397 http://wiki.github.com/celery/celery/ 398 399 400 .. _maintainers: 401 402 Maintainers 403 =========== 404 405 - `@ask`_ (primary maintainer) 406 - `@thedrow`_ 407 - `@chrisgogreen`_ 408 - `@PMickael`_ 409 - `@malinoff`_ 410 - And you? We really need more: https://github.com/celery/celery/issues/2534 411 412 .. _`@ask`: http://github.com/ask 413 .. _`@thedrow`: http://github.com/thedrow 414 .. _`@chrisgogreen`: http://github.com/chrisgogreen 415 .. _`@PMickael`: http://github.com/PMickael 416 .. _`@malinoff`: http://github.com/malinoff 417 418 419 .. _contributing-short: 420 421 Contributing 422 ============ 423 424 Development of `celery` happens at Github: http://github.com/celery/celery 425 426 You are highly encouraged to participate in the development 427 of `celery`. If you don't like Github (for some reason) you're welcome 428 to send regular patches. 429 430 Be sure to also read the `Contributing to Celery`_ section in the 431 documentation. 432 433 .. _`Contributing to Celery`: 434 http://docs.celeryproject.org/en/master/contributing.html 435 436 .. _license: 437 438 License 439 ======= 440 441 This software is licensed under the `New BSD License`. See the ``LICENSE`` 442 file in the top distribution directory for the full license text. 443 444 .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround 445 446 447 .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png 448 :alt: Bitdeli badge 449 :target: https://bitdeli.com/free 450 451 .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master 452 :target: https://travis-ci.org/celery/celery 453 .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg 454 :target: https://coveralls.io/r/celery/celery 455 [end of README.rst] [start of celery/app/defaults.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.app.defaults 4 ~~~~~~~~~~~~~~~~~~~ 5 6 Configuration introspection and defaults. 7 8 """ 9 from __future__ import absolute_import 10 11 import sys 12 13 from collections import deque, namedtuple 14 from datetime import timedelta 15 16 from celery.five import items 17 from celery.utils import strtobool 18 from celery.utils.functional import memoize 19 20 __all__ = ['Option', 'NAMESPACES', 'flatten', 'find'] 21 22 is_jython = sys.platform.startswith('java') 23 is_pypy = hasattr(sys, 'pypy_version_info') 24 25 DEFAULT_POOL = 'prefork' 26 if is_jython: 27 DEFAULT_POOL = 'threads' 28 elif is_pypy: 29 if sys.pypy_version_info[0:3] < (1, 5, 0): 30 DEFAULT_POOL = 'solo' 31 else: 32 DEFAULT_POOL = 'prefork' 33 34 DEFAULT_ACCEPT_CONTENT = ['json', 'pickle', 'msgpack', 'yaml'] 35 DEFAULT_PROCESS_LOG_FMT = """ 36 [%(asctime)s: %(levelname)s/%(processName)s] %(message)s 37 """.strip() 38 DEFAULT_LOG_FMT = '[%(asctime)s: %(levelname)s] %(message)s' 39 DEFAULT_TASK_LOG_FMT = """[%(asctime)s: %(levelname)s/%(processName)s] \ 40 %(task_name)s[%(task_id)s]: %(message)s""" 41 42 _BROKER_OLD = {'deprecate_by': '2.5', 'remove_by': '4.0', 43 'alt': 'BROKER_URL setting'} 44 _REDIS_OLD = {'deprecate_by': '2.5', 'remove_by': '4.0', 45 'alt': 'URL form of CELERY_RESULT_BACKEND'} 46 47 searchresult = namedtuple('searchresult', ('namespace', 'key', 'type')) 48 49 50 class Option(object): 51 alt = None 52 deprecate_by = None 53 remove_by = None 54 typemap = dict(string=str, int=int, float=float, any=lambda v: v, 55 bool=strtobool, dict=dict, tuple=tuple) 56 57 def __init__(self, default=None, *args, **kwargs): 58 self.default = default 59 self.type = kwargs.get('type') or 'string' 60 for attr, value in items(kwargs): 61 setattr(self, attr, value) 62 63 def to_python(self, value): 64 return self.typemap[self.type](value) 65 66 def __repr__(self): 67 return '<Option: type->{0} default->{1!r}>'.format(self.type, 68 self.default) 69 70 NAMESPACES = { 71 'BROKER': { 72 'URL': Option(None, type='string'), 73 'CONNECTION_TIMEOUT': Option(4, type='float'), 74 'CONNECTION_RETRY': Option(True, type='bool'), 75 'CONNECTION_MAX_RETRIES': Option(100, type='int'), 76 'FAILOVER_STRATEGY': Option(None, type='string'), 77 'HEARTBEAT': Option(None, type='int'), 78 'HEARTBEAT_CHECKRATE': Option(3.0, type='int'), 79 'LOGIN_METHOD': Option(None, type='string'), 80 'POOL_LIMIT': Option(10, type='int'), 81 'USE_SSL': Option(False, type='bool'), 82 'TRANSPORT': Option(type='string'), 83 'TRANSPORT_OPTIONS': Option({}, type='dict'), 84 'HOST': Option(type='string', **_BROKER_OLD), 85 'PORT': Option(type='int', **_BROKER_OLD), 86 'USER': Option(type='string', **_BROKER_OLD), 87 'PASSWORD': Option(type='string', **_BROKER_OLD), 88 'VHOST': Option(type='string', **_BROKER_OLD), 89 }, 90 'CASSANDRA': { 91 'COLUMN_FAMILY': Option(type='string'), 92 'DETAILED_MODE': Option(False, type='bool'), 93 'KEYSPACE': Option(type='string'), 94 'READ_CONSISTENCY': Option(type='string'), 95 'SERVERS': Option(type='list'), 96 'WRITE_CONSISTENCY': Option(type='string'), 97 }, 98 'CELERY': { 99 'ACCEPT_CONTENT': Option(DEFAULT_ACCEPT_CONTENT, type='list'), 100 'ACKS_LATE': Option(False, type='bool'), 101 'ALWAYS_EAGER': Option(False, type='bool'), 102 'ANNOTATIONS': Option(type='any'), 103 'BROADCAST_QUEUE': Option('celeryctl'), 104 'BROADCAST_EXCHANGE': Option('celeryctl'), 105 'BROADCAST_EXCHANGE_TYPE': Option('fanout'), 106 'CACHE_BACKEND': Option(), 107 'CACHE_BACKEND_OPTIONS': Option({}, type='dict'), 108 'CHORD_PROPAGATES': Option(True, type='bool'), 109 'COUCHBASE_BACKEND_SETTINGS': Option(None, type='dict'), 110 'CREATE_MISSING_QUEUES': Option(True, type='bool'), 111 'DEFAULT_RATE_LIMIT': Option(type='string'), 112 'DISABLE_RATE_LIMITS': Option(False, type='bool'), 113 'DEFAULT_ROUTING_KEY': Option('celery'), 114 'DEFAULT_QUEUE': Option('celery'), 115 'DEFAULT_EXCHANGE': Option('celery'), 116 'DEFAULT_EXCHANGE_TYPE': Option('direct'), 117 'DEFAULT_DELIVERY_MODE': Option(2, type='string'), 118 'EAGER_PROPAGATES_EXCEPTIONS': Option(False, type='bool'), 119 'ENABLE_UTC': Option(True, type='bool'), 120 'ENABLE_REMOTE_CONTROL': Option(True, type='bool'), 121 'EVENT_SERIALIZER': Option('json'), 122 'EVENT_QUEUE_EXPIRES': Option(60.0, type='float'), 123 'EVENT_QUEUE_TTL': Option(5.0, type='float'), 124 'IMPORTS': Option((), type='tuple'), 125 'INCLUDE': Option((), type='tuple'), 126 'IGNORE_RESULT': Option(False, type='bool'), 127 'MAX_CACHED_RESULTS': Option(100, type='int'), 128 'MESSAGE_COMPRESSION': Option(type='string'), 129 'MONGODB_BACKEND_SETTINGS': Option(type='dict'), 130 'REDIS_HOST': Option(type='string', **_REDIS_OLD), 131 'REDIS_PORT': Option(type='int', **_REDIS_OLD), 132 'REDIS_DB': Option(type='int', **_REDIS_OLD), 133 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 134 'REDIS_MAX_CONNECTIONS': Option(type='int'), 135 'RESULT_BACKEND': Option(type='string'), 136 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 137 'RESULT_DB_TABLENAMES': Option(type='dict'), 138 'RESULT_DBURI': Option(), 139 'RESULT_ENGINE_OPTIONS': Option(type='dict'), 140 'RESULT_EXCHANGE': Option('celeryresults'), 141 'RESULT_EXCHANGE_TYPE': Option('direct'), 142 'RESULT_SERIALIZER': Option('json'), 143 'RESULT_PERSISTENT': Option(None, type='bool'), 144 'RIAK_BACKEND_SETTINGS': Option(type='dict'), 145 'ROUTES': Option(type='any'), 146 'SEND_EVENTS': Option(False, type='bool'), 147 'SEND_TASK_ERROR_EMAILS': Option(False, type='bool'), 148 'SEND_TASK_SENT_EVENT': Option(False, type='bool'), 149 'STORE_ERRORS_EVEN_IF_IGNORED': Option(False, type='bool'), 150 'TASK_PROTOCOL': Option(1, type='int'), 151 'TASK_PUBLISH_RETRY': Option(True, type='bool'), 152 'TASK_PUBLISH_RETRY_POLICY': Option({ 153 'max_retries': 3, 154 'interval_start': 0, 155 'interval_max': 1, 156 'interval_step': 0.2}, type='dict'), 157 'TASK_RESULT_EXPIRES': Option(timedelta(days=1), type='float'), 158 'TASK_SERIALIZER': Option('json'), 159 'TIMEZONE': Option(type='string'), 160 'TRACK_STARTED': Option(False, type='bool'), 161 'REDIRECT_STDOUTS': Option(True, type='bool'), 162 'REDIRECT_STDOUTS_LEVEL': Option('WARNING'), 163 'QUEUES': Option(type='dict'), 164 'QUEUE_HA_POLICY': Option(None, type='string'), 165 'SECURITY_KEY': Option(type='string'), 166 'SECURITY_CERTIFICATE': Option(type='string'), 167 'SECURITY_CERT_STORE': Option(type='string'), 168 'WORKER_DIRECT': Option(False, type='bool'), 169 }, 170 'CELERYD': { 171 'AGENT': Option(None, type='string'), 172 'AUTOSCALER': Option('celery.worker.autoscale:Autoscaler'), 173 'AUTORELOADER': Option('celery.worker.autoreload:Autoreloader'), 174 'CONCURRENCY': Option(0, type='int'), 175 'TIMER': Option(type='string'), 176 'TIMER_PRECISION': Option(1.0, type='float'), 177 'FORCE_EXECV': Option(False, type='bool'), 178 'HIJACK_ROOT_LOGGER': Option(True, type='bool'), 179 'CONSUMER': Option('celery.worker.consumer:Consumer', type='string'), 180 'LOG_FORMAT': Option(DEFAULT_PROCESS_LOG_FMT), 181 'LOG_COLOR': Option(type='bool'), 182 'LOG_LEVEL': Option('WARN', deprecate_by='2.4', remove_by='4.0', 183 alt='--loglevel argument'), 184 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', 185 alt='--logfile argument'), 186 'MAX_TASKS_PER_CHILD': Option(type='int'), 187 'POOL': Option(DEFAULT_POOL), 188 'POOL_PUTLOCKS': Option(True, type='bool'), 189 'POOL_RESTARTS': Option(False, type='bool'), 190 'PREFETCH_MULTIPLIER': Option(4, type='int'), 191 'STATE_DB': Option(), 192 'TASK_LOG_FORMAT': Option(DEFAULT_TASK_LOG_FMT), 193 'TASK_SOFT_TIME_LIMIT': Option(type='float'), 194 'TASK_TIME_LIMIT': Option(type='float'), 195 'WORKER_LOST_WAIT': Option(10.0, type='float') 196 }, 197 'CELERYBEAT': { 198 'SCHEDULE': Option({}, type='dict'), 199 'SCHEDULER': Option('celery.beat:PersistentScheduler'), 200 'SCHEDULE_FILENAME': Option('celerybeat-schedule'), 201 'SYNC_EVERY': Option(0, type='int'), 202 'MAX_LOOP_INTERVAL': Option(0, type='float'), 203 'LOG_LEVEL': Option('INFO', deprecate_by='2.4', remove_by='4.0', 204 alt='--loglevel argument'), 205 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', 206 alt='--logfile argument'), 207 }, 208 'CELERYMON': { 209 'LOG_LEVEL': Option('INFO', deprecate_by='2.4', remove_by='4.0', 210 alt='--loglevel argument'), 211 'LOG_FILE': Option(deprecate_by='2.4', remove_by='4.0', 212 alt='--logfile argument'), 213 'LOG_FORMAT': Option(DEFAULT_LOG_FMT), 214 }, 215 'EMAIL': { 216 'HOST': Option('localhost'), 217 'PORT': Option(25, type='int'), 218 'HOST_USER': Option(), 219 'HOST_PASSWORD': Option(), 220 'TIMEOUT': Option(2, type='float'), 221 'USE_SSL': Option(False, type='bool'), 222 'USE_TLS': Option(False, type='bool'), 223 'CHARSET': Option('us-ascii'), 224 }, 225 'SERVER_EMAIL': Option('celery@localhost'), 226 'ADMINS': Option((), type='tuple'), 227 } 228 229 230 def flatten(d, ns=''): 231 stack = deque([(ns, d)]) 232 while stack: 233 name, space = stack.popleft() 234 for key, value in items(space): 235 if isinstance(value, dict): 236 stack.append((name + key + '_', value)) 237 else: 238 yield name + key, value 239 DEFAULTS = {key: value.default for key, value in flatten(NAMESPACES)} 240 241 242 def find_deprecated_settings(source): 243 from celery.utils import warn_deprecated 244 for name, opt in flatten(NAMESPACES): 245 if (opt.deprecate_by or opt.remove_by) and getattr(source, name, None): 246 warn_deprecated(description='The {0!r} setting'.format(name), 247 deprecation=opt.deprecate_by, 248 removal=opt.remove_by, 249 alternative='Use the {0.alt} instead'.format(opt)) 250 return source 251 252 253 @memoize(maxsize=None) 254 def find(name, namespace='celery'): 255 # - Try specified namespace first. 256 namespace = namespace.upper() 257 try: 258 return searchresult( 259 namespace, name.upper(), NAMESPACES[namespace][name.upper()], 260 ) 261 except KeyError: 262 # - Try all the other namespaces. 263 for ns, keys in items(NAMESPACES): 264 if ns.upper() == name.upper(): 265 return searchresult(None, ns, keys) 266 elif isinstance(keys, dict): 267 try: 268 return searchresult(ns, name.upper(), keys[name.upper()]) 269 except KeyError: 270 pass 271 # - See if name is a qualname last. 272 return searchresult(None, name.upper(), DEFAULTS[name.upper()]) 273 [end of celery/app/defaults.py] [start of celery/app/task.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.app.task 4 ~~~~~~~~~~~~~~~ 5 6 Task Implementation: Task request context, and the base task class. 7 8 """ 9 from __future__ import absolute_import 10 11 import sys 12 13 from billiard.einfo import ExceptionInfo 14 15 from celery import current_app, group 16 from celery import states 17 from celery._state import _task_stack 18 from celery.canvas import signature 19 from celery.exceptions import Ignore, MaxRetriesExceededError, Reject, Retry 20 from celery.five import class_property, items 21 from celery.result import EagerResult 22 from celery.utils import abstract 23 from celery.utils import uuid, maybe_reraise 24 from celery.utils.functional import mattrgetter, maybe_list 25 from celery.utils.imports import instantiate 26 from celery.utils.mail import ErrorMail 27 28 from .annotations import resolve_all as resolve_all_annotations 29 from .registry import _unpickle_task_v2 30 from .utils import appstr 31 32 __all__ = ['Context', 'Task'] 33 34 #: extracts attributes related to publishing a message from an object. 35 extract_exec_options = mattrgetter( 36 'queue', 'routing_key', 'exchange', 'priority', 'expires', 37 'serializer', 'delivery_mode', 'compression', 'time_limit', 38 'soft_time_limit', 'immediate', 'mandatory', # imm+man is deprecated 39 ) 40 41 # We take __repr__ very seriously around here ;) 42 R_BOUND_TASK = '<class {0.__name__} of {app}{flags}>' 43 R_UNBOUND_TASK = '<unbound {0.__name__}{flags}>' 44 R_SELF_TASK = '<@task {0.name} bound to other {0.__self__}>' 45 R_INSTANCE = '<@task: {0.name} of {app}{flags}>' 46 47 #: Here for backwards compatibility as tasks no longer use a custom metaclass. 48 TaskType = type 49 50 51 def _strflags(flags, default=''): 52 if flags: 53 return ' ({0})'.format(', '.join(flags)) 54 return default 55 56 57 def _reprtask(task, fmt=None, flags=None): 58 flags = list(flags) if flags is not None else [] 59 flags.append('v2 compatible') if task.__v2_compat__ else None 60 if not fmt: 61 fmt = R_BOUND_TASK if task._app else R_UNBOUND_TASK 62 return fmt.format( 63 task, flags=_strflags(flags), 64 app=appstr(task._app) if task._app else None, 65 ) 66 67 68 class Context(object): 69 # Default context 70 logfile = None 71 loglevel = None 72 hostname = None 73 id = None 74 args = None 75 kwargs = None 76 retries = 0 77 eta = None 78 expires = None 79 is_eager = False 80 headers = None 81 delivery_info = None 82 reply_to = None 83 root_id = None 84 parent_id = None 85 correlation_id = None 86 taskset = None # compat alias to group 87 group = None 88 chord = None 89 utc = None 90 called_directly = True 91 callbacks = None 92 errbacks = None 93 timelimit = None 94 _children = None # see property 95 _protected = 0 96 97 def __init__(self, *args, **kwargs): 98 self.update(*args, **kwargs) 99 100 def update(self, *args, **kwargs): 101 return self.__dict__.update(*args, **kwargs) 102 103 def clear(self): 104 return self.__dict__.clear() 105 106 def get(self, key, default=None): 107 return getattr(self, key, default) 108 109 def __repr__(self): 110 return '<Context: {0!r}>'.format(vars(self)) 111 112 @property 113 def children(self): 114 # children must be an empy list for every thread 115 if self._children is None: 116 self._children = [] 117 return self._children 118 119 120 class Task(object): 121 """Task base class. 122 123 When called tasks apply the :meth:`run` method. This method must 124 be defined by all tasks (that is unless the :meth:`__call__` method 125 is overridden). 126 127 """ 128 __trace__ = None 129 __v2_compat__ = False # set by old base in celery.task.base 130 131 ErrorMail = ErrorMail 132 MaxRetriesExceededError = MaxRetriesExceededError 133 134 #: Execution strategy used, or the qualified name of one. 135 Strategy = 'celery.worker.strategy:default' 136 137 #: This is the instance bound to if the task is a method of a class. 138 __self__ = None 139 140 #: The application instance associated with this task class. 141 _app = None 142 143 #: Name of the task. 144 name = None 145 146 #: If :const:`True` the task is an abstract base class. 147 abstract = True 148 149 #: Maximum number of retries before giving up. If set to :const:`None`, 150 #: it will **never** stop retrying. 151 max_retries = 3 152 153 #: Default time in seconds before a retry of the task should be 154 #: executed. 3 minutes by default. 155 default_retry_delay = 3 * 60 156 157 #: Rate limit for this task type. Examples: :const:`None` (no rate 158 #: limit), `'100/s'` (hundred tasks a second), `'100/m'` (hundred tasks 159 #: a minute),`'100/h'` (hundred tasks an hour) 160 rate_limit = None 161 162 #: If enabled the worker will not store task state and return values 163 #: for this task. Defaults to the :setting:`CELERY_IGNORE_RESULT` 164 #: setting. 165 ignore_result = None 166 167 #: If enabled the request will keep track of subtasks started by 168 #: this task, and this information will be sent with the result 169 #: (``result.children``). 170 trail = True 171 172 #: When enabled errors will be stored even if the task is otherwise 173 #: configured to ignore results. 174 store_errors_even_if_ignored = None 175 176 #: If enabled an email will be sent to :setting:`ADMINS` whenever a task 177 #: of this type fails. 178 send_error_emails = None 179 180 #: The name of a serializer that are registered with 181 #: :mod:`kombu.serialization.registry`. Default is `'pickle'`. 182 serializer = None 183 184 #: Hard time limit. 185 #: Defaults to the :setting:`CELERYD_TASK_TIME_LIMIT` setting. 186 time_limit = None 187 188 #: Soft time limit. 189 #: Defaults to the :setting:`CELERYD_TASK_SOFT_TIME_LIMIT` setting. 190 soft_time_limit = None 191 192 #: The result store backend used for this task. 193 backend = None 194 195 #: If disabled this task won't be registered automatically. 196 autoregister = True 197 198 #: If enabled the task will report its status as 'started' when the task 199 #: is executed by a worker. Disabled by default as the normal behaviour 200 #: is to not report that level of granularity. Tasks are either pending, 201 #: finished, or waiting to be retried. 202 #: 203 #: Having a 'started' status can be useful for when there are long 204 #: running tasks and there is a need to report which task is currently 205 #: running. 206 #: 207 #: The application default can be overridden using the 208 #: :setting:`CELERY_TRACK_STARTED` setting. 209 track_started = None 210 211 #: When enabled messages for this task will be acknowledged **after** 212 #: the task has been executed, and not *just before* which is the 213 #: default behavior. 214 #: 215 #: Please note that this means the task may be executed twice if the 216 #: worker crashes mid execution (which may be acceptable for some 217 #: applications). 218 #: 219 #: The application default can be overridden with the 220 #: :setting:`CELERY_ACKS_LATE` setting. 221 acks_late = None 222 223 #: Tuple of expected exceptions. 224 #: 225 #: These are errors that are expected in normal operation 226 #: and that should not be regarded as a real error by the worker. 227 #: Currently this means that the state will be updated to an error 228 #: state, but the worker will not log the event as an error. 229 throws = () 230 231 #: Default task expiry time. 232 expires = None 233 234 #: Task request stack, the current request will be the topmost. 235 request_stack = None 236 237 #: Some may expect a request to exist even if the task has not been 238 #: called. This should probably be deprecated. 239 _default_request = None 240 241 _exec_options = None 242 243 __bound__ = False 244 245 from_config = ( 246 ('send_error_emails', 'CELERY_SEND_TASK_ERROR_EMAILS'), 247 ('serializer', 'CELERY_TASK_SERIALIZER'), 248 ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), 249 ('track_started', 'CELERY_TRACK_STARTED'), 250 ('acks_late', 'CELERY_ACKS_LATE'), 251 ('ignore_result', 'CELERY_IGNORE_RESULT'), 252 ('store_errors_even_if_ignored', 253 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), 254 ) 255 256 #: ignored 257 accept_magic_kwargs = False 258 259 _backend = None # set by backend property. 260 261 __bound__ = False 262 263 # - Tasks are lazily bound, so that configuration is not set 264 # - until the task is actually used 265 266 @classmethod 267 def bind(self, app): 268 was_bound, self.__bound__ = self.__bound__, True 269 self._app = app 270 conf = app.conf 271 self._exec_options = None # clear option cache 272 273 for attr_name, config_name in self.from_config: 274 if getattr(self, attr_name, None) is None: 275 setattr(self, attr_name, conf[config_name]) 276 277 # decorate with annotations from config. 278 if not was_bound: 279 self.annotate() 280 281 from celery.utils.threads import LocalStack 282 self.request_stack = LocalStack() 283 284 # PeriodicTask uses this to add itself to the PeriodicTask schedule. 285 self.on_bound(app) 286 287 return app 288 289 @classmethod 290 def on_bound(self, app): 291 """This method can be defined to do additional actions when the 292 task class is bound to an app.""" 293 pass 294 295 @classmethod 296 def _get_app(self): 297 if self._app is None: 298 self._app = current_app 299 if not self.__bound__: 300 # The app property's __set__ method is not called 301 # if Task.app is set (on the class), so must bind on use. 302 self.bind(self._app) 303 return self._app 304 app = class_property(_get_app, bind) 305 306 @classmethod 307 def annotate(self): 308 for d in resolve_all_annotations(self.app.annotations, self): 309 for key, value in items(d): 310 if key.startswith('@'): 311 self.add_around(key[1:], value) 312 else: 313 setattr(self, key, value) 314 315 @classmethod 316 def add_around(self, attr, around): 317 orig = getattr(self, attr) 318 if getattr(orig, '__wrapped__', None): 319 orig = orig.__wrapped__ 320 meth = around(orig) 321 meth.__wrapped__ = orig 322 setattr(self, attr, meth) 323 324 def __call__(self, *args, **kwargs): 325 _task_stack.push(self) 326 self.push_request(args=args, kwargs=kwargs) 327 try: 328 # add self if this is a bound task 329 if self.__self__ is not None: 330 return self.run(self.__self__, *args, **kwargs) 331 return self.run(*args, **kwargs) 332 finally: 333 self.pop_request() 334 _task_stack.pop() 335 336 def __reduce__(self): 337 # - tasks are pickled into the name of the task only, and the reciever 338 # - simply grabs it from the local registry. 339 # - in later versions the module of the task is also included, 340 # - and the receiving side tries to import that module so that 341 # - it will work even if the task has not been registered. 342 mod = type(self).__module__ 343 mod = mod if mod and mod in sys.modules else None 344 return (_unpickle_task_v2, (self.name, mod), None) 345 346 def run(self, *args, **kwargs): 347 """The body of the task executed by workers.""" 348 raise NotImplementedError('Tasks must define the run method.') 349 350 def start_strategy(self, app, consumer, **kwargs): 351 return instantiate(self.Strategy, self, app, consumer, **kwargs) 352 353 def delay(self, *args, **kwargs): 354 """Star argument version of :meth:`apply_async`. 355 356 Does not support the extra options enabled by :meth:`apply_async`. 357 358 :param \*args: positional arguments passed on to the task. 359 :param \*\*kwargs: keyword arguments passed on to the task. 360 361 :returns :class:`celery.result.AsyncResult`: 362 363 """ 364 return self.apply_async(args, kwargs) 365 366 def apply_async(self, args=None, kwargs=None, task_id=None, producer=None, 367 link=None, link_error=None, shadow=None, **options): 368 """Apply tasks asynchronously by sending a message. 369 370 :keyword args: The positional arguments to pass on to the 371 task (a :class:`list` or :class:`tuple`). 372 373 :keyword kwargs: The keyword arguments to pass on to the 374 task (a :class:`dict`) 375 376 :keyword countdown: Number of seconds into the future that the 377 task should execute. Defaults to immediate 378 execution. 379 380 :keyword eta: A :class:`~datetime.datetime` object describing 381 the absolute time and date of when the task should 382 be executed. May not be specified if `countdown` 383 is also supplied. 384 385 :keyword expires: Either a :class:`int`, describing the number of 386 seconds, or a :class:`~datetime.datetime` object 387 that describes the absolute time and date of when 388 the task should expire. The task will not be 389 executed after the expiration time. 390 391 :keyword shadow: Override task name used in logs/monitoring 392 (default from :meth:`shadow_name`). 393 394 :keyword connection: Re-use existing broker connection instead 395 of establishing a new one. 396 397 :keyword retry: If enabled sending of the task message will be retried 398 in the event of connection loss or failure. Default 399 is taken from the :setting:`CELERY_TASK_PUBLISH_RETRY` 400 setting. Note that you need to handle the 401 producer/connection manually for this to work. 402 403 :keyword retry_policy: Override the retry policy used. See the 404 :setting:`CELERY_TASK_PUBLISH_RETRY_POLICY` 405 setting. 406 407 :keyword routing_key: Custom routing key used to route the task to a 408 worker server. If in combination with a 409 ``queue`` argument only used to specify custom 410 routing keys to topic exchanges. 411 412 :keyword queue: The queue to route the task to. This must be a key 413 present in :setting:`CELERY_QUEUES`, or 414 :setting:`CELERY_CREATE_MISSING_QUEUES` must be 415 enabled. See :ref:`guide-routing` for more 416 information. 417 418 :keyword exchange: Named custom exchange to send the task to. 419 Usually not used in combination with the ``queue`` 420 argument. 421 422 :keyword priority: The task priority, a number between 0 and 9. 423 Defaults to the :attr:`priority` attribute. 424 425 :keyword serializer: A string identifying the default 426 serialization method to use. Can be `pickle`, 427 `json`, `yaml`, `msgpack` or any custom 428 serialization method that has been registered 429 with :mod:`kombu.serialization.registry`. 430 Defaults to the :attr:`serializer` attribute. 431 432 :keyword compression: A string identifying the compression method 433 to use. Can be one of ``zlib``, ``bzip2``, 434 or any custom compression methods registered with 435 :func:`kombu.compression.register`. Defaults to 436 the :setting:`CELERY_MESSAGE_COMPRESSION` 437 setting. 438 :keyword link: A single, or a list of tasks to apply if the 439 task exits successfully. 440 :keyword link_error: A single, or a list of tasks to apply 441 if an error occurs while executing the task. 442 443 :keyword producer: :class:`kombu.Producer` instance to use. 444 445 :keyword add_to_parent: If set to True (default) and the task 446 is applied while executing another task, then the result 447 will be appended to the parent tasks ``request.children`` 448 attribute. Trailing can also be disabled by default using the 449 :attr:`trail` attribute 450 451 :keyword publisher: Deprecated alias to ``producer``. 452 453 :keyword headers: Message headers to be sent in the 454 task (a :class:`dict`) 455 456 :rtype :class:`celery.result.AsyncResult`: if 457 :setting:`CELERY_ALWAYS_EAGER` is not set, otherwise 458 :class:`celery.result.EagerResult`: 459 460 Also supports all keyword arguments supported by 461 :meth:`kombu.Producer.publish`. 462 463 .. note:: 464 If the :setting:`CELERY_ALWAYS_EAGER` setting is set, it will 465 be replaced by a local :func:`apply` call instead. 466 467 """ 468 try: 469 check_arguments = self.__header__ 470 except AttributeError: 471 pass 472 else: 473 check_arguments(*(args or ()), **(kwargs or {})) 474 475 app = self._get_app() 476 if app.conf.CELERY_ALWAYS_EAGER: 477 return self.apply(args, kwargs, task_id=task_id or uuid(), 478 link=link, link_error=link_error, **options) 479 # add 'self' if this is a "task_method". 480 if self.__self__ is not None: 481 args = args if isinstance(args, tuple) else tuple(args or ()) 482 args = (self.__self__,) + args 483 shadow = shadow or self.shadow_name(args, kwargs, options) 484 485 preopts = self._get_exec_options() 486 options = dict(preopts, **options) if options else preopts 487 return app.send_task( 488 self.name, args, kwargs, task_id=task_id, producer=producer, 489 link=link, link_error=link_error, result_cls=self.AsyncResult, 490 shadow=shadow, 491 **options 492 ) 493 494 def shadow_name(self, args, kwargs, options): 495 """Override for custom task name in worker logs/monitoring. 496 497 :param args: Task positional arguments. 498 :param kwargs: Task keyword arguments. 499 :param options: Task execution options. 500 501 **Example**: 502 503 .. code-block:: python 504 505 from celery.utils.imports import qualname 506 507 def shadow_name(task, args, kwargs, options): 508 return qualname(args[0]) 509 510 @app.task(shadow_name=shadow_name, serializer='pickle') 511 def apply_function_async(fun, *args, **kwargs): 512 return fun(*args, **kwargs) 513 514 """ 515 pass 516 517 def signature_from_request(self, request=None, args=None, kwargs=None, 518 queue=None, **extra_options): 519 request = self.request if request is None else request 520 args = request.args if args is None else args 521 kwargs = request.kwargs if kwargs is None else kwargs 522 limit_hard, limit_soft = request.timelimit or (None, None) 523 options = { 524 'task_id': request.id, 525 'link': request.callbacks, 526 'link_error': request.errbacks, 527 'group_id': request.group, 528 'chord': request.chord, 529 'soft_time_limit': limit_soft, 530 'time_limit': limit_hard, 531 'reply_to': request.reply_to, 532 'headers': request.headers, 533 } 534 options.update( 535 {'queue': queue} if queue else (request.delivery_info or {}), 536 ) 537 return self.signature( 538 args, kwargs, options, type=self, **extra_options 539 ) 540 subtask_from_request = signature_from_request 541 542 def retry(self, args=None, kwargs=None, exc=None, throw=True, 543 eta=None, countdown=None, max_retries=None, **options): 544 """Retry the task. 545 546 :param args: Positional arguments to retry with. 547 :param kwargs: Keyword arguments to retry with. 548 :keyword exc: Custom exception to report when the max restart 549 limit has been exceeded (default: 550 :exc:`~@MaxRetriesExceededError`). 551 552 If this argument is set and retry is called while 553 an exception was raised (``sys.exc_info()`` is set) 554 it will attempt to reraise the current exception. 555 556 If no exception was raised it will raise the ``exc`` 557 argument provided. 558 :keyword countdown: Time in seconds to delay the retry for. 559 :keyword eta: Explicit time and date to run the retry at 560 (must be a :class:`~datetime.datetime` instance). 561 :keyword max_retries: If set, overrides the default retry limit. 562 A value of :const:`None`, means "use the default", so if you want 563 infinite retries you would have to set the :attr:`max_retries` 564 attribute of the task to :const:`None` first. 565 :keyword time_limit: If set, overrides the default time limit. 566 :keyword soft_time_limit: If set, overrides the default soft 567 time limit. 568 :keyword \*\*options: Any extra options to pass on to 569 meth:`apply_async`. 570 :keyword throw: If this is :const:`False`, do not raise the 571 :exc:`~@Retry` exception, 572 that tells the worker to mark the task as being 573 retried. Note that this means the task will be 574 marked as failed if the task raises an exception, 575 or successful if it returns. 576 577 :raises celery.exceptions.Retry: To tell the worker that 578 the task has been re-sent for retry. This always happens, 579 unless the `throw` keyword argument has been explicitly set 580 to :const:`False`, and is considered normal operation. 581 582 **Example** 583 584 .. code-block:: pycon 585 586 >>> from imaginary_twitter_lib import Twitter 587 >>> from proj.celery import app 588 589 >>> @app.task(bind=True) 590 ... def tweet(self, auth, message): 591 ... twitter = Twitter(oauth=auth) 592 ... try: 593 ... twitter.post_status_update(message) 594 ... except twitter.FailWhale as exc: 595 ... # Retry in 5 minutes. 596 ... raise self.retry(countdown=60 * 5, exc=exc) 597 598 Although the task will never return above as `retry` raises an 599 exception to notify the worker, we use `raise` in front of the retry 600 to convey that the rest of the block will not be executed. 601 602 """ 603 request = self.request 604 retries = request.retries + 1 605 max_retries = self.max_retries if max_retries is None else max_retries 606 607 # Not in worker or emulated by (apply/always_eager), 608 # so just raise the original exception. 609 if request.called_directly: 610 maybe_reraise() # raise orig stack if PyErr_Occurred 611 raise exc or Retry('Task can be retried', None) 612 613 if not eta and countdown is None: 614 countdown = self.default_retry_delay 615 616 is_eager = request.is_eager 617 S = self.signature_from_request( 618 request, args, kwargs, 619 countdown=countdown, eta=eta, retries=retries, 620 **options 621 ) 622 623 if max_retries is not None and retries > max_retries: 624 if exc: 625 # first try to reraise the original exception 626 maybe_reraise() 627 # or if not in an except block then raise the custom exc. 628 raise exc 629 raise self.MaxRetriesExceededError( 630 "Can't retry {0}[{1}] args:{2} kwargs:{3}".format( 631 self.name, request.id, S.args, S.kwargs)) 632 633 ret = Retry(exc=exc, when=eta or countdown) 634 635 if is_eager: 636 # if task was executed eagerly using apply(), 637 # then the retry must also be executed eagerly. 638 S.apply().get() 639 if throw: 640 raise ret 641 return ret 642 643 try: 644 S.apply_async() 645 except Exception as exc: 646 raise Reject(exc, requeue=False) 647 if throw: 648 raise ret 649 return ret 650 651 def apply(self, args=None, kwargs=None, 652 link=None, link_error=None, **options): 653 """Execute this task locally, by blocking until the task returns. 654 655 :param args: positional arguments passed on to the task. 656 :param kwargs: keyword arguments passed on to the task. 657 :keyword throw: Re-raise task exceptions. Defaults to 658 the :setting:`CELERY_EAGER_PROPAGATES_EXCEPTIONS` 659 setting. 660 661 :rtype :class:`celery.result.EagerResult`: 662 663 """ 664 # trace imports Task, so need to import inline. 665 from celery.app.trace import build_tracer 666 667 app = self._get_app() 668 args = args or () 669 # add 'self' if this is a bound method. 670 if self.__self__ is not None: 671 args = (self.__self__,) + tuple(args) 672 kwargs = kwargs or {} 673 task_id = options.get('task_id') or uuid() 674 retries = options.get('retries', 0) 675 throw = app.either('CELERY_EAGER_PROPAGATES_EXCEPTIONS', 676 options.pop('throw', None)) 677 678 # Make sure we get the task instance, not class. 679 task = app._tasks[self.name] 680 681 request = {'id': task_id, 682 'retries': retries, 683 'is_eager': True, 684 'logfile': options.get('logfile'), 685 'loglevel': options.get('loglevel', 0), 686 'callbacks': maybe_list(link), 687 'errbacks': maybe_list(link_error), 688 'headers': options.get('headers'), 689 'delivery_info': {'is_eager': True}} 690 tb = None 691 tracer = build_tracer( 692 task.name, task, eager=True, 693 propagate=throw, app=self._get_app(), 694 ) 695 ret = tracer(task_id, args, kwargs, request) 696 retval = ret.retval 697 if isinstance(retval, ExceptionInfo): 698 retval, tb = retval.exception, retval.traceback 699 state = states.SUCCESS if ret.info is None else ret.info.state 700 return EagerResult(task_id, retval, state, traceback=tb) 701 702 def AsyncResult(self, task_id, **kwargs): 703 """Get AsyncResult instance for this kind of task. 704 705 :param task_id: Task id to get result for. 706 707 """ 708 return self._get_app().AsyncResult(task_id, backend=self.backend, 709 task_name=self.name, **kwargs) 710 711 def signature(self, args=None, *starargs, **starkwargs): 712 """Return :class:`~celery.signature` object for 713 this task, wrapping arguments and execution options 714 for a single task invocation.""" 715 starkwargs.setdefault('app', self.app) 716 return signature(self, args, *starargs, **starkwargs) 717 subtask = signature 718 719 def s(self, *args, **kwargs): 720 """``.s(*a, **k) -> .signature(a, k)``""" 721 return self.signature(args, kwargs) 722 723 def si(self, *args, **kwargs): 724 """``.si(*a, **k) -> .signature(a, k, immutable=True)``""" 725 return self.signature(args, kwargs, immutable=True) 726 727 def chunks(self, it, n): 728 """Creates a :class:`~celery.canvas.chunks` task for this task.""" 729 from celery import chunks 730 return chunks(self.s(), it, n, app=self.app) 731 732 def map(self, it): 733 """Creates a :class:`~celery.canvas.xmap` task from ``it``.""" 734 from celery import xmap 735 return xmap(self.s(), it, app=self.app) 736 737 def starmap(self, it): 738 """Creates a :class:`~celery.canvas.xstarmap` task from ``it``.""" 739 from celery import xstarmap 740 return xstarmap(self.s(), it, app=self.app) 741 742 def send_event(self, type_, **fields): 743 req = self.request 744 with self.app.events.default_dispatcher(hostname=req.hostname) as d: 745 return d.send(type_, uuid=req.id, **fields) 746 747 def replace(self, sig): 748 """Replace the current task, with a new task inheriting the 749 same task id. 750 751 :param sig: :class:`@signature` 752 753 Note: This will raise :exc:`~@Ignore`, so the best practice 754 is to always use ``raise self.replace(...)`` to convey 755 to the reader that the task will not continue after being replaced. 756 757 :param: Signature of new task. 758 759 """ 760 chord = self.request.chord 761 if isinstance(sig, group): 762 sig |= self.app.tasks['celery.accumulate'].s(index=0).set( 763 chord=chord, 764 ) 765 chord = None 766 sig.freeze(self.request.id, 767 group_id=self.request.group, 768 chord=chord, 769 root_id=self.request.root_id) 770 sig.delay() 771 raise Ignore('Chord member replaced by new task') 772 773 def add_to_chord(self, sig, lazy=False): 774 """Add signature to the chord the current task is a member of. 775 776 :param sig: Signature to extend chord with. 777 :param lazy: If enabled the new task will not actually be called, 778 and ``sig.delay()`` must be called manually. 779 780 Currently only supported by the Redis result backend when 781 ``?new_join=1`` is enabled. 782 783 """ 784 if not self.request.chord: 785 raise ValueError('Current task is not member of any chord') 786 result = sig.freeze(group_id=self.request.group, 787 chord=self.request.chord, 788 root_id=self.request.root_id) 789 self.backend.add_to_chord(self.request.group, result) 790 return sig.delay() if not lazy else sig 791 792 def update_state(self, task_id=None, state=None, meta=None): 793 """Update task state. 794 795 :keyword task_id: Id of the task to update, defaults to the 796 id of the current task 797 :keyword state: New state (:class:`str`). 798 :keyword meta: State metadata (:class:`dict`). 799 800 801 802 """ 803 if task_id is None: 804 task_id = self.request.id 805 self.backend.store_result(task_id, meta, state) 806 807 def on_success(self, retval, task_id, args, kwargs): 808 """Success handler. 809 810 Run by the worker if the task executes successfully. 811 812 :param retval: The return value of the task. 813 :param task_id: Unique id of the executed task. 814 :param args: Original arguments for the executed task. 815 :param kwargs: Original keyword arguments for the executed task. 816 817 The return value of this handler is ignored. 818 819 """ 820 pass 821 822 def on_retry(self, exc, task_id, args, kwargs, einfo): 823 """Retry handler. 824 825 This is run by the worker when the task is to be retried. 826 827 :param exc: The exception sent to :meth:`retry`. 828 :param task_id: Unique id of the retried task. 829 :param args: Original arguments for the retried task. 830 :param kwargs: Original keyword arguments for the retried task. 831 832 :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` 833 instance, containing the traceback. 834 835 The return value of this handler is ignored. 836 837 """ 838 pass 839 840 def on_failure(self, exc, task_id, args, kwargs, einfo): 841 """Error handler. 842 843 This is run by the worker when the task fails. 844 845 :param exc: The exception raised by the task. 846 :param task_id: Unique id of the failed task. 847 :param args: Original arguments for the task that failed. 848 :param kwargs: Original keyword arguments for the task 849 that failed. 850 851 :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` 852 instance, containing the traceback. 853 854 The return value of this handler is ignored. 855 856 """ 857 pass 858 859 def after_return(self, status, retval, task_id, args, kwargs, einfo): 860 """Handler called after the task returns. 861 862 :param status: Current task state. 863 :param retval: Task return value/exception. 864 :param task_id: Unique id of the task. 865 :param args: Original arguments for the task. 866 :param kwargs: Original keyword arguments for the task. 867 868 :keyword einfo: :class:`~billiard.einfo.ExceptionInfo` 869 instance, containing the traceback (if any). 870 871 The return value of this handler is ignored. 872 873 """ 874 pass 875 876 def send_error_email(self, context, exc, **kwargs): 877 if self.send_error_emails and \ 878 not getattr(self, 'disable_error_emails', None): 879 self.ErrorMail(self, **kwargs).send(context, exc) 880 881 def add_trail(self, result): 882 if self.trail: 883 self.request.children.append(result) 884 return result 885 886 def push_request(self, *args, **kwargs): 887 self.request_stack.push(Context(*args, **kwargs)) 888 889 def pop_request(self): 890 self.request_stack.pop() 891 892 def __repr__(self): 893 """`repr(task)`""" 894 return _reprtask(self, R_SELF_TASK if self.__self__ else R_INSTANCE) 895 896 def _get_request(self): 897 """Get current request object.""" 898 req = self.request_stack.top 899 if req is None: 900 # task was not called, but some may still expect a request 901 # to be there, perhaps that should be deprecated. 902 if self._default_request is None: 903 self._default_request = Context() 904 return self._default_request 905 return req 906 request = property(_get_request) 907 908 def _get_exec_options(self): 909 if self._exec_options is None: 910 self._exec_options = extract_exec_options(self) 911 return self._exec_options 912 913 @property 914 def backend(self): 915 backend = self._backend 916 if backend is None: 917 return self.app.backend 918 return backend 919 920 @backend.setter 921 def backend(self, value): # noqa 922 self._backend = value 923 924 @property 925 def __name__(self): 926 return self.__class__.__name__ 927 abstract.CallableTask.register(Task) 928 BaseTask = Task # compat alias 929 [end of celery/app/task.py] [start of celery/worker/request.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.worker.request 4 ~~~~~~~~~~~~~~~~~~~~~ 5 6 This module defines the :class:`Request` class, 7 which specifies how tasks are executed. 8 9 """ 10 from __future__ import absolute_import, unicode_literals 11 12 import logging 13 import socket 14 import sys 15 16 from datetime import datetime 17 from weakref import ref 18 19 from kombu.utils.encoding import safe_repr, safe_str 20 21 from celery import signals 22 from celery.app.trace import trace_task, trace_task_ret 23 from celery.exceptions import ( 24 Ignore, TaskRevokedError, InvalidTaskError, 25 SoftTimeLimitExceeded, TimeLimitExceeded, 26 WorkerLostError, Terminated, Retry, Reject, 27 ) 28 from celery.five import string 29 from celery.platforms import signals as _signals 30 from celery.utils.functional import noop 31 from celery.utils.log import get_logger 32 from celery.utils.timeutils import maybe_iso8601, timezone, maybe_make_aware 33 from celery.utils.serialization import get_pickled_exception 34 35 from . import state 36 37 __all__ = ['Request'] 38 39 IS_PYPY = hasattr(sys, 'pypy_version_info') 40 41 logger = get_logger(__name__) 42 debug, info, warn, error = (logger.debug, logger.info, 43 logger.warning, logger.error) 44 _does_info = False 45 _does_debug = False 46 47 48 def __optimize__(): 49 # this is also called by celery.app.trace.setup_worker_optimizations 50 global _does_debug 51 global _does_info 52 _does_debug = logger.isEnabledFor(logging.DEBUG) 53 _does_info = logger.isEnabledFor(logging.INFO) 54 __optimize__() 55 56 # Localize 57 tz_utc = timezone.utc 58 tz_or_local = timezone.tz_or_local 59 send_revoked = signals.task_revoked.send 60 61 task_accepted = state.task_accepted 62 task_ready = state.task_ready 63 revoked_tasks = state.revoked 64 65 66 class Request(object): 67 """A request for task execution.""" 68 acknowledged = False 69 time_start = None 70 worker_pid = None 71 time_limits = (None, None) 72 _already_revoked = False 73 _terminate_on_ack = None 74 _apply_result = None 75 _tzlocal = None 76 77 if not IS_PYPY: # pragma: no cover 78 __slots__ = ( 79 'app', 'type', 'name', 'id', 'on_ack', 'body', 80 'hostname', 'eventer', 'connection_errors', 'task', 'eta', 81 'expires', 'request_dict', 'on_reject', 'utc', 82 'content_type', 'content_encoding', 83 '__weakref__', '__dict__', 84 ) 85 86 def __init__(self, message, on_ack=noop, 87 hostname=None, eventer=None, app=None, 88 connection_errors=None, request_dict=None, 89 task=None, on_reject=noop, body=None, 90 headers=None, decoded=False, utc=True, 91 maybe_make_aware=maybe_make_aware, 92 maybe_iso8601=maybe_iso8601, **opts): 93 if headers is None: 94 headers = message.headers 95 if body is None: 96 body = message.body 97 self.app = app 98 self.message = message 99 self.body = body 100 self.utc = utc 101 if decoded: 102 self.content_type = self.content_encoding = None 103 else: 104 self.content_type, self.content_encoding = ( 105 message.content_type, message.content_encoding, 106 ) 107 108 self.id = headers['id'] 109 type = self.type = self.name = headers['task'] 110 if 'shadow' in headers: 111 self.name = headers['shadow'] 112 if 'timelimit' in headers: 113 self.time_limits = headers['timelimit'] 114 self.on_ack = on_ack 115 self.on_reject = on_reject 116 self.hostname = hostname or socket.gethostname() 117 self.eventer = eventer 118 self.connection_errors = connection_errors or () 119 self.task = task or self.app.tasks[type] 120 121 # timezone means the message is timezone-aware, and the only timezone 122 # supported at this point is UTC. 123 eta = headers.get('eta') 124 if eta is not None: 125 try: 126 eta = maybe_iso8601(eta) 127 except (AttributeError, ValueError, TypeError) as exc: 128 raise InvalidTaskError( 129 'invalid eta value {0!r}: {1}'.format(eta, exc)) 130 self.eta = maybe_make_aware(eta, self.tzlocal) 131 else: 132 self.eta = None 133 134 expires = headers.get('expires') 135 if expires is not None: 136 try: 137 expires = maybe_iso8601(expires) 138 except (AttributeError, ValueError, TypeError) as exc: 139 raise InvalidTaskError( 140 'invalid expires value {0!r}: {1}'.format(expires, exc)) 141 self.expires = maybe_make_aware(expires, self.tzlocal) 142 else: 143 self.expires = None 144 145 delivery_info = message.delivery_info or {} 146 properties = message.properties or {} 147 headers.update({ 148 'reply_to': properties.get('reply_to'), 149 'correlation_id': properties.get('correlation_id'), 150 'delivery_info': { 151 'exchange': delivery_info.get('exchange'), 152 'routing_key': delivery_info.get('routing_key'), 153 'priority': delivery_info.get('priority'), 154 'redelivered': delivery_info.get('redelivered'), 155 } 156 157 }) 158 self.request_dict = headers 159 160 @property 161 def delivery_info(self): 162 return self.request_dict['delivery_info'] 163 164 def execute_using_pool(self, pool, **kwargs): 165 """Used by the worker to send this task to the pool. 166 167 :param pool: A :class:`celery.concurrency.base.TaskPool` instance. 168 169 :raises celery.exceptions.TaskRevokedError: if the task was revoked 170 and ignored. 171 172 """ 173 task_id = self.id 174 task = self.task 175 if self.revoked(): 176 raise TaskRevokedError(task_id) 177 178 time_limit, soft_time_limit = self.time_limits 179 time_limit = time_limit or task.time_limit 180 soft_time_limit = soft_time_limit or task.soft_time_limit 181 result = pool.apply_async( 182 trace_task_ret, 183 args=(self.type, task_id, self.request_dict, self.body, 184 self.content_type, self.content_encoding), 185 accept_callback=self.on_accepted, 186 timeout_callback=self.on_timeout, 187 callback=self.on_success, 188 error_callback=self.on_failure, 189 soft_timeout=soft_time_limit, 190 timeout=time_limit, 191 correlation_id=task_id, 192 ) 193 # cannot create weakref to None 194 self._apply_result = ref(result) if result is not None else result 195 return result 196 197 def execute(self, loglevel=None, logfile=None): 198 """Execute the task in a :func:`~celery.app.trace.trace_task`. 199 200 :keyword loglevel: The loglevel used by the task. 201 :keyword logfile: The logfile used by the task. 202 203 """ 204 if self.revoked(): 205 return 206 207 # acknowledge task as being processed. 208 if not self.task.acks_late: 209 self.acknowledge() 210 211 request = self.request_dict 212 args, kwargs, embed = self.message.payload 213 request.update({'loglevel': loglevel, 'logfile': logfile, 214 'hostname': self.hostname, 'is_eager': False, 215 'args': args, 'kwargs': kwargs}, **embed or {}) 216 retval = trace_task(self.task, self.id, args, kwargs, request, 217 hostname=self.hostname, loader=self.app.loader, 218 app=self.app)[0] 219 self.acknowledge() 220 return retval 221 222 def maybe_expire(self): 223 """If expired, mark the task as revoked.""" 224 if self.expires: 225 now = datetime.now(self.expires.tzinfo) 226 if now > self.expires: 227 revoked_tasks.add(self.id) 228 return True 229 230 def terminate(self, pool, signal=None): 231 signal = _signals.signum(signal or 'TERM') 232 if self.time_start: 233 pool.terminate_job(self.worker_pid, signal) 234 self._announce_revoked('terminated', True, signal, False) 235 else: 236 self._terminate_on_ack = pool, signal 237 if self._apply_result is not None: 238 obj = self._apply_result() # is a weakref 239 if obj is not None: 240 obj.terminate(signal) 241 242 def _announce_revoked(self, reason, terminated, signum, expired): 243 task_ready(self) 244 self.send_event('task-revoked', 245 terminated=terminated, signum=signum, expired=expired) 246 if self.store_errors: 247 self.task.backend.mark_as_revoked(self.id, reason, request=self) 248 self.acknowledge() 249 self._already_revoked = True 250 send_revoked(self.task, request=self, 251 terminated=terminated, signum=signum, expired=expired) 252 253 def revoked(self): 254 """If revoked, skip task and mark state.""" 255 expired = False 256 if self._already_revoked: 257 return True 258 if self.expires: 259 expired = self.maybe_expire() 260 if self.id in revoked_tasks: 261 info('Discarding revoked task: %s[%s]', self.name, self.id) 262 self._announce_revoked( 263 'expired' if expired else 'revoked', False, None, expired, 264 ) 265 return True 266 return False 267 268 def send_event(self, type, **fields): 269 if self.eventer and self.eventer.enabled: 270 self.eventer.send(type, uuid=self.id, **fields) 271 272 def on_accepted(self, pid, time_accepted): 273 """Handler called when task is accepted by worker pool.""" 274 self.worker_pid = pid 275 self.time_start = time_accepted 276 task_accepted(self) 277 if not self.task.acks_late: 278 self.acknowledge() 279 self.send_event('task-started') 280 if _does_debug: 281 debug('Task accepted: %s[%s] pid:%r', self.name, self.id, pid) 282 if self._terminate_on_ack is not None: 283 self.terminate(*self._terminate_on_ack) 284 285 def on_timeout(self, soft, timeout): 286 """Handler called if the task times out.""" 287 task_ready(self) 288 if soft: 289 warn('Soft time limit (%ss) exceeded for %s[%s]', 290 soft, self.name, self.id) 291 exc = SoftTimeLimitExceeded(soft) 292 else: 293 error('Hard time limit (%ss) exceeded for %s[%s]', 294 timeout, self.name, self.id) 295 exc = TimeLimitExceeded(timeout) 296 297 if self.store_errors: 298 self.task.backend.mark_as_failure(self.id, exc, request=self) 299 300 if self.task.acks_late: 301 self.acknowledge() 302 303 def on_success(self, failed__retval__runtime, **kwargs): 304 """Handler called if the task was successfully processed.""" 305 failed, retval, runtime = failed__retval__runtime 306 if failed: 307 if isinstance(retval.exception, (SystemExit, KeyboardInterrupt)): 308 raise retval.exception 309 return self.on_failure(retval, return_ok=True) 310 task_ready(self) 311 312 if self.task.acks_late: 313 self.acknowledge() 314 315 self.send_event('task-succeeded', result=retval, runtime=runtime) 316 317 def on_retry(self, exc_info): 318 """Handler called if the task should be retried.""" 319 if self.task.acks_late: 320 self.acknowledge() 321 322 self.send_event('task-retried', 323 exception=safe_repr(exc_info.exception.exc), 324 traceback=safe_str(exc_info.traceback)) 325 326 def on_failure(self, exc_info, send_failed_event=True, return_ok=False): 327 """Handler called if the task raised an exception.""" 328 task_ready(self) 329 330 if isinstance(exc_info.exception, MemoryError): 331 raise MemoryError('Process got: %s' % (exc_info.exception,)) 332 elif isinstance(exc_info.exception, Reject): 333 return self.reject(requeue=exc_info.exception.requeue) 334 elif isinstance(exc_info.exception, Ignore): 335 return self.acknowledge() 336 337 exc = exc_info.exception 338 339 if isinstance(exc, Retry): 340 return self.on_retry(exc_info) 341 342 # These are special cases where the process would not have had 343 # time to write the result. 344 if self.store_errors: 345 if isinstance(exc, Terminated): 346 self._announce_revoked( 347 'terminated', True, string(exc), False) 348 send_failed_event = False # already sent revoked event 349 elif isinstance(exc, WorkerLostError) or not return_ok: 350 self.task.backend.mark_as_failure( 351 self.id, exc, request=self, 352 ) 353 # (acks_late) acknowledge after result stored. 354 if self.task.acks_late: 355 self.acknowledge() 356 357 if send_failed_event: 358 self.send_event( 359 'task-failed', 360 exception=safe_repr(get_pickled_exception(exc_info.exception)), 361 traceback=exc_info.traceback, 362 ) 363 364 if not return_ok: 365 error('Task handler raised error: %r', exc, 366 exc_info=exc_info.exc_info) 367 368 def acknowledge(self): 369 """Acknowledge task.""" 370 if not self.acknowledged: 371 self.on_ack(logger, self.connection_errors) 372 self.acknowledged = True 373 374 def reject(self, requeue=False): 375 if not self.acknowledged: 376 self.on_reject(logger, self.connection_errors, requeue) 377 self.acknowledged = True 378 379 def info(self, safe=False): 380 return {'id': self.id, 381 'name': self.name, 382 'type': self.type, 383 'body': self.body, 384 'hostname': self.hostname, 385 'time_start': self.time_start, 386 'acknowledged': self.acknowledged, 387 'delivery_info': self.delivery_info, 388 'worker_pid': self.worker_pid} 389 390 def __str__(self): 391 return ' '.join([ 392 self.humaninfo(), 393 ' eta:[{0}]'.format(self.eta) if self.eta else '', 394 ' expires:[{0}]'.format(self.expires) if self.expires else '', 395 ]) 396 shortinfo = __str__ 397 398 def humaninfo(self): 399 return '{0.name}[{0.id}]'.format(self) 400 401 def __repr__(self): 402 return '<{0}: {1}>'.format(type(self).__name__, self.humaninfo()) 403 404 @property 405 def tzlocal(self): 406 if self._tzlocal is None: 407 self._tzlocal = self.app.conf.CELERY_TIMEZONE 408 return self._tzlocal 409 410 @property 411 def store_errors(self): 412 return (not self.task.ignore_result or 413 self.task.store_errors_even_if_ignored) 414 415 @property 416 def task_id(self): 417 # XXX compat 418 return self.id 419 420 @task_id.setter # noqa 421 def task_id(self, value): 422 self.id = value 423 424 @property 425 def task_name(self): 426 # XXX compat 427 return self.name 428 429 @task_name.setter # noqa 430 def task_name(self, value): 431 self.name = value 432 433 @property 434 def reply_to(self): 435 # used by rpc backend when failures reported by parent process 436 return self.request_dict['reply_to'] 437 438 @property 439 def correlation_id(self): 440 # used similarly to reply_to 441 return self.request_dict['correlation_id'] 442 443 444 def create_request_cls(base, task, pool, hostname, eventer, 445 ref=ref, revoked_tasks=revoked_tasks, 446 task_ready=task_ready): 447 from celery.app.trace import trace_task_ret as trace 448 default_time_limit = task.time_limit 449 default_soft_time_limit = task.soft_time_limit 450 apply_async = pool.apply_async 451 acks_late = task.acks_late 452 events = eventer and eventer.enabled 453 454 class Request(base): 455 456 def execute_using_pool(self, pool, **kwargs): 457 task_id = self.id 458 if (self.expires or task_id in revoked_tasks) and self.revoked(): 459 raise TaskRevokedError(task_id) 460 461 time_limit, soft_time_limit = self.time_limits 462 time_limit = time_limit or default_time_limit 463 soft_time_limit = soft_time_limit or default_soft_time_limit 464 result = apply_async( 465 trace, 466 args=(self.type, task_id, self.request_dict, self.body, 467 self.content_type, self.content_encoding), 468 accept_callback=self.on_accepted, 469 timeout_callback=self.on_timeout, 470 callback=self.on_success, 471 error_callback=self.on_failure, 472 soft_timeout=soft_time_limit, 473 timeout=time_limit, 474 correlation_id=task_id, 475 ) 476 # cannot create weakref to None 477 self._apply_result = ref(result) if result is not None else result 478 return result 479 480 def on_success(self, failed__retval__runtime, **kwargs): 481 failed, retval, runtime = failed__retval__runtime 482 if failed: 483 if isinstance(retval.exception, ( 484 SystemExit, KeyboardInterrupt)): 485 raise retval.exception 486 return self.on_failure(retval, return_ok=True) 487 task_ready(self) 488 489 if acks_late: 490 self.acknowledge() 491 492 if events: 493 self.send_event( 494 'task-succeeded', result=retval, runtime=runtime, 495 ) 496 497 return Request 498 [end of celery/worker/request.py] </code> Here is an example of a patch file. It consists of changes to the code base. It specifies the file names, the line numbers of each change, and the removed and added lines. A single patch file can contain changes to multiple files. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch> I need you to solve the provided issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the format shown above. Respond below:
celery/celery
045b52f1450d6d5cc500e0057a4b498250dc5692
Message being acknowledged on WorkerLostError when CELERY_ACKS_LATE=True When using celery v3.0.24, with `CELERY_ACKS_LATE = True` , if the OOM killer kills the celery worker, then the worker acknowledges the message. As per [this](https://github.com/celery/celery/commit/e810420c) commit. The `exc_info.internal` comes in as `false`, which means it is not a internal error, due to which the message is acknowledged. The desirable behaviour, in such a case would be to not acknowledge the message (and be able to know, whether its a OOM error), so that some other worker can pick it up. As a workaround, I've commented out the [code](https://github.com/siddharth96/celery/commit/427695d1b23034dadda85fd7a48f7367831be4fa), where celery acknowledges the message, because in such a case, message will be lost.
This is deliberate as if a task is killed it may mean that the next invocation will also cause the same to happen. If the task is redelivered it may cause a loop where the same conditions occur again and again. Also, sadly you cannot distinguish processes killed by OOM from processes killed by other means, and if an administrator kills -9 a task going amok, you usually don't want that task to be called again. There could be a configuration option for not acking terminated tasks, but I'm not sure how useful that would be. A better solution could be to use `basic_reject(requeue=False)` instead of `basic_ack`, that way you can configure a dead letter queue so that the killed tasks will be sent to a queue for manual inspection. I must say, regardless of the status of this feature request, the documentation is misleading. Specifically, [this FAQ makes it seem that process failures would NOT acknowledge messages](http://celery.readthedocs.org/en/latest/faq.html#faq-acks-late-vs-retry). And [this FAQ boldface states](http://celery.readthedocs.org/en/latest/faq.html#id54) that in the event of a kill signal (9), that acks_late will allow the task to re-run (which again, is patently wrong based on this poorly documented behavior). Nowhere in the docs have I found that if the process _dies_, the message will be acknowledged, regardless of acks_late or not. (for instance, I have a set of 10k+ tasks, and some 1% of tasks wind up acknowledged but incomplete when a WorkerLostError is thrown in connection with the worker, although there are no other errors of any kind in any of my logs related to that task). TL;DR at the least, appropriately document the current state when describing the functionality and limitations of acks_late. A work-around would be helpful -- I'm not sure I understand the solution of using `basic_reject`, although I'll keep looking into it. The docs are referring to killing the worker process with KILL, not the child processes. The term worker will always refer to the worker instance, not the pool processes. The section within about acks_late is probably not very helpful and should be removed
2015-10-06T05:34:34Z
<patch> <patch> diff --git a/celery/app/defaults.py b/celery/app/defaults.py --- a/celery/app/defaults.py +++ b/celery/app/defaults.py @@ -132,6 +132,7 @@ def __repr__(self): 'REDIS_DB': Option(type='int', **_REDIS_OLD), 'REDIS_PASSWORD': Option(type='string', **_REDIS_OLD), 'REDIS_MAX_CONNECTIONS': Option(type='int'), + 'REJECT_ON_WORKER_LOST': Option(type='bool'), 'RESULT_BACKEND': Option(type='string'), 'RESULT_DB_SHORT_LIVED_SESSIONS': Option(False, type='bool'), 'RESULT_DB_TABLENAMES': Option(type='dict'), diff --git a/celery/app/task.py b/celery/app/task.py --- a/celery/app/task.py +++ b/celery/app/task.py @@ -220,6 +220,12 @@ class Task(object): #: :setting:`CELERY_ACKS_LATE` setting. acks_late = None + #: When CELERY_ACKS_LATE is set to True, the default behavior to + #: handle worker crash is to acknowledge the message. Setting + #: this to true allows the message to be rejected and requeued so + #: it will be executed again by another worker. + reject_on_worker_lost = None + #: Tuple of expected exceptions. #: #: These are errors that are expected in normal operation @@ -248,6 +254,7 @@ class Task(object): ('rate_limit', 'CELERY_DEFAULT_RATE_LIMIT'), ('track_started', 'CELERY_TRACK_STARTED'), ('acks_late', 'CELERY_ACKS_LATE'), + ('reject_on_worker_lost', 'CELERY_REJECT_ON_WORKER_LOST'), ('ignore_result', 'CELERY_IGNORE_RESULT'), ('store_errors_even_if_ignored', 'CELERY_STORE_ERRORS_EVEN_IF_IGNORED'), diff --git a/celery/worker/request.py b/celery/worker/request.py --- a/celery/worker/request.py +++ b/celery/worker/request.py @@ -326,7 +326,6 @@ def on_retry(self, exc_info): def on_failure(self, exc_info, send_failed_event=True, return_ok=False): """Handler called if the task raised an exception.""" task_ready(self) - if isinstance(exc_info.exception, MemoryError): raise MemoryError('Process got: %s' % (exc_info.exception,)) elif isinstance(exc_info.exception, Reject): @@ -352,7 +351,13 @@ def on_failure(self, exc_info, send_failed_event=True, return_ok=False): ) # (acks_late) acknowledge after result stored. if self.task.acks_late: - self.acknowledge() + reject_and_requeue = (self.task.reject_on_worker_lost and + isinstance(exc, WorkerLostError) and + self.delivery_info.get('redelivered', False) is False) + if reject_and_requeue: + self.reject(requeue=True) + else: + self.acknowledge() if send_failed_event: self.send_event( </patch> </s> </patch>
diff --git a/celery/tests/worker/test_request.py b/celery/tests/worker/test_request.py --- a/celery/tests/worker/test_request.py +++ b/celery/tests/worker/test_request.py @@ -325,6 +325,20 @@ def test_on_failure_Reject_rejects_with_requeue(self): req_logger, req.connection_errors, True, ) + def test_on_failure_WrokerLostError_rejects_with_requeue(self): + einfo = None + try: + raise WorkerLostError() + except: + einfo = ExceptionInfo(internal=True) + req = self.get_request(self.add.s(2, 2)) + req.task.acks_late = True + req.task.reject_on_worker_lost = True + req.delivery_info['redelivered'] = False + req.on_failure(einfo) + req.on_reject.assert_called_with(req_logger, + req.connection_errors, True) + def test_tzlocal_is_cached(self): req = self.get_request(self.add.s(2, 2)) req._tzlocal = 'foo'
1.0
NVIDIA__NeMo-473
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ``` </issue> <code> [start of README.rst] 1 .. image:: http://www.repostatus.org/badges/latest/active.svg 2 :target: http://www.repostatus.org/#active 3 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 4 5 .. image:: https://img.shields.io/badge/documentation-github.io-blue.svg 6 :target: https://nvidia.github.io/NeMo/ 7 :alt: NeMo documentation on GitHub pages 8 9 .. image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 10 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 11 :alt: NeMo core license and license for collections in this repo 12 13 .. image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 14 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python 15 :alt: Language grade: Python 16 17 .. image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 18 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ 19 :alt: Total alerts 20 21 .. image:: https://img.shields.io/badge/code%20style-black-000000.svg 22 :target: https://github.com/psf/black 23 :alt: Code style: black 24 25 26 27 NVIDIA Neural Modules: NeMo 28 =========================== 29 30 NeMo is a toolkit for defining and building `Conversational AI <https://developer.nvidia.com/conversational-ai#started>`_ applications. 31 32 Goal of the NeMo toolkit is to make it possible for researchers to easily compose complex neural network architectures for conversational AI using reusable components. Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 33 34 **Neural Modules** are conceptual blocks of neural networks that take *typed* inputs and produce *typed* outputs. Such modules typically represent data layers, encoders, decoders, language models, loss functions, or methods of combining activations. 35 36 The toolkit comes with extendable collections of pre-built modules for automatic speech recognition (ASR), natural language processing (NLP) and text synthesis (TTS). 37 38 **Introduction** 39 40 * Watch `this video <https://nvidia.github.io/NeMo/>`_ for a quick walk-through. 41 42 * Documentation (latest released version): https://nvidia.github.io/NeMo/ 43 44 * Read NVIDIA `Developer Blog for example applications <https://devblogs.nvidia.com/how-to-build-domain-specific-automatic-speech-recognition-models-on-gpus/>`_ 45 46 * Read NVIDIA `Developer Blog for Quartznet ASR model <https://devblogs.nvidia.com/develop-smaller-speech-recognition-models-with-nvidias-nemo-framework/>`_ 47 48 * Recommended version to install is **0.9.0** via ``pip install nemo-toolkit`` 49 50 * Recommended NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ 51 52 * Pretrained models are available on NVIDIA `NGC Model repository <https://ngc.nvidia.com/catalog/models?orderBy=modifiedDESC&query=nemo&quickFilter=models&filters=>`_ 53 54 55 Getting started 56 ~~~~~~~~~~~~~~~ 57 58 THE LATEST STABLE VERSION OF NeMo is **0.9.0** (Available via PIP). 59 60 **Requirements** 61 62 1) Python 3.6 or 3.7 63 2) PyTorch 1.4.* with GPU support 64 3) (optional, for best performance) NVIDIA APEX. Install from here: https://github.com/NVIDIA/apex 65 66 **NeMo Docker Container** 67 NVIDIA `NGC NeMo Toolkit container <https://ngc.nvidia.com/catalog/containers/nvidia:nemo>`_ is now available. 68 69 * Pull the docker: ``docker pull nvcr.io/nvidia/nemo:v0.9`` 70 * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/nemo:v0.9`` 71 72 If you are using the NVIDIA `NGC PyTorch container <https://ngc.nvidia.com/catalog/containers/nvidia:pytorch>`_ follow these instructions 73 74 * Pull the docker: ``docker pull nvcr.io/nvidia/pytorch:20.01-py3`` 75 * Run: ``docker run --runtime=nvidia -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/pytorch:20.01-py3`` 76 * ``apt-get update && apt-get install -y libsndfile1`` 77 * ``pip install nemo_toolkit`` NeMo core 78 * ``pip install nemo_asr`` NeMo ASR (Speech Recognition) collection 79 * ``pip install nemo_nlp`` NeMo NLP (Natural Language Processing) collection 80 * ``pip install nemo_tts`` NeMo TTS (Speech Synthesis) collection 81 82 See `examples/start_here` to get started with the simplest example. The folder `examples` contains several examples to get you started with various tasks in NLP and ASR. 83 84 **Tutorials** 85 86 * `Speech recognition <https://nvidia.github.io/NeMo/asr/intro.html>`_ 87 * `Natural language processing <https://nvidia.github.io/NeMo/nlp/intro.html>`_ 88 * `Speech Synthesis <https://nvidia.github.io/NeMo/tts/intro.html>`_ 89 90 91 DEVELOPMENT 92 ~~~~~~~~~~~ 93 If you'd like to use master branch and/or develop NeMo you can run "reinstall.sh" script. 94 95 `Documentation (master branch) <http://nemo-master-docs.s3-website.us-east-2.amazonaws.com/>`_. 96 97 **Installing From Github** 98 99 If you prefer to use NeMo's latest development version (from GitHub) follow the steps below: 100 101 1) Clone the repository ``git clone https://github.com/NVIDIA/NeMo.git`` 102 2) Go to NeMo folder and re-install the toolkit with collections: 103 104 .. code-block:: bash 105 106 ./reinstall.sh 107 108 **Style tests** 109 110 .. code-block:: bash 111 112 python setup.py style # Checks overall project code style and output issues with diff. 113 python setup.py style --fix # Tries to fix error in-place. 114 python setup.py style --scope=tests # Operates within certain scope (dir of file). 115 116 **Unittests** 117 118 This command runs unittests: 119 120 .. code-block:: bash 121 122 ./reinstall.sh 123 python pytest tests 124 125 126 Citation 127 ~~~~~~~~ 128 129 If you are using NeMo please cite the following publication 130 131 .. code-block:: tex 132 133 @misc{nemo2019, 134 title={NeMo: a toolkit for building AI applications using Neural Modules}, 135 author={Oleksii Kuchaiev and Jason Li and Huyen Nguyen and Oleksii Hrinchuk and Ryan Leary and Boris Ginsburg and Samuel Kriman and Stanislav Beliaev and Vitaly Lavrukhin and Jack Cook and Patrice Castonguay and Mariya Popova and Jocelyn Huang and Jonathan M. Cohen}, 136 year={2019}, 137 eprint={1909.09577}, 138 archivePrefix={arXiv}, 139 primaryClass={cs.LG} 140 } 141 142 [end of README.rst] [start of nemo/backends/pytorch/actions.py] 1 # Copyright (c) 2019 NVIDIA Corporation 2 import copy 3 import importlib 4 import itertools 5 import json 6 import os 7 from collections import defaultdict 8 from contextlib import ExitStack 9 from pathlib import Path 10 from typing import List, Optional 11 12 import torch 13 import torch.distributed as dist 14 import torch.nn as nn 15 import torch.optim as optim 16 from torch.nn.parallel import DistributedDataParallel as DDP 17 18 from nemo import logging 19 from nemo.backends.pytorch.module_wrapper import TrainableNeuralModuleWrapper 20 from nemo.backends.pytorch.nm import DataLayerNM, TrainableNM 21 from nemo.backends.pytorch.optimizers import AdamW, Novograd, master_params 22 from nemo.core import DeploymentFormat, DeviceType, NeuralModule, NmTensor 23 from nemo.core.callbacks import ActionCallback, EvaluatorCallback, SimpleLossLoggerCallback 24 from nemo.core.neural_factory import Actions, ModelMode, Optimization 25 from nemo.core.neural_types import * 26 from nemo.utils.helpers import get_checkpoint_from_dir 27 28 # these imports will happen on as-needed basis 29 amp = None 30 # convert_syncbn = None 31 # create_syncbn_process_group = None 32 LARC = None 33 FusedLAMB = None 34 FusedAdam = None 35 FusedNovoGrad = None 36 37 AmpOptimizations = { 38 Optimization.mxprO0: "O0", 39 Optimization.mxprO1: "O1", 40 Optimization.mxprO2: "O2", 41 Optimization.mxprO3: "O3", 42 } 43 44 _float_2_half_req = { 45 Optimization.mxprO1, 46 Optimization.mxprO2, 47 Optimization.mxprO3, 48 } 49 50 51 class PtActions(Actions): 52 def __init__( 53 self, local_rank=None, global_rank=None, tb_writer=None, optimization_level=Optimization.mxprO0, 54 ): 55 need_apex = local_rank is not None or optimization_level != Optimization.mxprO0 56 if need_apex: 57 try: 58 apex = importlib.import_module('apex') 59 if optimization_level != Optimization.mxprO0: 60 global amp 61 amp = importlib.import_module('apex.amp') 62 if local_rank is not None: 63 # global convert_syncbn 64 # global create_syncbn_process_group 65 global LARC 66 global FusedLAMB 67 global FusedAdam 68 global FusedNovoGrad 69 parallel = importlib.import_module('apex.parallel') 70 apex_optimizer = importlib.import_module('apex.optimizers') 71 # convert_syncbn = parallel.convert_syncbn_model 72 # create_syncbn_process_group = parallel.create_syncbn_process_group 73 LARC = parallel.LARC 74 FusedLAMB = apex_optimizer.FusedLAMB 75 FusedAdam = apex_optimizer.FusedAdam 76 FusedNovoGrad = apex_optimizer.FusedNovoGrad 77 78 except ImportError: 79 raise ImportError( 80 "NVIDIA Apex is necessary for distributed training and" 81 "mixed precision training. It only works on GPUs." 82 "Please install Apex from " 83 "https://www.github.com/nvidia/apex" 84 ) 85 86 super(PtActions, self).__init__( 87 local_rank=local_rank, global_rank=global_rank, optimization_level=optimization_level, 88 ) 89 90 # will be [unique_instance_id -> (NMModule, PTModule)] 91 self.module_reference_table = {} 92 self.step = 0 93 self.epoch_num = 0 94 self.optimizers = [] 95 self.tb_writer = tb_writer 96 self._modules = set() 97 self.cache = None 98 self.amp_initialized = False 99 100 @property 101 def modules(self): 102 return self._modules 103 104 def __get_top_sorted_modules_and_dataloader(self, hook): 105 """ 106 Constructs DAG leading to hook and creates its topological order. 107 It also populates self.module_reference_table. 108 Args: 109 hook: an NmTensor or a list of NmTensors representing leaf nodes 110 in DAG 111 112 Returns: 113 list of modules with their call arguments and outputs, and dataset 114 """ 115 116 def create_node(producer, producer_args): 117 if producer_args is None: 118 return tuple((producer, ())) 119 else: 120 return tuple((producer, tuple([(k, v) for k, v in producer_args.items()]),)) 121 122 def is_in_degree_zero(node, processed_nodes): 123 """A node has in degree of zero""" 124 if node[1] == (): 125 return True 126 for portname, nmtensor in node[1]: 127 nd = create_node(nmtensor.producer, nmtensor.producer_args) 128 if nd not in processed_nodes: 129 return False 130 return True 131 132 hooks = hook if isinstance(hook, list) else [hook] 133 134 # ensures that no tensors are processed twice 135 processed_nmtensors = set() 136 137 indices_to_remove = [] 138 # Check for duplicates in hook 139 for i, nmtensor in enumerate(hook): 140 if nmtensor in processed_nmtensors: 141 indices_to_remove.append(i) 142 else: 143 processed_nmtensors.add(nmtensor) 144 145 for i in reversed(indices_to_remove): 146 hook.pop(i) 147 148 _top_sorted_modules = [] 149 all_nodes = {} 150 151 # extract all nodes to all_nodes set 152 hooks_lst = list(hooks) 153 while len(hooks_lst) > 0: 154 # take nmtensor from the end of the list 155 nmtensor = hooks_lst.pop() 156 node = create_node(nmtensor.producer, nmtensor.producer_args) 157 # Store nmtensor as an output of its producer 158 # first make sure all keys are present per output port 159 # and nm is inside all_nodes 160 if node not in all_nodes: 161 all_nodes[node] = {k: None for k in nmtensor.producer.output_ports} 162 # second, populate output port with current nmtensor 163 # where applicable 164 all_nodes[node][nmtensor.name] = nmtensor 165 processed_nmtensors.add(nmtensor) 166 if nmtensor.producer_args is not None and nmtensor.producer_args != {}: 167 for _, new_nmtensor in nmtensor.producer_args.items(): 168 if new_nmtensor not in processed_nmtensors: 169 # put in the start of list 170 hooks_lst.insert(0, new_nmtensor) 171 172 all_node_with_output = [] 173 # Iterate over all_nodes to create new nodes that include its output 174 # now all nodes have (module, input tensors, output tensors) 175 for node in all_nodes: 176 all_node_with_output.append(tuple((node[0], node[1], all_nodes[node]))) 177 178 processed_nodes = [] 179 while len(all_node_with_output) > 0: 180 for node in all_node_with_output.copy(): 181 # if node's in_degree is zero it can be added to 182 # _top_sorted_modules 183 # this will also reduce in_degree of its children 184 if is_in_degree_zero(node, processed_nodes): 185 _top_sorted_modules.append(node) 186 processed_nodes.append((node[0], node[1])) 187 all_node_with_output.remove(node) 188 189 # Create top_sorted_modules aka callchain 190 top_sorted_modules = [] 191 for i, m in enumerate(_top_sorted_modules): 192 top_sorted_modules.append((m[0], dict(m[1]), m[2])) 193 # Ensure that there is only one dataset in callchain 194 if i > 0 and isinstance(m[0], DataLayerNM): 195 raise ValueError("There were more than one DataLayer NeuralModule inside your DAG.") 196 197 if not isinstance(top_sorted_modules[0][0], DataLayerNM): 198 raise ValueError("The first module in your DAG was not a DataLayer NeuralModule.") 199 200 tdataset = top_sorted_modules[0][0].dataset 201 202 # populate self.module_reference_table 203 for m in top_sorted_modules: 204 if m[0].factory is None and self._local_rank is not None: 205 raise ValueError( 206 "Neural module {0} was created without " 207 "NeuralModuleFactory, but you are trying to" 208 "run in distributed mode. Please instantiate" 209 "NeuralModuleFactory first and pass its " 210 "instance as `factory` parameter to all your" 211 "Neural Module objects." 212 "".format(str(m[0])) 213 ) 214 key = m[0].unique_instance_id 215 if key not in self.module_reference_table: 216 if isinstance(m[0], TrainableNeuralModuleWrapper): 217 self.module_reference_table[key] = (m[0], m[0]._pt_module) 218 else: 219 self.module_reference_table[key] = (m[0], m[0]) 220 221 return top_sorted_modules, tdataset 222 223 def create_optimizer(self, optimizer, things_to_optimize, optimizer_params=None): 224 """ 225 Wrapper function around __setup_optimizer() 226 227 Args: 228 optimizer : A instantiated PyTorch optimizer or string. For 229 currently supported strings, see __setup_optimizer(). 230 things_to_optimize (list): Must be a list of Neural Modules and/or 231 parameters. If a Neural Module is passed, all trainable 232 parameters are extracted and passed to the optimizer. 233 optimizer_params (dict): Optional parameters dictionary. 234 235 Returns: 236 Optimizer 237 """ 238 239 optimizer_instance = None 240 optimizer_class = None 241 if isinstance(optimizer, str): 242 optimizer_class = optimizer 243 elif isinstance(optimizer, torch.optim.Optimizer): 244 optimizer_instance = optimizer 245 else: 246 raise ValueError("`optimizer` must be a string or an instance of torch.optim.Optimizer") 247 248 modules_to_optimize = [] 249 tensors_to_optimize = [] 250 if not isinstance(things_to_optimize, list): 251 things_to_optimize = [things_to_optimize] 252 for thing in things_to_optimize: 253 if isinstance(thing, NeuralModule): 254 modules_to_optimize.append(thing) 255 elif isinstance(thing, NmTensor): 256 tensors_to_optimize.append(thing) 257 else: 258 raise ValueError( 259 "{} passed to create_optimizer() was neither a neural module nor a neural module tensor" 260 ) 261 262 if tensors_to_optimize: 263 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(tensors_to_optimize) 264 265 for module in call_chain: 266 if module[0] not in modules_to_optimize: 267 modules_to_optimize.append(module[0]) 268 269 # Extract trainable weights which will be optimized 270 params_list = [p.parameters() for p in modules_to_optimize if isinstance(p, TrainableNM) or p.is_trainable()] 271 params_to_optimize = itertools.chain(*params_list) 272 273 if optimizer_params is None: 274 optimizer_params = {} 275 # Init amp 276 optimizer = self.__setup_optimizer( 277 optimizer_instance=optimizer_instance, 278 optimizer_class=optimizer_class, 279 optimization_params=optimizer_params, 280 params_to_optimize=params_to_optimize, 281 ) 282 283 self.optimizers.append(optimizer) 284 return optimizer 285 286 @staticmethod 287 def __setup_optimizer( 288 optimizer_instance, optimizer_class, optimization_params, params_to_optimize, 289 ): 290 291 if optimizer_instance is None: 292 # Setup optimizer instance, by default it is SGD 293 lr = optimization_params["lr"] 294 if optimizer_class.lower() == "sgd": 295 optimizer = optim.SGD( 296 params_to_optimize, 297 lr=lr, 298 momentum=optimization_params.get("momentum", 0.9), 299 weight_decay=optimization_params.get("weight_decay", 0.0), 300 ) 301 elif optimizer_class.lower() == "adam": 302 optimizer = optim.Adam( 303 params=params_to_optimize, lr=lr, betas=optimization_params.get("betas", (0.9, 0.999)), 304 ) 305 elif optimizer_class.lower() == "fused_adam": 306 optimizer = FusedAdam(params=params_to_optimize, lr=lr) 307 elif optimizer_class.lower() == "adam_w": 308 optimizer = AdamW( 309 params=params_to_optimize, 310 lr=lr, 311 weight_decay=optimization_params.get("weight_decay", 0.0), 312 betas=optimization_params.get("betas", (0.9, 0.999)), 313 ) 314 elif optimizer_class.lower() == "novograd": 315 optimizer = Novograd( 316 params_to_optimize, 317 lr=lr, 318 weight_decay=optimization_params.get("weight_decay", 0.0), 319 luc=optimization_params.get("luc", False), 320 luc_trust=optimization_params.get("luc_eta", 1e-3), 321 betas=optimization_params.get("betas", (0.95, 0.25)), 322 ) 323 elif optimizer_class.lower() == "fused_novograd": 324 optimizer = FusedNovoGrad( 325 params_to_optimize, 326 lr=lr, 327 weight_decay=optimization_params.get("weight_decay", 0.0), 328 reg_inside_moment=True, 329 grad_averaging=False, 330 betas=optimization_params.get("betas", (0.95, 0.25)), 331 ) 332 elif optimizer_class.lower() == "fused_lamb": 333 optimizer = FusedLAMB(params_to_optimize, lr=lr,) 334 else: 335 raise ValueError("Unknown optimizer class: {0}".format(optimizer_class)) 336 337 if optimization_params.get("larc", False): 338 logging.info("Enabling larc") 339 optimizer = LARC(optimizer, trust_coefficient=optimization_params.get("larc_eta", 2e-2),) 340 else: 341 logging.info("Optimizer instance: {0} is provided.") 342 if optimizer_class is not None and optimizer_class != "": 343 logging.warning("Ignoring `optimizer_class` parameter because `optimizer_instance` is provided") 344 if optimization_params is not None and optimization_params != {}: 345 logging.warning( 346 "Ignoring `optimization_params` parameter for " 347 "optimizer because `optimizer_instance` is provided" 348 ) 349 optimizer = optimizer_instance 350 return optimizer 351 352 def __initialize_amp( 353 self, optimizer, optim_level, amp_max_loss_scale=2.0 ** 24, amp_min_loss_scale=1.0, 354 ): 355 if optim_level not in AmpOptimizations: 356 raise ValueError(f"__initialize_amp() was called with unknown optim_level={optim_level}") 357 # in this case, nothing to do here 358 if optim_level == Optimization.mxprO0: 359 return optimizer 360 361 if len(self.modules) < 1: 362 raise ValueError("There were no modules to initialize") 363 pt_modules = [] 364 for module in self.modules: 365 if isinstance(module, nn.Module): 366 pt_modules.append(module) 367 elif isinstance(module, TrainableNeuralModuleWrapper): 368 pt_modules.append(module._pt_module) 369 370 _, optimizer = amp.initialize( 371 max_loss_scale=amp_max_loss_scale, 372 min_loss_scale=amp_min_loss_scale, 373 models=pt_modules, 374 optimizers=optimizer, 375 opt_level=AmpOptimizations[optim_level], 376 ) 377 self.amp_initialized = True 378 return optimizer 379 380 def __nm_graph_forward_pass( 381 self, call_chain, registered_tensors, mode=ModelMode.train, use_cache=False, 382 ): 383 for ind in range(1, len(call_chain)): 384 if use_cache: 385 in_cache = True 386 for tensor in call_chain[ind][2].values(): 387 if tensor is None: 388 # NM has an output tensor that is not used in the 389 # current call chain, so we don't care if it's not in 390 # cache 391 continue 392 if tensor.unique_name not in registered_tensors: 393 in_cache = False 394 if in_cache: 395 continue 396 call_args = call_chain[ind][1] 397 # module = call_chain[ind][0] 398 m_id = call_chain[ind][0].unique_instance_id 399 pmodule = self.module_reference_table[m_id][1] 400 401 # if self._local_rank is not None: 402 # if isinstance(pmodule, DDP): 403 # if disable_allreduce: 404 # pmodule.disable_allreduce() 405 # else: 406 # pmodule.enable_allreduce() 407 408 if mode == ModelMode.train: 409 # if module.is_trainable(): 410 if isinstance(pmodule, nn.Module): 411 pmodule.train() 412 elif mode == ModelMode.eval: 413 # if module.is_trainable(): 414 if isinstance(pmodule, nn.Module): 415 pmodule.eval() 416 else: 417 raise ValueError("Unknown ModelMode") 418 # prepare call signature for `module` 419 call_set = {} 420 for tensor_name, nmtensor in call_args.items(): 421 # _add_uuid_2_name(nmtensor.name, nmtensor.producer._uuid) 422 key = nmtensor.unique_name 423 call_set[tensor_name] = registered_tensors[key] 424 # actual PyTorch module call with signature 425 if isinstance(self.module_reference_table[m_id][0], TrainableNeuralModuleWrapper,): 426 new_tensors = pmodule(**call_set) 427 else: 428 new_tensors = pmodule(force_pt=True, **call_set) 429 430 if not isinstance(new_tensors, List): 431 if not isinstance(new_tensors, tuple): 432 new_tensors = [new_tensors] 433 else: 434 new_tensors = list(new_tensors) 435 for t_tensor, nm_tensor in zip(new_tensors, call_chain[ind][2].values()): 436 if nm_tensor is None: 437 continue 438 t_name = nm_tensor.unique_name 439 if t_name not in registered_tensors: 440 registered_tensors[t_name] = t_tensor 441 else: 442 raise ValueError("A NMTensor was produced twice in " f"the same DAG. {t_name}") 443 444 @staticmethod 445 def pad_tensor(t: torch.Tensor, target_size: torch.Size): 446 padded_shape = target_size.cpu().data.numpy().tolist() 447 padded_t = torch.zeros(padded_shape).cuda().type_as(t) 448 t_size = t.size() 449 if len(t_size) == 0: 450 padded_t = t 451 elif len(t_size) == 1: 452 padded_t[: t_size[0]] = t 453 elif len(t_size) == 2: 454 padded_t[: t_size[0], : t_size[1]] = t 455 elif len(t_size) == 3: 456 padded_t[: t_size[0], : t_size[1], : t_size[2]] = t 457 elif len(t_size) == 4: 458 padded_t[: t_size[0], : t_size[1], : t_size[2], : t_size[3]] = t 459 else: 460 raise NotImplementedError 461 return padded_t 462 463 @staticmethod 464 def depad_tensor(t: torch.Tensor, original_size: torch.Size): 465 t_size = original_size 466 if len(t_size) == 0: 467 depadded_t = t 468 elif len(t_size) == 1: 469 depadded_t = t[: t_size[0]] 470 elif len(t_size) == 2: 471 depadded_t = t[: t_size[0], : t_size[1]] 472 elif len(t_size) == 3: 473 depadded_t = t[: t_size[0], : t_size[1], : t_size[2]] 474 elif len(t_size) == 4: 475 depadded_t = t[: t_size[0], : t_size[1], : t_size[2], : t_size[3]] 476 else: 477 raise NotImplementedError 478 return depadded_t 479 480 def _eval(self, tensors_2_evaluate, callback, step, verbose=False): 481 """ 482 Evaluation process. 483 WARNING THIS function assumes that all tensors_2_evaluate are based 484 on a single datalayer 485 Args: 486 tensors_2_evaluate: list of NmTensors to evaluate 487 callback: instance of EvaluatorCallback 488 step: current training step, used for logging 489 490 Returns: 491 None 492 """ 493 with torch.no_grad(): 494 # each call chain corresponds to a tensor in tensors_2_evaluate 495 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors_2_evaluate) 496 # "Retrieve" data layer from call chain. 497 dl_nm = call_chain[0][0] 498 499 # Prepare eval_dataloader 500 # For distributed training it should have disjoint subsets of 501 # all data on every worker 502 is_distributed = False 503 world_size = None 504 if dl_nm.placement == DeviceType.AllGpu: 505 assert dist.is_initialized() 506 is_distributed = True 507 world_size = torch.distributed.get_world_size() 508 # logging.info( 509 # "Doing distributed evaluation. Rank {0} of {1}".format( 510 # self.local_rank, world_size 511 # ) 512 # ) 513 if dl_nm.dataset is not None: 514 sampler = torch.utils.data.distributed.DistributedSampler( 515 dataset=dl_nm.dataset, shuffle=dl_nm.shuffle 516 ) 517 eval_dataloader = torch.utils.data.DataLoader( 518 dataset=dl_nm.dataset, 519 sampler=sampler, 520 num_workers=dl_nm.num_workers, 521 batch_size=dl_nm.batch_size, 522 shuffle=False, 523 ) 524 else: 525 eval_dataloader = dl_nm.data_iterator 526 527 if hasattr(eval_dataloader, 'sampler'): 528 eval_dataloader.sampler.set_epoch(0) 529 else: # Not distributed 530 if dl_nm.dataset is not None: 531 # Todo: remove local_parameters 532 eval_dataloader = torch.utils.data.DataLoader( 533 dataset=dl_nm.dataset, 534 sampler=None, # not distributed sampler 535 num_workers=dl_nm.num_workers, 536 batch_size=dl_nm.batch_size, 537 shuffle=dl_nm.shuffle, 538 ) 539 else: 540 eval_dataloader = dl_nm.data_iterator 541 # after this eval_dataloader is ready to be used 542 # reset global_var_dict - results of evaluation will be stored 543 # there 544 545 callback.clear_global_var_dict() 546 dl_device = dl_nm._device 547 548 # Evaluation mini-batch for loop 549 num_batches = None 550 if hasattr(eval_dataloader, "__len__"): 551 num_batches = len(eval_dataloader) 552 for epoch_i, data in enumerate(eval_dataloader, 0): 553 if ( 554 verbose 555 and num_batches is not None 556 and (num_batches < 10 or (epoch_i % int(num_batches / 10) == 0)) 557 ): 558 logging.info(f"Evaluating batch {epoch_i} out of {num_batches}") 559 tensors = [] 560 if isinstance(data, torch.Tensor): 561 data = (data,) 562 for d in data: 563 if isinstance(d, torch.Tensor): 564 tensors.append(d.to(dl_device)) 565 else: 566 tensors.append(d) 567 568 registered_e_tensors = { 569 t.unique_name: d for t, d in zip(call_chain[0][2].values(), tensors) if t is not None 570 } 571 self.__nm_graph_forward_pass( 572 call_chain=call_chain, registered_tensors=registered_e_tensors, mode=ModelMode.eval, 573 ) 574 575 if not is_distributed or self.global_rank == 0: 576 values_dict = {} 577 # If distributed. For the outer loop, we need to ensure that 578 # all processes loop through the elements in the same order 579 for t2e in tensors_2_evaluate: 580 key = t2e.unique_name 581 if key not in registered_e_tensors.keys(): 582 logging.info("WARNING: Tensor {} was not found during eval".format(key)) 583 continue 584 if is_distributed: 585 # where we will all_gather results from all workers 586 tensors_list = [] 587 # where we will all_gather tensor sizes 588 tensor_on_worker = registered_e_tensors[key] 589 if tensor_on_worker.shape != torch.Size([]): 590 tensor_on_worker_size_as_tensor = torch.tensor(tensor_on_worker.shape).cuda() 591 sizes = [] 592 for ind in range(world_size): 593 sizes.append(torch.empty_like(tensor_on_worker_size_as_tensor)) 594 dist.all_gather(sizes, tensor_on_worker_size_as_tensor) 595 mx_dim, _ = torch.max(torch.stack(sizes), dim=0) 596 else: # this is a singleton. For example, loss value 597 sizes = [torch.Size([])] * world_size 598 mx_dim = None 599 for ind in range(world_size): 600 # we have to use max shape for all_gather 601 if mx_dim is None: # singletons 602 tensors_list.append(torch.tensor(2).cuda().type_as(tensor_on_worker)) 603 else: # non-singletons 604 tensors_list.append( 605 torch.empty(mx_dim.cpu().data.numpy().tolist()).cuda().type_as(tensor_on_worker) 606 ) 607 608 if mx_dim is not None: 609 t_to_send = self.pad_tensor(tensor_on_worker, mx_dim) 610 else: 611 t_to_send = tensor_on_worker 612 dist.all_gather(tensors_list, t_to_send) 613 tensors_list = [self.depad_tensor(t, size) for t, size in zip(tensors_list, sizes)] 614 if self.global_rank == 0: 615 values_dict["IS_FROM_DIST_EVAL"] = True 616 values_dict[key] = tensors_list 617 else: # NON-DISTRIBUTED TRAINING 618 values_dict["IS_FROM_DIST_EVAL"] = False 619 values_dict[key] = [registered_e_tensors[key]] 620 if callback.user_iter_callback and (self.global_rank is None or self.global_rank == 0): 621 # values_dict will contain results from all workers 622 callback.user_iter_callback(values_dict, callback._global_var_dict) 623 624 # final aggregation (over minibatches) and logging of results 625 # should happend only on one worker 626 if callback.user_done_callback and (self.global_rank is None or self.global_rank == 0): 627 vals_to_log = callback.user_done_callback(callback._global_var_dict) 628 # log results to Tensorboard or Weights & Biases 629 if vals_to_log is not None: 630 if hasattr(callback, 'swriter') and callback.swriter is not None: 631 if hasattr(callback, 'tb_writer_func') and callback.tb_writer_func is not None: 632 callback.tb_writer_func(callback.swriter, vals_to_log, step) 633 else: 634 for key, val in vals_to_log.items(): 635 callback.swriter.add_scalar(key, val, step) 636 if hasattr(callback, 'wandb_log'): 637 callback.wandb_log(vals_to_log) 638 639 def _infer( 640 self, tensors_to_return, verbose=False, cache=False, use_cache=False, offload_to_cpu=True, 641 ): 642 """ 643 Does the same as _eval() just with tensors instead of eval callback. 644 """ 645 # Checking that cache is used properly 646 if cache and use_cache: 647 raise ValueError( 648 "cache and use_cache were both set. However cache must first be created prior to using it." 649 ) 650 if cache: 651 if self.cache is not None: 652 raise ValueError("cache was set but was not empty") 653 self.cache = [] 654 if use_cache: 655 if not self.cache: 656 raise ValueError("use_cache was set, but cache was empty") 657 658 with torch.no_grad(): 659 # each call chain corresponds to a tensor in tensors_2_evaluate 660 dl_nm = None 661 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors_to_return) 662 dl_nm = call_chain[0][0] 663 664 # Prepare eval_dataloader 665 # For distributed training it should have disjoint subsets of 666 # all data on every worker 667 is_distributed = False 668 world_size = None 669 if dl_nm.placement == DeviceType.AllGpu: 670 if self.cache or use_cache: 671 raise NotImplementedError("Caching is not available for distributed training.") 672 assert dist.is_initialized() 673 is_distributed = True 674 world_size = torch.distributed.get_world_size() 675 # logging.info( 676 # "Doing distributed evaluation. Rank {0} of {1}".format( 677 # self.local_rank, world_size 678 # ) 679 # ) 680 if dl_nm.dataset is not None: 681 sampler = torch.utils.data.distributed.DistributedSampler( 682 dataset=dl_nm.dataset, shuffle=dl_nm.shuffle 683 ) 684 eval_dataloader = torch.utils.data.DataLoader( 685 dataset=dl_nm.dataset, 686 sampler=sampler, 687 num_workers=dl_nm.num_workers, 688 batch_size=dl_nm.batch_size, 689 shuffle=False, 690 ) 691 else: 692 eval_dataloader = dl_nm.data_iterator 693 eval_dataloader.sampler.set_epoch(0) 694 elif not use_cache: # Not distributed and not using cache 695 # Dataloaders are only used if use_cache is False 696 # When caching, the DAG must cache all outputs from dataloader 697 if dl_nm.dataset is not None: 698 # Todo: remove local_parameters 699 eval_dataloader = torch.utils.data.DataLoader( 700 dataset=dl_nm.dataset, 701 sampler=None, # not distributed sampler 702 num_workers=dl_nm.num_workers, 703 batch_size=dl_nm.batch_size, 704 shuffle=dl_nm.shuffle, 705 ) 706 else: 707 eval_dataloader = dl_nm.data_iterator 708 # after this eval_dataloader is ready to be used 709 # reset global_var_dict - results of evaluation will be stored 710 # there 711 712 if not is_distributed or self.global_rank == 0: 713 values_dict = {} 714 for t in tensors_to_return: 715 values_dict[t.unique_name] = [] 716 dl_device = dl_nm._device 717 718 # Evaluation mini-batch for loop 719 if use_cache: 720 num_batches = len(self.cache) 721 loop_iterator = self.cache 722 else: 723 num_batches = len(eval_dataloader) 724 loop_iterator = eval_dataloader 725 726 for epoch_i, data in enumerate(loop_iterator, 0): 727 logging.debug(torch.cuda.memory_allocated()) 728 if verbose and (num_batches < 10 or (epoch_i % int(num_batches / 10) == 0)): 729 logging.info(f"Evaluating batch {epoch_i} out of {num_batches}") 730 tensors = [] 731 if use_cache: 732 registered_e_tensors = data 733 # delete tensors_to_return 734 for t in tensors_to_return: 735 if t.unique_name in registered_e_tensors: 736 del registered_e_tensors[t.unique_name] 737 # Need to check for device type mismatch 738 for t in registered_e_tensors: 739 registered_e_tensors[t].to(dl_device) 740 else: 741 if isinstance(data, torch.Tensor): 742 data = (data,) 743 for d in data: 744 if isinstance(d, torch.Tensor): 745 tensors.append(d.to(dl_device)) 746 else: 747 tensors.append(d) 748 749 registered_e_tensors = { 750 t.unique_name: d for t, d in zip(call_chain[0][2].values(), tensors) if t is not None 751 } 752 self.__nm_graph_forward_pass( 753 call_chain=call_chain, 754 registered_tensors=registered_e_tensors, 755 mode=ModelMode.eval, 756 use_cache=use_cache, 757 ) 758 759 # if offload_to_cpu: 760 # # Take all cuda tensors and save them to value_dict as 761 # # cpu tensors to save GPU memory 762 # for name, tensor in registered_e_tensors.items(): 763 # if isinstance(tensor, torch.Tensor): 764 # registered_e_tensors[name] = tensor.cpu() 765 if cache: 766 self.append_to_cache(registered_e_tensors, offload_to_cpu) 767 768 # If distributed. For the outer loop, we need to ensure that 769 # all processes loop through the elements in the same order 770 for t2e in tensors_to_return: 771 key = t2e.unique_name 772 if key not in registered_e_tensors.keys(): 773 logging.info("WARNING: Tensor {} was not found during eval".format(key)) 774 continue 775 if is_distributed: 776 # where we will all_gather results from all workers 777 tensors_list = [] 778 # where we will all_gather tensor sizes 779 tensor_on_worker = registered_e_tensors[key] 780 if tensor_on_worker.shape != torch.Size([]): 781 tensor_on_worker_size_as_tensor = torch.tensor(tensor_on_worker.shape).cuda() 782 sizes = [] 783 for ind in range(world_size): 784 sizes.append(torch.empty_like(tensor_on_worker_size_as_tensor)) 785 dist.all_gather(sizes, tensor_on_worker_size_as_tensor) 786 mx_dim, _ = torch.max(torch.stack(sizes), dim=0) 787 else: # this is a singleton. For example, loss value 788 sizes = [torch.Size([])] * world_size 789 mx_dim = None 790 for ind in range(world_size): 791 # we have to use max shape for all_gather 792 if mx_dim is None: # singletons 793 tensors_list.append(torch.tensor(2).cuda().type_as(tensor_on_worker)) 794 else: # non-singletons 795 tensors_list.append( 796 torch.empty(mx_dim.cpu().data.numpy().tolist()).cuda().type_as(tensor_on_worker) 797 ) 798 799 if mx_dim is not None: 800 t_to_send = self.pad_tensor(tensor_on_worker, mx_dim) 801 else: 802 t_to_send = tensor_on_worker 803 dist.all_gather(tensors_list, t_to_send) 804 tensors_list = [self.depad_tensor(t, size) for t, size in zip(tensors_list, sizes)] 805 if offload_to_cpu: 806 tensors_list = [t.cpu() for t in tensors_list] 807 if self.global_rank == 0: 808 values_dict[key] += tensors_list 809 else: # NON-DISTRIBUTED TRAINING 810 tensor = registered_e_tensors[key] 811 if offload_to_cpu and isinstance(tensor, torch.Tensor): 812 tensor = tensor.cpu() 813 values_dict[key] += [tensor] 814 815 if not is_distributed or self.global_rank == 0: 816 inferred_tensors = [] 817 for t in tensors_to_return: 818 inferred_tensors.append(values_dict[t.unique_name]) 819 return inferred_tensors 820 821 # For all other ranks 822 return None 823 824 def append_to_cache(self, registered_tensors: dict, offload_to_cpu): 825 """Simpler helper function to add results of __nm_graph_forward_pass to 826 current cache. 827 """ 828 if offload_to_cpu: 829 for t in registered_tensors: 830 registered_tensors[t] = registered_tensors[t].cpu() 831 self.cache.append(registered_tensors) 832 833 def clear_cache(self): 834 """ Simple helpful function to clear cache by setting self.cache to 835 None 836 """ 837 self.cache = None 838 839 def save_state_to(self, path: str): 840 """ 841 Saves current state such as step, epoch and optimizer parameters 842 Args: 843 path: 844 845 Returns: 846 847 """ 848 state = { 849 "step": self.step, 850 "epoch_num": self.epoch_num, 851 "optimizer_state": [opt.state_dict() for opt in self.optimizers], 852 } 853 torch.save(state, path) 854 855 def restore_state_from(self, path: str): 856 """ 857 Restores state such as step, epoch and optimizer parameters 858 Args: 859 path: 860 861 Returns: 862 863 """ 864 if os.path.isfile(path): 865 # map_location could be cuda:<device_id> but cpu seems to be more 866 # general since we are also saving step and epoch_num 867 # load_state_dict should move the variables to the relevant device 868 checkpoint = torch.load(path, map_location="cpu") 869 self.step = checkpoint["step"] 870 self.epoch_num = checkpoint["epoch_num"] 871 if checkpoint["optimizer_state"]: 872 for opt, opt_chkpt in zip(self.optimizers, checkpoint["optimizer_state"]): 873 opt.load_state_dict(opt_chkpt) 874 else: 875 raise FileNotFoundError("Could not find checkpoint file: {0}".format(path)) 876 877 @staticmethod 878 def _check_all_tensors(list_of_tensors): 879 """Method that checks if the passed list contains all NmTensors 880 """ 881 if not isinstance(list_of_tensors, list): 882 return False 883 for tensor in list_of_tensors: 884 if not isinstance(tensor, NmTensor): 885 return False 886 return True 887 888 @staticmethod 889 def _check_tuples(list_of_tuples): 890 """Method that checks if the passed tuple contains an optimizer in the 891 first element, and a list of NmTensors in the second. 892 """ 893 for tup in list_of_tuples: 894 if not (isinstance(tup[0], torch.optim.Optimizer) and PtActions._check_all_tensors(tup[1])): 895 return False 896 return True 897 898 def _get_all_modules(self, training_loop, callbacks, logging_callchain=None): 899 """Gets all neural modules that will be used by train() and eval() via 900 EvaluatorCallbacks. Saves all modules to self.modules 901 """ 902 # If there is a SimpleLossLoggerCallback, create an logger_callchain 903 # with all callchains from training_loop and 904 # SimpleLossLoggerCallback.tensors 905 if logging_callchain: 906 for module in logging_callchain: 907 self.modules.add(module[0]) 908 909 # Else grab all callchains from training_loop 910 else: 911 for step in training_loop: 912 for module in step[2]: 913 self.modules.add(module[0]) 914 915 # Lastly, grab all eval modules 916 if callbacks is not None: 917 for callback in callbacks: 918 if isinstance(callback, EvaluatorCallback): 919 (callchain, _,) = self.__get_top_sorted_modules_and_dataloader(hook=callback.eval_tensors) 920 for module in callchain: 921 self.modules.add(module[0]) 922 923 @staticmethod 924 def __module_export(module, output, d_format: DeploymentFormat, input_example=None, output_example=None): 925 # Check if output already exists 926 destination = Path(output) 927 if destination.exists(): 928 raise FileExistsError(f"Destination {output} already exists. " f"Aborting export.") 929 930 input_names = list(module.input_ports.keys()) 931 output_names = list(module.output_ports.keys()) 932 dynamic_axes = defaultdict(list) 933 934 def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defaultdict): 935 if ntype.axes: 936 for ind, axis in enumerate(ntype.axes): 937 if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: 938 dynamic_axes[port_name].append(ind) 939 940 # This is a hack for Jasper to Jarvis export -- need re-design for this 941 inputs_to_drop = set() 942 outputs_to_drop = set() 943 if type(module).__name__ == "JasperEncoder": 944 logging.info( 945 "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " 946 "deployment" 947 ) 948 inputs_to_drop.add("length") 949 outputs_to_drop.add("encoded_lengths") 950 951 # for input_ports 952 for port_name, ntype in module.input_ports.items(): 953 if port_name in inputs_to_drop: 954 input_names.remove(port_name) 955 continue 956 __extract_dynamic_axes(port_name, ntype, dynamic_axes) 957 # for output_ports 958 for port_name, ntype in module.output_ports.items(): 959 if port_name in outputs_to_drop: 960 output_names.remove(port_name) 961 continue 962 __extract_dynamic_axes(port_name, ntype, dynamic_axes) 963 964 if len(dynamic_axes) == 0: 965 dynamic_axes = None 966 967 # Make a deep copy of init parameters. 968 init_params_copy = copy.deepcopy(module._init_params) 969 970 # Remove NeMo-related things from the module 971 # We need to change __call__ method. Note that this will change the 972 # whole class, not just this object! Which is why we need to repair it 973 # in the finally block 974 type(module).__call__ = torch.nn.Module.__call__ 975 976 # Reset standard instance field - making the file (probably) lighter. 977 module._init_params = None 978 module._placement = None 979 module._factory = None 980 module._device = None 981 982 module.eval() 983 try: 984 if d_format == DeploymentFormat.TORCHSCRIPT: 985 if input_example is None: 986 # Route 1 - via torch.jit.script 987 traced_m = torch.jit.script(module) 988 traced_m.save(output) 989 else: 990 # Route 2 - via tracing 991 traced_m = torch.jit.trace(module, input_example) 992 traced_m.save(output) 993 elif d_format == DeploymentFormat.ONNX or d_format == DeploymentFormat.TRTONNX: 994 if input_example is None: 995 raise ValueError(f'Example input is None, but ONNX tracing was' f' attempted') 996 if output_example is None: 997 if isinstance(input_example, tuple): 998 output_example = module.forward(*input_example) 999 else: 1000 output_example = module.forward(input_example) 1001 with torch.jit.optimized_execution(True): 1002 jitted_model = torch.jit.trace(module, input_example) 1003 1004 torch.onnx.export( 1005 jitted_model, 1006 input_example, 1007 output, 1008 input_names=input_names, 1009 output_names=output_names, 1010 verbose=False, 1011 export_params=True, 1012 do_constant_folding=True, 1013 dynamic_axes=dynamic_axes, 1014 opset_version=11, 1015 example_outputs=output_example, 1016 ) 1017 # fn = output + ".readable" 1018 # with open(fn, 'w') as f: 1019 # tempModel = onnx.load(output) 1020 # onnx.save(tempModel, output + ".copy") 1021 # onnx.checker.check_model(tempModel) 1022 # pgraph = onnx.helper.printable_graph(tempModel.graph) 1023 # f.write(pgraph) 1024 1025 elif d_format == DeploymentFormat.PYTORCH: 1026 torch.save(module.state_dict(), output) 1027 with open(output + ".json", 'w') as outfile: 1028 json.dump(init_params_copy, outfile) 1029 1030 else: 1031 raise NotImplementedError(f"Not supported deployment format: {d_format}") 1032 except Exception as e: # nopep8 1033 logging.error(f'module export failed for {module} ' f'with exception {e}') 1034 finally: 1035 1036 def __old_call__(self, force_pt=False, *input, **kwargs): 1037 pt_call = len(input) > 0 or force_pt 1038 if pt_call: 1039 return nn.Module.__call__(self, *input, **kwargs) 1040 else: 1041 return NeuralModule.__call__(self, **kwargs) 1042 1043 type(module).__call__ = __old_call__ 1044 1045 @staticmethod 1046 def deployment_export(module, output: str, d_format: DeploymentFormat, input_example=None, output_example=None): 1047 """Exports Neural Module instance for deployment. 1048 1049 Args: 1050 module: neural module to export 1051 output (str): where export results should be saved 1052 d_format (DeploymentFormat): which deployment format to use 1053 input_example: sometimes tracing will require input examples 1054 output_example: Should match inference on input_example 1055 amp_max_loss_scale (float): Max value for amp loss scaling. 1056 Defaults to 2.0**24. 1057 """ 1058 1059 with torch.no_grad(): 1060 PtActions.__module_export( 1061 module=module, 1062 output=output, 1063 d_format=d_format, 1064 input_example=input_example, 1065 output_example=output_example, 1066 ) 1067 1068 def train( 1069 self, 1070 tensors_to_optimize, 1071 optimizer=None, 1072 optimization_params=None, 1073 callbacks: Optional[List[ActionCallback]] = None, 1074 lr_policy=None, 1075 batches_per_step=None, 1076 stop_on_nan_loss=False, 1077 synced_batchnorm=False, 1078 synced_batchnorm_groupsize=0, 1079 gradient_predivide=False, 1080 amp_max_loss_scale=2.0 ** 24, 1081 ): 1082 if gradient_predivide: 1083 logging.error( 1084 "gradient_predivide is currently disabled, and is under consideration for removal in future versions. " 1085 "If this functionality is needed, please raise a github issue." 1086 ) 1087 if not optimization_params: 1088 optimization_params = {} 1089 num_epochs = optimization_params.get("num_epochs", None) 1090 max_steps = optimization_params.get("max_steps", None) 1091 if num_epochs is None and max_steps is None: 1092 raise ValueError("You must specify either max_steps or num_epochs") 1093 grad_norm_clip = optimization_params.get('grad_norm_clip', None) 1094 1095 if batches_per_step is None: 1096 batches_per_step = 1 1097 # this is necessary because we average gradients over batch 1098 bps_scale = torch.FloatTensor([1.0 / batches_per_step]).squeeze() 1099 1100 if tensors_to_optimize is None: 1101 # This is Evaluation Mode 1102 self._init_callbacks(callbacks) 1103 # Do action start callbacks 1104 self._perform_on_action_end(callbacks=callbacks) 1105 return 1106 # Check if tensors_to_optimize is just a list of NmTensors 1107 elif tensors_to_optimize is not None and ( 1108 isinstance(tensors_to_optimize[0], NmTensor) and PtActions._check_all_tensors(tensors_to_optimize) 1109 ): 1110 # Parse graph into a topologically sorted sequence of neural 1111 # modules' calls 1112 (opt_call_chain, t_dataset,) = self.__get_top_sorted_modules_and_dataloader(hook=tensors_to_optimize) 1113 # Extract trainable weights which will be optimized 1114 params_list = [ 1115 p[0].parameters() for p in opt_call_chain if isinstance(p[0], TrainableNM) or p[0].is_trainable() 1116 ] 1117 params_to_optimize = itertools.chain(*params_list) 1118 1119 # Setup optimizer instance. By default it is SGD 1120 optimizer_instance = None 1121 optimizer_class = None 1122 if isinstance(optimizer, str): 1123 optimizer_class = optimizer 1124 elif isinstance(optimizer, torch.optim.Optimizer): 1125 optimizer_instance = optimizer 1126 else: 1127 raise ValueError("optimizer was not understood") 1128 optimizer = self.__setup_optimizer( 1129 optimizer_instance=optimizer_instance, 1130 optimizer_class=optimizer_class, 1131 optimization_params=optimization_params, 1132 params_to_optimize=params_to_optimize, 1133 ) 1134 1135 training_loop = [(optimizer, tensors_to_optimize, opt_call_chain)] 1136 1137 self.optimizers.append(optimizer) 1138 assert ( 1139 len(self.optimizers) == 1 1140 ), "There was more than one optimizer, was create_optimizer() called before train()?" 1141 1142 elif PtActions._check_tuples(tensors_to_optimize): 1143 if batches_per_step != 1: 1144 raise ValueError("Gradient accumlation with multiple optimizers is not supported") 1145 datasets = [] 1146 training_loop = [] 1147 for step in tensors_to_optimize: 1148 (step_call_chain, dataset,) = self.__get_top_sorted_modules_and_dataloader(hook=step[1]) 1149 datasets.append(dataset) 1150 training_loop.append((step[0], step[1], step_call_chain)) 1151 1152 t_dataset = datasets[0] 1153 for dataset in datasets: 1154 if type(dataset) is not type(t_dataset): 1155 raise ValueError("There were two training datasets, we only support 1.") 1156 else: 1157 raise ValueError("tensors_to_optimize was not understood") 1158 1159 logging_callchain = None 1160 # callbacks setup 1161 if callbacks is not None: 1162 for callback in callbacks: 1163 if not isinstance(callback, ActionCallback): 1164 raise ValueError("A callback was received that was not a child of ActionCallback") 1165 elif isinstance(callback, SimpleLossLoggerCallback): 1166 if logging_callchain: 1167 raise ValueError("We only support one logger callback but more than one were found") 1168 logger_step_freq = callback._step_freq 1169 logging_tensors = callback.tensors 1170 all_tensors = logging_tensors 1171 for step in training_loop: 1172 all_tensors = all_tensors + step[1] 1173 (logging_callchain, _,) = self.__get_top_sorted_modules_and_dataloader(hook=all_tensors) 1174 1175 self._get_all_modules(training_loop, callbacks, logging_callchain) 1176 1177 # Intialize Amp if needed 1178 if self._optim_level in AmpOptimizations: 1179 # Store mapping of self.optimizers to optimizer in callchain 1180 training_loop_opts = [] 1181 for opt in training_loop: 1182 training_loop_opts.append(self.optimizers.index(opt[0])) 1183 self.optimizers = self.__initialize_amp( 1184 optimizer=self.optimizers, 1185 optim_level=self._optim_level, 1186 amp_max_loss_scale=amp_max_loss_scale, 1187 amp_min_loss_scale=optimization_params.get('amp_min_loss_scale', 1.0), 1188 ) 1189 # Use stored mapping to map amp_init opts to training loop 1190 for i, step in enumerate(training_loop): 1191 training_loop[i] = ( 1192 self.optimizers[training_loop_opts[i]], 1193 step[1], 1194 step[2], 1195 ) 1196 1197 dataNM = training_loop[0][2][0][0] 1198 if dataNM.placement == DeviceType.AllGpu: 1199 # if len(training_loop) > 1: 1200 # raise NotImplementedError( 1201 # "Distributed training does nor work with multiple " 1202 # "optimizers") 1203 logging.info("Doing distributed training") 1204 if t_dataset is not None: 1205 train_sampler = torch.utils.data.distributed.DistributedSampler( 1206 dataset=t_dataset, shuffle=dataNM.shuffle 1207 ) 1208 train_dataloader = torch.utils.data.DataLoader( 1209 dataset=t_dataset, 1210 sampler=train_sampler, 1211 num_workers=dataNM.num_workers, 1212 batch_size=dataNM.batch_size, 1213 shuffle=False, 1214 ) 1215 else: 1216 train_dataloader = dataNM.data_iterator 1217 if hasattr(train_dataloader, 'sampler'): 1218 train_sampler = train_dataloader.sampler 1219 else: 1220 train_sampler = None 1221 1222 for train_iter in training_loop: 1223 call_chain = train_iter[2] 1224 for i in range(1, len(call_chain) - 1): 1225 key = call_chain[i][0].unique_instance_id 1226 pmodule = self.module_reference_table[key][1] 1227 if not isinstance(pmodule, DDP) and isinstance(pmodule, torch.nn.Module): 1228 # gpf = 1 1229 # if gradient_predivide: 1230 # gpf = dist.get_world_size() 1231 # pmodule = DDP(pmodule, gradient_predivide_factor=gpf) # Old Apex Method 1232 1233 # Per pytorch docs, convert sync bn prior to DDP 1234 if synced_batchnorm: 1235 world_size = dist.get_world_size() 1236 sync_batchnorm_group = None 1237 if synced_batchnorm_groupsize > 0: 1238 if world_size % synced_batchnorm_groupsize != 0: 1239 raise ValueError( 1240 f"Synchronized batch norm group size ({synced_batchnorm_groupsize}) must be 0" 1241 f" or divide total number of GPUs ({world_size})." 1242 ) 1243 # Find ranks of other nodes in the same batchnorm group 1244 rank = torch.distributed.get_rank() 1245 group = rank // synced_batchnorm_groupsize 1246 group_rank_ids = range( 1247 group * synced_batchnorm_groupsize, (group + 1) * synced_batchnorm_groupsize 1248 ) 1249 sync_batchnorm_group = torch.distributed.new_group(group_rank_ids) 1250 1251 pmodule = nn.SyncBatchNorm.convert_sync_batchnorm( 1252 pmodule, process_group=sync_batchnorm_group 1253 ) 1254 1255 # By default, disable broadcast_buffers. This disables batch norm synchronization on forward 1256 # pass 1257 pmodule = DDP( 1258 pmodule, device_ids=[self.local_rank], broadcast_buffers=False, find_unused_parameters=True 1259 ) 1260 1261 # # Convert batchnorm modules to synced if applicable 1262 # if synced_batchnorm and isinstance(pmodule, torch.nn.Module): 1263 # world_size = dist.get_world_size() 1264 # if synced_batchnorm_groupsize > 0 and world_size % synced_batchnorm_groupsize != 0: 1265 # raise ValueError( 1266 # f"Synchronized batch norm group size" 1267 # f" ({synced_batchnorm_groupsize}) must be 0" 1268 # f" or divide total number of GPUs" 1269 # f" ({world_size})." 1270 # ) 1271 # process_group = create_syncbn_process_group(synced_batchnorm_groupsize) 1272 # pmodule = convert_syncbn(pmodule, process_group=process_group) 1273 1274 self.module_reference_table[key] = ( 1275 self.module_reference_table[key][0], 1276 pmodule, 1277 ) 1278 # single GPU/CPU training 1279 else: 1280 if t_dataset is not None: 1281 train_sampler = None 1282 train_dataloader = torch.utils.data.DataLoader( 1283 dataset=t_dataset, 1284 sampler=None, 1285 num_workers=dataNM.num_workers, 1286 batch_size=dataNM.batch_size, 1287 shuffle=dataNM.shuffle, 1288 ) 1289 else: 1290 train_dataloader = dataNM.data_iterator 1291 train_sampler = None 1292 1293 self._init_callbacks(callbacks) 1294 # Do action start callbacks 1295 self._perform_on_action_start(callbacks=callbacks) 1296 1297 # MAIN TRAINING LOOP 1298 # iteration over epochs 1299 while num_epochs is None or self.epoch_num < num_epochs: 1300 if train_sampler is not None: 1301 train_sampler.set_epoch(self.epoch_num) 1302 if max_steps is not None and self.step >= max_steps: 1303 break 1304 1305 # Register epochs start with callbacks 1306 self._perform_on_epoch_start(callbacks=callbacks) 1307 1308 # iteration over batches in epoch 1309 batch_counter = 0 1310 for _, data in enumerate(train_dataloader, 0): 1311 if max_steps is not None and self.step >= max_steps: 1312 break 1313 1314 if batch_counter == 0: 1315 # Started step, zero gradients 1316 curr_optimizer = training_loop[self.step % len(training_loop)][0] 1317 curr_optimizer.zero_grad() 1318 # Register iteration start with callbacks 1319 self._perform_on_iteration_start(callbacks=callbacks) 1320 1321 # set learning rate policy 1322 if lr_policy is not None: 1323 adjusted_lr = lr_policy(optimization_params["lr"], self.step, self.epoch_num) 1324 for param_group in curr_optimizer.param_groups: 1325 param_group["lr"] = adjusted_lr 1326 if self.tb_writer is not None: 1327 value = curr_optimizer.param_groups[0]['lr'] 1328 self.tb_writer.add_scalar('param/lr', value, self.step) 1329 if callbacks is not None: 1330 for callback in callbacks: 1331 callback.learning_rate = curr_optimizer.param_groups[0]['lr'] 1332 1333 # registered_tensors will contain created tensors 1334 # named by output port and uuid of module which created them 1335 # Get and properly name tensors returned by data layer 1336 curr_call_chain = training_loop[self.step % len(training_loop)][2] 1337 dl_device = curr_call_chain[0][0]._device 1338 if logging_callchain and self.step % logger_step_freq == 0: 1339 curr_call_chain = logging_callchain 1340 tensors = [] 1341 if isinstance(data, torch.Tensor): 1342 data = (data,) 1343 for d in data: 1344 if isinstance(d, torch.Tensor): 1345 tensors.append(d.to(dl_device)) 1346 else: 1347 tensors.append(d) 1348 1349 registered_tensors = { 1350 t.unique_name: d for t, d in zip(curr_call_chain[0][2].values(), tensors) if t is not None 1351 } 1352 disable_allreduce = batch_counter < (batches_per_step - 1) 1353 self.__nm_graph_forward_pass( 1354 call_chain=curr_call_chain, registered_tensors=registered_tensors, 1355 ) 1356 1357 curr_tensors_to_optimize = training_loop[self.step % len(training_loop)][1] 1358 final_loss = 0 1359 nan = False 1360 for tensor in curr_tensors_to_optimize: 1361 if ( 1362 torch.isnan(registered_tensors[tensor.unique_name]).any() 1363 or torch.isinf(registered_tensors[tensor.unique_name]).any() 1364 ): 1365 if stop_on_nan_loss: 1366 raise ValueError('Loss is NaN or inf - exiting') 1367 logging.warning('Loss is NaN or inf') 1368 curr_optimizer.zero_grad() 1369 nan = True 1370 break 1371 final_loss += registered_tensors[tensor.unique_name] 1372 if nan: 1373 continue 1374 if self._optim_level in AmpOptimizations and self._optim_level != Optimization.mxprO0: 1375 with amp.scale_loss(final_loss, curr_optimizer, delay_unscale=disable_allreduce) as scaled_loss: 1376 if torch.isnan(scaled_loss).any() or torch.isinf(scaled_loss).any(): 1377 if stop_on_nan_loss: 1378 raise ValueError('Loss is NaN or inf -' ' exiting') 1379 logging.warning('WARNING: Loss is NaN or inf') 1380 curr_optimizer.zero_grad() 1381 continue 1382 if disable_allreduce: 1383 with ExitStack() as stack: 1384 for mod in self.get_DDP_modules(curr_call_chain): 1385 stack.enter_context(mod.no_sync()) 1386 scaled_loss.backward(bps_scale.to(scaled_loss.get_device())) 1387 else: 1388 scaled_loss.backward(bps_scale.to(scaled_loss.get_device())) 1389 # no AMP optimizations needed 1390 else: 1391 # multi-GPU, float32 1392 if self._local_rank is not None: 1393 if disable_allreduce: 1394 with ExitStack() as stack: 1395 for mod in self.get_DDP_modules(curr_call_chain): 1396 stack.enter_context(mod.no_sync()) 1397 final_loss.backward(bps_scale.to(final_loss.get_device())) 1398 else: 1399 final_loss.backward(bps_scale.to(final_loss.get_device())) 1400 # single device (CPU or GPU) 1401 else: 1402 # Fix (workaround?) enabling to backpropagate gradiens on CPUs. 1403 if final_loss.get_device() < 0: 1404 final_loss.backward(bps_scale) 1405 else: 1406 final_loss.backward(bps_scale.to(final_loss.get_device())) 1407 1408 batch_counter += 1 1409 1410 if batch_counter == batches_per_step: 1411 # Ended step. Do optimizer update 1412 if grad_norm_clip is not None: 1413 torch.nn.utils.clip_grad_norm_(master_params(curr_optimizer), grad_norm_clip) 1414 curr_optimizer.step() 1415 batch_counter = 0 1416 # Register iteration end with callbacks 1417 self._update_callbacks( 1418 callbacks=callbacks, registered_tensors=registered_tensors, 1419 ) 1420 self._perform_on_iteration_end(callbacks=callbacks) 1421 self.step += 1 1422 # End of epoch for loop 1423 # Register epochs end with callbacks 1424 self._perform_on_epoch_end(callbacks=callbacks) 1425 self.epoch_num += 1 1426 self._perform_on_action_end(callbacks=callbacks) 1427 1428 def infer( 1429 self, 1430 tensors, 1431 checkpoint_dir=None, 1432 ckpt_pattern='', 1433 verbose=True, 1434 cache=False, 1435 use_cache=False, 1436 offload_to_cpu=True, 1437 modules_to_restore=None, 1438 ): 1439 """See NeuralModuleFactory.infer() 1440 """ 1441 1442 call_chain, _ = self.__get_top_sorted_modules_and_dataloader(hook=tensors) 1443 if checkpoint_dir: 1444 # Find all modules that need to be restored 1445 if modules_to_restore is None: 1446 modules_to_restore = [] 1447 modules_to_restore_name = [] 1448 for op in call_chain: 1449 if op[0].num_weights > 0: 1450 modules_to_restore.append(op[0]) 1451 1452 if not isinstance(modules_to_restore, list): 1453 modules_to_restore = [modules_to_restore] 1454 modules_to_restore_name = [] 1455 for mod in modules_to_restore: 1456 if not isinstance(mod, NeuralModule): 1457 raise ValueError("Found something that was not a Neural Module inside modules_to_restore") 1458 elif mod.num_weights == 0: 1459 raise ValueError("Found a Neural Module with 0 weights inside modules_to_restore") 1460 modules_to_restore_name.append(str(mod)) 1461 1462 module_checkpoints = get_checkpoint_from_dir(modules_to_restore_name, checkpoint_dir, ckpt_pattern) 1463 1464 for mod, checkpoint in zip(modules_to_restore, module_checkpoints): 1465 logging.info(f"Restoring {mod} from {checkpoint}") 1466 mod.restore_from(checkpoint, self._local_rank) 1467 1468 # Init Amp 1469 if ( 1470 self._optim_level in AmpOptimizations 1471 and self._optim_level != Optimization.mxprO0 1472 and not self.amp_initialized 1473 ): 1474 pt_modules = [] 1475 for i in range(len(call_chain)): 1476 if isinstance(call_chain[i][0], nn.Module): 1477 pt_modules.append(call_chain[i][0]) 1478 elif isinstance(call_chain[i][0], TrainableNeuralModuleWrapper): 1479 pt_modules.append(call_chain[i][0]._pt_module) 1480 1481 amp.initialize( 1482 min_loss_scale=1.0, models=pt_modules, optimizers=None, opt_level=AmpOptimizations[self._optim_level], 1483 ) 1484 self.amp_initialized = True 1485 1486 # Run infer 1487 return self._infer( 1488 tensors_to_return=tensors, 1489 verbose=verbose, 1490 cache=cache, 1491 use_cache=use_cache, 1492 offload_to_cpu=offload_to_cpu, 1493 ) 1494 1495 def get_DDP_modules(self, call_chain): 1496 modules = [] 1497 for ind in range(1, len(call_chain)): 1498 m_id = call_chain[ind][0].unique_instance_id 1499 module = self.module_reference_table[m_id][1] 1500 if isinstance(module, DDP): 1501 modules.append(module) 1502 1503 return modules 1504 [end of nemo/backends/pytorch/actions.py] [start of nemo/collections/asr/jasper.py] 1 # Copyright (c) 2019 NVIDIA Corporation 2 from typing import Optional 3 4 import torch 5 import torch.nn as nn 6 import torch.nn.functional as F 7 8 import nemo 9 from .parts.jasper import JasperBlock, init_weights, jasper_activations 10 from nemo.backends.pytorch.nm import TrainableNM 11 from nemo.core.neural_types import * 12 from nemo.utils.decorators import add_port_docs 13 14 logging = nemo.logging 15 16 17 class JasperEncoder(TrainableNM): 18 """ 19 Jasper Encoder creates the pre-processing (prologue), Jasper convolution 20 block, and the first 3 post-processing (epilogue) layers as described in 21 Jasper (https://arxiv.org/abs/1904.03288) 22 23 Args: 24 jasper (list): A list of dictionaries. Each element in the list 25 represents the configuration of one Jasper Block. Each element 26 should contain:: 27 28 { 29 # Required parameters 30 'filters' (int) # Number of output channels, 31 'repeat' (int) # Number of sub-blocks, 32 'kernel' (int) # Size of conv kernel, 33 'stride' (int) # Conv stride 34 'dilation' (int) # Conv dilation 35 'dropout' (float) # Dropout probability 36 'residual' (bool) # Whether to use residual or not. 37 # Optional parameters 38 'residual_dense' (bool) # Whether to use Dense Residuals 39 # or not. 'residual' must be True for 'residual_dense' 40 # to be enabled. 41 # Defaults to False. 42 'separable' (bool) # Whether to use separable convolutions. 43 # Defaults to False 44 'groups' (int) # Number of groups in each conv layer. 45 # Defaults to 1 46 'heads' (int) # Sharing of separable filters 47 # Defaults to -1 48 'tied' (bool) # Whether to use the same weights for all 49 # sub-blocks. 50 # Defaults to False 51 'se' (bool) # Whether to add Squeeze and Excitation 52 # sub-blocks. 53 # Defaults to False 54 'se_reduction_ratio' (int) # The reduction ratio of the Squeeze 55 # sub-module. 56 # Must be an integer > 1. 57 # Defaults to 16 58 'kernel_size_factor' (float) # Conv kernel size multiplier 59 # Can be either an int or float 60 # Kernel size is recomputed as below: 61 # new_kernel_size = int(max(1, (kernel_size * kernel_width))) 62 # to prevent kernel sizes than 1. 63 # Note: If rescaled kernel size is an even integer, 64 # adds 1 to the rescaled kernel size to allow "same" 65 # padding. 66 } 67 68 activation (str): Activation function used for each sub-blocks. Can be 69 one of ["hardtanh", "relu", "selu"]. 70 feat_in (int): Number of channels being input to this module 71 normalization_mode (str): Normalization to be used in each sub-block. 72 Can be one of ["batch", "layer", "instance", "group"] 73 Defaults to "batch". 74 residual_mode (str): Type of residual connection. 75 Can be "add" or "max". 76 Defaults to "add". 77 norm_groups (int): Number of groups for "group" normalization type. 78 If set to -1, number of channels is used. 79 Defaults to -1. 80 conv_mask (bool): Controls the use of sequence length masking prior 81 to convolutions. 82 Defaults to True. 83 frame_splicing (int): Defaults to 1. 84 init_mode (str): Describes how neural network parameters are 85 initialized. Options are ['xavier_uniform', 'xavier_normal', 86 'kaiming_uniform','kaiming_normal']. 87 Defaults to "xavier_uniform". 88 """ 89 90 length: Optional[torch.Tensor] 91 92 @property 93 @add_port_docs() 94 def input_ports(self): 95 """Returns definitions of module input ports. 96 """ 97 return { 98 # "audio_signal": NeuralType( 99 # {0: AxisType(BatchTag), 1: AxisType(SpectrogramSignalTag), 2: AxisType(ProcessedTimeTag),} 100 # ), 101 # "length": NeuralType({0: AxisType(BatchTag)}), 102 "audio_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 103 "length": NeuralType(tuple('B'), LengthsType()), 104 } 105 106 @property 107 @add_port_docs() 108 def output_ports(self): 109 """Returns definitions of module output ports. 110 """ 111 return { 112 # "outputs": NeuralType( 113 # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag),} 114 # ), 115 # "encoded_lengths": NeuralType({0: AxisType(BatchTag)}), 116 "outputs": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), 117 "encoded_lengths": NeuralType(tuple('B'), LengthsType()), 118 } 119 120 @property 121 def disabled_deployment_input_ports(self): 122 return set(["length"]) 123 124 @property 125 def disabled_deployment_output_ports(self): 126 return set(["encoded_lengths"]) 127 128 def prepare_for_deployment(self): 129 m_count = 0 130 for m in self.modules(): 131 if type(m).__name__ == "MaskedConv1d": 132 m.use_mask = False 133 m_count += 1 134 logging.warning(f"Turned off {m_count} masked convolutions") 135 136 def __init__( 137 self, 138 jasper, 139 activation, 140 feat_in, 141 normalization_mode="batch", 142 residual_mode="add", 143 norm_groups=-1, 144 conv_mask=True, 145 frame_splicing=1, 146 init_mode='xavier_uniform', 147 ): 148 super().__init__() 149 150 activation = jasper_activations[activation]() 151 feat_in = feat_in * frame_splicing 152 153 residual_panes = [] 154 encoder_layers = [] 155 self.dense_residual = False 156 for lcfg in jasper: 157 dense_res = [] 158 if lcfg.get('residual_dense', False): 159 residual_panes.append(feat_in) 160 dense_res = residual_panes 161 self.dense_residual = True 162 groups = lcfg.get('groups', 1) 163 separable = lcfg.get('separable', False) 164 heads = lcfg.get('heads', -1) 165 se = lcfg.get('se', False) 166 se_reduction_ratio = lcfg.get('se_reduction_ratio', 16) 167 kernel_size_factor = lcfg.get('kernel_size_factor', 1.0) 168 encoder_layers.append( 169 JasperBlock( 170 feat_in, 171 lcfg['filters'], 172 repeat=lcfg['repeat'], 173 kernel_size=lcfg['kernel'], 174 stride=lcfg['stride'], 175 dilation=lcfg['dilation'], 176 dropout=lcfg['dropout'], 177 residual=lcfg['residual'], 178 groups=groups, 179 separable=separable, 180 heads=heads, 181 residual_mode=residual_mode, 182 normalization=normalization_mode, 183 norm_groups=norm_groups, 184 activation=activation, 185 residual_panes=dense_res, 186 conv_mask=conv_mask, 187 se=se, 188 se_reduction_ratio=se_reduction_ratio, 189 kernel_size_factor=kernel_size_factor, 190 ) 191 ) 192 feat_in = lcfg['filters'] 193 194 self.encoder = nn.Sequential(*encoder_layers) 195 self.apply(lambda x: init_weights(x, mode=init_mode)) 196 self.to(self._device) 197 198 def forward(self, audio_signal, length=None): 199 # type: (Tensor, Optional[Tensor]) -> Tensor, Optional[Tensor] 200 201 s_input, length = self.encoder(([audio_signal], length)) 202 if length is None: 203 return s_input[-1] 204 return s_input[-1], length 205 206 207 class JasperDecoderForCTC(TrainableNM): 208 """ 209 Jasper Decoder creates the final layer in Jasper that maps from the outputs 210 of Jasper Encoder to the vocabulary of interest. 211 212 Args: 213 feat_in (int): Number of channels being input to this module 214 num_classes (int): Number of characters in ASR model's vocab/labels. 215 This count should not include the CTC blank symbol. 216 init_mode (str): Describes how neural network parameters are 217 initialized. Options are ['xavier_uniform', 'xavier_normal', 218 'kaiming_uniform','kaiming_normal']. 219 Defaults to "xavier_uniform". 220 """ 221 222 @property 223 @add_port_docs() 224 def input_ports(self): 225 """Returns definitions of module input ports. 226 """ 227 return { 228 # "encoder_output": NeuralType( 229 # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag),} 230 # ) 231 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()) 232 } 233 234 @property 235 @add_port_docs() 236 def output_ports(self): 237 """Returns definitions of module output ports. 238 """ 239 # return {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag), 2: AxisType(ChannelTag),})} 240 return {"output": NeuralType(('B', 'T', 'D'), LogprobsType())} 241 242 def __init__(self, feat_in, num_classes, init_mode="xavier_uniform"): 243 super().__init__() 244 245 self._feat_in = feat_in 246 # Add 1 for blank char 247 self._num_classes = num_classes + 1 248 249 self.decoder_layers = nn.Sequential(nn.Conv1d(self._feat_in, self._num_classes, kernel_size=1, bias=True)) 250 self.apply(lambda x: init_weights(x, mode=init_mode)) 251 self.to(self._device) 252 253 def forward(self, encoder_output): 254 return F.log_softmax(self.decoder_layers(encoder_output).transpose(1, 2), dim=-1) 255 256 257 class JasperDecoderForClassification(TrainableNM): 258 """ 259 Jasper Decoder creates the final layer in Jasper that maps from the outputs 260 of Jasper Encoder to one class label. 261 262 Args: 263 feat_in (int): Number of channels being input to this module 264 num_classes (int): Number of characters in ASR model's vocab/labels. 265 This count should not include the CTC blank symbol. 266 init_mode (str): Describes how neural network parameters are 267 initialized. Options are ['xavier_uniform', 'xavier_normal', 268 'kaiming_uniform','kaiming_normal']. 269 Defaults to "xavier_uniform". 270 """ 271 272 @property 273 def input_ports(self): 274 """Returns definitions of module input ports. 275 """ 276 return { 277 # "encoder_output": NeuralType( 278 # {0: AxisType(BatchTag), 1: AxisType(EncodedRepresentationTag), 2: AxisType(ProcessedTimeTag)} 279 # ) 280 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()) 281 } 282 283 @property 284 def output_ports(self): 285 """Returns definitions of module output ports. 286 """ 287 # return {"logits": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} 288 return {"logits": NeuralType(('B', 'D'), LogitsType())} 289 290 def __init__( 291 self, *, feat_in, num_classes, init_mode="xavier_uniform", return_logits=True, pooling_type='avg', **kwargs 292 ): 293 TrainableNM.__init__(self, **kwargs) 294 295 self._feat_in = feat_in 296 self._return_logits = return_logits 297 self._num_classes = num_classes 298 299 if pooling_type == 'avg': 300 self.pooling = nn.AdaptiveAvgPool1d(1) 301 elif pooling_type == 'max': 302 self.pooling = nn.AdaptiveMaxPool1d(1) 303 else: 304 raise ValueError('Pooling type chosen is not valid. Must be either `avg` or `max`') 305 306 self.decoder_layers = nn.Sequential(nn.Linear(self._feat_in, self._num_classes, bias=True)) 307 self.apply(lambda x: init_weights(x, mode=init_mode)) 308 self.to(self._device) 309 310 def forward(self, encoder_output): 311 batch, in_channels, timesteps = encoder_output.size() 312 313 encoder_output = self.pooling(encoder_output).view(batch, in_channels) # [B, C] 314 logits = self.decoder_layers(encoder_output) # [B, num_classes] 315 316 if self._return_logits: 317 return logits 318 319 return F.softmax(logits, dim=-1) 320 [end of nemo/collections/asr/jasper.py] [start of nemo/core/neural_factory.py] 1 # ! /usr/bin/python 2 # -*- coding: utf-8 -*- 3 4 # Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. 5 # 6 # Licensed under the Apache License, Version 2.0 (the "License"); 7 # you may not use this file except in compliance with the License. 8 # You may obtain a copy of the License at 9 # 10 # http://www.apache.org/licenses/LICENSE-2.0 11 # 12 # Unless required by applicable law or agreed to in writing, software 13 # distributed under the License is distributed on an "AS IS" BASIS, 14 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 15 # See the License for the specific language governing permissions and 16 # limitations under the License. 17 18 __all__ = [ 19 'Backend', 20 'ModelMode', 21 'Optimization', 22 'DeviceType', 23 'Actions', 24 'NeuralModuleFactory', 25 'DeploymentFormat', 26 ] 27 28 import random 29 from abc import ABC, abstractmethod 30 from enum import Enum 31 from typing import List, Optional 32 33 import numpy as np 34 35 import nemo 36 from ..utils import ExpManager 37 from .callbacks import ActionCallback, EvaluatorCallback 38 from .neural_types import * 39 from nemo.utils.decorators import deprecated 40 41 logging = nemo.logging 42 43 44 class DeploymentFormat(Enum): 45 """Which format to use when exporting a Neural Module for deployment""" 46 47 AUTO = 0 48 PYTORCH = 1 49 TORCHSCRIPT = 2 50 ONNX = 3 51 TRTONNX = 4 52 53 54 class Backend(Enum): 55 """Supported backends. For now, it is only PyTorch.""" 56 57 PyTorch = 1 58 NotSupported = 2 59 60 61 class ModelMode(Enum): 62 """Training Mode or Evaluation/Inference""" 63 64 train = 0 65 eval = 1 66 67 68 class Optimization(Enum): 69 """Various levels of Apex/amp Optimization. 70 WARNING: This might have effect on model accuracy.""" 71 72 mxprO0 = 0 73 mxprO1 = 1 74 mxprO2 = 2 75 mxprO3 = 3 76 77 78 class DeviceType(Enum): 79 """Device types where Neural Modules can be placed.""" 80 81 GPU = 1 82 CPU = 2 83 AllGpu = 3 84 85 86 class Actions(ABC): 87 """Basic actions allowed on graphs of Neural Modules""" 88 89 def __init__(self, local_rank, global_rank, optimization_level=Optimization.mxprO0): 90 self._local_rank = local_rank 91 self._global_rank = global_rank 92 self._optim_level = optimization_level 93 self.step = None 94 self.epoch_num = None 95 96 @property 97 def local_rank(self): 98 """Local rank during distributed execution. None if single GPU/CPU 99 100 Returns: 101 (int) rank or worker or None if not in distributed model 102 """ 103 return self._local_rank 104 105 @property 106 def global_rank(self): 107 """Global rank during distributed execution. None if single GPU/CPU 108 109 Returns: 110 (int) rank or worker or None if not in distributed model 111 """ 112 return self._global_rank 113 114 @abstractmethod 115 def train( 116 self, 117 tensors_to_optimize: List[NmTensor], 118 callbacks: Optional[List[ActionCallback]], 119 lr_policy=None, 120 batches_per_step=None, 121 stop_on_nan_loss=False, 122 ): 123 """This action executes training and (optionally) evaluation. 124 125 Args: 126 tensors_to_optimize: which tensors to optimize. Typically this is 127 single loss tesnor. 128 callbacks: list of callback objects 129 lr_policy: function which should take (initial_lr, step, epoch) and 130 return learning rate 131 batches_per_step: number of mini-batches to process before one 132 optimizer step. (default: None, same as 1). Use this 133 to simulate larger batch sizes on hardware which could not fit 134 larger batch in memory otherwise. Effectively, this will make 135 "algorithmic" batch size per GPU/worker = batches_per_step* 136 batch_size 137 stop_on_nan_loss: (default: False) If set to True, the training 138 will stop if loss=nan. If set to False, the training will 139 continue, but the gradients will be zeroed before next 140 mini-batch. 141 142 Returns: 143 None 144 """ 145 pass 146 147 @abstractmethod 148 def infer(self, tensors: List[NmTensor]): 149 """This action executes inference. Nothing is optimized. 150 Args: 151 tensors: which tensors to evaluate. 152 153 Returns: 154 None 155 """ 156 pass 157 158 @abstractmethod 159 def save_state_to(self, path: str): 160 """ 161 Saves current state such as step, epoch and optimizer parameters 162 Args: 163 path: 164 165 Returns: 166 167 """ 168 pass 169 170 @abstractmethod 171 def restore_state_from(self, path: str): 172 """ 173 Restores state such as step, epoch and optimizer parameters 174 Args: 175 path: 176 177 Returns: 178 179 """ 180 pass 181 182 @abstractmethod 183 def create_optimizer(self, optimizer, things_to_optimize, optimizer_params): 184 """ 185 Creates an optimizer object to be use in the train() method. 186 187 Args: 188 optimizer: Specifies which optimizer to use. 189 things_to_optimize: A list of neural modules or tensors to be 190 optimized. 191 optimizer_params: Specifies the parameters of the optimizer 192 193 Returns: 194 Optimizer 195 """ 196 pass 197 198 def _perform_on_iteration_start(self, callbacks): 199 # TODO: Most of these checks can be relaxed since we enforce callbacks 200 # to be a list of ActionCallback objects 201 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 202 for callback in callbacks: 203 callback.on_iteration_start() 204 205 def _perform_on_iteration_end(self, callbacks): 206 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 207 for callback in callbacks: 208 callback.on_iteration_end() 209 210 def _perform_on_action_start(self, callbacks): 211 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 212 for callback in callbacks: 213 callback.on_action_start() 214 215 def _perform_on_action_end(self, callbacks): 216 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 217 for callback in callbacks: 218 callback.on_action_end() 219 220 def _perform_on_epoch_start(self, callbacks): 221 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 222 for callback in callbacks: 223 callback.on_epoch_start() 224 225 def _perform_on_epoch_end(self, callbacks): 226 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 227 for callback in callbacks: 228 callback.on_epoch_end() 229 230 def _init_callbacks(self, callbacks): 231 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 232 for callback in callbacks: 233 callback.action = self 234 235 def _update_callbacks( 236 self, callbacks=None, registered_tensors=None, 237 ): 238 # if self.local_rank is None or self.local_rank == 0: 239 if callbacks is not None and isinstance(callbacks, List) and len(callbacks) > 0: 240 for callback in callbacks: 241 callback._registered_tensors = registered_tensors 242 243 244 def _str_to_opt_level(opt_str: str) -> Optimization: 245 number = int(opt_str[1:]) 246 if number not in Optimization._value2member_map_: 247 raise ValueError(f"Unknown optimization value {opt_str}") 248 return Optimization(number) 249 250 251 class NeuralModuleFactory(object): 252 _DEFAULT = None 253 254 """ 255 Neural Module Factory instance is used to create neural modules and 256 trainers 257 258 Args: 259 backend (Backend): Currently only Backend.PyTorch is supported 260 local_rank (int): Process rank. Should be set by distributed runner 261 optimization_level (Optimization): Level of optimization to use. Will 262 be passed to neural modules and actions created by this factory. 263 placement (DeviceType: where to place NeuralModule instances by default 264 cudnn_benchmark (bool): (default False) If set to True it will use 265 cudnnFind method to find the best kernels instead of using 266 heuristics. If the shapes of your inputs are constant this 267 should help, for various shapes it can slow things down. Give it 268 few iterations to warmup if set to True. Currently only supported 269 by PyTorch backend. 270 random_seed (int): (default None) Sets random seed to control for 271 randomness. This should be used for debugging purposes as it might 272 have negative impact on performance. Can't be used when 273 `cudnn_benchmark=True`. 274 master_process (bool): (default True) Flag for master process 275 indication 276 set_default (bool): (default True) True if should set this instance as 277 default factory for modules instantiating. 278 """ 279 280 def __init__( 281 self, 282 backend=Backend.PyTorch, 283 local_rank=None, 284 optimization_level=Optimization.mxprO0, 285 placement=None, 286 cudnn_benchmark=False, 287 random_seed=None, 288 set_default=True, 289 log_dir=None, 290 checkpoint_dir=None, 291 tensorboard_dir=None, 292 create_tb_writer=False, 293 files_to_copy=None, 294 add_time_to_log_dir=False, 295 ): 296 self._local_rank = local_rank 297 self._global_rank = None 298 299 if isinstance(optimization_level, str): 300 optimization_level = _str_to_opt_level(optimization_level) 301 self._optim_level = optimization_level 302 303 if placement is None: 304 if local_rank is not None: 305 device = DeviceType.AllGpu 306 else: 307 device = DeviceType.GPU 308 309 self._placement = device 310 else: 311 self._placement = placement 312 313 self._backend = backend 314 self._world_size = 1 315 broadcast_func = None 316 if backend == Backend.PyTorch: 317 # TODO: Move all framework specific code from this file 318 import torch 319 320 if self._placement != DeviceType.CPU: 321 if not torch.cuda.is_available(): 322 raise ValueError( 323 "You requested to use GPUs but CUDA is " 324 "not installed. You can try running using" 325 " CPU-only. To do this, instantiate your" 326 " factory with placement=DeviceType.CPU" 327 "\n" 328 "Note that this is slow and is not " 329 "well supported." 330 ) 331 332 torch.backends.cudnn.benchmark = cudnn_benchmark 333 if random_seed is not None and cudnn_benchmark: 334 raise ValueError("cudnn_benchmark can not be set to True when random_seed is not None.") 335 if random_seed is not None: 336 torch.backends.cudnn.deterministic = True 337 torch.backends.cudnn.benchmark = False 338 torch.manual_seed(random_seed) 339 np.random.seed(random_seed) 340 random.seed(random_seed) 341 342 if self._local_rank is not None: 343 torch.distributed.init_process_group(backend="nccl", init_method="env://") 344 345 cuda_set = True 346 # Try to set cuda device. This should fail if self._local_rank 347 # is greater than the number of available GPUs 348 try: 349 torch.cuda.set_device(self._local_rank) 350 except RuntimeError: 351 # Note in this case, all tensors are now sent to GPU 0 352 # who could crash because of OOM. Thus init_process_group() 353 # must be done before any cuda tensors are allocated 354 cuda_set = False 355 cuda_set_t = torch.cuda.IntTensor([cuda_set]) 356 357 # Do an all_reduce to ensure all workers obtained a GPU 358 # For the strangest reason, BAND doesn't work so I am resorting 359 # to MIN. 360 torch.distributed.all_reduce(cuda_set_t, op=torch.distributed.ReduceOp.MIN) 361 if cuda_set_t.item() == 0: 362 raise RuntimeError( 363 "There was an error initializing distributed training." 364 " Perhaps you specified more gpus than you have " 365 "available" 366 ) 367 368 del cuda_set_t 369 torch.cuda.empty_cache() 370 # Remove test tensor from memory 371 372 self._world_size = torch.distributed.get_world_size() 373 self._global_rank = torch.distributed.get_rank() 374 375 def torch_broadcast_wrapper(str_len=None, string=None, src=0): 376 """Wrapper function to broadcast string values across all 377 workers 378 """ 379 # Create byte cuda torch tensor 380 if string is not None: 381 string_tensor = torch.tensor(list(string.encode()), dtype=torch.uint8).cuda() 382 else: 383 string_tensor = torch.tensor([0] * str_len, dtype=torch.uint8).cuda() 384 # Run broadcast 385 torch.distributed.broadcast(string_tensor, src) 386 # turn byte tensor back to string 387 return_string = string_tensor.cpu().numpy() 388 return_string = b''.join(return_string).decode() 389 return return_string 390 391 broadcast_func = torch_broadcast_wrapper 392 else: 393 raise NotImplementedError("Only Pytorch backend is currently supported.") 394 395 # Create ExpManager 396 # if log_dir is None, only create logger 397 self._exp_manager = ExpManager( 398 work_dir=log_dir, 399 ckpt_dir=checkpoint_dir, 400 use_tb=create_tb_writer, 401 tb_dir=tensorboard_dir, 402 local_rank=local_rank, 403 global_rank=self._global_rank, 404 files_to_copy=files_to_copy, 405 add_time=add_time_to_log_dir, 406 exist_ok=True, 407 broadcast_func=broadcast_func, 408 ) 409 self._tb_writer = self._exp_manager.tb_writer 410 411 # Create trainer 412 self._trainer = self._get_trainer(tb_writer=self._tb_writer) 413 414 if set_default: 415 NeuralModuleFactory.set_default_factory(self) 416 417 @classmethod 418 def get_default_factory(cls): 419 return cls._DEFAULT 420 421 @classmethod 422 def set_default_factory(cls, factory): 423 cls._DEFAULT = factory 424 425 @classmethod 426 def reset_default_factory(cls): 427 cls._DEFAULT = None 428 429 @staticmethod 430 def __name_import(name): 431 components = name.split(".") 432 mod = __import__(components[0]) 433 for comp in components[1:]: 434 mod = getattr(mod, comp) 435 return mod 436 437 @deprecated(version=0.11) 438 def __get_pytorch_module(self, name, collection, params, pretrained): 439 # TK: "factory" is not passed as parameter anymore. 440 # params["factory"] = self 441 442 if collection == "toys" or collection == "tutorials" or collection == "other": 443 constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.tutorials." + name) 444 elif collection == "nemo_nlp": 445 constructor = NeuralModuleFactory.__name_import("nemo_nlp." + name) 446 if name == "BERT" and pretrained is True: 447 params["pretrained"] = True 448 elif collection == "nemo_asr": 449 constructor = NeuralModuleFactory.__name_import("nemo_asr." + name) 450 elif collection == "nemo_lpr": 451 constructor = NeuralModuleFactory.__name_import("nemo_lpr." + name) 452 elif collection == 'common': 453 constructor = NeuralModuleFactory.__name_import('nemo.backends.pytorch.common.' + name) 454 elif collection == "torchvision": 455 import torchvision.models as tv_models 456 import nemo.backends.pytorch.module_wrapper as mw 457 import torch.nn as nn 458 459 if name == "ImageFolderDataLayer": 460 constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.torchvision.data." + name) 461 instance = constructor(**params) 462 return instance 463 else: 464 _nm_name = name.lower() 465 if _nm_name == "resnet18": 466 input_ports = { 467 "x": NeuralType( 468 { 469 0: AxisType(BatchTag), 470 1: AxisType(ChannelTag), 471 2: AxisType(HeightTag, 224), 472 3: AxisType(WidthTag, 224), 473 } 474 ) 475 } 476 output_ports = {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} 477 478 pt_model = tv_models.resnet18(pretrained=pretrained) 479 num_classes = params.get("num_classes", None) 480 if num_classes is not None: 481 pt_model.fc = nn.Linear(512, params["num_classes"]) 482 return mw.TrainableNeuralModuleWrapper( 483 pt_nn_module=pt_model, input_ports_dict=input_ports, output_ports_dict=output_ports, 484 ) 485 elif _nm_name == "resnet50": 486 input_ports = { 487 "x": NeuralType( 488 { 489 0: AxisType(BatchTag), 490 1: AxisType(ChannelTag), 491 2: AxisType(HeightTag, 224), 492 3: AxisType(WidthTag, 224), 493 } 494 ) 495 } 496 output_ports = {"output": NeuralType({0: AxisType(BatchTag), 1: AxisType(ChannelTag)})} 497 498 pt_model = tv_models.resnet50(pretrained=pretrained) 499 num_classes = params.get("num_classes", None) 500 if num_classes is not None: 501 pt_model.fc = nn.Linear(2048, params["num_classes"]) 502 return mw.TrainableNeuralModuleWrapper( 503 pt_nn_module=pt_model, input_ports_dict=input_ports, output_ports_dict=output_ports, 504 ) 505 else: 506 collection_path = "nemo.collections." + collection + "." + name 507 constructor = NeuralModuleFactory.__name_import(collection_path) 508 if name == "BERT" and pretrained is True: 509 params["pretrained"] = True 510 511 # TK: "placement" is not passed as parameter anymore. 512 # if "placement" not in params: 513 # params["placement"] = self._placement 514 instance = constructor(**params) 515 return instance 516 517 @deprecated(version=0.11) 518 def get_module(self, name, collection, params, pretrained=False): 519 """ 520 Creates NeuralModule instance 521 522 Args: 523 name (str): name of NeuralModule which instance should be returned. 524 params (dict): local parameters which should be passed to 525 NeuralModule's constructor. 526 collection (str): in which collection to look for 527 `neural_module_name` 528 pretrained (bool): return pre-trained instance or randomly 529 initialized (default) 530 531 Returns: 532 NeuralModule instance 533 """ 534 535 # TK: "optimization_level" is not passed as parameter anymore. 536 # if params is not None and "optimization_level" in params: 537 # if params["optimization_level"] != self._optim_level: 538 # logging.warning( 539 # "Module's {0} requested optimization level {1} is" 540 # "different from the one specified by factory - {2}." 541 # "Using: {3} for this module".format( 542 # name, params["optimization_level"], self._optim_level, params["optimization_level"], 543 # ) 544 # ) 545 # else: 546 # if params is None: 547 # params = {} 548 # params["optimization_level"] = self._optim_level 549 550 if self._backend == Backend.PyTorch: 551 return self.__get_pytorch_module(name=name, collection=collection, params=params, pretrained=pretrained,) 552 else: 553 return None 554 555 def create_optimizer(self, optimizer, things_to_optimize, optimizer_params): 556 return self._trainer.create_optimizer( 557 optimizer=optimizer, things_to_optimize=things_to_optimize, optimizer_params=optimizer_params, 558 ) 559 560 def train( 561 self, 562 tensors_to_optimize, 563 optimizer=None, 564 optimization_params=None, 565 callbacks: Optional[List[ActionCallback]] = None, 566 lr_policy=None, 567 batches_per_step=None, 568 stop_on_nan_loss=False, 569 synced_batchnorm=False, 570 synced_batchnorm_groupsize=0, 571 gradient_predivide=False, 572 amp_max_loss_scale=2.0 ** 24, 573 reset=False, 574 ): 575 if reset: 576 self.reset_trainer() 577 return self._trainer.train( 578 tensors_to_optimize=tensors_to_optimize, 579 optimizer=optimizer, 580 optimization_params=optimization_params, 581 callbacks=callbacks, 582 lr_policy=lr_policy, 583 batches_per_step=batches_per_step, 584 stop_on_nan_loss=stop_on_nan_loss, 585 synced_batchnorm=synced_batchnorm, 586 synced_batchnorm_groupsize=synced_batchnorm_groupsize, 587 gradient_predivide=gradient_predivide, 588 amp_max_loss_scale=amp_max_loss_scale, 589 ) 590 591 def eval(self, callbacks: List[EvaluatorCallback]): 592 if callbacks is None or len(callbacks) == 0: 593 raise ValueError(f"You need to provide at lease one evaluation" f"callback to eval") 594 for callback in callbacks: 595 if not isinstance(callback, EvaluatorCallback): 596 raise TypeError(f"All callbacks passed to the eval action must" f"be inherited from EvaluatorCallback") 597 self.train( 598 tensors_to_optimize=None, optimizer='sgd', callbacks=callbacks, optimization_params={'num_epochs': 1}, 599 ) 600 601 def deployment_export( 602 self, module, output: str, d_format: DeploymentFormat, input_example=None, output_example=None 603 ): 604 """Exports Neural Module instance for deployment. 605 606 Args: 607 module: neural module to export 608 output (str): where export results should be saved 609 d_format (DeploymentFormat): which deployment format to use 610 input_example: sometimes tracing will require input examples 611 output_example: Should match inference on input_example 612 """ 613 module.prepare_for_deployment() 614 615 return self._trainer.deployment_export( 616 module=module, 617 output=output, 618 d_format=d_format, 619 input_example=input_example, 620 output_example=output_example, 621 ) 622 623 def infer( 624 self, 625 tensors: List[NmTensor], 626 checkpoint_dir=None, 627 ckpt_pattern='', 628 verbose=True, 629 cache=False, 630 use_cache=False, 631 offload_to_cpu=True, 632 modules_to_restore=None, 633 ): 634 """Runs inference to obtain values for tensors 635 636 Args: 637 tensors (list[NmTensor]): List of NeMo tensors that we want to get 638 values of. 639 checkpoint_dir (str): Path to checkpoint directory. Default is None 640 which does not load checkpoints. 641 ckpt_pattern (str): Pattern used to check for checkpoints inside 642 checkpoint_dir. Default is '' which matches any checkpoints 643 inside checkpoint_dir. 644 verbose (bool): Controls printing. Defaults to True. 645 cache (bool): If True, cache all `tensors` and intermediate tensors 646 so that future calls that have use_cache set will avoid 647 computation. Defaults to False. 648 use_cache (bool): Values from `tensors` will be always re-computed. 649 It will re-use intermediate tensors from the DAG leading to 650 `tensors`. If you want something to be re-computed, put it into 651 `tensors` list. Defaults to False. 652 offload_to_cpu (bool): If True, all evaluated tensors are moved to 653 cpu memory after each inference batch. Defaults to True. 654 modules_to_restore (list): Defaults to None, in which case all 655 NMs inside callchain with weights will be restored. If 656 specified only the modules inside this list will be restored. 657 658 Returns: 659 List of evaluated tensors. Each element in the list is also a list 660 where each element is now a batch of tensor values. 661 """ 662 return self._trainer.infer( 663 tensors=tensors, 664 checkpoint_dir=checkpoint_dir, 665 ckpt_pattern=ckpt_pattern, 666 verbose=verbose, 667 cache=cache, 668 use_cache=use_cache, 669 offload_to_cpu=offload_to_cpu, 670 modules_to_restore=modules_to_restore, 671 ) 672 673 def clear_cache(self): 674 """Helper function to clean inference cache.""" 675 self._trainer.clear_cache() 676 677 @deprecated(version="future") 678 def _get_trainer(self, tb_writer=None): 679 if self._backend == Backend.PyTorch: 680 constructor = NeuralModuleFactory.__name_import("nemo.backends.pytorch.PtActions") 681 instance = constructor( 682 local_rank=self._local_rank, 683 global_rank=self._global_rank, 684 tb_writer=tb_writer, 685 optimization_level=self._optim_level, 686 ) 687 return instance 688 else: 689 raise ValueError("Only PyTorch backend is currently supported.") 690 691 @deprecated( 692 version="future", 693 explanation="Please use .train(...), .eval(...), .infer(...) and " 694 f".create_optimizer(...) of the NeuralModuleFactory instance directly.", 695 ) 696 def get_trainer(self, tb_writer=None): 697 if self._trainer: 698 logging.warning( 699 "The trainer instance was created during initialization of " 700 "Neural factory, using the already created instance." 701 ) 702 return self._trainer 703 return self._get_trainer(tb_writer) 704 705 def reset_trainer(self): 706 del self._trainer 707 self._trainer = self._get_trainer(tb_writer=self._tb_writer) 708 709 def sync_all_processes(self, status=True): 710 """ Helper function for testing that allows proccess 0 to inform all 711 other processes of failures. Does nothing if not using distributed 712 training. Usage example can be seen in examples/asr/jasper_an4.py 713 714 Args: 715 status (bool): Defaults to True. If any proccess passes False, it 716 will trigger a graceful exit on all other processes. It is 717 assumed that the process that passed False will print an error 718 message on its own and exit 719 """ 720 if self._world_size == 1: 721 logging.info("sync_all_processes does nothing if there is one process") 722 return 723 if self._backend == Backend.PyTorch: 724 import torch 725 726 status_tensor = torch.cuda.IntTensor([status]) 727 torch.distributed.all_reduce(status_tensor, op=torch.distributed.ReduceOp.MIN) 728 if status_tensor.item() == 0: 729 logging.error("At least one process had a failure") 730 if status: 731 raise ValueError( 732 f"Process with global rank {self._global_rank} entered" 733 " sync_all_processes with a passing status, but " 734 "another process indicated a failure" 735 ) 736 737 @property 738 def world_size(self): 739 return self._world_size 740 741 @property 742 def tb_writer(self): 743 return self._tb_writer 744 745 @property 746 def placement(self): 747 return self._placement 748 749 @property 750 def optim_level(self): 751 return self._optim_level 752 753 @property 754 @deprecated(version=0.11, explanation="Please use ``nemo.logging instead``") 755 def logger(self): 756 return nemo.logging 757 758 @property 759 def checkpoint_dir(self): 760 return self._exp_manager.ckpt_dir 761 762 @property 763 def work_dir(self): 764 return self._exp_manager.work_dir 765 766 @property 767 def global_rank(self): 768 return self._global_rank 769 [end of nemo/core/neural_factory.py] [start of nemo/core/neural_modules.py] 1 # ! /usr/bin/python 2 # -*- coding: utf-8 -*- 3 4 # Copyright (c) 2019-, NVIDIA CORPORATION. All rights reserved. 5 # 6 # Licensed under the Apache License, Version 2.0 (the "License"); 7 # you may not use this file except in compliance with the License. 8 # You may obtain a copy of the License at 9 # 10 # http://www.apache.org/licenses/LICENSE-2.0 11 # 12 # Unless required by applicable law or agreed to in writing, software 13 # distributed under the License is distributed on an "AS IS" BASIS, 14 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 15 # See the License for the specific language governing permissions and 16 # limitations under the License. 17 18 """This file contains NeuralModule and NmTensor classes.""" 19 __all__ = ['WeightShareTransform', 'NeuralModule'] 20 21 import collections 22 import uuid 23 from abc import ABC, abstractmethod 24 from collections import namedtuple 25 from enum import Enum 26 from inspect import getargvalues, getfullargspec, stack 27 from os import path 28 from typing import Dict, List, Optional, Set, Tuple 29 30 from ruamel.yaml import YAML 31 32 from .neural_types import ( 33 CanNotInferResultNeuralType, 34 NeuralPortNameMismatchError, 35 NeuralPortNmTensorMismatchError, 36 NeuralType, 37 NeuralTypeComparisonResult, 38 NmTensor, 39 ) 40 from nemo import logging 41 from nemo.core import NeuralModuleFactory 42 from nemo.package_info import __version__ as nemo_version 43 from nemo.utils.decorators.deprecated import deprecated 44 45 YAML = YAML(typ='safe') 46 47 48 class WeightShareTransform(Enum): 49 """When sharing parameters, what kind of transform to apply.""" 50 51 SAME = 0 52 TRANSPOSE = 1 53 54 55 PretrainedModelInfo = namedtuple( 56 "PretrainedModleInfo", ("pretrained_model_name", "description", "parameters", "location"), 57 ) 58 59 60 class NeuralModule(ABC): 61 """Abstract class that every Neural Module must inherit from. 62 """ 63 64 def __init__(self): 65 66 # Get default factory. 67 self._factory = NeuralModuleFactory.get_default_factory() 68 69 # Set module properties from factory else use defaults 70 self._placement = self._factory.placement 71 # If one needs to change that should override it manually. 72 73 # Optimization level. 74 self._opt_level = self._factory.optim_level 75 76 # Get object UUID. 77 self._uuid = str(uuid.uuid4()) 78 79 # Retrieve dictionary of parameters (keys, values) passed to init. 80 self._init_params = self.__extract_init_params() 81 82 # Pint the types of the values. 83 # for key, value in self._init_params.items(): 84 # print("{}: {} ({})".format(key, value, type(value))) 85 86 # Validate the parameters. 87 # self._validate_params(self._init_params) 88 89 @property 90 def init_params(self) -> Optional[Dict]: 91 """ 92 Property returning parameters used to instantiate the module. 93 94 Returns: 95 Dictionary containing parameters used to instantiate the module. 96 """ 97 return self._init_params 98 99 def __extract_init_params(self): 100 """ 101 Retrieves the dictionary of of parameters (keys, values) passed to constructor of a class derived 102 (also indirectly) from the Neural Module class. 103 104 Returns: 105 Dictionary containing parameters passed to init(). 106 """ 107 # Get names of arguments of the original module init method. 108 init_keys = getfullargspec(type(self).__init__).args 109 110 # Remove self. 111 if "self" in init_keys: 112 init_keys.remove("self") 113 114 # Create list of params. 115 init_params = {}.fromkeys(init_keys) 116 117 # Retrieve values of those params from the call list. 118 for frame in stack()[1:]: 119 localvars = getargvalues(frame[0]).locals 120 # print("localvars: ", localvars) 121 for key in init_keys: 122 # Found the variable! 123 if key in localvars.keys(): 124 # Save the value. 125 init_params[key] = localvars[key] 126 127 # Return parameters. 128 return init_params 129 130 def __validate_params(self, params): 131 """ 132 Checks whether dictionary contains parameters being primitive types (string, int, float etc.) 133 or (lists of)+ primitive types. 134 135 Args: 136 params: dictionary of parameters. 137 138 Returns: 139 True if all parameters were ok, False otherwise. 140 """ 141 ok = True 142 143 # Iterate over parameters and check them one by one. 144 for key, variable in params.items(): 145 if not self.__is_of_allowed_type(variable): 146 logging.warning( 147 "Parameter '{}' contains a variable '{}' of type '{}' which is not allowed.".format( 148 key, variable, type(variable) 149 ) 150 ) 151 ok = False 152 153 # Return the result. 154 return ok 155 156 def __is_of_allowed_type(self, var): 157 """ 158 A recursive function that checks if a given variable is of allowed type. 159 160 Args: 161 pretrained_model_name (str): name of pretrained model to use in order. 162 163 Returns: 164 True if all parameters were ok, False otherwise. 165 """ 166 # Special case: None is also allowed. 167 if var is None: 168 return True 169 170 var_type = type(var) 171 172 # If this is list - check its elements. 173 if var_type == list: 174 for list_var in var: 175 if not self.__is_of_allowed_type(list_var): 176 return False 177 178 # If this is dict - check its elements. 179 elif var_type == dict: 180 for _, dict_var in var.items(): 181 if not self.__is_of_allowed_type(dict_var): 182 return False 183 184 elif var_type not in (str, int, float, bool): 185 return False 186 187 # Well, seems that everything is ok. 188 return True 189 190 def _create_config_header(self): 191 """ A protected method that create a header stored later in the configuration file. """ 192 193 # Get module "full specification". 194 module_full_spec = str(self.__module__) + "." + str(self.__class__.__qualname__) 195 module_class_name = type(self).__name__ 196 # print(module_full_spec) 197 198 # Check whether module belongs to a collection. 199 spec_list = module_full_spec.split(".") 200 201 # Do not check Neural Modules from unit tests. 202 if spec_list[0] == "tests": 203 # Set collection variables. 204 collection_type = "tests" 205 collection_version = None 206 else: 207 # Check if component belongs to any collection 208 if len(spec_list) < 3 or (spec_list[0] != "nemo" and spec_list[1] != "collection"): 209 logging.warning( 210 "Module `{}` does not belong to any collection. This won't be allowed in the next release.".format( 211 module_class_name 212 ) 213 ) 214 collection_type = "unknown" 215 collection_version = None 216 else: 217 # Ok, set collection. 218 collection_type = spec_list[2] 219 collection_version = None 220 # TODO: to be SET! 221 # print(getattr("nemo.collections.nlp", __version__)) 222 223 # Create a "header" with module "specification". 224 header = { 225 "nemo_core_version": nemo_version, 226 "collection_type": collection_type, 227 "collection_version": collection_version, 228 # "class": module_class_name, # Operating only on full_spec now. 229 "full_spec": module_full_spec, 230 } 231 return header 232 233 def export_to_config(self, config_file): 234 """ 235 A function that exports module "configuration" (i.e. init parameters) to a YAML file. 236 Raises a ValueError exception in case then parameters coudn't be exported. 237 238 Args: 239 config_file: path (absolute or relative) and name of the config file (YML) 240 """ 241 # Check if generic export will work. 242 if not self.__validate_params(self._init_params): 243 raise ValueError( 244 "Generic configuration export enables to use of parameters of primitive types (string, int, float) " 245 F"or (lists of/dicts of) primitive types. Please implement your own custom `export_to_config()` and " 246 F"`import_from_config()` methods for your custom Module class." 247 ) 248 249 # Greate an absolute path. 250 abs_path_file = path.expanduser(config_file) 251 252 # Create the dictionary to be exported. 253 to_export = {} 254 255 # Add "header" with module "specification". 256 to_export["header"] = self._create_config_header() 257 258 # Add init parameters. 259 to_export["init_params"] = self._init_params 260 # print(to_export) 261 262 # All parameters are ok, let's export. 263 with open(abs_path_file, 'w') as outfile: 264 YAML.dump(to_export, outfile) 265 266 logging.info( 267 "Configuration of module {} ({}) exported to {}".format(self._uuid, type(self).__name__, abs_path_file) 268 ) 269 270 @classmethod 271 def _validate_config_file(cls, config_file, section_name=None): 272 """ 273 Class method validating whether the config file has a proper content (sections, specification etc.). 274 Raises an ImportError exception when config file is invalid or 275 incompatible (when called from a particular class). 276 277 Args: 278 config_file: path (absolute or relative) and name of the config file (YML) 279 280 section_name: section in the configuration file storing module configuration (optional, DEFAULT: None) 281 282 Returns: 283 A loaded configuration file (dictionary). 284 """ 285 # Greate an absolute path. 286 abs_path_file = path.expanduser(config_file) 287 288 # Open the config file. 289 with open(abs_path_file, 'r') as stream: 290 loaded_config = YAML.load(stream) 291 292 # Check section. 293 if section_name is not None: 294 if section_name not in loaded_config: 295 raise ImportError( 296 "The loaded config `{}` doesn't contain the indicated `{}` section".format( 297 config_file, section_name 298 ) 299 ) 300 # Section exists - use only it for configuration. 301 loaded_config = loaded_config[section_name] 302 303 # Make sure that the config is valid. 304 if "header" not in loaded_config: 305 raise ImportError("The loaded config `{}` doesn't contain the `header` section".format(config_file)) 306 307 if "init_params" not in loaded_config: 308 raise ImportError("The loaded config `{}` doesn't contain the `init_params` section".format(config_file)) 309 310 # Parse the "full specification". 311 spec_list = loaded_config["header"]["full_spec"].split(".") 312 313 # Check if config contains data of a compatible class. 314 if cls.__name__ != "NeuralModule" and spec_list[-1] != cls.__name__: 315 txt = "The loaded file `{}` contains configuration of ".format(config_file) 316 txt = txt + "`{}` thus cannot be used for instantiation of an object of type `{}`".format( 317 spec_list[-1], cls.__name__ 318 ) 319 raise ImportError(txt) 320 321 # Success - return configuration. 322 return loaded_config 323 324 @classmethod 325 def import_from_config(cls, config_file, section_name=None, overwrite_params={}): 326 """ 327 Class method importing the configuration file. 328 Raises an ImportError exception when config file is invalid or 329 incompatible (when called from a particular class). 330 331 Args: 332 config_file: path (absolute or relative) and name of the config file (YML) 333 334 section_name: section in the configuration file storing module configuration (optional, DEFAULT: None) 335 336 overwrite_params: Dictionary containing parameters that will be added to or overwrite (!) the default 337 parameters loaded from the configuration file 338 339 Returns: 340 Instance of the created NeuralModule object. 341 """ 342 # Validate the content of the configuration file (its header). 343 loaded_config = cls._validate_config_file(config_file, section_name) 344 345 # Parse the "full specification". 346 spec_list = loaded_config["header"]["full_spec"].split(".") 347 348 # Get object class from "full specification". 349 mod_obj = __import__(spec_list[0]) 350 for spec in spec_list[1:]: 351 mod_obj = getattr(mod_obj, spec) 352 # print(mod_obj) 353 354 # Get init parameters. 355 init_params = loaded_config["init_params"] 356 # Update parameters with additional ones. 357 init_params.update(overwrite_params) 358 359 # Create and return the object. 360 obj = mod_obj(**init_params) 361 logging.info( 362 "Instantiated a new Neural Module of type `{}` using configuration loaded from the `{}` file".format( 363 spec_list[-1], config_file 364 ) 365 ) 366 return obj 367 368 @deprecated(version=0.11) 369 @staticmethod 370 def create_ports(**kwargs): 371 """ Deprecated method, to be remoted in the next release.""" 372 raise Exception( 373 'Deprecated method. Please implement ``inputs`` and ``outputs`` \ 374 properties to define module ports instead' 375 ) 376 377 @property 378 @abstractmethod 379 def input_ports(self) -> Optional[Dict[str, NeuralType]]: 380 """Returns definitions of module input ports 381 382 Returns: 383 A (dict) of module's input ports names to NeuralTypes mapping 384 """ 385 386 @property 387 @abstractmethod 388 def output_ports(self) -> Optional[Dict[str, NeuralType]]: 389 """Returns definitions of module output ports 390 391 Returns: 392 A (dict) of module's output ports names to NeuralTypes mapping 393 """ 394 395 @property 396 def disabled_deployment_input_ports(self) -> Optional[Set[str]]: 397 """Returns names of input ports that will not be included in an export 398 399 Returns: 400 A (set) of module's input port names that are not exportable 401 """ 402 return set([]) 403 404 @property 405 def disabled_deployment_output_ports(self) -> Optional[Set[str]]: 406 """Returns names of output ports that will not be included in an export 407 408 Returns: 409 A (set) of module's output port names that are not exportable 410 """ 411 return set([]) 412 413 def prepare_for_deployment(self) -> None: 414 """Patch the module if required to prepare for deployment 415 416 """ 417 return 418 419 @staticmethod 420 def pretrained_storage(): 421 return '' 422 423 def __call__(self, **kwargs): 424 """This method allows objects to be called with their port names 425 426 Args: 427 kwargs: Input ports and their values. For example: 428 ... 429 mymodule1 = Subclass1_of_NeuralModule(...) 430 mymodule2 = Subclass2_of_NeuralModule(...) 431 ... 432 out_port1, out_port2 = mymodule1(input_port1=value1, 433 input_port2=value2, 434 input_port3=value3) 435 out_port11 = mymodule2(input_port1=out_port2) 436 ... 437 438 Returns: 439 NmTensor object or tuple of NmTensor objects 440 """ 441 # Get input and output ports definitions. 442 input_port_defs = self.input_ports 443 output_port_defs = self.output_ports 444 445 first_input_nmtensor_type = None 446 input_nmtensors_are_of_same_type = True 447 for port_name, tgv in kwargs.items(): 448 # make sure that passed arguments correspond to input port names 449 if port_name not in input_port_defs.keys(): 450 raise NeuralPortNameMismatchError("Wrong input port name: {0}".format(port_name)) 451 452 input_port = input_port_defs[port_name] 453 type_comatibility = input_port.compare(tgv) 454 if ( 455 type_comatibility != NeuralTypeComparisonResult.SAME 456 and type_comatibility != NeuralTypeComparisonResult.GREATER 457 ): 458 raise NeuralPortNmTensorMismatchError( 459 "\n\nIn {0}. \n" 460 "Port: {1} and a NmTensor it was fed are \n" 461 "of incompatible neural types:\n\n{2} \n\n and \n\n{3}" 462 "\n\nType comparison result: {4}".format( 463 self.__class__.__name__, port_name, input_port_defs[port_name], tgv, type_comatibility, 464 ) 465 ) 466 467 # if first_input_nmtensor_type is None: 468 # first_input_nmtensor_type = NeuralType(tgv._axis2type) 469 # else: 470 # if first_input_nmtensor_type._axis2type is None: 471 # input_nmtensors_are_of_same_type = True 472 # else: 473 # input_nmtensors_are_of_same_type = first_input_nmtensor_type.compare( 474 # tgv 475 # ) == NeuralTypeComparisonResult.SAME and len(first_input_nmtensor_type._axis2type) 476 # if not ( 477 # type_comatibility == NeuralTypeComparisonResult.SAME 478 # or type_comatibility == NeuralTypeComparisonResult.GREATER 479 # ): 480 # raise NeuralPortNmTensorMismatchError( 481 # "\n\nIn {0}. \n" 482 # "Port: {1} and a NmTensor it was fed are \n" 483 # "of incompatible neural types:\n\n{2} \n\n and \n\n{3}" 484 # "\n\nType comparison result: {4}".format( 485 # self.__class__.__name__, port_name, input_port_defs[port_name], tgv, type_comatibility, 486 # ) 487 # ) 488 # if type_comatibility == NeuralTypeComparisonResult.LESS: 489 # print('Types were raised') 490 491 if len(output_port_defs) == 1: 492 out_name = list(output_port_defs)[0] 493 out_type = output_port_defs[out_name] 494 if out_type is None: 495 if input_nmtensors_are_of_same_type: 496 out_type = first_input_nmtensor_type 497 else: 498 raise CanNotInferResultNeuralType( 499 "Can't infer output neural type. Likely your inputs are of different type." 500 ) 501 return NmTensor(producer=self, producer_args=kwargs, name=out_name, ntype=out_type,) 502 else: 503 result = [] 504 for out_port, n_type in output_port_defs.items(): 505 out_type = n_type 506 if out_type is None: 507 if input_nmtensors_are_of_same_type: 508 out_type = first_input_nmtensor_type 509 else: 510 raise CanNotInferResultNeuralType( 511 "Can't infer output neural type. Likely your inputs are of different type." 512 ) 513 result.append(NmTensor(producer=self, producer_args=kwargs, name=out_port, ntype=out_type,)) 514 515 # Creating ad-hoc class for returning from module's forward pass. 516 output_class_name = f'{self.__class__.__name__}Output' 517 field_names = list(output_port_defs) 518 result_type = collections.namedtuple(typename=output_class_name, field_names=field_names,) 519 520 # Tie tuple of output tensors with corresponding names. 521 result = result_type(*result) 522 523 return result 524 525 def __str__(self): 526 return self.__class__.__name__ 527 528 @abstractmethod 529 def get_weights(self) -> Optional[Dict[(str, bool)]]: 530 """Returns NeuralModule's weights copy. 531 532 Returns: 533 Dictionary of name -> (weights, trainable)""" 534 pass 535 536 @abstractmethod 537 def set_weights( 538 self, 539 name2weight: Dict[(str, Tuple[str, bool])], 540 name2name_and_transform: Dict[(str, Tuple[str, WeightShareTransform])] = None, 541 ): 542 """Sets weight from given values. For every named weight in 543 name2weight, 544 if weight with the same name is found in the model, it will be set to 545 found value. 546 547 WARNING: This will NOT tie weights. It will copy values. 548 549 If ``name2name_and_transform`` is provided then if will set weights 550 using 551 name mapping and transform. For example, suppose ``objec1.X = 3x5 552 weight``. 553 Then, if ``name2name_and_transform['X']=('Y', 554 WeightShareTransform.TRANSPOSE)`` 555 and ``Y`` is 5x3 weight and ``name2weight['Y']=Y. Then: 556 ``object1.set_weights(name2weight, name2name_and_transform)`` will 557 set object1.X=transpose(Y). 558 559 Args: 560 name2weight (dict): dictionary of name to (weight, trainable). 561 Typically this is output of get_weights method. 562 name2name_and_transform: mapping from name -> (name, transform) 563 """ 564 pass 565 566 @staticmethod 567 def list_pretrained_models() -> Optional[List[PretrainedModelInfo]]: 568 """List all available pre-trained models (e.g. weights) for this NM. 569 570 Returns: 571 A list of PretrainedModelInfo tuples. 572 The pretrained_model_name field of the tuple can be used to 573 retrieve pre-trained model's weights (pass it as 574 pretrained_model_name argument to the module's constructor) 575 """ 576 return None 577 578 def get_config_dict_and_checkpoint(self, pretrained_model_name): 579 """WARNING: This part is work in progress""" 580 return None 581 582 @abstractmethod 583 def tie_weights_with( 584 self, 585 module, 586 weight_names=List[str], 587 name2name_and_transform: Dict[(str, Tuple[str, WeightShareTransform])] = None, 588 ): 589 """Ties weights between self and module. For every weight name in 590 weight_names, if weight with the same name is found in self, it will 591 be tied 592 with a same weight from ``module``. 593 594 WARNING: Once weights are tied, updates to one weights's weights 595 will affect 596 other module's weights. 597 598 599 If ``name2name_and_transform`` is provided then if will set weights 600 using 601 name mapping and transform. For example, suppose ``objec1.X = 3x5 602 weights`` 603 and ``object2.Y = 5x3 weights``. Then these weights can be tied like 604 this: 605 606 .. code-block:: python 607 608 object1.tie_weights_with(object2, weight_names=['X'], 609 name2name_and_transform = 610 { 'X': ('Y', WeightShareTransform.TRANSPOSE)}) 611 612 613 Args: 614 module: with which module to tie weights 615 weight_names (List[str]): list of self weights' names 616 name2name_and_transform: mapping from name -> (name, transform) 617 """ 618 pass 619 620 def is_trainable(self) -> bool: 621 """ 622 Checks if NeuralModule is trainable. 623 A NeuralModule is trainable IFF it contains at least one trainable 624 weight 625 626 Returns: 627 True if module has trainable weights, False otherwise 628 """ 629 weights = self.get_weights() 630 if weights is None: 631 return False 632 for name, w in weights.items(): 633 if w[1]: 634 return True 635 return False 636 637 @abstractmethod 638 def save_to(self, path: str): 639 """Save module state to file. 640 641 Args: 642 path (string): path to while where to save. 643 """ 644 pass 645 646 @abstractmethod 647 def restore_from(self, path: str): 648 """Restore module's state from file. 649 650 Args: 651 path (string): path to where to restore from. 652 """ 653 pass 654 655 @abstractmethod 656 def freeze(self, weights: Set[str] = None): 657 """Freeze weights 658 659 Args: 660 weights (set): set of weight names to freeze 661 If None, all weights are freezed. 662 """ 663 pass 664 665 @abstractmethod 666 def unfreeze(self, weights: Set[str] = None): 667 """Unfreeze weights 668 669 Args: 670 weights (set): set of weight names to unfreeze 671 If None, all weights are unfreezed. 672 """ 673 pass 674 675 @property 676 def placement(self): 677 """Module's placement. Currently CPU or GPU. 678 DataParallel and ModelParallel will come later. 679 680 Returns: 681 (DeviceType) Device where NM's weights are located 682 """ 683 return self._placement 684 685 @property 686 @deprecated(version=0.11) 687 def local_parameters(self) -> Optional[Dict]: 688 """Get module's parameters 689 690 Returns: 691 module's parameters 692 """ 693 return self._init_params 694 # return self._local_parameters 695 696 @property 697 def unique_instance_id(self): 698 """A unique instance id for this object 699 700 Returns: 701 A uniq uuid which can be used to identify this object 702 """ 703 return self._uuid 704 705 @property 706 def factory(self): 707 """ Neural module factory which created this module 708 Returns: NeuralModuleFactory instance or None 709 """ 710 return self._factory 711 712 @property 713 @abstractmethod 714 def num_weights(self): 715 """Number of module's weights 716 """ 717 pass 718 [end of nemo/core/neural_modules.py] </code> Here is an example of a patch file. It consists of changes to the code base. It specifies the file names, the line numbers of each change, and the removed and added lines. A single patch file can contain changes to multiple files. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch> I need you to solve the provided issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the format shown above. Respond below:
NVIDIA/NeMo
ba4616f1f011d599de87f0cb3315605e715d402a
Jasper Encoder Export failed The export of Jasper Encoder is failing. I am using the core API [deployment_export](https://nvidia.github.io/NeMo/api-docs/nemo.html#nemo.core.neural_factory.NeuralModuleFactory.deployment_export) like in the script: https://github.com/NVIDIA/NeMo/blob/403238f82d26879ba5fca53fbf75b3cdc70fb49b/scripts/export_jasper_to_onnx.py#L92 I believe the issue (as shown below) is that the` input_example` provided does not match the `output_example`. ``` /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) ``` **What is the correct `input_example` and `output_example` to export JasperEncoder?** The full output can be seen here: ``` adrianaf@2a520c7abb1e:/tmp/NeMo$ ! python /tmp/NeMo/scripts/export_jasper_to_onnx.py --config /raid/datasets/asr/data/config_files/WSJ-test_acoustic_quartznet15x5.yaml --nn_encoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperEncoder-STEP-247400.pt --nn_decoder /home/adrianaf/projects/nemo_asr_app/models/quartznet15x5/JasperDecoderForCTC-STEP-247400.pt --onnx_encoder /raid/datasets/asr/data/models/ONNX/pre-trained_encoder.onnx --onnx_decoder /raid/datasets/asr/data/models/ONNX/pre-trained_decoder.onnx /opt/conda/lib/python3.6/site-packages/torchvision/io/_video_opt.py:17: UserWarning: video reader based on ffmpeg c++ ops not available warnings.warn("video reader based on ffmpeg c++ ops not available") /tmp/NeMo/nemo/collections/asr/audio_preprocessing.py:48: UserWarning: Could not import torchaudio. Some features might not work. warnings.warn('Could not import torchaudio. Some features might not work.') [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:48] Loading config file... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:52] Determining model shape... [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:60] Num encoder input features: 64 [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:61] Num decoder input features: 1024 [NeMo W 2020-02-23 19:09:42 deprecated:68] Function ``_get_trainer`` is deprecated. It is going to be removed in the future version. [NeMo I 2020-02-23 19:09:42 export_jasper_to_onnx:65] Initializing models... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:76] Loading checkpoints... [NeMo I 2020-02-23 19:09:45 export_jasper_to_onnx:91] Exporting encoder... [NeMo W 2020-02-23 19:09:45 neural_factory:627] Turned off 170 masked convolutions [NeMo I 2020-02-23 19:09:45 actions:937] Module is JasperEncoder. We are removing input and output length ports since they are not needed for deployment [NeMo W 2020-02-23 19:09:46 deprecated:68] Function ``local_parameters`` is deprecated. It is going to be removed in the 0.11 version. /opt/conda/lib/python3.6/site-packages/torch/jit/__init__.py:1023: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error: Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 870, 67] (0.6547648906707764 vs. 0.6546438932418823) and 812 other locations (0.00%) check_tolerance, _force_outplace, True, _module_class) [NeMo E 2020-02-23 19:10:07 actions:1023] module export failed for JasperEncoder with exception number of output names provided (2) exceeded number of outputs (1) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:98] Exporting decoder... graph(%encoder_output : Float(1, 1024, 128), %1 : Float(29, 1024, 1), %2 : Float(29)): %3 : Float(1, 29, 128) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%encoder_output, %1, %2), scope: JasperDecoderForCTC/Sequential[decoder_layers]/Conv1d[0] # /opt/conda/lib/python3.6/site-packages/torch/nn/modules/conv.py:202:0 %4 : Float(1, 128, 29) = onnx::Transpose[perm=[0, 2, 1]](%3), scope: JasperDecoderForCTC # /tmp/NeMo/nemo/collections/asr/jasper.py:235:0 %output : Float(1, 128, 29) = onnx::LogSoftmax[axis=2](%4), scope: JasperDecoderForCTC # /opt/conda/lib/python3.6/site-packages/torch/nn/functional.py:1317:0 return (%output) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input encoder_output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) /opt/conda/lib/python3.6/site-packages/torch/onnx/utils.py:774: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input output 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key)) [NeMo I 2020-02-23 19:10:07 export_jasper_to_onnx:105] Export completed successfully. ```
2020-03-10T03:03:23Z
<patch> <patch> diff --git a/nemo/backends/pytorch/actions.py b/nemo/backends/pytorch/actions.py --- a/nemo/backends/pytorch/actions.py +++ b/nemo/backends/pytorch/actions.py @@ -937,26 +937,16 @@ def __extract_dynamic_axes(port_name: str, ntype: NeuralType, dynamic_axes: defa if axis.kind == AxisKind.Batch or axis.kind == AxisKind.Time: dynamic_axes[port_name].append(ind) - # This is a hack for Jasper to Jarvis export -- need re-design for this - inputs_to_drop = set() - outputs_to_drop = set() - if type(module).__name__ == "JasperEncoder": - logging.info( - "Module is JasperEncoder. We are removing input and output length ports since they are not needed for " - "deployment" - ) - inputs_to_drop.add("length") - outputs_to_drop.add("encoded_lengths") - + # extract dynamic axes and remove unnecessary inputs/outputs # for input_ports for port_name, ntype in module.input_ports.items(): - if port_name in inputs_to_drop: + if port_name in module._disabled_deployment_input_ports: input_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) # for output_ports for port_name, ntype in module.output_ports.items(): - if port_name in outputs_to_drop: + if port_name in module._disabled_deployment_output_ports: output_names.remove(port_name) continue __extract_dynamic_axes(port_name, ntype, dynamic_axes) diff --git a/nemo/collections/asr/jasper.py b/nemo/collections/asr/jasper.py --- a/nemo/collections/asr/jasper.py +++ b/nemo/collections/asr/jasper.py @@ -118,14 +118,14 @@ def output_ports(self): } @property - def disabled_deployment_input_ports(self): + def _disabled_deployment_input_ports(self): return set(["length"]) @property - def disabled_deployment_output_ports(self): + def _disabled_deployment_output_ports(self): return set(["encoded_lengths"]) - def prepare_for_deployment(self): + def _prepare_for_deployment(self): m_count = 0 for m in self.modules(): if type(m).__name__ == "MaskedConv1d": diff --git a/nemo/core/neural_factory.py b/nemo/core/neural_factory.py --- a/nemo/core/neural_factory.py +++ b/nemo/core/neural_factory.py @@ -610,7 +610,7 @@ def deployment_export( input_example: sometimes tracing will require input examples output_example: Should match inference on input_example """ - module.prepare_for_deployment() + module._prepare_for_deployment() return self._trainer.deployment_export( module=module, diff --git a/nemo/core/neural_modules.py b/nemo/core/neural_modules.py --- a/nemo/core/neural_modules.py +++ b/nemo/core/neural_modules.py @@ -393,7 +393,7 @@ def output_ports(self) -> Optional[Dict[str, NeuralType]]: """ @property - def disabled_deployment_input_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_input_ports(self) -> Optional[Set[str]]: """Returns names of input ports that will not be included in an export Returns: @@ -402,7 +402,7 @@ def disabled_deployment_input_ports(self) -> Optional[Set[str]]: return set([]) @property - def disabled_deployment_output_ports(self) -> Optional[Set[str]]: + def _disabled_deployment_output_ports(self) -> Optional[Set[str]]: """Returns names of output ports that will not be included in an export Returns: @@ -410,7 +410,7 @@ def disabled_deployment_output_ports(self) -> Optional[Set[str]]: """ return set([]) - def prepare_for_deployment(self) -> None: + def _prepare_for_deployment(self) -> None: """Patch the module if required to prepare for deployment """ </patch> </s> </patch>
diff --git a/tests/unit/core/test_deploy_export.py b/tests/unit/core/test_deploy_export.py --- a/tests/unit/core/test_deploy_export.py +++ b/tests/unit/core/test_deploy_export.py @@ -46,9 +46,11 @@ import nemo.collections.nlp.nm.trainables.common.token_classification_nm from nemo import logging +TRT_ONNX_DISABLED = False + # Check if the required libraries and runtimes are installed. +# Only initialize GPU after this runner is activated. try: - # Only initialize GPU after this runner is activated. import pycuda.autoinit # This import causes pycuda to automatically manage CUDA context creation and cleanup. @@ -63,16 +65,17 @@ ) from .tensorrt_runner import TensorRTRunnerV2 except: - # Skip tests. - pytestmark = pytest.mark.skip + TRT_ONNX_DISABLED = True @pytest.mark.usefixtures("neural_factory") class TestDeployExport(TestCase): - def setUp(self): - logging.setLevel(logging.WARNING) - device = nemo.core.DeviceType.GPU - self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) + # def setUp(self): + # super().setUp() + + # logging.setLevel(logging.WARNING) + # device = nemo.core.DeviceType.GPU + # self.nf = nemo.core.NeuralModuleFactory(backend=nemo.core.Backend.PyTorch, placement=device) def __test_export_route(self, module, out_name, mode, input_example=None): out = Path(out_name) @@ -112,7 +115,13 @@ def __test_export_route(self, module, out_name, mode, input_example=None): loader_cache = DataLoaderCache(data_loader) profile_shapes = OrderedDict() names = list(module.input_ports) + list(module.output_ports) - + names = list( + filter( + lambda x: x + not in (module._disabled_deployment_input_ports | module._disabled_deployment_output_ports), + names, + ) + ) if isinstance(input_example, tuple): si = [tuple(input_example[i].shape) for i in range(len(input_example))] elif isinstance(input_example, OrderedDict): @@ -152,7 +161,7 @@ def __test_export_route(self, module, out_name, mode, input_example=None): input_names = list(input_metadata.keys()) for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: + if input_name in module._disabled_deployment_input_ports: continue inputs[input_name] = ( input_example[input_name].cpu().numpy() @@ -209,8 +218,8 @@ def __test_export_route(self, module, out_name, mode, input_example=None): ort_inputs = ort_session.get_inputs() for i in range(len(input_names)): input_name = input_names[i] - if input_name in module.disabled_deployment_input_ports: - input_name = ort_inputs[i].name + if input_name in module._disabled_deployment_input_ports: + continue inputs[input_name] = ( input_example[input_name].cpu().numpy() if isinstance(input_example, OrderedDict) @@ -263,9 +272,10 @@ def __test_export_route(self, module, out_name, mode, input_example=None): def __test_export_route_all(self, module, out_name, input_example=None): if input_example is not None: - self.__test_export_route( - module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example - ) + if not TRT_ONNX_DISABLED: + self.__test_export_route( + module, out_name + '.trt.onnx', nemo.core.DeploymentFormat.TRTONNX, input_example=input_example + ) self.__test_export_route(module, out_name + '.onnx', nemo.core.DeploymentFormat.ONNX, input_example) self.__test_export_route(module, out_name + '.pt', nemo.core.DeploymentFormat.PYTORCH, input_example) self.__test_export_route(module, out_name + '.ts', nemo.core.DeploymentFormat.TORCHSCRIPT, input_example) @@ -323,9 +333,7 @@ def test_jasper_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="jasper_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randn(256).cuda()), + module=jasper_encoder, out_name="jasper_encoder", input_example=torch.randn(16, 64, 256).cuda(), ) @pytest.mark.unit @@ -343,7 +351,5 @@ def test_quartz_encoder(self): ) self.__test_export_route_all( - module=jasper_encoder, - out_name="quartz_encoder", - input_example=(torch.randn(16, 64, 256).cuda(), torch.randint(20, (16,)).cuda()), + module=jasper_encoder, out_name="quartz_encoder", input_example=torch.randn(16, 64, 256).cuda(), )
1.0
NVIDIA__NeMo-3632
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> ./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`. </issue> <code> [start of README.rst] 1 2 |status| |documentation| |license| |lgtm_grade| |lgtm_alerts| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |lgtm_grade| image:: https://img.shields.io/lgtm/grade/python/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 17 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/context:python 18 :alt: Language grade: Python 19 20 .. |lgtm_alerts| image:: https://img.shields.io/lgtm/alerts/g/NVIDIA/NeMo.svg?logo=lgtm&logoWidth=18 21 :target: https://lgtm.com/projects/g/NVIDIA/NeMo/alerts/ 22 :alt: Total alerts 23 24 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 25 :target: https://github.com/psf/black 26 :alt: Code style: black 27 28 .. _main-readme: 29 30 **NVIDIA NeMo** 31 =============== 32 33 Introduction 34 ------------ 35 36 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), natural language processing (NLP), and text-to-speech synthesis (TTS). 37 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 38 39 `Pre-trained NeMo models. <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_ 40 41 `Introductory video. <https://www.youtube.com/embed/wBgpMf_KQVw>`_ 42 43 Key Features 44 ------------ 45 46 * Speech processing 47 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 48 * Supported models: Jasper, QuartzNet, CitriNet, Conformer-CTC, Conformer-Transducer, ContextNet, ... 49 * Supports CTC and Transducer/RNNT losses/decoders 50 * Beam Search decoding 51 * `Language Modelling for ASR <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 52 * Streaming and Buffered ASR (CTC/Transdcer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/main/examples/asr/asr_chunked_inference>`_ 53 * `Speech Classification and Speech Command Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition) 54 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 55 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 56 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 57 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 58 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 59 * Natural Language Processing 60 * `Compatible with Hugging Face Transformers and NVIDIA Megatron <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html>`_ 61 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation.html>`_ 62 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 63 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 64 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 65 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 66 * `BERT pre-training <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/bert_pretraining.html>`_ 67 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 68 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 69 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 70 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 71 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 72 * `Neural Duplex Text Normalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization.html>`_ 73 * `Prompt Tuning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/megatron_finetuning.html#prompt-tuning>`_ 74 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 75 * `Speech synthesis (TTS) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 76 * Spectrogram generation: Tacotron2, GlowTTS, TalkNet, FastPitch, FastSpeech2, Mixer-TTS, Mixer-TTS-X 77 * Vocoders: WaveGlow, SqueezeWave, UniGlow, MelGAN, HiFiGAN, UnivNet 78 * End-to-end speech generation: FastPitch_HifiGan_E2E, FastSpeech2_HifiGan_E2E 79 * `NGC collection of pre-trained TTS models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 80 * `Tools <https://github.com/NVIDIA/NeMo/tree/main/tools>`_ 81 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/text_processing_deployment.html>`_ 82 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 83 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 84 85 86 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 87 88 Requirements 89 ------------ 90 91 1) Python 3.6, 3.7 or 3.8 92 2) Pytorch 1.10.0 or above 93 3) NVIDIA GPU for training 94 95 Documentation 96 ------------- 97 98 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 99 :alt: Documentation Status 100 :scale: 100% 101 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 102 103 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 104 :alt: Documentation Status 105 :scale: 100% 106 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 107 108 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 109 | Version | Status | Description | 110 +=========+=============+==========================================================================================================================================+ 111 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 112 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 113 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 114 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 115 116 Tutorials 117 --------- 118 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 119 120 Getting help with NeMo 121 ---------------------- 122 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 123 124 125 Installation 126 ------------ 127 128 Pip 129 ~~~ 130 Use this installation mode if you want the latest released version. 131 132 .. code-block:: bash 133 134 apt-get update && apt-get install -y libsndfile1 ffmpeg 135 pip install Cython 136 pip install nemo_toolkit['all'] 137 138 .. note:: 139 140 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 141 142 Pip from source 143 ~~~~~~~~~~~~~~~ 144 Use this installation mode if you want the a version from particular GitHub branch (e.g main). 145 146 .. code-block:: bash 147 148 apt-get update && apt-get install -y libsndfile1 ffmpeg 149 pip install Cython 150 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 151 152 153 From source 154 ~~~~~~~~~~~ 155 Use this installation mode if you are contributing to NeMo. 156 157 .. code-block:: bash 158 159 apt-get update && apt-get install -y libsndfile1 ffmpeg 160 git clone https://github.com/NVIDIA/NeMo 161 cd NeMo 162 ./reinstall.sh 163 164 .. note:: 165 166 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 167 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 168 169 RNNT 170 ~~~~ 171 Note that RNNT requires numba to be installed from conda. 172 173 .. code-block:: bash 174 175 conda remove numba 176 pip uninstall numba 177 conda install -c conda-forge numba 178 179 Megatron GPT 180 ~~~~~~~~~~~~ 181 Megatron GPT training requires NVIDIA Apex to be installed. 182 183 .. code-block:: bash 184 185 git clone https://github.com/NVIDIA/apex 186 cd apex 187 git checkout c8bcc98176ad8c3a0717082600c70c907891f9cb 188 pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" ./ 189 190 Docker containers: 191 ~~~~~~~~~~~~~~~~~~ 192 To build a nemo container with Dockerfile from a branch, please run 193 194 .. code-block:: bash 195 196 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 197 198 199 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 22.01-py3 and then installing from GitHub. 200 201 .. code-block:: bash 202 203 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 204 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 205 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:22.01-py3 206 207 Examples 208 -------- 209 210 Many examples can be found under `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 211 212 213 Contributing 214 ------------ 215 216 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 217 218 Publications 219 ------------ 220 221 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/blob/main/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 222 223 Citation 224 -------- 225 226 .. code-block:: bash 227 228 @article{kuchaiev2019nemo, 229 title={Nemo: a toolkit for building ai applications using neural modules}, 230 author={Kuchaiev, Oleksii and Li, Jason and Nguyen, Huyen and Hrinchuk, Oleksii and Leary, Ryan and Ginsburg, Boris and Kriman, Samuel and Beliaev, Stanislav and Lavrukhin, Vitaly and Cook, Jack and others}, 231 journal={arXiv preprint arXiv:1909.09577}, 232 year={2019} 233 } 234 235 License 236 ------- 237 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 238 [end of README.rst] [start of /dev/null] 1 [end of /dev/null] [start of nemo_text_processing/text_normalization/__init__.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from nemo.utils import logging 16 17 try: 18 import pynini 19 20 PYNINI_AVAILABLE = True 21 except (ModuleNotFoundError, ImportError): 22 logging.warning( 23 "`pynini` is not installed ! \n" 24 "Please run the `nemo_text_processing/setup.sh` script" 25 "prior to usage of this toolkit." 26 ) 27 28 PYNINI_AVAILABLE = False 29 [end of nemo_text_processing/text_normalization/__init__.py] [start of nemo_text_processing/text_normalization/en/graph_utils.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 import os 17 import string 18 from pathlib import Path 19 from typing import Dict 20 21 from nemo_text_processing.text_normalization.en.utils import get_abs_path 22 23 try: 24 import pynini 25 from pynini import Far 26 from pynini.export import export 27 from pynini.examples import plurals 28 from pynini.lib import byte, pynutil, utf8 29 30 NEMO_CHAR = utf8.VALID_UTF8_CHAR 31 32 NEMO_DIGIT = byte.DIGIT 33 NEMO_LOWER = pynini.union(*string.ascii_lowercase).optimize() 34 NEMO_UPPER = pynini.union(*string.ascii_uppercase).optimize() 35 NEMO_ALPHA = pynini.union(NEMO_LOWER, NEMO_UPPER).optimize() 36 NEMO_ALNUM = pynini.union(NEMO_DIGIT, NEMO_ALPHA).optimize() 37 NEMO_HEX = pynini.union(*string.hexdigits).optimize() 38 NEMO_NON_BREAKING_SPACE = u"\u00A0" 39 NEMO_SPACE = " " 40 NEMO_WHITE_SPACE = pynini.union(" ", "\t", "\n", "\r", u"\u00A0").optimize() 41 NEMO_NOT_SPACE = pynini.difference(NEMO_CHAR, NEMO_WHITE_SPACE).optimize() 42 NEMO_NOT_QUOTE = pynini.difference(NEMO_CHAR, r'"').optimize() 43 44 NEMO_PUNCT = pynini.union(*map(pynini.escape, string.punctuation)).optimize() 45 NEMO_GRAPH = pynini.union(NEMO_ALNUM, NEMO_PUNCT).optimize() 46 47 NEMO_SIGMA = pynini.closure(NEMO_CHAR) 48 49 delete_space = pynutil.delete(pynini.closure(NEMO_WHITE_SPACE)) 50 insert_space = pynutil.insert(" ") 51 delete_extra_space = pynini.cross(pynini.closure(NEMO_WHITE_SPACE, 1), " ") 52 delete_preserve_order = pynini.closure( 53 pynutil.delete(" preserve_order: true") 54 | (pynutil.delete(" field_order: \"") + NEMO_NOT_QUOTE + pynutil.delete("\"")) 55 ) 56 57 suppletive = pynini.string_file(get_abs_path("data/suppletive.tsv")) 58 # _v = pynini.union("a", "e", "i", "o", "u") 59 _c = pynini.union( 60 "b", "c", "d", "f", "g", "h", "j", "k", "l", "m", "n", "p", "q", "r", "s", "t", "v", "w", "x", "y", "z" 61 ) 62 _ies = NEMO_SIGMA + _c + pynini.cross("y", "ies") 63 _es = NEMO_SIGMA + pynini.union("s", "sh", "ch", "x", "z") + pynutil.insert("es") 64 _s = NEMO_SIGMA + pynutil.insert("s") 65 66 graph_plural = plurals._priority_union( 67 suppletive, plurals._priority_union(_ies, plurals._priority_union(_es, _s, NEMO_SIGMA), NEMO_SIGMA), NEMO_SIGMA 68 ).optimize() 69 70 SINGULAR_TO_PLURAL = graph_plural 71 PLURAL_TO_SINGULAR = pynini.invert(graph_plural) 72 TO_LOWER = pynini.union(*[pynini.cross(x, y) for x, y in zip(string.ascii_uppercase, string.ascii_lowercase)]) 73 TO_UPPER = pynini.invert(TO_LOWER) 74 75 PYNINI_AVAILABLE = True 76 except (ModuleNotFoundError, ImportError): 77 # Create placeholders 78 NEMO_CHAR = None 79 80 NEMO_DIGIT = None 81 NEMO_LOWER = None 82 NEMO_UPPER = None 83 NEMO_ALPHA = None 84 NEMO_ALNUM = None 85 NEMO_HEX = None 86 NEMO_NON_BREAKING_SPACE = u"\u00A0" 87 NEMO_SPACE = " " 88 NEMO_WHITE_SPACE = None 89 NEMO_NOT_SPACE = None 90 NEMO_NOT_QUOTE = None 91 92 NEMO_PUNCT = None 93 NEMO_GRAPH = None 94 95 NEMO_SIGMA = None 96 97 delete_space = None 98 insert_space = None 99 delete_extra_space = None 100 delete_preserve_order = None 101 102 suppletive = None 103 # _v = pynini.union("a", "e", "i", "o", "u") 104 _c = None 105 _ies = None 106 _es = None 107 _s = None 108 109 graph_plural = None 110 111 SINGULAR_TO_PLURAL = None 112 PLURAL_TO_SINGULAR = None 113 TO_LOWER = None 114 TO_UPPER = None 115 116 PYNINI_AVAILABLE = False 117 118 119 def generator_main(file_name: str, graphs: Dict[str, 'pynini.FstLike']): 120 """ 121 Exports graph as OpenFst finite state archive (FAR) file with given file name and rule name. 122 123 Args: 124 file_name: exported file name 125 graphs: Mapping of a rule name and Pynini WFST graph to be exported 126 """ 127 exporter = export.Exporter(file_name) 128 for rule, graph in graphs.items(): 129 exporter[rule] = graph.optimize() 130 exporter.close() 131 print(f'Created {file_name}') 132 133 134 def get_plurals(fst): 135 """ 136 Given singular returns plurals 137 138 Args: 139 fst: Fst 140 141 Returns plurals to given singular forms 142 """ 143 return SINGULAR_TO_PLURAL @ fst 144 145 146 def get_singulars(fst): 147 """ 148 Given plural returns singulars 149 150 Args: 151 fst: Fst 152 153 Returns singulars to given plural forms 154 """ 155 return PLURAL_TO_SINGULAR @ fst 156 157 158 def convert_space(fst) -> 'pynini.FstLike': 159 """ 160 Converts space to nonbreaking space. 161 Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" 162 This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. 163 164 Args: 165 fst: input fst 166 167 Returns output fst where breaking spaces are converted to non breaking spaces 168 """ 169 return fst @ pynini.cdrewrite(pynini.cross(NEMO_SPACE, NEMO_NON_BREAKING_SPACE), "", "", NEMO_SIGMA) 170 171 172 class GraphFst: 173 """ 174 Base class for all grammar fsts. 175 176 Args: 177 name: name of grammar class 178 kind: either 'classify' or 'verbalize' 179 deterministic: if True will provide a single transduction option, 180 for False multiple transduction are generated (used for audio-based normalization) 181 """ 182 183 def __init__(self, name: str, kind: str, deterministic: bool = True): 184 self.name = name 185 self.kind = str 186 self._fst = None 187 self.deterministic = deterministic 188 189 self.far_path = Path(os.path.dirname(__file__) + '/grammars/' + kind + '/' + name + '.far') 190 if self.far_exist(): 191 self._fst = Far(self.far_path, mode="r", arc_type="standard", far_type="default").get_fst() 192 193 def far_exist(self) -> bool: 194 """ 195 Returns true if FAR can be loaded 196 """ 197 return self.far_path.exists() 198 199 @property 200 def fst(self) -> 'pynini.FstLike': 201 return self._fst 202 203 @fst.setter 204 def fst(self, fst): 205 self._fst = fst 206 207 def add_tokens(self, fst) -> 'pynini.FstLike': 208 """ 209 Wraps class name around to given fst 210 211 Args: 212 fst: input fst 213 214 Returns: 215 Fst: fst 216 """ 217 return pynutil.insert(f"{self.name} {{ ") + fst + pynutil.insert(" }") 218 219 def delete_tokens(self, fst) -> 'pynini.FstLike': 220 """ 221 Deletes class name wrap around output of given fst 222 223 Args: 224 fst: input fst 225 226 Returns: 227 Fst: fst 228 """ 229 res = ( 230 pynutil.delete(f"{self.name}") 231 + delete_space 232 + pynutil.delete("{") 233 + delete_space 234 + fst 235 + delete_space 236 + pynutil.delete("}") 237 ) 238 return res @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) 239 [end of nemo_text_processing/text_normalization/en/graph_utils.py] [start of nemo_text_processing/text_normalization/en/taggers/punctuation.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 import sys 17 from unicodedata import category 18 19 from nemo_text_processing.text_normalization.en.graph_utils import GraphFst 20 21 try: 22 import pynini 23 from pynini.lib import pynutil 24 25 PYNINI_AVAILABLE = False 26 except (ModuleNotFoundError, ImportError): 27 PYNINI_AVAILABLE = False 28 29 30 class PunctuationFst(GraphFst): 31 """ 32 Finite state transducer for classifying punctuation 33 e.g. a, -> tokens { name: "a" } tokens { name: "," } 34 35 Args: 36 deterministic: if True will provide a single transduction option, 37 for False multiple transduction are generated (used for audio-based normalization) 38 39 """ 40 41 def __init__(self, deterministic: bool = True): 42 super().__init__(name="punctuation", kind="classify", deterministic=deterministic) 43 44 s = "!#%&\'()*+,-./:;<=>?@^_`{|}~\"" 45 46 punct_unicode = [chr(i) for i in range(sys.maxunicode) if category(chr(i)).startswith("P")] 47 punct_unicode.remove('[') 48 punct_unicode.remove(']') 49 punct = pynini.union(*s) | pynini.union(*punct_unicode) 50 51 self.graph = punct 52 self.fst = (pynutil.insert("name: \"") + self.graph + pynutil.insert("\"")).optimize() 53 [end of nemo_text_processing/text_normalization/en/taggers/punctuation.py] [start of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 17 from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space 18 19 try: 20 import pynini 21 from pynini.lib import pynutil 22 23 PYNINI_AVAILABLE = True 24 except (ModuleNotFoundError, ImportError): 25 PYNINI_AVAILABLE = False 26 27 28 class WhiteListFst(GraphFst): 29 """ 30 Finite state transducer for verbalizing whitelist 31 e.g. tokens { name: "misses" } } -> misses 32 33 Args: 34 deterministic: if True will provide a single transduction option, 35 for False multiple transduction are generated (used for audio-based normalization) 36 """ 37 38 def __init__(self, deterministic: bool = True): 39 super().__init__(name="whitelist", kind="verbalize", deterministic=deterministic) 40 graph = ( 41 pynutil.delete("name:") 42 + delete_space 43 + pynutil.delete("\"") 44 + pynini.closure(NEMO_CHAR - " ", 1) 45 + pynutil.delete("\"") 46 ) 47 graph = graph @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) 48 self.fst = graph.optimize() 49 [end of nemo_text_processing/text_normalization/en/verbalizers/whitelist.py] [start of nemo_text_processing/text_normalization/en/verbalizers/word.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space 17 18 try: 19 import pynini 20 from pynini.lib import pynutil 21 22 PYNINI_AVAILABLE = True 23 except (ModuleNotFoundError, ImportError): 24 PYNINI_AVAILABLE = False 25 26 27 class WordFst(GraphFst): 28 """ 29 Finite state transducer for verbalizing word 30 e.g. tokens { name: "sleep" } -> sleep 31 32 Args: 33 deterministic: if True will provide a single transduction option, 34 for False multiple transduction are generated (used for audio-based normalization) 35 """ 36 37 def __init__(self, deterministic: bool = True): 38 super().__init__(name="word", kind="verbalize", deterministic=deterministic) 39 chars = pynini.closure(NEMO_CHAR - " ", 1) 40 char = pynutil.delete("name:") + delete_space + pynutil.delete("\"") + chars + pynutil.delete("\"") 41 graph = char @ pynini.cdrewrite(pynini.cross(u"\u00A0", " "), "", "", NEMO_SIGMA) 42 43 self.fst = graph.optimize() 44 [end of nemo_text_processing/text_normalization/en/verbalizers/word.py] [start of nemo_text_processing/text_normalization/normalize.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import itertools 16 import os 17 import re 18 from argparse import ArgumentParser 19 from collections import OrderedDict 20 from math import factorial 21 from typing import Dict, List, Union 22 23 from nemo_text_processing.text_normalization.data_loader_utils import get_installation_msg, pre_process 24 from nemo_text_processing.text_normalization.token_parser import PRESERVE_ORDER_KEY, TokenParser 25 from tqdm import tqdm 26 27 try: 28 import pynini 29 30 PYNINI_AVAILABLE = True 31 32 except (ModuleNotFoundError, ImportError): 33 PYNINI_AVAILABLE = False 34 35 try: 36 from nemo.collections.common.tokenizers.moses_tokenizers import MosesProcessor 37 from nemo.collections.nlp.data.text_normalization.utils import post_process_punct 38 39 NLP_AVAILABLE = True 40 except (ModuleNotFoundError, ImportError): 41 NLP_AVAILABLE = False 42 43 44 SPACE_DUP = re.compile(' {2,}') 45 46 47 class Normalizer: 48 """ 49 Normalizer class that converts text from written to spoken form. 50 Useful for TTS preprocessing. 51 52 Args: 53 input_case: expected input capitalization 54 lang: language specifying the TN rules, by default: English 55 cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. 56 overwrite_cache: set to True to overwrite .far files 57 whitelist: path to a file with whitelist replacements 58 """ 59 60 def __init__( 61 self, 62 input_case: str, 63 lang: str = 'en', 64 deterministic: bool = True, 65 cache_dir: str = None, 66 overwrite_cache: bool = False, 67 whitelist: str = None, 68 ): 69 assert input_case in ["lower_cased", "cased"] 70 71 if not PYNINI_AVAILABLE: 72 raise ImportError(get_installation_msg()) 73 74 if lang == 'en' and deterministic: 75 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify import ClassifyFst 76 from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst 77 elif lang == 'en' and not deterministic: 78 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify_with_audio import ClassifyFst 79 from nemo_text_processing.text_normalization.en.verbalizers.verbalize_final import VerbalizeFinalFst 80 elif lang == 'ru': 81 # Ru TN only support non-deterministic cases and produces multiple normalization options 82 # use normalize_with_audio.py 83 from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst 84 from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst 85 elif lang == 'de': 86 # Ru TN only support non-deterministic cases and produces multiple normalization options 87 # use normalize_with_audio.py 88 from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst 89 from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst 90 self.tagger = ClassifyFst( 91 input_case=input_case, 92 deterministic=deterministic, 93 cache_dir=cache_dir, 94 overwrite_cache=overwrite_cache, 95 whitelist=whitelist, 96 ) 97 self.verbalizer = VerbalizeFinalFst(deterministic=deterministic) 98 self.parser = TokenParser() 99 self.lang = lang 100 101 if NLP_AVAILABLE: 102 self.processor = MosesProcessor(lang_id=lang) 103 else: 104 self.processor = None 105 print("NeMo NLP is not available. Moses de-tokenization will be skipped.") 106 107 def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: 108 """ 109 NeMo text normalizer 110 111 Args: 112 texts: list of input strings 113 verbose: whether to print intermediate meta information 114 115 Returns converted list input strings 116 """ 117 res = [] 118 for input in tqdm(texts): 119 try: 120 text = self.normalize(input, verbose=verbose, punct_post_process=punct_post_process) 121 except: 122 print(input) 123 raise Exception 124 res.append(text) 125 return res 126 127 def _estimate_number_of_permutations_in_nested_dict( 128 self, token_group: Dict[str, Union[OrderedDict, str, bool]] 129 ) -> int: 130 num_perms = 1 131 for k, inner in token_group.items(): 132 if isinstance(inner, dict): 133 num_perms *= self._estimate_number_of_permutations_in_nested_dict(inner) 134 num_perms *= factorial(len(token_group)) 135 return num_perms 136 137 def _split_tokens_to_reduce_number_of_permutations( 138 self, tokens: List[dict], max_number_of_permutations_per_split: int = 729 139 ) -> List[List[dict]]: 140 """ 141 Splits a sequence of tokens in a smaller sequences of tokens in a way that maximum number of composite 142 tokens permutations does not exceed ``max_number_of_permutations_per_split``. 143 144 For example, 145 146 .. code-block:: python 147 tokens = [ 148 {"tokens": {"date": {"year": "twenty eighteen", "month": "december", "day": "thirty one"}}}, 149 {"tokens": {"date": {"year": "twenty eighteen", "month": "january", "day": "eight"}}}, 150 ] 151 split = normalizer._split_tokens_to_reduce_number_of_permutations( 152 tokens, max_number_of_permutations_per_split=6 153 ) 154 assert split == [ 155 [{"tokens": {"date": {"year": "twenty eighteen", "month": "december", "day": "thirty one"}}}], 156 [{"tokens": {"date": {"year": "twenty eighteen", "month": "january", "day": "eight"}}}], 157 ] 158 159 Date tokens contain 3 items each which gives 6 permutations for every date. Since there are 2 dates, total 160 number of permutations would be ``6 * 6 == 36``. Parameter ``max_number_of_permutations_per_split`` equals 6, 161 so input sequence of tokens is split into 2 smaller sequences. 162 163 Args: 164 tokens (:obj:`List[dict]`): a list of dictionaries, possibly nested. 165 max_number_of_permutations_per_split (:obj:`int`, `optional`, defaults to :obj:`243`): a maximum number 166 of permutations which can be generated from input sequence of tokens. 167 168 Returns: 169 :obj:`List[List[dict]]`: a list of smaller sequences of tokens resulting from ``tokens`` split. 170 """ 171 splits = [] 172 prev_end_of_split = 0 173 current_number_of_permutations = 1 174 for i, token_group in enumerate(tokens): 175 n = self._estimate_number_of_permutations_in_nested_dict(token_group) 176 if n * current_number_of_permutations > max_number_of_permutations_per_split: 177 splits.append(tokens[prev_end_of_split:i]) 178 prev_end_of_split = i 179 current_number_of_permutations = 1 180 if n > max_number_of_permutations_per_split: 181 raise ValueError( 182 f"Could not split token list with respect to condition that every split can generate number of " 183 f"permutations less or equal to " 184 f"`max_number_of_permutations_per_split={max_number_of_permutations_per_split}`. " 185 f"There is an unsplittable token group that generates more than " 186 f"{max_number_of_permutations_per_split} permutations. Try to increase " 187 f"`max_number_of_permutations_per_split` parameter." 188 ) 189 current_number_of_permutations *= n 190 splits.append(tokens[prev_end_of_split:]) 191 assert sum([len(s) for s in splits]) == len(tokens) 192 return splits 193 194 def normalize( 195 self, text: str, verbose: bool = False, punct_pre_process: bool = False, punct_post_process: bool = False 196 ) -> str: 197 """ 198 Main function. Normalizes tokens from written to spoken form 199 e.g. 12 kg -> twelve kilograms 200 201 Args: 202 text: string that may include semiotic classes 203 verbose: whether to print intermediate meta information 204 punct_pre_process: whether to perform punctuation pre-processing, for example, [25] -> [ 25 ] 205 punct_post_process: whether to normalize punctuation 206 207 Returns: spoken form 208 """ 209 original_text = text 210 if punct_pre_process: 211 text = pre_process(text) 212 text = text.strip() 213 if not text: 214 if verbose: 215 print(text) 216 return text 217 text = pynini.escape(text) 218 tagged_lattice = self.find_tags(text) 219 tagged_text = self.select_tag(tagged_lattice) 220 if verbose: 221 print(tagged_text) 222 self.parser(tagged_text) 223 tokens = self.parser.parse() 224 split_tokens = self._split_tokens_to_reduce_number_of_permutations(tokens) 225 output = "" 226 for s in split_tokens: 227 tags_reordered = self.generate_permutations(s) 228 verbalizer_lattice = None 229 for tagged_text in tags_reordered: 230 tagged_text = pynini.escape(tagged_text) 231 232 verbalizer_lattice = self.find_verbalizer(tagged_text) 233 if verbalizer_lattice.num_states() != 0: 234 break 235 if verbalizer_lattice is None: 236 raise ValueError(f"No permutations were generated from tokens {s}") 237 output += ' ' + self.select_verbalizer(verbalizer_lattice) 238 output = SPACE_DUP.sub(' ', output[1:]) 239 if punct_post_process: 240 # do post-processing based on Moses detokenizer 241 if self.processor: 242 output = self.processor.moses_detokenizer.detokenize([output], unescape=False) 243 output = post_process_punct(input=original_text, normalized_text=output) 244 else: 245 print("NEMO_NLP collection is not available: skipping punctuation post_processing") 246 return output 247 248 def _permute(self, d: OrderedDict) -> List[str]: 249 """ 250 Creates reorderings of dictionary elements and serializes as strings 251 252 Args: 253 d: (nested) dictionary of key value pairs 254 255 Return permutations of different string serializations of key value pairs 256 """ 257 l = [] 258 if PRESERVE_ORDER_KEY in d.keys(): 259 d_permutations = [d.items()] 260 else: 261 d_permutations = itertools.permutations(d.items()) 262 for perm in d_permutations: 263 subl = [""] 264 for k, v in perm: 265 if isinstance(v, str): 266 subl = ["".join(x) for x in itertools.product(subl, [f"{k}: \"{v}\" "])] 267 elif isinstance(v, OrderedDict): 268 rec = self._permute(v) 269 subl = ["".join(x) for x in itertools.product(subl, [f" {k} {{ "], rec, [f" }} "])] 270 elif isinstance(v, bool): 271 subl = ["".join(x) for x in itertools.product(subl, [f"{k}: true "])] 272 else: 273 raise ValueError() 274 l.extend(subl) 275 return l 276 277 def generate_permutations(self, tokens: List[dict]): 278 """ 279 Generates permutations of string serializations of list of dictionaries 280 281 Args: 282 tokens: list of dictionaries 283 284 Returns string serialization of list of dictionaries 285 """ 286 287 def _helper(prefix: str, tokens: List[dict], idx: int): 288 """ 289 Generates permutations of string serializations of given dictionary 290 291 Args: 292 tokens: list of dictionaries 293 prefix: prefix string 294 idx: index of next dictionary 295 296 Returns string serialization of dictionary 297 """ 298 if idx == len(tokens): 299 yield prefix 300 return 301 token_options = self._permute(tokens[idx]) 302 for token_option in token_options: 303 yield from _helper(prefix + token_option, tokens, idx + 1) 304 305 return _helper("", tokens, 0) 306 307 def find_tags(self, text: str) -> 'pynini.FstLike': 308 """ 309 Given text use tagger Fst to tag text 310 311 Args: 312 text: sentence 313 314 Returns: tagged lattice 315 """ 316 lattice = text @ self.tagger.fst 317 return lattice 318 319 def select_tag(self, lattice: 'pynini.FstLike') -> str: 320 """ 321 Given tagged lattice return shortest path 322 323 Args: 324 tagged_text: tagged text 325 326 Returns: shortest path 327 """ 328 tagged_text = pynini.shortestpath(lattice, nshortest=1, unique=True).string() 329 return tagged_text 330 331 def find_verbalizer(self, tagged_text: str) -> 'pynini.FstLike': 332 """ 333 Given tagged text creates verbalization lattice 334 This is context-independent. 335 336 Args: 337 tagged_text: input text 338 339 Returns: verbalized lattice 340 """ 341 lattice = tagged_text @ self.verbalizer.fst 342 return lattice 343 344 def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: 345 """ 346 Given verbalized lattice return shortest path 347 348 Args: 349 lattice: verbalization lattice 350 351 Returns: shortest path 352 """ 353 output = pynini.shortestpath(lattice, nshortest=1, unique=True).string() 354 return output 355 356 357 def parse_args(): 358 parser = ArgumentParser() 359 parser.add_argument("input_string", help="input string", type=str) 360 parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) 361 parser.add_argument( 362 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 363 ) 364 parser.add_argument("--verbose", help="print info for debugging", action='store_true') 365 parser.add_argument( 366 "--punct_post_process", help="set to True to enable punctuation post processing", action="store_true" 367 ) 368 parser.add_argument( 369 "--punct_pre_process", help="set to True to enable punctuation pre processing", action="store_true" 370 ) 371 parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") 372 parser.add_argument("--whitelist", help="path to a file with with whitelist", default=None, type=str) 373 parser.add_argument( 374 "--cache_dir", 375 help="path to a dir with .far grammar file. Set to None to avoid using cache", 376 default=None, 377 type=str, 378 ) 379 return parser.parse_args() 380 381 382 if __name__ == "__main__": 383 args = parse_args() 384 whitelist = os.path.abspath(args.whitelist) if args.whitelist else None 385 normalizer = Normalizer( 386 input_case=args.input_case, 387 cache_dir=args.cache_dir, 388 overwrite_cache=args.overwrite_cache, 389 whitelist=whitelist, 390 lang=args.language, 391 ) 392 print( 393 normalizer.normalize( 394 args.input_string, 395 verbose=args.verbose, 396 punct_pre_process=args.punct_pre_process, 397 punct_post_process=args.punct_post_process, 398 ) 399 ) 400 [end of nemo_text_processing/text_normalization/normalize.py] [start of nemo_text_processing/text_normalization/normalize_with_audio.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import json 16 import os 17 import time 18 from argparse import ArgumentParser 19 from glob import glob 20 from typing import List, Tuple 21 22 from joblib import Parallel, delayed 23 from nemo_text_processing.text_normalization.normalize import Normalizer 24 from tqdm import tqdm 25 26 try: 27 from nemo.collections.asr.metrics.wer import word_error_rate 28 from nemo.collections.asr.models import ASRModel 29 30 ASR_AVAILABLE = True 31 except (ModuleNotFoundError, ImportError): 32 ASR_AVAILABLE = False 33 34 try: 35 import pynini 36 from pynini.lib import rewrite 37 38 PYNINI_AVAILABLE = True 39 except (ModuleNotFoundError, ImportError): 40 PYNINI_AVAILABLE = False 41 42 try: 43 from nemo.collections.nlp.data.text_normalization.utils import post_process_punct 44 from nemo_text_processing.text_normalization.data_loader_utils import pre_process 45 46 NLP_AVAILABLE = True 47 except (ModuleNotFoundError, ImportError): 48 NLP_AVAILABLE = False 49 50 """ 51 The script provides multiple normalization options and chooses the best one that minimizes CER of the ASR output 52 (most of the semiotic classes use deterministic=False flag). 53 54 To run this script with a .json manifest file, the manifest file should contain the following fields: 55 "audio_data" - path to the audio file 56 "text" - raw text 57 "pred_text" - ASR model prediction 58 59 See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions 60 61 When the manifest is ready, run: 62 python normalize_with_audio.py \ 63 --audio_data PATH/TO/MANIFEST.JSON \ 64 --language en 65 66 67 To run with a single audio file, specify path to audio and text with: 68 python normalize_with_audio.py \ 69 --audio_data PATH/TO/AUDIO.WAV \ 70 --language en \ 71 --text raw text OR PATH/TO/.TXT/FILE 72 --model QuartzNet15x5Base-En \ 73 --verbose 74 75 To see possible normalization options for a text input without an audio file (could be used for debugging), run: 76 python python normalize_with_audio.py --text "RAW TEXT" 77 78 Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference 79 """ 80 81 82 class NormalizerWithAudio(Normalizer): 83 """ 84 Normalizer class that converts text from written to spoken form. 85 Useful for TTS preprocessing. 86 87 Args: 88 input_case: expected input capitalization 89 lang: language 90 cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. 91 overwrite_cache: set to True to overwrite .far files 92 whitelist: path to a file with whitelist replacements 93 """ 94 95 def __init__( 96 self, 97 input_case: str, 98 lang: str = 'en', 99 cache_dir: str = None, 100 overwrite_cache: bool = False, 101 whitelist: str = None, 102 ): 103 104 super().__init__( 105 input_case=input_case, 106 lang=lang, 107 deterministic=False, 108 cache_dir=cache_dir, 109 overwrite_cache=overwrite_cache, 110 whitelist=whitelist, 111 ) 112 113 def normalize(self, text: str, n_tagged: int, punct_post_process: bool = True, verbose: bool = False,) -> str: 114 """ 115 Main function. Normalizes tokens from written to spoken form 116 e.g. 12 kg -> twelve kilograms 117 118 Args: 119 text: string that may include semiotic classes 120 n_tagged: number of tagged options to consider, -1 - to get all possible tagged options 121 punct_post_process: whether to normalize punctuation 122 verbose: whether to print intermediate meta information 123 124 Returns: 125 normalized text options (usually there are multiple ways of normalizing a given semiotic class) 126 """ 127 original_text = text 128 129 if self.lang == "en": 130 text = pre_process(text) 131 text = text.strip() 132 if not text: 133 if verbose: 134 print(text) 135 return text 136 text = pynini.escape(text) 137 138 if n_tagged == -1: 139 if self.lang == "en": 140 try: 141 tagged_texts = rewrite.rewrites(text, self.tagger.fst_no_digits) 142 except pynini.lib.rewrite.Error: 143 tagged_texts = rewrite.rewrites(text, self.tagger.fst) 144 else: 145 tagged_texts = rewrite.rewrites(text, self.tagger.fst) 146 else: 147 if self.lang == "en": 148 try: 149 tagged_texts = rewrite.top_rewrites(text, self.tagger.fst_no_digits, nshortest=n_tagged) 150 except pynini.lib.rewrite.Error: 151 tagged_texts = rewrite.top_rewrites(text, self.tagger.fst, nshortest=n_tagged) 152 else: 153 tagged_texts = rewrite.top_rewrites(text, self.tagger.fst, nshortest=n_tagged) 154 155 # non-deterministic Eng normalization uses tagger composed with verbalizer, no permutation in between 156 if self.lang == "en": 157 normalized_texts = tagged_texts 158 else: 159 normalized_texts = [] 160 for tagged_text in tagged_texts: 161 self._verbalize(tagged_text, normalized_texts, verbose=verbose) 162 163 if len(normalized_texts) == 0: 164 raise ValueError() 165 166 if punct_post_process: 167 # do post-processing based on Moses detokenizer 168 if self.processor: 169 normalized_texts = [self.processor.detokenize([t]) for t in normalized_texts] 170 normalized_texts = [ 171 post_process_punct(input=original_text, normalized_text=t) for t in normalized_texts 172 ] 173 174 normalized_texts = set(normalized_texts) 175 return normalized_texts 176 177 def _verbalize(self, tagged_text: str, normalized_texts: List[str], verbose: bool = False): 178 """ 179 Verbalizes tagged text 180 181 Args: 182 tagged_text: text with tags 183 normalized_texts: list of possible normalization options 184 verbose: if true prints intermediate classification results 185 """ 186 187 def get_verbalized_text(tagged_text): 188 return rewrite.rewrites(tagged_text, self.verbalizer.fst) 189 190 self.parser(tagged_text) 191 tokens = self.parser.parse() 192 tags_reordered = self.generate_permutations(tokens) 193 for tagged_text_reordered in tags_reordered: 194 try: 195 tagged_text_reordered = pynini.escape(tagged_text_reordered) 196 normalized_texts.extend(get_verbalized_text(tagged_text_reordered)) 197 if verbose: 198 print(tagged_text_reordered) 199 200 except pynini.lib.rewrite.Error: 201 continue 202 203 def select_best_match( 204 self, 205 normalized_texts: List[str], 206 input_text: str, 207 pred_text: str, 208 verbose: bool = False, 209 remove_punct: bool = False, 210 ): 211 """ 212 Selects the best normalization option based on the lowest CER 213 214 Args: 215 normalized_texts: normalized text options 216 input_text: input text 217 pred_text: ASR model transcript of the audio file corresponding to the normalized text 218 verbose: whether to print intermediate meta information 219 remove_punct: whether to remove punctuation before calculating CER 220 221 Returns: 222 normalized text with the lowest CER and CER value 223 """ 224 if pred_text == "": 225 return input_text, 1000 226 227 normalized_texts_cer = calculate_cer(normalized_texts, pred_text, remove_punct) 228 normalized_texts_cer = sorted(normalized_texts_cer, key=lambda x: x[1]) 229 normalized_text, cer = normalized_texts_cer[0] 230 231 if verbose: 232 print('-' * 30) 233 for option in normalized_texts: 234 print(option) 235 print('-' * 30) 236 return normalized_text, cer 237 238 239 def calculate_cer(normalized_texts: List[str], pred_text: str, remove_punct=False) -> List[Tuple[str, float]]: 240 """ 241 Calculates character error rate (CER) 242 243 Args: 244 normalized_texts: normalized text options 245 pred_text: ASR model output 246 247 Returns: normalized options with corresponding CER 248 """ 249 normalized_options = [] 250 for text in normalized_texts: 251 text_clean = text.replace('-', ' ').lower() 252 if remove_punct: 253 for punct in "!?:;,.-()*+-/<=>@^_": 254 text_clean = text_clean.replace(punct, "") 255 cer = round(word_error_rate([pred_text], [text_clean], use_cer=True) * 100, 2) 256 normalized_options.append((text, cer)) 257 return normalized_options 258 259 260 def get_asr_model(asr_model): 261 """ 262 Returns ASR Model 263 264 Args: 265 asr_model: NeMo ASR model 266 """ 267 if os.path.exists(args.model): 268 asr_model = ASRModel.restore_from(asr_model) 269 elif args.model in ASRModel.get_available_model_names(): 270 asr_model = ASRModel.from_pretrained(asr_model) 271 else: 272 raise ValueError( 273 f'Provide path to the pretrained checkpoint or choose from {ASRModel.get_available_model_names()}' 274 ) 275 return asr_model 276 277 278 def parse_args(): 279 parser = ArgumentParser() 280 parser.add_argument("--text", help="input string or path to a .txt file", default=None, type=str) 281 parser.add_argument( 282 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 283 ) 284 parser.add_argument( 285 "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str 286 ) 287 parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") 288 parser.add_argument( 289 '--model', type=str, default='QuartzNet15x5Base-En', help='Pre-trained model name or path to model checkpoint' 290 ) 291 parser.add_argument( 292 "--n_tagged", 293 type=int, 294 default=30, 295 help="number of tagged options to consider, -1 - return all possible tagged options", 296 ) 297 parser.add_argument("--verbose", help="print info for debugging", action="store_true") 298 parser.add_argument( 299 "--no_remove_punct_for_cer", 300 help="Set to True to NOT remove punctuation before calculating CER", 301 action="store_true", 302 ) 303 parser.add_argument( 304 "--no_punct_post_process", help="set to True to disable punctuation post processing", action="store_true" 305 ) 306 parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") 307 parser.add_argument("--whitelist", help="path to a file with with whitelist", default=None, type=str) 308 parser.add_argument( 309 "--cache_dir", 310 help="path to a dir with .far grammar file. Set to None to avoid using cache", 311 default=None, 312 type=str, 313 ) 314 parser.add_argument("--n_jobs", default=-2, type=int, help="The maximum number of concurrently running jobs") 315 parser.add_argument("--batch_size", default=200, type=int, help="Number of examples for each process") 316 return parser.parse_args() 317 318 319 def _normalize_line(normalizer: NormalizerWithAudio, n_tagged, verbose, line: str, remove_punct, punct_post_process): 320 line = json.loads(line) 321 pred_text = line["pred_text"] 322 323 normalized_texts = normalizer.normalize( 324 text=line["text"], verbose=verbose, n_tagged=n_tagged, punct_post_process=punct_post_process, 325 ) 326 327 normalized_text, cer = normalizer.select_best_match( 328 normalized_texts=normalized_texts, 329 input_text=line["text"], 330 pred_text=pred_text, 331 verbose=verbose, 332 remove_punct=remove_punct, 333 ) 334 line["nemo_normalized"] = normalized_text 335 line["CER_nemo_normalized"] = cer 336 return line 337 338 339 def normalize_manifest( 340 normalizer, 341 audio_data: str, 342 n_jobs: int, 343 n_tagged: int, 344 remove_punct: bool, 345 punct_post_process: bool, 346 batch_size: int, 347 ): 348 """ 349 Args: 350 args.audio_data: path to .json manifest file. 351 """ 352 353 def __process_batch(batch_idx, batch, dir_name): 354 normalized_lines = [ 355 _normalize_line( 356 normalizer, 357 n_tagged, 358 verbose=False, 359 line=line, 360 remove_punct=remove_punct, 361 punct_post_process=punct_post_process, 362 ) 363 for line in tqdm(batch) 364 ] 365 366 with open(f"{dir_name}/{batch_idx}.json", "w") as f_out: 367 for line in normalized_lines: 368 f_out.write(json.dumps(line, ensure_ascii=False) + '\n') 369 370 print(f"Batch -- {batch_idx} -- is complete") 371 return normalized_lines 372 373 manifest_out = audio_data.replace('.json', '_normalized.json') 374 with open(audio_data, 'r') as f: 375 lines = f.readlines() 376 377 print(f'Normalizing {len(lines)} lines of {audio_data}...') 378 379 # to save intermediate results to a file 380 batch = min(len(lines), batch_size) 381 382 tmp_dir = manifest_out.replace(".json", "_parts") 383 os.makedirs(tmp_dir, exist_ok=True) 384 385 Parallel(n_jobs=n_jobs)( 386 delayed(__process_batch)(idx, lines[i : i + batch], tmp_dir) 387 for idx, i in enumerate(range(0, len(lines), batch)) 388 ) 389 390 # aggregate all intermediate files 391 with open(manifest_out, "w") as f_out: 392 for batch_f in sorted(glob(f"{tmp_dir}/*.json")): 393 with open(batch_f, "r") as f_in: 394 lines = f_in.read() 395 f_out.write(lines) 396 397 print(f'Normalized version saved at {manifest_out}') 398 399 400 if __name__ == "__main__": 401 args = parse_args() 402 403 if not ASR_AVAILABLE and args.audio_data: 404 raise ValueError("NeMo ASR collection is not installed.") 405 start = time.time() 406 args.whitelist = os.path.abspath(args.whitelist) if args.whitelist else None 407 if args.text is not None: 408 normalizer = NormalizerWithAudio( 409 input_case=args.input_case, 410 lang=args.language, 411 cache_dir=args.cache_dir, 412 overwrite_cache=args.overwrite_cache, 413 whitelist=args.whitelist, 414 ) 415 416 if os.path.exists(args.text): 417 with open(args.text, 'r') as f: 418 args.text = f.read().strip() 419 normalized_texts = normalizer.normalize( 420 text=args.text, 421 verbose=args.verbose, 422 n_tagged=args.n_tagged, 423 punct_post_process=not args.no_punct_post_process, 424 ) 425 426 if args.audio_data: 427 asr_model = get_asr_model(args.model) 428 pred_text = asr_model.transcribe([args.audio_data])[0] 429 normalized_text, cer = normalizer.select_best_match( 430 normalized_texts=normalized_texts, 431 pred_text=pred_text, 432 input_text=args.text, 433 verbose=args.verbose, 434 remove_punct=not args.no_remove_punct_for_cer, 435 ) 436 print(f"Transcript: {pred_text}") 437 print(f"Normalized: {normalized_text}") 438 else: 439 print("Normalization options:") 440 for norm_text in normalized_texts: 441 print(norm_text) 442 elif not os.path.exists(args.audio_data): 443 raise ValueError(f"{args.audio_data} not found.") 444 elif args.audio_data.endswith('.json'): 445 normalizer = NormalizerWithAudio( 446 input_case=args.input_case, 447 lang=args.language, 448 cache_dir=args.cache_dir, 449 overwrite_cache=args.overwrite_cache, 450 whitelist=args.whitelist, 451 ) 452 normalize_manifest( 453 normalizer=normalizer, 454 audio_data=args.audio_data, 455 n_jobs=args.n_jobs, 456 n_tagged=args.n_tagged, 457 remove_punct=not args.no_remove_punct_for_cer, 458 punct_post_process=not args.no_punct_post_process, 459 batch_size=args.batch_size, 460 ) 461 else: 462 raise ValueError( 463 "Provide either path to .json manifest in '--audio_data' OR " 464 + "'--audio_data' path to audio file and '--text' path to a text file OR" 465 "'--text' string text (for debugging without audio)" 466 ) 467 print(f'Execution time: {round((time.time() - start)/60, 2)} min.') 468 [end of nemo_text_processing/text_normalization/normalize_with_audio.py] [start of tools/text_processing_deployment/pynini_export.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # Copyright 2015 and onwards Google, Inc. 3 # 4 # Licensed under the Apache License, Version 2.0 (the "License"); 5 # you may not use this file except in compliance with the License. 6 # You may obtain a copy of the License at 7 # 8 # http://www.apache.org/licenses/LICENSE-2.0 9 # 10 # Unless required by applicable law or agreed to in writing, software 11 # distributed under the License is distributed on an "AS IS" BASIS, 12 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 # See the License for the specific language governing permissions and 14 # limitations under the License. 15 16 17 import os 18 import time 19 from argparse import ArgumentParser 20 21 from nemo.utils import logging 22 23 try: 24 import pynini 25 from nemo_text_processing.text_normalization.en.graph_utils import generator_main 26 27 PYNINI_AVAILABLE = True 28 except (ModuleNotFoundError, ImportError): 29 30 logging.warning( 31 "`pynini` is not installed ! \n" 32 "Please run the `nemo_text_processing/setup.sh` script" 33 "prior to usage of this toolkit." 34 ) 35 36 PYNINI_AVAILABLE = False 37 38 39 # This script exports compiled grammars inside nemo_text_processing into OpenFst finite state archive files 40 # tokenize_and_classify.far and verbalize.far for production purposes 41 42 43 def itn_grammars(**kwargs): 44 d = {} 45 d['classify'] = { 46 'TOKENIZE_AND_CLASSIFY': ITNClassifyFst( 47 cache_dir=kwargs["cache_dir"], overwrite_cache=kwargs["overwrite_cache"] 48 ).fst 49 } 50 d['verbalize'] = {'ALL': ITNVerbalizeFst().fst, 'REDUP': pynini.accep("REDUP")} 51 return d 52 53 54 def tn_grammars(**kwargs): 55 d = {} 56 d['classify'] = { 57 'TOKENIZE_AND_CLASSIFY': TNClassifyFst( 58 input_case=kwargs["input_case"], 59 deterministic=True, 60 cache_dir=kwargs["cache_dir"], 61 overwrite_cache=kwargs["overwrite_cache"], 62 ).fst 63 } 64 d['verbalize'] = {'ALL': TNVerbalizeFst(deterministic=True).fst, 'REDUP': pynini.accep("REDUP")} 65 return d 66 67 68 def export_grammars(output_dir, grammars): 69 """ 70 Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. 71 72 Args: 73 output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. 74 grammars: grammars to be exported 75 """ 76 77 for category, graphs in grammars.items(): 78 out_dir = os.path.join(output_dir, category) 79 if not os.path.exists(out_dir): 80 os.makedirs(out_dir) 81 time.sleep(1) 82 if category == "classify": 83 category = "tokenize_and_classify" 84 generator_main(f"{out_dir}/{category}.far", graphs) 85 86 87 def parse_args(): 88 parser = ArgumentParser() 89 parser.add_argument("--output_dir", help="output directory for grammars", required=True, type=str) 90 parser.add_argument( 91 "--language", help="language", choices=["en", "de", "es", "ru", 'fr', 'vi'], type=str, default='en' 92 ) 93 parser.add_argument( 94 "--grammars", help="grammars to be exported", choices=["tn_grammars", "itn_grammars"], type=str, required=True 95 ) 96 parser.add_argument( 97 "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str 98 ) 99 parser.add_argument("--overwrite_cache", help="set to True to re-create .far grammar files", action="store_true") 100 parser.add_argument( 101 "--cache_dir", 102 help="path to a dir with .far grammar file. Set to None to avoid using cache", 103 default=None, 104 type=str, 105 ) 106 return parser.parse_args() 107 108 109 if __name__ == '__main__': 110 args = parse_args() 111 112 if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': 113 raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') 114 115 if args.language == 'en': 116 from nemo_text_processing.inverse_text_normalization.en.taggers.tokenize_and_classify import ( 117 ClassifyFst as ITNClassifyFst, 118 ) 119 from nemo_text_processing.inverse_text_normalization.en.verbalizers.verbalize import ( 120 VerbalizeFst as ITNVerbalizeFst, 121 ) 122 from nemo_text_processing.text_normalization.en.taggers.tokenize_and_classify import ( 123 ClassifyFst as TNClassifyFst, 124 ) 125 from nemo_text_processing.text_normalization.en.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst 126 elif args.language == 'de': 127 from nemo_text_processing.inverse_text_normalization.de.taggers.tokenize_and_classify import ( 128 ClassifyFst as ITNClassifyFst, 129 ) 130 from nemo_text_processing.inverse_text_normalization.de.verbalizers.verbalize import ( 131 VerbalizeFst as ITNVerbalizeFst, 132 ) 133 from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ( 134 ClassifyFst as TNClassifyFst, 135 ) 136 from nemo_text_processing.text_normalization.de.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst 137 elif args.language == 'ru': 138 from nemo_text_processing.inverse_text_normalization.ru.taggers.tokenize_and_classify import ( 139 ClassifyFst as ITNClassifyFst, 140 ) 141 from nemo_text_processing.inverse_text_normalization.ru.verbalizers.verbalize import ( 142 VerbalizeFst as ITNVerbalizeFst, 143 ) 144 elif args.language == 'es': 145 from nemo_text_processing.inverse_text_normalization.es.taggers.tokenize_and_classify import ( 146 ClassifyFst as ITNClassifyFst, 147 ) 148 from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( 149 VerbalizeFst as ITNVerbalizeFst, 150 ) 151 elif args.language == 'fr': 152 from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( 153 ClassifyFst as ITNClassifyFst, 154 ) 155 from nemo_text_processing.inverse_text_normalization.fr.verbalizers.verbalize import ( 156 VerbalizeFst as ITNVerbalizeFst, 157 ) 158 elif args.language == 'vi': 159 from nemo_text_processing.inverse_text_normalization.vi.taggers.tokenize_and_classify import ( 160 ClassifyFst as ITNClassifyFst, 161 ) 162 from nemo_text_processing.inverse_text_normalization.vi.verbalizers.verbalize import ( 163 VerbalizeFst as ITNVerbalizeFst, 164 ) 165 166 output_dir = os.path.join(args.output_dir, args.language) 167 export_grammars( 168 output_dir=output_dir, 169 grammars=locals()[args.grammars]( 170 input_case=args.input_case, cache_dir=args.cache_dir, overwrite_cache=args.overwrite_cache 171 ), 172 ) 173 [end of tools/text_processing_deployment/pynini_export.py] </code> Here is an example of a patch file. It consists of changes to the code base. It specifies the file names, the line numbers of each change, and the removed and added lines. A single patch file can contain changes to multiple files. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch> I need you to solve the provided issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the format shown above. Respond below:
NVIDIA/NeMo
022f0292aecbc98d591d49423d5045235394f793
./reinstall.sh crashes due to not being able to uninstall llvmlite Starting off of `nemo:1.5.1` container, cloning the NeMo repo to a folder inside of it and calling `./reinstall.sh` fails with ``` ERROR: Cannot uninstall 'llvmlite'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall. ``` `pip install -e` on the other hand succeeds installing `nemo:1.7.0rc0` and `numpy:1.22.2`, the rest of the packages remain untouched. It seems that `./reinstall.sh` which used to work fine, a week or so ago when following the same procedure to upgrade to `nemo:1.6.0rc` redeveloped issue #841. The solution remains the same, first call ``` pip install --ignore-installed llvmlite ``` followed by `./reinstall.sh`. In this case, apart `llvml`, the following packages are updated ``` ftfy-6.0.3 nemo-toolkit-1.7.0rc0 numba-0.55.1 pytorch-lightning-1.5.9 sacrebleu-2.0.0 setuptools-59.5.0 ``` Interestingly `numpy` in this case is left at `1.21.5`.
2022-02-09T05:12:31Z
<patch> <patch> diff --git a/nemo_text_processing/text_normalization/__init__.py b/nemo_text_processing/text_normalization/__init__.py --- a/nemo_text_processing/text_normalization/__init__.py +++ b/nemo_text_processing/text_normalization/__init__.py @@ -21,7 +21,7 @@ except (ModuleNotFoundError, ImportError): logging.warning( "`pynini` is not installed ! \n" - "Please run the `nemo_text_processing/setup.sh` script" + "Please run the `nemo_text_processing/setup.sh` script " "prior to usage of this toolkit." ) diff --git a/nemo_text_processing/text_normalization/en/graph_utils.py b/nemo_text_processing/text_normalization/en/graph_utils.py --- a/nemo_text_processing/text_normalization/en/graph_utils.py +++ b/nemo_text_processing/text_normalization/en/graph_utils.py @@ -159,7 +159,7 @@ def convert_space(fst) -> 'pynini.FstLike': """ Converts space to nonbreaking space. Used only in tagger grammars for transducing token values within quotes, e.g. name: "hello kitty" - This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. + This is making transducer significantly slower, so only use when there could be potential spaces within quotes, otherwise leave it. Args: fst: input fst @@ -208,9 +208,9 @@ def add_tokens(self, fst) -> 'pynini.FstLike': """ Wraps class name around to given fst - Args: + Args: fst: input fst - + Returns: Fst: fst """ diff --git a/nemo_text_processing/text_normalization/en/taggers/punctuation.py b/nemo_text_processing/text_normalization/en/taggers/punctuation.py --- a/nemo_text_processing/text_normalization/en/taggers/punctuation.py +++ b/nemo_text_processing/text_normalization/en/taggers/punctuation.py @@ -22,7 +22,7 @@ import pynini from pynini.lib import pynutil - PYNINI_AVAILABLE = False + PYNINI_AVAILABLE = True except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py --- a/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/whitelist.py @@ -12,8 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -21,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/en/verbalizers/word.py b/nemo_text_processing/text_normalization/en/verbalizers/word.py --- a/nemo_text_processing/text_normalization/en/verbalizers/word.py +++ b/nemo_text_processing/text_normalization/en/verbalizers/word.py @@ -12,7 +12,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - from nemo_text_processing.text_normalization.en.graph_utils import NEMO_CHAR, NEMO_SIGMA, GraphFst, delete_space try: @@ -20,6 +19,7 @@ from pynini.lib import pynutil PYNINI_AVAILABLE = True + except (ModuleNotFoundError, ImportError): PYNINI_AVAILABLE = False diff --git a/nemo_text_processing/text_normalization/es/__init__.py b/nemo_text_processing/text_normalization/es/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/__init__.py @@ -0,0 +1,15 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +LOCALIZATION = "eu" # Set to am for alternate formatting diff --git a/nemo_text_processing/text_normalization/es/data/__init__.py b/nemo_text_processing/text_normalization/es/data/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/dates/__init__.py b/nemo_text_processing/text_normalization/es/data/dates/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/dates/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/electronic/__init__.py b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/electronic/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/fractions/__init__.py b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/fractions/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/measures/__init__.py b/nemo_text_processing/text_normalization/es/data/measures/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/measures/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/money/__init__.py b/nemo_text_processing/text_normalization/es/data/money/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/money/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/numbers/__init__.py b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/numbers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/ordinals/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/roman/__init__.py b/nemo_text_processing/text_normalization/es/data/roman/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/roman/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/data/time/__init__.py b/nemo_text_processing/text_normalization/es/data/time/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/data/time/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/graph_utils.py b/nemo_text_processing/text_normalization/es/graph_utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/graph_utils.py @@ -0,0 +1,179 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, NEMO_SPACE +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + digits = pynini.project(pynini.string_file(get_abs_path("data/numbers/digit.tsv")), "input") + tens = pynini.project(pynini.string_file(get_abs_path("data/numbers/ties.tsv")), "input") + teens = pynini.project(pynini.string_file(get_abs_path("data/numbers/teen.tsv")), "input") + twenties = pynini.project(pynini.string_file(get_abs_path("data/numbers/twenties.tsv")), "input") + hundreds = pynini.project(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv")), "input") + + accents = pynini.string_map([("รก", "a"), ("รฉ", "e"), ("รญ", "i"), ("รณ", "o"), ("รบ", "u")]) + + if LOCALIZATION == "am": # Setting localization for central and northern america formatting + cardinal_separator = pynini.string_map([",", NEMO_SPACE]) + decimal_separator = pynini.accep(".") + else: + cardinal_separator = pynini.string_map([".", NEMO_SPACE]) + decimal_separator = pynini.accep(",") + + ones = pynini.union("un", "รบn") + fem_ones = pynini.union(pynini.cross("un", "una"), pynini.cross("รบn", "una"), pynini.cross("uno", "una")) + one_to_one_hundred = pynini.union(digits, tens, teens, twenties, tens + pynini.accep(" y ") + digits) + fem_hundreds = hundreds @ pynini.cdrewrite(pynini.cross("ientos", "ientas"), "", "", NEMO_SIGMA) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digits = None + tens = None + teens = None + twenties = None + hundreds = None + + accents = None + + cardinal_separator = None + decimal_separator = None + + ones = None + fem_ones = None + one_to_one_hundred = None + fem_hundreds = None + + PYNINI_AVAILABLE = False + + +def strip_accent(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Converts all accented vowels to non-accented equivalents + + Args: + fst: Any fst. Composes vowel conversion onto fst's output strings + """ + return fst @ pynini.cdrewrite(accents, "", "", NEMO_SIGMA) + + +def shift_cardinal_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Applies gender conversion rules to a cardinal string. These include: rendering all masculine forms of "uno" (including apocopated forms) as "una" and + Converting all gendered numbers in the hundreds series (200,300,400...) to feminine equivalent (e.g. "doscientos" -> "doscientas"). Converssion only applies + to value place for <1000 and multiple of 1000. (e.g. "doscientos mil doscientos" -> "doscientas mil doscientas".) For place values greater than the thousands, there + is no gender shift as the higher powers of ten ("millones", "billones") are masculine nouns and any conversion would be formally + ungrammatical. + e.g. + "doscientos" -> "doscientas" + "doscientos mil" -> "doscientas mil" + "doscientos millones" -> "doscientos millones" + "doscientos mil millones" -> "doscientos mil millones" + "doscientos millones doscientos mil doscientos" -> "doscientos millones doscientas mil doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + before_mil = ( + NEMO_SPACE + + (pynini.accep("mil") | pynini.accep("milรฉsimo")) + + pynini.closure(NEMO_SPACE + hundreds, 0, 1) + + pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + + pynini.union(pynini.accep("[EOS]"), pynini.accep("\""), decimal_separator) + ) + before_double_digits = pynini.closure(NEMO_SPACE + one_to_one_hundred, 0, 1) + pynini.union( + pynini.accep("[EOS]"), pynini.accep("\"") + ) + + fem_allign = pynini.cdrewrite(fem_hundreds, "", before_mil, NEMO_SIGMA) # doscientas mil dosciento + fem_allign @= pynini.cdrewrite(fem_hundreds, "", before_double_digits, NEMO_SIGMA) # doscientas mil doscienta + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union("[EOS]", "\"", decimal_separator), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def shift_number_gender(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Performs gender conversion on all verbalized numbers in output. All values in the hundreds series (200,300,400) are changed to + feminine gender (e.g. "doscientos" -> "doscientas") and all forms of "uno" (including apocopated forms) are converted to "una". + This has no boundary restriction and will perform shift across all values in output string. + e.g. + "doscientos" -> "doscientas" + "doscientos millones" -> "doscientas millones" + "doscientos millones doscientos" -> "doscientas millones doscientas" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + fem_allign = pynini.cdrewrite(fem_hundreds, "", "", NEMO_SIGMA) + fem_allign @= pynini.cdrewrite( + fem_ones, "", pynini.union(NEMO_SPACE, pynini.accep("[EOS]"), pynini.accep("\"")), NEMO_SIGMA + ) # If before a quote or EOS, we know it's the end of a string + + return fst @ fem_allign + + +def strip_cardinal_apocope(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Reverts apocope on cardinal strings in line with formation rules. e.g. "un" -> "uno". Due to cardinal formation rules, this in effect only + affects strings where the final value is a variation of "un". + e.g. + "un" -> "uno" + "veintiรบn" -> "veintiuno" + + Args: + fst: Any fst. Composes conversion onto fst's output strings + """ + # Since cardinals use apocope by default for large values (e.g. "millรณn"), this only needs to act on the last instance of one + strip = pynini.cross("un", "uno") | pynini.cross("รบn", "uno") + strip = pynini.cdrewrite(strip, "", pynini.union("[EOS]", "\""), NEMO_SIGMA) + return fst @ strip + + +def roman_to_int(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Alters given fst to convert Roman integers (lower and upper cased) into Arabic numerals. Valid for values up to 1000. + e.g. + "V" -> "5" + "i" -> "1" + + Args: + fst: Any fst. Composes fst onto Roman conversion outputs. + """ + + def _load_roman(file: str): + roman = load_labels(get_abs_path(file)) + roman_numerals = [(x, y) for x, y in roman] + [(x.upper(), y) for x, y in roman] + return pynini.string_map(roman_numerals) + + digit = _load_roman("data/roman/digit.tsv") + ties = _load_roman("data/roman/ties.tsv") + hundreds = _load_roman("data/roman/hundreds.tsv") + + graph = ( + digit + | ties + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + | ( + hundreds + + (ties | pynutil.add_weight(pynutil.insert("0"), 0.01)) + + (digit | pynutil.add_weight(pynutil.insert("0"), 0.01)) + ) + ).optimize() + + return graph @ fst diff --git a/nemo_text_processing/text_normalization/es/taggers/__init__.py b/nemo_text_processing/text_normalization/es/taggers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/taggers/cardinal.py b/nemo_text_processing/text_normalization/es/taggers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/cardinal.py @@ -0,0 +1,190 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import cardinal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + teen = pynini.invert(pynini.string_file(get_abs_path("data/numbers/teen.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/ties.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/numbers/twenties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/numbers/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + zero = None + digit = None + teen = None + ties = None + twenties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def filter_punctuation(fst: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Helper function for parsing number strings. Converts common cardinal strings (groups of three digits delineated by 'cardinal_separator' - see graph_utils) + and converts to a string of digits: + "1 000" -> "1000" + "1.000.000" -> "1000000" + Args: + fst: Any pynini.FstLike object. Function composes fst onto string parser fst + + Returns: + fst: A pynini.FstLike object + """ + exactly_three_digits = NEMO_DIGIT ** 3 # for blocks of three + up_to_three_digits = pynini.closure(NEMO_DIGIT, 1, 3) # for start of string + + cardinal_string = pynini.closure( + NEMO_DIGIT, 1 + ) # For string w/o punctuation (used for page numbers, thousand series) + + cardinal_string |= ( + up_to_three_digits + + pynutil.delete(cardinal_separator) + + pynini.closure(exactly_three_digits + pynutil.delete(cardinal_separator)) + + exactly_three_digits + ) + + return cardinal_string @ fst + + +class CardinalFst(GraphFst): + """ + Finite state transducer for classifying cardinals, e.g. + "1000" -> cardinal { integer: "mil" } + "2.000.000" -> cardinal { integer: "dos millones" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="classify", deterministic=deterministic) + + # Any single digit + graph_digit = digit + digits_no_one = (NEMO_DIGIT - "1") @ graph_digit + + # Any double digit + graph_tens = teen + graph_tens |= ties + (pynutil.delete('0') | (pynutil.insert(" y ") + graph_digit)) + graph_tens |= twenties + + self.tens = graph_tens.optimize() + + self.two_digit_non_zero = pynini.union( + graph_digit, graph_tens, (pynini.cross("0", NEMO_SPACE) + graph_digit) + ).optimize() + + # Three digit strings + graph_hundreds = hundreds + pynini.union( + pynutil.delete("00"), (insert_space + graph_tens), (pynini.cross("0", NEMO_SPACE) + graph_digit) + ) + graph_hundreds |= pynini.cross("100", "cien") + graph_hundreds |= ( + pynini.cross("1", "ciento") + insert_space + pynini.union(graph_tens, pynutil.delete("0") + graph_digit) + ) + + self.hundreds = graph_hundreds.optimize() + + # For all three digit strings with leading zeroes (graph appends '0's to manage place in string) + graph_hundreds_component = pynini.union(graph_hundreds, pynutil.delete("0") + graph_tens) + + graph_hundreds_component_at_least_one_none_zero_digit = graph_hundreds_component | ( + pynutil.delete("00") + graph_digit + ) + graph_hundreds_component_at_least_one_none_zero_digit_no_one = graph_hundreds_component | ( + pynutil.delete("00") + digits_no_one + ) + + graph_thousands_component_at_least_one_none_zero_digit = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_thousands_component_at_least_one_none_zero_digit_no_one = pynini.union( + pynutil.delete("000") + graph_hundreds_component_at_least_one_none_zero_digit_no_one, + graph_hundreds_component_at_least_one_none_zero_digit_no_one + + pynutil.insert(" mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + pynini.cross("001", "mil") + + ((insert_space + graph_hundreds_component_at_least_one_none_zero_digit) | pynutil.delete("000")), + ) + + graph_million = pynutil.add_weight(pynini.cross("000001", "un millรณn"), -0.001) + graph_million |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" millones") + graph_million |= pynutil.delete("000000") + graph_million += insert_space + + graph_billion = pynutil.add_weight(pynini.cross("000001", "un billรณn"), -0.001) + graph_billion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" billones") + graph_billion |= pynutil.delete("000000") + graph_billion += insert_space + + graph_trillion = pynutil.add_weight(pynini.cross("000001", "un trillรณn"), -0.001) + graph_trillion |= graph_thousands_component_at_least_one_none_zero_digit_no_one + pynutil.insert(" trillones") + graph_trillion |= pynutil.delete("000000") + graph_trillion += insert_space + + graph = ( + graph_trillion + + graph_billion + + graph_million + + (graph_thousands_component_at_least_one_none_zero_digit | pynutil.delete("000000")) + ) + + self.graph = ( + ((NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 0)) + @ pynini.cdrewrite(pynini.closure(pynutil.insert("0")), "[BOS]", "", NEMO_SIGMA) + @ NEMO_DIGIT ** 24 + @ graph + @ pynini.cdrewrite(delete_space, "[BOS]", "", NEMO_SIGMA) + @ pynini.cdrewrite(delete_space, "", "[EOS]", NEMO_SIGMA) + @ pynini.cdrewrite( + pynini.cross(pynini.closure(NEMO_WHITE_SPACE, 2), NEMO_SPACE), NEMO_ALPHA, NEMO_ALPHA, NEMO_SIGMA + ) + ) + self.graph |= zero + + self.graph = filter_punctuation(self.graph).optimize() + + optional_minus_graph = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + final_graph = optional_minus_graph + pynutil.insert("integer: \"") + self.graph + pynutil.insert("\"") + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/date.py b/nemo_text_processing/text_normalization/es/taggers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/date.py @@ -0,0 +1,107 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_DIGIT, NEMO_SPACE, GraphFst, delete_extra_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + articles = pynini.union("de", "del", "el", "del aรฑo") + delete_leading_zero = (pynutil.delete("0") | (NEMO_DIGIT - "0")) + NEMO_DIGIT + month_numbers = pynini.string_file(get_abs_path("data/dates/months.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + articles = None + delete_leading_zero = None + month_numbers = None + + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for classifying date, e.g. + "01.04.2010" -> date { day: "un" month: "enero" year: "dos mil diez" preserve_order: true } + "marzo 4 2000" -> date { month: "marzo" day: "cuatro" year: "dos mil" } + "1990-20-01" -> date { year: "mil novecientos noventa" day: "veinte" month: "enero" } + + Args: + cardinal: cardinal GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool): + super().__init__(name="date", kind="classify", deterministic=deterministic) + + number_to_month = month_numbers.optimize() + month_graph = pynini.project(number_to_month, "output") + + numbers = cardinal.graph + optional_leading_zero = delete_leading_zero | NEMO_DIGIT + + # 01, 31, 1 + digit_day = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 32)]) @ numbers + day = (pynutil.insert("day: \"") + digit_day + pynutil.insert("\"")).optimize() + + digit_month = optional_leading_zero @ pynini.union(*[str(x) for x in range(1, 13)]) + number_to_month = digit_month @ number_to_month + + month_name = (pynutil.insert("month: \"") + month_graph + pynutil.insert("\"")).optimize() + month_number = (pynutil.insert("month: \"") + number_to_month + pynutil.insert("\"")).optimize() + + # prefer cardinal over year + year = (NEMO_DIGIT - "0") + pynini.closure(NEMO_DIGIT, 1, 3) # 90, 990, 1990 + year @= numbers + self.year = year + + year_only = pynutil.insert("year: \"") + year + pynutil.insert("\"") + year_with_articles = ( + pynutil.insert("year: \"") + pynini.closure(articles + NEMO_SPACE, 0, 1) + year + pynutil.insert("\"") + ) + + graph_dmy = ( + day + + pynini.closure(pynutil.delete(" de")) + + NEMO_SPACE + + month_name + + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + graph_mdy = ( # English influences on language + month_name + delete_extra_space + day + pynini.closure(NEMO_SPACE + year_with_articles, 0, 1) + ) + + separators = [".", "-", "/"] + for sep in separators: + year_optional = pynini.closure(pynini.cross(sep, NEMO_SPACE) + year_only, 0, 1) + new_graph = day + pynini.cross(sep, NEMO_SPACE) + month_number + year_optional + graph_dmy |= new_graph + if not deterministic: + new_graph = month_number + pynini.cross(sep, NEMO_SPACE) + day + year_optional + graph_mdy |= new_graph + + dash = "-" + day_optional = pynini.closure(pynini.cross(dash, NEMO_SPACE) + day, 0, 1) + graph_ymd = NEMO_DIGIT ** 4 @ year_only + pynini.cross(dash, NEMO_SPACE) + month_number + day_optional + + final_graph = graph_dmy + pynutil.insert(" preserve_order: true") + final_graph |= graph_ymd + final_graph |= graph_mdy + + self.final_graph = final_graph.optimize() + self.fst = self.add_tokens(self.final_graph).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/decimals.py b/nemo_text_processing/text_normalization/es/taggers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/decimals.py @@ -0,0 +1,138 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + cardinal_separator, + decimal_separator, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + quantities = pynini.string_file(get_abs_path("data/numbers/quantities.tsv")) + digit = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + quantities = None + digit = None + zero = None + + PYNINI_AVAILABLE = False + + +def get_quantity(decimal_graph: 'pynini.FstLike', cardinal_graph: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Returns FST that transforms either a cardinal or decimal followed by a quantity into a numeral, + e.g. 2 millones -> integer_part: "dos" quantity: "millones" + e.g. 2,4 millones -> integer_part: "dos" fractional_part: "quatro" quantity: "millones" + e.g. 2,400 millones -> integer_part: "dos mil cuatrocientos" fractional_part: "quatro" quantity: "millones" + + Args: + decimal_graph: DecimalFST + cardinal_graph: CardinalFST + """ + numbers = pynini.closure(NEMO_DIGIT, 1, 6) @ cardinal_graph + numbers = pynini.cdrewrite(pynutil.delete(cardinal_separator), "", "", NEMO_SIGMA) @ numbers + + res = ( + pynutil.insert("integer_part: \"") + + numbers # The cardinal we're passing only produces 'un' for one, so gender agreement is safe (all quantities are masculine). Limit to 10^6 power. + + pynutil.insert("\"") + + NEMO_SPACE + + pynutil.insert("quantity: \"") + + quantities + + pynutil.insert("\"") + ) + res |= decimal_graph + NEMO_SPACE + pynutil.insert("quantity: \"") + quantities + pynutil.insert("\"") + return res + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + -11,4006 billones -> decimal { negative: "true" integer_part: "once" fractional_part: "cuatro cero cero seis" quantity: "billones" preserve_order: true } + 1 billรณn -> decimal { integer_part: "un" quantity: "billรณn" preserve_order: true } + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + graph_digit = digit | zero + + if not deterministic: + graph = pynini.union(graph_digit, cardinal.hundreds, cardinal.tens) + graph += pynini.closure(insert_space + graph) + + else: + # General pattern seems to be 1-3 digits: map as cardinal, default to digits otherwise \ + graph = pynini.union( + graph_digit, + cardinal.tens, + cardinal.hundreds, + graph_digit + pynini.closure(insert_space + graph_digit, 3), + zero + + pynini.closure(insert_space + zero) + + pynini.closure(insert_space + graph_digit), # For cases such as "1,010" + ) + + # Need to strip apocope everywhere BUT end of string + reverse_apocope = pynini.string_map([("un", "uno"), ("รบn", "uno")]) + apply_reverse_apocope = pynini.cdrewrite(reverse_apocope, "", NEMO_SPACE, NEMO_SIGMA) + graph @= apply_reverse_apocope + + # Technically decimals should be space delineated groups of three, e.g. (1,333 333). This removes any possible spaces + strip_formatting = pynini.cdrewrite(delete_space, "", "", NEMO_SIGMA) + graph = strip_formatting @ graph + + self.graph = graph.optimize() + + graph_separator = pynutil.delete(decimal_separator) + optional_graph_negative = pynini.closure(pynutil.insert("negative: ") + pynini.cross("-", "\"true\" "), 0, 1) + + self.graph_fractional = pynutil.insert("fractional_part: \"") + self.graph + pynutil.insert("\"") + + # Integer graph maintains apocope except for ones place + graph_integer = ( + strip_cardinal_apocope(cardinal.graph) + if deterministic + else pynini.union(cardinal.graph, strip_cardinal_apocope(cardinal.graph)) + ) # Gives us forms w/ and w/o apocope + self.graph_integer = pynutil.insert("integer_part: \"") + graph_integer + pynutil.insert("\"") + final_graph_wo_sign = self.graph_integer + graph_separator + insert_space + self.graph_fractional + + self.final_graph_wo_negative = ( + final_graph_wo_sign | get_quantity(final_graph_wo_sign, cardinal.graph).optimize() + ) + final_graph = optional_graph_negative + self.final_graph_wo_negative + + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/electronic.py b/nemo_text_processing/text_normalization/es/taggers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/electronic.py @@ -0,0 +1,84 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_ALPHA, NEMO_DIGIT, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + common_domains = [x[0] for x in load_labels(get_abs_path("data/electronic/domain.tsv"))] + symbols = [x[0] for x in load_labels(get_abs_path("data/electronic/symbols.tsv"))] + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + common_domains = None + symbols = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for classifying electronic: email addresses + e.g. "abc@hotmail.com" -> electronic { username: "abc" domain: "hotmail.com" preserve_order: true } + e.g. "www.abc.com/123" -> electronic { protocol: "www." domain: "abc.com/123" preserve_order: true } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="classify", deterministic=deterministic) + + dot = pynini.accep(".") + accepted_common_domains = pynini.union(*common_domains) + accepted_symbols = pynini.union(*symbols) - dot + accepted_characters = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols) + acceepted_characters_with_dot = pynini.closure(NEMO_ALPHA | NEMO_DIGIT | accepted_symbols | dot) + + # email + username = ( + pynutil.insert("username: \"") + + acceepted_characters_with_dot + + pynutil.insert("\"") + + pynini.cross('@', ' ') + ) + domain_graph = accepted_characters + dot + accepted_characters + domain_graph = pynutil.insert("domain: \"") + domain_graph + pynutil.insert("\"") + domain_common_graph = ( + pynutil.insert("domain: \"") + + accepted_characters + + accepted_common_domains + + pynini.closure((accepted_symbols | dot) + pynini.closure(accepted_characters, 1), 0, 1) + + pynutil.insert("\"") + ) + graph = (username + domain_graph) | domain_common_graph + + # url + protocol_start = pynini.accep("https://") | pynini.accep("http://") + protocol_end = ( + pynini.accep("www.") + if deterministic + else pynini.accep("www.") | pynini.cross("www.", "doble ve doble ve doble ve.") + ) + protocol = protocol_start | protocol_end | (protocol_start + protocol_end) + protocol = pynutil.insert("protocol: \"") + protocol + pynutil.insert("\"") + graph |= protocol + insert_space + (domain_graph | domain_common_graph) + self.graph = graph + + final_graph = self.add_tokens(self.graph + pynutil.insert(" preserve_order: true")) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/fraction.py b/nemo_text_processing/text_normalization/es/taggers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/fraction.py @@ -0,0 +1,124 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + ordinal_exceptions = pynini.string_file(get_abs_path("data/fractions/ordinal_exceptions.tsv")) + higher_powers_of_ten = pynini.string_file(get_abs_path("data/fractions/powers_of_ten.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + ordinal_exceptions = None + higher_powers_of_ten = None + + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for classifying fraction + "23 4/5" -> + tokens { fraction { integer: "veintitrรฉs" numerator: "cuatro" denominator: "quinto" mophosyntactic_features: "ordinal" } } + + Args: + cardinal: CardinalFst + ordinal: OrdinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, ordinal: GraphFst, deterministic: bool = True): + super().__init__(name="fraction", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + ordinal_graph = ordinal.graph + + # 2-10 are all ordinals + three_to_ten = pynini.string_map(["2", "3", "4", "5", "6", "7", "8", "9", "10",]) + block_three_to_ten = pynutil.delete(three_to_ten) # To block cardinal productions + if not deterministic: # Multiples of tens are sometimes rendered as ordinals + three_to_ten |= pynini.string_map(["20", "30", "40", "50", "60", "70", "80", "90",]) + graph_three_to_ten = three_to_ten @ ordinal_graph + graph_three_to_ten @= pynini.cdrewrite(ordinal_exceptions, "", "", NEMO_SIGMA) + + # Higher powers of tens (and multiples) are converted to ordinals. + hundreds = pynini.string_map(["100", "200", "300", "400", "500", "600", "700", "800", "900",]) + graph_hundreds = hundreds @ ordinal_graph + + multiples_of_thousand = ordinal.multiples_of_thousand # So we can have X milรฉsimos + + graph_higher_powers_of_ten = ( + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + + pynini.closure("mil ", 0, 1) + + pynini.closure(ordinal.one_to_one_thousand + NEMO_SPACE, 0, 1) + ) # x millones / x mil millones / x mil z millones + graph_higher_powers_of_ten += higher_powers_of_ten + graph_higher_powers_of_ten = cardinal_graph @ graph_higher_powers_of_ten + graph_higher_powers_of_ten @= pynini.cdrewrite( + pynutil.delete("un "), pynini.accep("[BOS]"), pynini.project(higher_powers_of_ten, "output"), NEMO_SIGMA + ) # we drop 'un' from these ordinals (millionths, not one-millionths) + + graph_higher_powers_of_ten = multiples_of_thousand | graph_hundreds | graph_higher_powers_of_ten + block_higher_powers_of_ten = pynutil.delete( + pynini.project(graph_higher_powers_of_ten, "input") + ) # For cardinal graph + + graph_fractions_ordinals = graph_higher_powers_of_ten | graph_three_to_ten + graph_fractions_ordinals += pynutil.insert( + "\" morphosyntactic_features: \"ordinal\"" + ) # We note the root for processing later + + # Blocking the digits and hundreds from Cardinal graph + graph_fractions_cardinals = pynini.cdrewrite( + block_three_to_ten | block_higher_powers_of_ten, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fractions_cardinals @= NEMO_CHAR.plus @ pynini.cdrewrite( + pynutil.delete("0"), pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA + ) # Empty characters become '0' for NEMO_CHAR fst, so ned to block + graph_fractions_cardinals @= cardinal_graph + graph_fractions_cardinals += pynutil.insert( + "\" morphosyntactic_features: \"add_root\"" + ) # blocking these entries to reduce erroneous possibilities in debugging + + if deterministic: + graph_fractions_cardinals = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ graph_fractions_cardinals + ) # Past hundreds the conventional scheme can be hard to read. For determinism we stop here + + graph_denominator = pynini.union( + graph_fractions_ordinals, + graph_fractions_cardinals, + pynutil.add_weight(cardinal_graph + pynutil.insert("\""), 0.001), + ) # Last form is simply recording the cardinal. Weighting so last resort + + integer = pynutil.insert("integer_part: \"") + cardinal_graph + pynutil.insert("\"") + NEMO_SPACE + numerator = ( + pynutil.insert("numerator: \"") + cardinal_graph + (pynini.cross("/", "\" ") | pynini.cross(" / ", "\" ")) + ) + denominator = pynutil.insert("denominator: \"") + graph_denominator + + self.graph = pynini.closure(integer, 0, 1) + numerator + denominator + + final_graph = self.add_tokens(self.graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/measure.py b/nemo_text_processing/text_normalization/es/taggers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/measure.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_NON_BREAKING_SPACE, + NEMO_SPACE, + GraphFst, + convert_space, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit = pynini.string_file(get_abs_path("data/measures/measurements.tsv")) + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit = None + unit_plural_fem = None + unit_plural_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for classifying measure, e.g. + "2,4 g" -> measure { cardinal { integer_part: "dos" fractional_part: "cuatro" units: "gramos" preserve_order: true } } + "1 g" -> measure { cardinal { integer: "un" units: "gramo" preserve_order: true } } + "1 millรณn g" -> measure { cardinal { integer: "un quantity: "millรณn" units: "gramos" preserve_order: true } } + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + This class also converts words containing numbers and letters + e.g. "a-8" โ€”> "a ocho" + e.g. "1,2-a" โ€”> "uno coma dos a" + + + Args: + cardinal: CardinalFst + decimal: DecimalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, fraction: GraphFst, deterministic: bool = True): + super().__init__(name="measure", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + + unit_singular = unit + unit_plural = unit_singular @ (unit_plural_fem | unit_plural_masc) + + graph_unit_singular = convert_space(unit_singular) + graph_unit_plural = convert_space(unit_plural) + + optional_graph_negative = pynini.closure("-", 0, 1) + + graph_unit_denominator = ( + pynini.cross("/", "por") + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_singular + ) + + optional_unit_denominator = pynini.closure( + pynutil.insert(NEMO_NON_BREAKING_SPACE) + graph_unit_denominator, 0, 1, + ) + + unit_plural = ( + pynutil.insert("units: \"") + + ((graph_unit_plural + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + unit_singular_graph = ( + pynutil.insert("units: \"") + + ((graph_unit_singular + optional_unit_denominator) | graph_unit_denominator) + + pynutil.insert("\"") + ) + + subgraph_decimal = decimal.fst + insert_space + pynini.closure(NEMO_SPACE, 0, 1) + unit_plural + + subgraph_cardinal = ( + (optional_graph_negative + (pynini.closure(NEMO_DIGIT) - "1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_plural + ) + + subgraph_cardinal |= ( + (optional_graph_negative + pynini.accep("1")) @ cardinal.fst + + insert_space + + pynini.closure(delete_space, 0, 1) + + unit_singular_graph + ) + + subgraph_fraction = fraction.fst + insert_space + pynini.closure(delete_space, 0, 1) + unit_plural + + decimal_times = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_times = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.insert("\" } units: \"") + + pynini.union('x', 'X') + + pynutil.insert("\"") + ) + + cardinal_dash_alpha = ( + pynutil.insert("cardinal { integer: \"") + + strip_cardinal_apocope(cardinal_graph) + + pynutil.delete('-') + + pynutil.insert("\" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + decimal_dash_alpha = ( + pynutil.insert("decimal { ") + + decimal.final_graph_wo_negative + + pynutil.delete('-') + + pynutil.insert(" } units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.insert("\"") + ) + + alpha_dash_cardinal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" cardinal { integer: \"") + + cardinal_graph + + pynutil.insert("\" } preserve_order: true") + ) + + alpha_dash_decimal = ( + pynutil.insert("units: \"") + + pynini.closure(NEMO_ALPHA, 1) + + pynutil.delete('-') + + pynutil.insert("\"") + + pynutil.insert(" decimal { ") + + decimal.final_graph_wo_negative + + pynutil.insert(" } preserve_order: true") + ) + + final_graph = ( + subgraph_decimal + | subgraph_cardinal + | subgraph_fraction + | cardinal_dash_alpha + | alpha_dash_cardinal + | decimal_dash_alpha + | decimal_times + | cardinal_times + | alpha_dash_decimal + ) + final_graph += pynutil.insert(" preserve_order: true") + final_graph = self.add_tokens(final_graph) + + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/money.py b/nemo_text_processing/text_normalization/es/taggers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/money.py @@ -0,0 +1,194 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_ALPHA, + NEMO_DIGIT, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import decimal_separator +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + maj_singular_labels = load_labels(get_abs_path("data/money/currency_major.tsv")) + maj_singular = pynini.string_file((get_abs_path("data/money/currency_major.tsv"))) + min_singular = pynini.string_file(get_abs_path("data/money/currency_minor.tsv")) + fem_plural = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc_plural = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + maj_singular_labels = None + min_singular = None + maj_singular = None + fem_plural = None + masc_plural = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for classifying money, e.g. + "โ‚ฌ1" -> money { currency_maj: "euro" integer_part: "un"} + "โ‚ฌ1,000" -> money { currency_maj: "euro" integer_part: "un" } + "โ‚ฌ1,001" -> money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un" } + "ยฃ1,4" -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true } + -> money { integer_part: "una" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "penique" preserve_order: true } + "0,01 ยฃ" -> money { fractional_part: "un" currency_min: "penique" preserve_order: true } + "0,02 ยฃ" -> money { fractional_part: "dos" currency_min: "peniques" preserve_order: true } + "ยฃ0,01 million" -> money { currency_maj: "libra" integer_part: "cero" fractional_part: "cero un" quantity: "million" } + + Args: + cardinal: CardinalFst + decimal: DecimalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="classify", deterministic=deterministic) + cardinal_graph = cardinal.graph + graph_decimal_final = decimal.final_graph_wo_negative + + maj_singular_graph = maj_singular + min_singular_graph = min_singular + maj_plural_graph = maj_singular @ (fem_plural | masc_plural) + min_plural_graph = min_singular @ (fem_plural | masc_plural) + + graph_maj_singular = pynutil.insert("currency_maj: \"") + maj_singular_graph + pynutil.insert("\"") + graph_maj_plural = pynutil.insert("currency_maj: \"") + maj_plural_graph + pynutil.insert("\"") + + graph_integer_one = pynutil.insert("integer_part: \"") + pynini.cross("1", "un") + pynutil.insert("\"") + + decimal_with_quantity = (NEMO_SIGMA + NEMO_ALPHA) @ graph_decimal_final + + graph_decimal_plural = pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, # 1,05 $ + ) + graph_decimal_plural = ( + (NEMO_SIGMA - "1") + decimal_separator + NEMO_SIGMA + ) @ graph_decimal_plural # Can't have "un euros" + + graph_decimal_singular = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_decimal_final, # $1,05 + graph_decimal_final + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, # 1,05 $ + ) + graph_decimal_singular = (pynini.accep("1") + decimal_separator + NEMO_SIGMA) @ graph_decimal_singular + + graph_decimal = pynini.union( + graph_decimal_singular, + graph_decimal_plural, + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + decimal_with_quantity, + ) + + graph_integer = ( + pynutil.insert("integer_part: \"") + ((NEMO_SIGMA - "1") @ cardinal_graph) + pynutil.insert("\"") + ) + + graph_integer_only = pynini.union( + graph_maj_singular + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer_one, + graph_integer_one + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_singular, + ) + graph_integer_only |= pynini.union( + graph_maj_plural + pynini.closure(delete_space, 0, 1) + insert_space + graph_integer, + graph_integer + pynini.closure(delete_space, 0, 1) + insert_space + graph_maj_plural, + ) + + graph = graph_integer_only | graph_decimal + + # remove trailing zeros of non zero number in the first 2 digits and fill up to 2 digits + # e.g. 2000 -> 20, 0200->02, 01 -> 01, 10 -> 10 + # not accepted: 002, 00, 0, + two_digits_fractional_part = ( + pynini.closure(NEMO_DIGIT) + (NEMO_DIGIT - "0") + pynini.closure(pynutil.delete("0")) + ) @ ( + (pynutil.delete("0") + (NEMO_DIGIT - "0")) + | ((NEMO_DIGIT - "0") + pynutil.insert("0")) + | ((NEMO_DIGIT - "0") + NEMO_DIGIT) + ) + + graph_min_singular = pynutil.insert("currency_min: \"") + min_singular_graph + pynutil.insert("\"") + graph_min_plural = pynutil.insert("currency_min: \"") + min_plural_graph + pynutil.insert("\"") + + # format ** euro ** cent + decimal_graph_with_minor = None + for curr_symbol, _ in maj_singular_labels: + preserve_order = pynutil.insert(" preserve_order: true") + + integer_plus_maj = pynini.union( + graph_integer + insert_space + pynutil.insert(curr_symbol) @ graph_maj_plural, + graph_integer_one + insert_space + pynutil.insert(curr_symbol) @ graph_maj_singular, + ) + # non zero integer part + integer_plus_maj = (pynini.closure(NEMO_DIGIT) - "0") @ integer_plus_maj + + graph_fractional_one = ( + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ pynini.cross("1", "un") + + pynutil.insert("\"") + ) + + graph_fractional = ( + two_digits_fractional_part @ (pynini.closure(NEMO_DIGIT, 1, 2) - "1") @ cardinal.two_digit_non_zero + ) + graph_fractional = pynutil.insert("fractional_part: \"") + graph_fractional + pynutil.insert("\"") + + fractional_plus_min = pynini.union( + graph_fractional + insert_space + pynutil.insert(curr_symbol) @ graph_min_plural, + graph_fractional_one + insert_space + pynutil.insert(curr_symbol) @ graph_min_singular, + ) + + decimal_graph_with_minor_curr = ( + integer_plus_maj + pynini.cross(decimal_separator, NEMO_SPACE) + fractional_plus_min + ) + decimal_graph_with_minor_curr |= pynutil.add_weight( + integer_plus_maj + + pynini.cross(decimal_separator, NEMO_SPACE) + + pynutil.insert("fractional_part: \"") + + two_digits_fractional_part @ cardinal.two_digit_non_zero + + pynutil.insert("\""), + weight=0.0001, + ) + + decimal_graph_with_minor_curr |= pynutil.delete("0,") + fractional_plus_min + decimal_graph_with_minor_curr = pynini.union( + pynutil.delete(curr_symbol) + + pynini.closure(delete_space, 0, 1) + + decimal_graph_with_minor_curr + + preserve_order, + decimal_graph_with_minor_curr + + preserve_order + + pynini.closure(delete_space, 0, 1) + + pynutil.delete(curr_symbol), + ) + + decimal_graph_with_minor = ( + decimal_graph_with_minor_curr + if decimal_graph_with_minor is None + else pynini.union(decimal_graph_with_minor, decimal_graph_with_minor_curr) + ) + + final_graph = graph | pynutil.add_weight(decimal_graph_with_minor, -0.001) + + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/ordinal.py b/nemo_text_processing/text_normalization/es/taggers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/ordinal.py @@ -0,0 +1,186 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import roman_to_int, strip_accent +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/digit.tsv"))) + teens = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/teen.tsv"))) + twenties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/twenties.tsv"))) + ties = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/ties.tsv"))) + hundreds = pynini.invert(pynini.string_file(get_abs_path("data/ordinals/hundreds.tsv"))) + + PYNINI_AVAILABLE = True + +except (ImportError, ModuleNotFoundError): + digit = None + teens = None + twenties = None + ties = None + hundreds = None + + PYNINI_AVAILABLE = False + + +def get_one_to_one_thousand(cardinal: 'pynini.FstLike') -> 'pynini.FstLike': + """ + Produces an acceptor for verbalizations of all numbers from 1 to 1000. Needed for ordinals and fractions. + + Args: + cardinal: CardinalFst + + Returns: + fst: A pynini.FstLike object + """ + numbers = pynini.string_map([str(_) for _ in range(1, 1000)]) @ cardinal + return pynini.project(numbers, "output").optimize() + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for classifying ordinal + "21.ยบ" -> ordinal { integer: "vigรฉsimo primero" morphosyntactic_features: "gender_masc" } + This class converts ordinal up to the millionth (millonรฉsimo) order (exclusive). + + This FST also records the ending of the ordinal (called "morphosyntactic_features"): + either as gender_masc, gender_fem, or apocope. Also introduces plural feature for non-deterministic graphs. + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="ordinal", kind="classify") + cardinal_graph = cardinal.graph + + graph_digit = digit.optimize() + graph_teens = teens.optimize() + graph_ties = ties.optimize() + graph_twenties = twenties.optimize() + graph_hundreds = hundreds.optimize() + + if not deterministic: + # Some alternative derivations + graph_ties = graph_ties | pynini.cross("sesenta", "setuagรฉsimo") + + graph_teens = graph_teens | pynini.cross("once", "decimoprimero") + graph_teens |= pynini.cross("doce", "decimosegundo") + + graph_digit = graph_digit | pynini.cross("nueve", "nono") + graph_digit |= pynini.cross("siete", "sรฉtimo") + + graph_tens_component = ( + graph_teens + | (graph_ties + pynini.closure(pynini.cross(" y ", NEMO_SPACE) + graph_digit, 0, 1)) + | graph_twenties + ) + + graph_hundred_component = pynini.union( + graph_hundreds + pynini.closure(NEMO_SPACE + pynini.union(graph_tens_component, graph_digit), 0, 1), + graph_tens_component, + graph_digit, + ) + + # Need to go up to thousands for fractions + self.one_to_one_thousand = get_one_to_one_thousand(cardinal_graph) + + thousands = pynini.cross("mil", "milรฉsimo") + + graph_thousands = ( + strip_accent(self.one_to_one_thousand) + NEMO_SPACE + thousands + ) # Cardinals become prefix for thousands series. Snce accent on the powers of ten we strip accent from leading words + graph_thousands @= pynini.cdrewrite(delete_space, "", "milรฉsimo", NEMO_SIGMA) # merge as a prefix + graph_thousands |= thousands + + self.multiples_of_thousand = (cardinal_graph @ graph_thousands).optimize() + + if ( + not deterministic + ): # Formally the words preceding the power of ten should be a prefix, but some maintain word boundaries. + graph_thousands |= (self.one_to_one_thousand @ graph_hundred_component) + NEMO_SPACE + thousands + + graph_thousands += pynini.closure(NEMO_SPACE + graph_hundred_component, 0, 1) + + ordinal_graph = graph_thousands | graph_hundred_component + ordinal_graph = cardinal_graph @ ordinal_graph + + if not deterministic: + # The 10's and 20's series can also be two words + split_words = pynini.cross("decimo", "dรฉcimo ") | pynini.cross("vigesimo", "vigรฉsimo ") + split_words = pynini.cdrewrite(split_words, "", NEMO_CHAR, NEMO_SIGMA) + ordinal_graph |= ordinal_graph @ split_words + + # If "octavo" is preceeded by the "o" within string, it needs deletion + ordinal_graph @= pynini.cdrewrite(pynutil.delete("o"), "", "octavo", NEMO_SIGMA) + + self.graph = ordinal_graph.optimize() + + masc = pynini.accep("gender_masc") + fem = pynini.accep("gender_fem") + apocope = pynini.accep("apocope") + + delete_period = pynini.closure(pynutil.delete("."), 0, 1) # Sometimes the period is omitted f + + accept_masc = delete_period + pynini.cross("ยบ", masc) + accep_fem = delete_period + pynini.cross("ยช", fem) + accep_apocope = delete_period + pynini.cross("แต‰สณ", apocope) + + # Managing Romanization + graph_roman = pynutil.insert("integer: \"") + roman_to_int(ordinal_graph) + pynutil.insert("\"") + if not deterministic: + # Introduce plural + plural = pynini.closure(pynutil.insert("/plural"), 0, 1) + accept_masc += plural + accep_fem += plural + + # Romanizations have no morphology marker, so in non-deterministic case we provide option for all + insert_morphology = pynutil.insert(pynini.union(masc, fem)) + plural + insert_morphology |= pynutil.insert(apocope) + insert_morphology = ( + pynutil.insert(" morphosyntactic_features: \"") + insert_morphology + pynutil.insert("\"") + ) + + graph_roman += insert_morphology + + else: + # We assume masculine gender as default + graph_roman += pynutil.insert(" morphosyntactic_features: \"gender_masc\"") + + # Rest of graph + convert_abbreviation = accept_masc | accep_fem | accep_apocope + + graph = ( + pynutil.insert("integer: \"") + + ordinal_graph + + pynutil.insert("\"") + + pynutil.insert(" morphosyntactic_features: \"") + + convert_abbreviation + + pynutil.insert("\"") + ) + graph = pynini.union(graph, graph_roman) + + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/telephone.py b/nemo_text_processing/text_normalization/es/taggers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/telephone.py @@ -0,0 +1,156 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_SIGMA, GraphFst, insert_space +from nemo_text_processing.text_normalization.es.graph_utils import ones +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + graph_digit = pynini.string_file(get_abs_path("data/numbers/digit.tsv")) + graph_ties = pynini.string_file(get_abs_path("data/numbers/ties.tsv")) + graph_teen = pynini.string_file(get_abs_path("data/numbers/teen.tsv")) + graph_twenties = pynini.string_file(get_abs_path("data/numbers/twenties.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + graph_digit = None + graph_ties = None + graph_teen = None + graph_twenties = None + + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for classifying telephone numbers, e.g. + 123-123-5678 -> { number_part: "uno dos tres uno dos tres cinco seis siete ocho" }. + In Spanish, digits are generally read individually, or as 2-digit numbers, + eg. "123" = "uno dos tres", + "1234" = "doce treinta y cuatro". + This will verbalize sequences of 10 (3+3+4 e.g. 123-456-7890). + 9 (3+3+3 e.g. 123-456-789) and 8 (4+4 e.g. 1234-5678) digits. + + (we ignore more complicated cases such as "doscientos y dos" or "tres nueves"). + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="classify") + + # create `single_digits` and `double_digits` graphs as these will be + # the building blocks of possible telephone numbers + single_digits = pynini.invert(graph_digit).optimize() | pynini.cross("0", "cero") + + double_digits = pynini.union( + graph_twenties, + graph_teen, + (graph_ties + pynutil.delete("0")), + (graph_ties + insert_space + pynutil.insert("y") + insert_space + graph_digit), + ) + double_digits = pynini.invert(double_digits) + + # define `ten_digit_graph`, `nine_digit_graph`, `eight_digit_graph` + # which produces telephone numbers spoken (1) only with single digits, + # or (2) spoken with double digits (and sometimes single digits) + + # 10-digit option (1): all single digits + ten_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + # 9-digit option (1): all single digits + nine_digit_graph = ( + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 2, 2) + + single_digits + ) + + # 8-digit option (1): all single digits + eight_digit_graph = ( + pynini.closure(single_digits + insert_space, 4, 4) + + pynutil.delete("-") + + pynini.closure(single_digits + insert_space, 3, 3) + + single_digits + ) + + if not deterministic: + # 10-digit option (2): (1+2) + (1+2) + (2+2) digits + ten_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + # 9-digit option (2): (1+2) + (1+2) + (1+2) digits + nine_digit_graph |= ( + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + single_digits + + insert_space + + double_digits + ) + + # 8-digit option (2): (2+2) + (2+2) digits + eight_digit_graph |= ( + double_digits + + insert_space + + double_digits + + insert_space + + pynutil.delete("-") + + double_digits + + insert_space + + double_digits + ) + + number_part = pynini.union(ten_digit_graph, nine_digit_graph, eight_digit_graph) + number_part @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", "", NEMO_SIGMA) + + number_part = pynutil.insert("number_part: \"") + number_part + pynutil.insert("\"") + + graph = number_part + final_graph = self.add_tokens(graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/time.py b/nemo_text_processing/text_normalization/es/taggers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/time.py @@ -0,0 +1,218 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_DIGIT, + NEMO_SIGMA, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + time_zone_graph = pynini.string_file(get_abs_path("data/time/time_zone.tsv")) + suffix = pynini.string_file(get_abs_path("data/time/time_suffix.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + time_zone_graph = None + suffix = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for classifying time, e.g. + "02:15 est" -> time { hours: "dos" minutes: "quince" zone: "e s t"} + "2 h" -> time { hours: "dos" } + "9 h" -> time { hours: "nueve" } + "02:15:10 h" -> time { hours: "dos" minutes: "quince" seconds: "diez"} + + Args: + cardinal: CardinalFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, cardinal: GraphFst, deterministic: bool = True): + super().__init__(name="time", kind="classify", deterministic=deterministic) + + delete_time_delimiter = pynutil.delete(pynini.union(".", ":")) + + one = pynini.string_map([("un", "una"), ("รบn", "una")]) + change_one = pynini.cdrewrite(one, "", "", NEMO_SIGMA) + cardinal_graph = cardinal.graph @ change_one + + day_suffix = pynutil.insert("suffix: \"") + suffix + pynutil.insert("\"") + day_suffix = delete_space + insert_space + day_suffix + + delete_hora_suffix = delete_space + insert_space + pynutil.delete("h") + delete_minute_suffix = delete_space + insert_space + pynutil.delete("min") + delete_second_suffix = delete_space + insert_space + pynutil.delete("s") + + labels_hour_24 = [ + str(x) for x in range(0, 25) + ] # Can see both systems. Twelve hour requires am/pm for ambiguity resolution + labels_hour_12 = [str(x) for x in range(1, 13)] + labels_minute_single = [str(x) for x in range(1, 10)] + labels_minute_double = [str(x) for x in range(10, 60)] + + delete_leading_zero_to_double_digit = ( + pynini.closure(pynutil.delete("0") | (NEMO_DIGIT - "0"), 0, 1) + NEMO_DIGIT + ) + + graph_24 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_24) + ) + graph_12 = ( + pynini.closure(NEMO_DIGIT, 1, 2) @ delete_leading_zero_to_double_digit @ pynini.union(*labels_hour_12) + ) + + graph_hour_24 = graph_24 @ cardinal_graph + graph_hour_12 = graph_12 @ cardinal_graph + + graph_minute_single = delete_leading_zero_to_double_digit @ pynini.union(*labels_minute_single) + graph_minute_double = pynini.union(*labels_minute_double) + + graph_minute = pynini.union(graph_minute_single, graph_minute_double) @ cardinal_graph + + final_graph_hour_only_24 = ( + pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + delete_hora_suffix + ) + final_graph_hour_only_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + day_suffix + + final_graph_hour_24 = pynutil.insert("hours: \"") + graph_hour_24 + pynutil.insert("\"") + final_graph_hour_12 = pynutil.insert("hours: \"") + graph_hour_12 + pynutil.insert("\"") + + final_graph_minute = pynutil.insert("minutes: \"") + graph_minute + pynutil.insert("\"") + final_graph_second = pynutil.insert("seconds: \"") + graph_minute + pynutil.insert("\"") + final_time_zone_optional = pynini.closure( + delete_space + insert_space + pynutil.insert("zone: \"") + time_zone_graph + pynutil.insert("\""), 0, 1, + ) + + # 02.30 h + graph_hm = ( + final_graph_hour_24 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 h + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_24 + + delete_hora_suffix + + delete_space + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + delete_minute_suffix + + pynini.closure( + delete_space + + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)) + + delete_second_suffix, + 0, + 1, + ) # For seconds + + final_time_zone_optional + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_12 + + delete_time_delimiter + + (pynutil.delete("00") | (insert_space + final_graph_minute)) + + pynini.closure( + delete_time_delimiter + (pynini.cross("00", " seconds: \"0\"") | (insert_space + final_graph_second)), + 0, + 1, + ) # For seconds 2.30.35 a. m. + + day_suffix + + final_time_zone_optional + ) + + graph_h = ( + pynini.union(final_graph_hour_only_24, final_graph_hour_only_12) + final_time_zone_optional + ) # Should always have a time indicator, else we'll pass to cardinals + + if not deterministic: + # This includes alternate vocalization (hour menos min, min para hour), here we shift the times and indicate a `style` tag + hour_shift_24 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_24.tsv"))) + hour_shift_12 = pynini.invert(pynini.string_file(get_abs_path("data/time/hour_to_12.tsv"))) + minute_shift = pynini.string_file(get_abs_path("data/time/minute_to.tsv")) + + graph_hour_to_24 = graph_24 @ hour_shift_24 @ cardinal_graph + graph_hour_to_12 = graph_12 @ hour_shift_12 @ cardinal_graph + + graph_minute_to = pynini.union(graph_minute_single, graph_minute_double) @ minute_shift @ cardinal_graph + + final_graph_hour_to_24 = pynutil.insert("hours: \"") + graph_hour_to_24 + pynutil.insert("\"") + final_graph_hour_to_12 = pynutil.insert("hours: \"") + graph_hour_to_12 + pynutil.insert("\"") + + final_graph_minute_to = pynutil.insert("minutes: \"") + graph_minute_to + pynutil.insert("\"") + + graph_menos = pynutil.insert(" style: \"1\"") + graph_para = pynutil.insert(" style: \"2\"") + + final_graph_style = graph_menos | graph_para + + # 02.30 h (omitting seconds since a bit awkward) + graph_hm |= ( + final_graph_hour_to_24 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + pynini.closure(delete_hora_suffix, 0, 1) # 2.30 is valid if unambiguous + + final_time_zone_optional + + final_graph_style + ) + + # 2 h 30 min + graph_hm |= ( + final_graph_hour_to_24 + + delete_hora_suffix + + delete_space + + insert_space + + final_graph_minute_to + + delete_minute_suffix + + final_time_zone_optional + + final_graph_style + ) + + # 2.30 a. m. (Only for 12 hour clock) + graph_hm |= ( + final_graph_hour_to_12 + + delete_time_delimiter + + insert_space + + final_graph_minute_to + + day_suffix + + final_time_zone_optional + + final_graph_style + ) + + final_graph = graph_hm | graph_h + if deterministic: + final_graph = final_graph + pynutil.insert(" preserve_order: true") + final_graph = final_graph.optimize() + final_graph = self.add_tokens(final_graph) + self.fst = final_graph.optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/tokenize_and_classify.py @@ -0,0 +1,157 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os + +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_space, + generator_main, +) +from nemo_text_processing.text_normalization.en.taggers.punctuation import PunctuationFst +from nemo_text_processing.text_normalization.es.taggers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.taggers.date import DateFst +from nemo_text_processing.text_normalization.es.taggers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.taggers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.taggers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.taggers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.taggers.money import MoneyFst +from nemo_text_processing.text_normalization.es.taggers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.taggers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.taggers.time import TimeFst +from nemo_text_processing.text_normalization.es.taggers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.taggers.word import WordFst + +from nemo.utils import logging + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class ClassifyFst(GraphFst): + """ + Final class that composes all other classification grammars. This class can process an entire sentence, that is lower cased. + For deployment, this grammar will be compiled and exported to OpenFst Finate State aRchive (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + cache_dir: path to a dir with .far grammar file. Set to None to avoid using cache. + overwrite_cache: set to True to overwrite .far files + whitelist: path to a file with whitelist replacements + """ + + def __init__( + self, + input_case: str, + deterministic: bool = False, + cache_dir: str = None, + overwrite_cache: bool = False, + whitelist: str = None, + ): + super().__init__(name="tokenize_and_classify", kind="classify", deterministic=deterministic) + far_file = None + if cache_dir is not None and cache_dir != "None": + os.makedirs(cache_dir, exist_ok=True) + whitelist_file = os.path.basename(whitelist) if whitelist else "" + far_file = os.path.join( + cache_dir, f"_{input_case}_es_tn_{deterministic}_deterministic{whitelist_file}.far" + ) + if not overwrite_cache and far_file and os.path.exists(far_file): + self.fst = pynini.Far(far_file, mode="r")["tokenize_and_classify"] + logging.info(f"ClassifyFst.fst was restored from {far_file}.") + else: + logging.info(f"Creating ClassifyFst grammars. This might take some time...") + + self.cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = self.cardinal.fst + + self.ordinal = OrdinalFst(cardinal=self.cardinal, deterministic=deterministic) + ordinal_graph = self.ordinal.fst + + self.decimal = DecimalFst(cardinal=self.cardinal, deterministic=deterministic) + decimal_graph = self.decimal.fst + + self.fraction = FractionFst(cardinal=self.cardinal, ordinal=self.ordinal, deterministic=deterministic) + fraction_graph = self.fraction.fst + self.measure = MeasureFst( + cardinal=self.cardinal, decimal=self.decimal, fraction=self.fraction, deterministic=deterministic + ) + measure_graph = self.measure.fst + self.date = DateFst(cardinal=self.cardinal, deterministic=deterministic) + date_graph = self.date.fst + word_graph = WordFst(deterministic=deterministic).fst + self.time = TimeFst(self.cardinal, deterministic=deterministic) + time_graph = self.time.fst + self.telephone = TelephoneFst(deterministic=deterministic) + telephone_graph = self.telephone.fst + self.electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = self.electronic.fst + self.money = MoneyFst(cardinal=self.cardinal, decimal=self.decimal, deterministic=deterministic) + money_graph = self.money.fst + self.whitelist = WhiteListFst(input_case=input_case, deterministic=deterministic, input_file=whitelist) + whitelist_graph = self.whitelist.fst + punct_graph = PunctuationFst(deterministic=deterministic).fst + + classify = ( + pynutil.add_weight(whitelist_graph, 1.01) + | pynutil.add_weight(time_graph, 1.09) + | pynutil.add_weight(measure_graph, 1.08) + | pynutil.add_weight(cardinal_graph, 1.1) + | pynutil.add_weight(fraction_graph, 1.09) + | pynutil.add_weight(date_graph, 1.1) + | pynutil.add_weight(ordinal_graph, 1.1) + | pynutil.add_weight(decimal_graph, 1.1) + | pynutil.add_weight(money_graph, 1.1) + | pynutil.add_weight(telephone_graph, 1.1) + | pynutil.add_weight(electronic_graph, 1.1) + | pynutil.add_weight(word_graph, 200) + ) + punct = pynutil.insert("tokens { ") + pynutil.add_weight(punct_graph, weight=2.1) + pynutil.insert(" }") + punct = pynini.closure( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct), + 1, + ) + token = pynutil.insert("tokens { ") + classify + pynutil.insert(" }") + token_plus_punct = ( + pynini.closure(punct + pynutil.insert(" ")) + token + pynini.closure(pynutil.insert(" ") + punct) + ) + + graph = token_plus_punct + pynini.closure( + ( + pynini.compose(pynini.closure(NEMO_WHITE_SPACE, 1), delete_extra_space) + | (pynutil.insert(" ") + punct + pynutil.insert(" ")) + ) + + token_plus_punct + ) + + graph = delete_space + graph + delete_space + graph |= punct + + self.fst = graph.optimize() + + if far_file: + generator_main(far_file, {"tokenize_and_classify": self.fst}) + logging.info(f"ClassifyFst grammars are saved to {far_file}.") diff --git a/nemo_text_processing/text_normalization/es/taggers/whitelist.py b/nemo_text_processing/text_normalization/es/taggers/whitelist.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/whitelist.py @@ -0,0 +1,69 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, convert_space +from nemo_text_processing.text_normalization.es.utils import get_abs_path, load_labels + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WhiteListFst(GraphFst): + """ + Finite state transducer for classifying whitelist, e.g. + "sr." -> tokens { name: "seรฑor" } + This class has highest priority among all classifier grammars. Whitelisted tokens are defined and loaded from "data/whitelist.tsv". + + Args: + input_case: accepting either "lower_cased" or "cased" input. + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + input_file: path to a file with whitelist replacements + """ + + def __init__(self, input_case: str, deterministic: bool = True, input_file: str = None): + super().__init__(name="whitelist", kind="classify", deterministic=deterministic) + + def _get_whitelist_graph(input_case, file): + whitelist = load_labels(file) + if input_case == "lower_cased": + whitelist = [[x[0].lower()] + x[1:] for x in whitelist] + graph = pynini.string_map(whitelist) + return graph + + graph = _get_whitelist_graph(input_case, get_abs_path("data/whitelist.tsv")) + if not deterministic and input_case != "lower_cased": + graph |= pynutil.add_weight( + _get_whitelist_graph("lower_cased", get_abs_path("data/whitelist.tsv")), weight=0.0001 + ) + + if input_file: + whitelist_provided = _get_whitelist_graph(input_case, input_file) + if not deterministic: + graph |= whitelist_provided + else: + graph = whitelist_provided + + if not deterministic: + units_graph = _get_whitelist_graph(input_case, file=get_abs_path("data/measures/measurements.tsv")) + graph |= units_graph + + self.graph = graph + self.final_graph = convert_space(self.graph).optimize() + self.fst = (pynutil.insert("name: \"") + self.final_graph + pynutil.insert("\"")).optimize() diff --git a/nemo_text_processing/text_normalization/es/taggers/word.py b/nemo_text_processing/text_normalization/es/taggers/word.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/taggers/word.py @@ -0,0 +1,39 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_SPACE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class WordFst(GraphFst): + """ + Finite state transducer for classifying word. + e.g. dormir -> tokens { name: "dormir" } + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="word", kind="classify") + word = pynutil.insert("name: \"") + pynini.closure(NEMO_NOT_SPACE, 1) + pynutil.insert("\"") + self.fst = word.optimize() diff --git a/nemo_text_processing/text_normalization/es/utils.py b/nemo_text_processing/text_normalization/es/utils.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/utils.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import csv +import os + + +def get_abs_path(rel_path): + """ + Get absolute path + + Args: + rel_path: relative path to this file + + Returns absolute path + """ + return os.path.dirname(os.path.abspath(__file__)) + '/' + rel_path + + +def load_labels(abs_path): + """ + loads relative path file as dictionary + + Args: + abs_path: absolute path + + Returns dictionary of mappings + """ + label_tsv = open(abs_path) + labels = list(csv.reader(label_tsv, delimiter="\t")) + return labels diff --git a/nemo_text_processing/text_normalization/es/verbalizers/__init__.py b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. diff --git a/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/cardinal.py @@ -0,0 +1,57 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_cardinal_gender, strip_cardinal_apocope + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class CardinalFst(GraphFst): + """ + Finite state transducer for verbalizing cardinals + e.g. cardinal { integer: "dos" } -> "dos" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="cardinal", kind="verbalize", deterministic=deterministic) + optional_sign = pynini.closure(pynini.cross("negative: \"true\" ", "menos "), 0, 1) + self.optional_sign = optional_sign + + integer = pynini.closure(NEMO_NOT_QUOTE, 1) + self.integer = pynutil.delete(" \"") + integer + pynutil.delete("\"") + + integer = pynutil.delete("integer:") + self.integer + self.numbers = integer + graph = optional_sign + self.numbers + + if not deterministic: + # For alternate renderings + no_adjust = graph + fem_adjust = shift_cardinal_gender(graph) + apocope_adjust = strip_cardinal_apocope(graph) + graph = no_adjust | fem_adjust | apocope_adjust + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/date.py b/nemo_text_processing/text_normalization/es/verbalizers/date.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/date.py @@ -0,0 +1,86 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import strip_cardinal_apocope +from nemo_text_processing.text_normalization.es.taggers.date import articles + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DateFst(GraphFst): + """ + Finite state transducer for verbalizing date, e.g. + date { day: "treinta y uno" month: "marzo" year: "dos mil" } -> "treinta y uno de marzo de dos mil" + date { day: "uno" month: "mayo" year: "del mil novecientos noventa" } -> "primero de mayo del mil novecientos noventa" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="date", kind="verbalize", deterministic=deterministic) + + day_cardinal = pynutil.delete("day: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + day = strip_cardinal_apocope(day_cardinal) + + primero = pynini.cdrewrite(pynini.cross("uno", "primero"), "[BOS]", "[EOS]", NEMO_SIGMA) + day = ( + (day @ primero) if deterministic else pynini.union(day, day @ primero) + ) # Primero for first day is traditional, but will vary depending on region + + month = pynutil.delete("month: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + year = ( + pynutil.delete("year: \"") + + articles + + NEMO_SPACE + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # Insert preposition if wasn't originally with the year. This would mean a space was present + year = pynutil.add_weight(year, -0.001) + year |= ( + pynutil.delete("year: \"") + + pynutil.insert("de ") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + # day month year + graph_dmy = day + pynini.cross(NEMO_SPACE, " de ") + month + pynini.closure(pynini.accep(" ") + year, 0, 1) + + graph_mdy = month + NEMO_SPACE + day + pynini.closure(NEMO_SPACE + year, 0, 1) + if deterministic: + graph_mdy += pynutil.delete(" preserve_order: true") # Only accepts this if was explicitly passed + + self.graph = graph_dmy | graph_mdy + final_graph = self.graph + delete_preserve_order + + delete_tokens = self.delete_tokens(final_graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/decimals.py b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/decimals.py @@ -0,0 +1,87 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es import LOCALIZATION +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class DecimalFst(GraphFst): + """ + Finite state transducer for classifying decimal, e.g. + decimal { negative: "true" integer_part: "dos" fractional_part: "cuatro cero" quantity: "billones" } -> menos dos coma quatro cero billones + decimal { integer_part: "un" quantity: "billรณn" } -> un billรณn + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="decimal", kind="classify", deterministic=deterministic) + + self.optional_sign = pynini.closure(pynini.cross("negative: \"true\"", "menos ") + delete_space, 0, 1) + self.integer = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + self.fractional_default = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + conjunction = pynutil.insert(" punto ") if LOCALIZATION == "am" else pynutil.insert(" coma ") + if not deterministic: + conjunction |= pynutil.insert(pynini.union(" con ", " y ")) + self.fractional_default |= strip_cardinal_apocope(self.fractional_default) + self.fractional = conjunction + self.fractional_default + + self.quantity = ( + delete_space + + insert_space + + pynutil.delete("quantity: \"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + self.optional_quantity = pynini.closure(self.quantity, 0, 1) + + graph = self.optional_sign + pynini.union( + (self.integer + self.quantity), (self.integer + delete_space + self.fractional + self.optional_quantity) + ) + + self.numbers = graph.optimize() + self.numbers_no_quantity = self.integer + delete_space + self.fractional + self.optional_quantity + + if not deterministic: + graph |= self.optional_sign + ( + shift_cardinal_gender(self.integer + delete_space) + shift_number_gender(self.fractional) + ) + + graph += delete_preserve_order + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/electronic.py b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/electronic.py @@ -0,0 +1,91 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + digit_no_zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/digit.tsv"))) + zero = pynini.invert(pynini.string_file(get_abs_path("data/numbers/zero.tsv"))) + + graph_symbols = pynini.string_file(get_abs_path("data/electronic/symbols.tsv")) + server_common = pynini.string_file(get_abs_path("data/electronic/server_name.tsv")) + domain_common = pynini.string_file(get_abs_path("data/electronic/domain.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + digit_no_zero = None + zero = None + + graph_symbols = None + server_common = None + domain_common = None + + PYNINI_AVAILABLE = False + + +class ElectronicFst(GraphFst): + """ + Finite state transducer for verbalizing electronic + e.g. electronic { username: "abc" domain: "hotmail.com" } -> "a b c arroba hotmail punto com" + -> "a b c arroba h o t m a i l punto c o m" + -> "a b c arroba hotmail punto c o m" + -> "a b c at h o t m a i l punto com" + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="electronic", kind="verbalize", deterministic=deterministic) + + graph_digit_no_zero = ( + digit_no_zero @ pynini.cdrewrite(pynini.cross("un", "uno"), "", "", NEMO_SIGMA).optimize() + ) + graph_digit = graph_digit_no_zero | zero + + def add_space_after_char(): + return pynini.closure(NEMO_NOT_QUOTE - pynini.accep(" ") + insert_space) + ( + NEMO_NOT_QUOTE - pynini.accep(" ") + ) + + verbalize_characters = pynini.cdrewrite(graph_symbols | graph_digit, "", "", NEMO_SIGMA) + + user_name = pynutil.delete("username: \"") + add_space_after_char() + pynutil.delete("\"") + user_name @= verbalize_characters + + convert_defaults = pynutil.add_weight(NEMO_NOT_QUOTE, weight=0.0001) | domain_common | server_common + domain = convert_defaults + pynini.closure(insert_space + convert_defaults) + domain @= verbalize_characters + + domain = pynutil.delete("domain: \"") + domain + pynutil.delete("\"") + protocol = ( + pynutil.delete("protocol: \"") + + add_space_after_char() @ pynini.cdrewrite(graph_symbols, "", "", NEMO_SIGMA) + + pynutil.delete("\"") + ) + self.graph = (pynini.closure(protocol + pynini.accep(" "), 0, 1) + domain) | ( + user_name + pynini.accep(" ") + pynutil.insert("arroba ") + domain + ) + delete_tokens = self.delete_tokens(self.graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/fraction.py b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/fraction.py @@ -0,0 +1,184 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_CHAR, + NEMO_NOT_QUOTE, + NEMO_NOT_SPACE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + accents, + shift_cardinal_gender, + strip_cardinal_apocope, +) + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class FractionFst(GraphFst): + """ + Finite state transducer for verbalizing fraction + e.g. tokens { fraction { integer: "treinta y tres" numerator: "cuatro" denominator: "quinto" } } -> + treinta y tres y cuatro quintos + + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="fraction", kind="verbalize", deterministic=deterministic) + + # Derivational strings append 'avo' as a suffix. Adding space for processing aid + fraction_stem = pynutil.insert(" avo") + plural = pynutil.insert("s") + + integer = ( + pynutil.delete("integer_part: \"") + + strip_cardinal_apocope(pynini.closure(NEMO_NOT_QUOTE)) + + pynutil.delete("\"") + ) + + numerator_one = pynutil.delete("numerator: \"") + pynini.accep("un") + pynutil.delete("\" ") + numerator = ( + pynutil.delete("numerator: \"") + + pynini.difference(pynini.closure(NEMO_NOT_QUOTE), "un") + + pynutil.delete("\" ") + ) + + denominator_add_stem = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + + fraction_stem + + pynutil.delete("\" morphosyntactic_features: \"add_root\"") + ) + denominator_ordinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\" morphosyntactic_features: \"ordinal\"") + ) + denominator_cardinal = pynutil.delete("denominator: \"") + ( + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + ) + + denominator_singular = pynini.union(denominator_add_stem, denominator_ordinal) + denominator_plural = denominator_singular + plural + + if not deterministic: + # Occasional exceptions + denominator_singular |= denominator_add_stem @ pynini.string_map( + [("once avo", "undรฉcimo"), ("doce avo", "duodรฉcimo")] + ) + + # Merging operations + merge = pynini.cdrewrite( + pynini.cross(" y ", "i"), "", "", NEMO_SIGMA + ) # The denominator must be a single word, with the conjunction "y" replaced by i + merge @= pynini.cdrewrite(delete_space, "", pynini.difference(NEMO_CHAR, "parte"), NEMO_SIGMA) + + # The merger can produce duplicate vowels. This is not allowed in orthography + delete_duplicates = pynini.string_map([("aa", "a"), ("oo", "o")]) # Removes vowels + delete_duplicates = pynini.cdrewrite(delete_duplicates, "", "", NEMO_SIGMA) + + remove_accents = pynini.cdrewrite( + accents, + pynini.union(NEMO_SPACE, pynini.accep("[BOS]")) + pynini.closure(NEMO_NOT_SPACE), + pynini.closure(NEMO_NOT_SPACE) + pynini.union("avo", "ava", "รฉsimo", "รฉsima"), + NEMO_SIGMA, + ) + merge_into_single_word = merge @ remove_accents @ delete_duplicates + + fraction_default = numerator + delete_space + insert_space + (denominator_plural @ merge_into_single_word) + fraction_with_one = ( + numerator_one + delete_space + insert_space + (denominator_singular @ merge_into_single_word) + ) + + fraction_with_cardinal = strip_cardinal_apocope(numerator | numerator_one) + fraction_with_cardinal += ( + delete_space + pynutil.insert(" sobre ") + strip_cardinal_apocope(denominator_cardinal) + ) + + conjunction = pynutil.insert(" y ") + + if not deterministic: + # There is an alternative rendering where ordinals act as adjectives for 'parte'. This requires use of the feminine + # Other rules will manage use of "un" at end, so just worry about endings + exceptions = pynini.string_map([("tercia", "tercera")]) + apply_exceptions = pynini.cdrewrite(exceptions, "", "", NEMO_SIGMA) + vowel_change = pynini.cdrewrite(pynini.cross("o", "a"), "", pynini.accep("[EOS]"), NEMO_SIGMA) + + denominator_singular_fem = shift_cardinal_gender(denominator_singular) @ vowel_change @ apply_exceptions + denominator_plural_fem = denominator_singular_fem + plural + + numerator_one_fem = shift_cardinal_gender(numerator_one) + numerator_fem = shift_cardinal_gender(numerator) + + fraction_with_cardinal |= ( + (numerator_one_fem | numerator_fem) + + delete_space + + pynutil.insert(" sobre ") + + shift_cardinal_gender(denominator_cardinal) + ) + + # Still need to manage stems + merge_stem = pynini.cdrewrite( + delete_space, "", pynini.union("avo", "ava", "avos", "avas"), NEMO_SIGMA + ) # For managing alternative spacing + merge_stem @= remove_accents @ delete_duplicates + + fraction_with_one_fem = numerator_one_fem + delete_space + insert_space + fraction_with_one_fem += pynini.union( + denominator_singular_fem @ merge_stem, denominator_singular_fem @ merge_into_single_word + ) # Both forms exists + fraction_with_one_fem @= pynini.cdrewrite(pynini.cross("una media", "media"), "", "", NEMO_SIGMA) + fraction_with_one_fem += pynutil.insert(" parte") + + fraction_default_fem = numerator_fem + delete_space + insert_space + fraction_default_fem += pynini.union( + denominator_plural_fem @ merge_stem, denominator_plural_fem @ merge_into_single_word + ) + fraction_default_fem += pynutil.insert(" partes") + + fraction_default |= ( + numerator + delete_space + insert_space + denominator_plural @ merge_stem + ) # Case of no merger + fraction_default |= fraction_default_fem + + fraction_with_one |= numerator_one + delete_space + insert_space + denominator_singular @ merge_stem + fraction_with_one |= fraction_with_one_fem + + # Integers are influenced by dominant noun, need to allow feminine forms as well + integer |= shift_cardinal_gender(integer) + + # Remove 'un medio' + fraction_with_one @= pynini.cdrewrite(pynini.cross("un medio", "medio"), "", "", NEMO_SIGMA) + + integer = pynini.closure(integer + delete_space + conjunction, 0, 1) + + fraction = fraction_with_one | fraction_default | fraction_with_cardinal + + graph = integer + fraction + + self.graph = graph + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/measure.py b/nemo_text_processing/text_normalization/es/verbalizers/measure.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/measure.py @@ -0,0 +1,110 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + NEMO_WHITE_SPACE, + GraphFst, + delete_extra_space, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ones, shift_cardinal_gender +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + unit_plural_fem = pynini.string_file(get_abs_path("data/measures/measurements_plural_fem.tsv")) + unit_plural_masc = pynini.string_file(get_abs_path("data/measures/measurements_plural_masc.tsv")) + + unit_singular_fem = pynini.project(unit_plural_fem, "input") + unit_singular_masc = pynini.project(unit_plural_masc, "input") + + unit_plural_fem = pynini.project(unit_plural_fem, "output") + unit_plural_masc = pynini.project(unit_plural_masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + unit_plural_fem = None + unit_plural_masc = None + + unit_singular_fem = None + unit_singular_masc = None + + PYNINI_AVAILABLE = False + + +class MeasureFst(GraphFst): + """ + Finite state transducer for verbalizing measure, e.g. + measure { cardinal { integer: "dos" units: "gramos" } } -> "dos gramos" + measure { cardinal { integer_part: "dos" quantity: "millones" units: "gramos" } } -> "dos millones de gramos" + + Args: + decimal: DecimalFst + cardinal: CardinalFst + fraction: FractionFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, cardinal: GraphFst, fraction: GraphFst, deterministic: bool): + super().__init__(name="measure", kind="verbalize", deterministic=deterministic) + + graph_decimal = decimal.fst + graph_cardinal = cardinal.fst + graph_fraction = fraction.fst + + unit_masc = (unit_plural_masc | unit_singular_masc) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_masc |= "por" + pynini.closure(NEMO_NOT_QUOTE, 1) + unit_masc = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_masc) + pynutil.delete("\"") + + unit_fem = (unit_plural_fem | unit_singular_fem) + pynini.closure( + NEMO_WHITE_SPACE + "por" + pynini.closure(NEMO_NOT_QUOTE, 1), 0, 1 + ) + unit_fem = pynutil.delete("units: \"") + (pynini.closure(NEMO_NOT_QUOTE) @ unit_fem) + pynutil.delete("\"") + + graph_masc = (graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_masc + graph_fem = ( + shift_cardinal_gender(graph_cardinal | graph_decimal | graph_fraction) + NEMO_WHITE_SPACE + unit_fem + ) + graph = graph_masc | graph_fem + + graph = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph + ) # billones de xyz + + graph @= pynini.cdrewrite(pynini.cross(ones, "uno"), "", NEMO_WHITE_SPACE + "por", NEMO_SIGMA) + + # To manage alphanumeric combonations ("a-8, 5x"), we let them use a weighted default path. + alpha_num_unit = pynutil.delete("units: \"") + pynini.closure(NEMO_NOT_QUOTE) + pynutil.delete("\"") + graph_alpha_num = pynini.union( + (graph_cardinal | graph_decimal) + NEMO_SPACE + alpha_num_unit, + alpha_num_unit + delete_extra_space + (graph_cardinal | graph_decimal), + ) + + graph |= pynutil.add_weight(graph_alpha_num, 0.01) + + graph += delete_preserve_order + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/money.py b/nemo_text_processing/text_normalization/es/verbalizers/money.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/money.py @@ -0,0 +1,195 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + NEMO_SPACE, + GraphFst, + delete_preserve_order, +) +from nemo_text_processing.text_normalization.es.graph_utils import ( + shift_cardinal_gender, + shift_number_gender, + strip_cardinal_apocope, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + fem = pynini.string_file((get_abs_path("data/money/currency_plural_fem.tsv"))) + masc = pynini.string_file((get_abs_path("data/money/currency_plural_masc.tsv"))) + + fem_singular = pynini.project(fem, "input") + masc_singular = pynini.project(masc, "input") + + fem_plural = pynini.project(fem, "output") + masc_plural = pynini.project(masc, "output") + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + fem_plural = None + masc_plural = None + + fem_singular = None + masc_singular = None + + PYNINI_AVAILABLE = False + + +class MoneyFst(GraphFst): + """ + Finite state transducer for verbalizing money, e.g. + money { currency_maj: "euro" integer_part: "un"} -> "un euro" + money { currency_maj: "euro" integer_part: "un" fractional_part: "cero cero un"} -> "uno coma cero cero uno euros" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" preserve_order: true} -> "una libra cuarenta" + money { integer_part: "un" currency_maj: "libra" fractional_part: "cuarenta" currency_min: "peniques" preserve_order: true} -> "una libra con cuarenta peniques" + money { fractional_part: "un" currency_min: "penique" preserve_order: true} -> "un penique" + + Args: + decimal: GraphFst + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, decimal: GraphFst, deterministic: bool = True): + super().__init__(name="money", kind="verbalize", deterministic=deterministic) + + maj_singular_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + maj_singular_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + maj_plural_masc = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + maj_plural_fem = ( + pynutil.delete("currency_maj: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + maj_masc = maj_plural_masc | maj_singular_masc # Tagger kept quantity resolution stable + maj_fem = maj_plural_fem | maj_singular_fem + + min_singular_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_singular) + + pynutil.delete("\"") + ) + min_singular_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_singular) + + pynutil.delete("\"") + ) + + min_plural_masc = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ masc_plural) + + pynutil.delete("\"") + ) + min_plural_fem = ( + pynutil.delete("currency_min: \"") + + (pynini.closure(NEMO_NOT_QUOTE, 1) @ fem_plural) + + pynutil.delete("\"") + ) + + min_masc = min_plural_masc | min_singular_masc + min_fem = min_plural_fem | min_singular_fem + + fractional_part = ( + pynutil.delete("fractional_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + ) + + integer_part = pynutil.delete("integer_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + optional_add_and = pynini.closure(pynutil.insert(pynini.union("con ", "y ")), 0, 1) + + # *** currency_maj + graph_integer_masc = integer_part + NEMO_SPACE + maj_masc + graph_integer_fem = shift_cardinal_gender(integer_part) + NEMO_SPACE + maj_fem + graph_integer = graph_integer_fem | graph_integer_masc + + # *** currency_maj + (***) | ((con) *** current_min) + graph_integer_with_minor_masc = ( + integer_part + + NEMO_SPACE + + maj_masc + + NEMO_SPACE + + pynini.union( + optional_add_and + strip_cardinal_apocope(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor_fem = ( + shift_cardinal_gender(integer_part) + + NEMO_SPACE + + maj_fem + + NEMO_SPACE + + pynini.union( + optional_add_and + shift_cardinal_gender(fractional_part), + (optional_add_and + fractional_part + NEMO_SPACE + min_masc), + (optional_add_and + shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem), + ) # Could be minor currency that is different gender + + delete_preserve_order + ) + + graph_integer_with_minor = graph_integer_with_minor_fem | graph_integer_with_minor_masc + + # *** coma *** currency_maj + graph_decimal_masc = decimal.numbers + NEMO_SPACE + maj_masc + + # Need to fix some of the inner parts, so don't use decimal here (note: quantities covered by masc) + graph_decimal_fem = ( + pynini.accep("integer_part: \"") + + shift_cardinal_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SPACE + + pynini.accep("fractional_part: \"") + + shift_number_gender(pynini.closure(NEMO_NOT_QUOTE, 1)) + + pynini.accep("\"") + + NEMO_SIGMA + ) + graph_decimal_fem @= decimal.numbers_no_quantity + graph_decimal_fem += NEMO_SPACE + maj_fem + + graph_decimal = graph_decimal_fem | graph_decimal_masc + graph_decimal = ( + pynini.cdrewrite( + pynutil.insert(" de"), "quantity: \"" + pynini.closure(NEMO_NOT_QUOTE, 1), "\"", NEMO_SIGMA + ) + @ graph_decimal + ) # formally it's millones/billones de *** + + # *** current_min + graph_minor_masc = fractional_part + NEMO_SPACE + min_masc + delete_preserve_order + graph_minor_fem = shift_cardinal_gender(fractional_part) + NEMO_SPACE + min_fem + delete_preserve_order + graph_minor = graph_minor_fem | graph_minor_masc + + graph = graph_integer | graph_integer_with_minor | graph_decimal | graph_minor + + delete_tokens = self.delete_tokens(graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/ordinal.py @@ -0,0 +1,76 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, NEMO_SIGMA, NEMO_SPACE, GraphFst +from nemo_text_processing.text_normalization.es.graph_utils import shift_number_gender + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class OrdinalFst(GraphFst): + """ + Finite state transducer for verbalizing ordinals + e.g. ordinal { integer: "tercer" } } -> "tercero" + -> "tercera" + -> "tercer" + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="ordinal", kind="verbalize", deterministic=deterministic) + + graph = pynutil.delete("integer: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + + # masculne gender we leave as is + graph_masc = graph + pynutil.delete(" morphosyntactic_features: \"gender_masc") + + # shift gender + graph_fem_ending = graph @ pynini.cdrewrite( + pynini.cross("o", "a"), "", NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + graph_fem = shift_number_gender(graph_fem_ending) + pynutil.delete(" morphosyntactic_features: \"gender_fem") + + # Apocope just changes tercero and primero. May occur if someone wrote 11.er (uncommon) + graph_apocope = ( + pynini.cross("tercero", "tercer") + | pynini.cross("primero", "primer") + | pynini.cross("undรฉcimo", "decimoprimer") + ) # In case someone wrote 11.er with deterministic + graph_apocope = (graph @ pynini.cdrewrite(graph_apocope, "", "", NEMO_SIGMA)) + pynutil.delete( + " morphosyntactic_features: \"apocope" + ) + + graph = graph_apocope | graph_masc | graph_fem + + if not deterministic: + # Plural graph + graph_plural = pynini.cdrewrite( + pynutil.insert("s"), pynini.union("o", "a"), NEMO_SPACE | pynini.accep("[EOS]"), NEMO_SIGMA + ) + + graph |= (graph @ graph_plural) + pynutil.delete("/plural") + + self.graph = graph + pynutil.delete("\"") + + delete_tokens = self.delete_tokens(self.graph) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/telephone.py b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/telephone.py @@ -0,0 +1,42 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import NEMO_NOT_QUOTE, GraphFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class TelephoneFst(GraphFst): + """ + Finite state transducer for verbalizing telephone, e.g. + telephone { number_part: "uno dos tres uno dos tres cinco seis siete ocho" } + -> uno dos tres uno dos tres cinco seis siete ocho + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="telephone", kind="verbalize") + + number_part = pynutil.delete("number_part: \"") + pynini.closure(NEMO_NOT_QUOTE, 1) + pynutil.delete("\"") + delete_tokens = self.delete_tokens(number_part) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/time.py b/nemo_text_processing/text_normalization/es/verbalizers/time.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/time.py @@ -0,0 +1,269 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import ( + NEMO_NOT_QUOTE, + NEMO_SIGMA, + GraphFst, + delete_preserve_order, + delete_space, + insert_space, +) +from nemo_text_processing.text_normalization.es.utils import get_abs_path + +try: + import pynini + from pynini.lib import pynutil + + alt_minutes = pynini.string_file(get_abs_path("data/time/alt_minutes.tsv")) + + morning_times = pynini.string_file(get_abs_path("data/time/morning_times.tsv")) + afternoon_times = pynini.string_file(get_abs_path("data/time/afternoon_times.tsv")) + evening_times = pynini.string_file(get_abs_path("data/time/evening_times.tsv")) + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + alt_minutes = None + + morning_times = None + afternoon_times = None + evening_times = None + + PYNINI_AVAILABLE = False + + +class TimeFst(GraphFst): + """ + Finite state transducer for verbalizing time, e.g. + time { hours: "doce" minutes: "media" suffix: "a m" } -> doce y media de la noche + time { hours: "doce" } -> twelve o'clock + + Args: + deterministic: if True will provide a single transduction option, + for False multiple transduction are generated (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="time", kind="verbalize", deterministic=deterministic) + + change_minutes = pynini.cdrewrite(alt_minutes, pynini.accep("[BOS]"), pynini.accep("[EOS]"), NEMO_SIGMA) + + morning_phrases = pynini.cross("am", "de la maรฑana") + afternoon_phrases = pynini.cross("pm", "de la tarde") + evening_phrases = pynini.cross("pm", "de la noche") + + # For the 12's + mid_times = pynini.accep("doce") + mid_phrases = ( + pynini.string_map([("pm", "del mediodรญa"), ("am", "de la noche")]) + if deterministic + else pynini.string_map( + [ + ("pm", "de la maรฑana"), + ("pm", "del dรญa"), + ("pm", "del mediodรญa"), + ("am", "de la noche"), + ("am", "de la medianoche"), + ] + ) + ) + + hour = ( + pynutil.delete("hours:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = ( + pynutil.delete("minutes:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + minute = (minute @ change_minutes) if deterministic else pynini.union(minute, minute @ change_minutes) + + suffix = ( + pynutil.delete("suffix:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + zone = ( + pynutil.delete("zone:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + optional_zone = pynini.closure(delete_space + insert_space + zone, 0, 1) + second = ( + pynutil.delete("seconds:") + + delete_space + + pynutil.delete("\"") + + pynini.closure(NEMO_NOT_QUOTE, 1) + + pynutil.delete("\"") + ) + + graph_hms = ( + hour + + pynutil.insert(" horas ") + + delete_space + + minute + + pynutil.insert(" minutos y ") + + delete_space + + second + + pynutil.insert(" segundos") + ) + + graph_hm = hour + delete_space + pynutil.insert(" y ") + minute + graph_hm |= pynini.union( + (hour @ morning_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases), + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases), + (hour @ evening_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases), + (hour @ mid_times) + + delete_space + + pynutil.insert(" y ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases), + ) + + graph_h = pynini.union( + hour, + (hour @ morning_times) + delete_space + insert_space + (suffix @ morning_phrases), + (hour @ afternoon_times) + delete_space + insert_space + (suffix @ afternoon_phrases), + (hour @ evening_times) + delete_space + insert_space + (suffix @ evening_phrases), + (hour @ mid_times) + delete_space + insert_space + (suffix @ mid_phrases), + ) + + graph = (graph_hms | graph_hm | graph_h) + optional_zone + + if not deterministic: + graph_style_1 = pynutil.delete(" style: \"1\"") + graph_style_2 = pynutil.delete(" style: \"2\"") + + graph_menos = hour + delete_space + pynutil.insert(" menos ") + minute + graph_style_1 + graph_menos |= ( + (hour @ morning_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ afternoon_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ evening_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_1 + ) + graph_menos |= ( + (hour @ mid_times) + + delete_space + + pynutil.insert(" menos ") + + minute + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_1 + ) + graph_menos += optional_zone + + graph_para = minute + pynutil.insert(" para las ") + delete_space + hour + graph_style_2 + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ morning_times) + + delete_space + + insert_space + + (suffix @ morning_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ afternoon_times) + + delete_space + + insert_space + + (suffix @ afternoon_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ evening_times) + + delete_space + + insert_space + + (suffix @ evening_phrases) + + graph_style_2 + ) + graph_para |= ( + minute + + pynutil.insert(" para las ") + + delete_space + + (hour @ mid_times) + + delete_space + + insert_space + + (suffix @ mid_phrases) + + graph_style_2 + ) + graph_para += optional_zone + graph_para @= pynini.cdrewrite( + pynini.cross(" las ", " la "), "para", "una", NEMO_SIGMA + ) # Need agreement with one + + graph |= graph_menos | graph_para + delete_tokens = self.delete_tokens(graph + delete_preserve_order) + self.fst = delete_tokens.optimize() diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize.py @@ -0,0 +1,73 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst +from nemo_text_processing.text_normalization.en.verbalizers.whitelist import WhiteListFst +from nemo_text_processing.text_normalization.es.verbalizers.cardinal import CardinalFst +from nemo_text_processing.text_normalization.es.verbalizers.date import DateFst +from nemo_text_processing.text_normalization.es.verbalizers.decimals import DecimalFst +from nemo_text_processing.text_normalization.es.verbalizers.electronic import ElectronicFst +from nemo_text_processing.text_normalization.es.verbalizers.fraction import FractionFst +from nemo_text_processing.text_normalization.es.verbalizers.measure import MeasureFst +from nemo_text_processing.text_normalization.es.verbalizers.money import MoneyFst +from nemo_text_processing.text_normalization.es.verbalizers.ordinal import OrdinalFst +from nemo_text_processing.text_normalization.es.verbalizers.telephone import TelephoneFst +from nemo_text_processing.text_normalization.es.verbalizers.time import TimeFst + + +class VerbalizeFst(GraphFst): + """ + Composes other verbalizer grammars. + For deployment, this grammar will be compiled and exported to OpenFst Finate State Archiv (FAR) File. + More details to deployment at NeMo/tools/text_processing_deployment. + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize", kind="verbalize", deterministic=deterministic) + cardinal = CardinalFst(deterministic=deterministic) + cardinal_graph = cardinal.fst + ordinal = OrdinalFst(deterministic=deterministic) + ordinal_graph = ordinal.fst + decimal = DecimalFst(deterministic=deterministic) + decimal_graph = decimal.fst + fraction = FractionFst(deterministic=deterministic) + fraction_graph = fraction.fst + date = DateFst(deterministic=deterministic) + date_graph = date.fst + measure = MeasureFst(cardinal=cardinal, decimal=decimal, fraction=fraction, deterministic=deterministic) + measure_graph = measure.fst + electronic = ElectronicFst(deterministic=deterministic) + electronic_graph = electronic.fst + whitelist_graph = WhiteListFst(deterministic=deterministic).fst + money_graph = MoneyFst(decimal=decimal, deterministic=deterministic).fst + telephone_graph = TelephoneFst(deterministic=deterministic).fst + time_graph = TimeFst(deterministic=deterministic).fst + + graph = ( + cardinal_graph + | measure_graph + | decimal_graph + | ordinal_graph + | date_graph + | electronic_graph + | money_graph + | fraction_graph + | whitelist_graph + | telephone_graph + | time_graph + ) + self.fst = graph diff --git a/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py new file mode 100644 --- /dev/null +++ b/nemo_text_processing/text_normalization/es/verbalizers/verbalize_final.py @@ -0,0 +1,52 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from nemo_text_processing.text_normalization.en.graph_utils import GraphFst, delete_extra_space, delete_space +from nemo_text_processing.text_normalization.en.verbalizers.word import WordFst +from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst + +try: + import pynini + from pynini.lib import pynutil + + PYNINI_AVAILABLE = True + +except (ModuleNotFoundError, ImportError): + PYNINI_AVAILABLE = False + + +class VerbalizeFinalFst(GraphFst): + """ + Finite state transducer that verbalizes an entire sentence + + Args: + deterministic: if True will provide a single transduction option, + for False multiple options (used for audio-based normalization) + """ + + def __init__(self, deterministic: bool = True): + super().__init__(name="verbalize_final", kind="verbalize", deterministic=deterministic) + verbalize = VerbalizeFst(deterministic=deterministic).fst + word = WordFst(deterministic=deterministic).fst + types = verbalize | word + graph = ( + pynutil.delete("tokens") + + delete_space + + pynutil.delete("{") + + delete_space + + types + + delete_space + + pynutil.delete("}") + ) + graph = delete_space + pynini.closure(graph + delete_extra_space) + graph + delete_space + self.fst = graph diff --git a/nemo_text_processing/text_normalization/normalize.py b/nemo_text_processing/text_normalization/normalize.py --- a/nemo_text_processing/text_normalization/normalize.py +++ b/nemo_text_processing/text_normalization/normalize.py @@ -46,8 +46,8 @@ class Normalizer: """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -83,10 +83,11 @@ def __init__( from nemo_text_processing.text_normalization.ru.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.ru.verbalizers.verbalize_final import VerbalizeFinalFst elif lang == 'de': - # Ru TN only support non-deterministic cases and produces multiple normalization options - # use normalize_with_audio.py from nemo_text_processing.text_normalization.de.taggers.tokenize_and_classify import ClassifyFst from nemo_text_processing.text_normalization.de.verbalizers.verbalize_final import VerbalizeFinalFst + elif lang == 'es': + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ClassifyFst + from nemo_text_processing.text_normalization.es.verbalizers.verbalize_final import VerbalizeFinalFst self.tagger = ClassifyFst( input_case=input_case, deterministic=deterministic, @@ -106,7 +107,7 @@ def __init__( def normalize_list(self, texts: List[str], verbose=False, punct_post_process: bool = False) -> List[str]: """ - NeMo text normalizer + NeMo text normalizer Args: texts: list of input strings @@ -357,7 +358,7 @@ def select_verbalizer(self, lattice: 'pynini.FstLike') -> str: def parse_args(): parser = ArgumentParser() parser.add_argument("input_string", help="input string", type=str) - parser.add_argument("--language", help="language", choices=["en", "de"], default="en", type=str) + parser.add_argument("--language", help="language", choices=["en", "de", "es"], default="en", type=str) parser.add_argument( "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) diff --git a/nemo_text_processing/text_normalization/normalize_with_audio.py b/nemo_text_processing/text_normalization/normalize_with_audio.py --- a/nemo_text_processing/text_normalization/normalize_with_audio.py +++ b/nemo_text_processing/text_normalization/normalize_with_audio.py @@ -55,15 +55,15 @@ "audio_data" - path to the audio file "text" - raw text "pred_text" - ASR model prediction - + See https://github.com/NVIDIA/NeMo/blob/main/examples/asr/transcribe_speech.py on how to add ASR predictions - + When the manifest is ready, run: python normalize_with_audio.py \ --audio_data PATH/TO/MANIFEST.JSON \ - --language en - - + --language en + + To run with a single audio file, specify path to audio and text with: python normalize_with_audio.py \ --audio_data PATH/TO/AUDIO.WAV \ @@ -71,18 +71,18 @@ --text raw text OR PATH/TO/.TXT/FILE --model QuartzNet15x5Base-En \ --verbose - + To see possible normalization options for a text input without an audio file (could be used for debugging), run: python python normalize_with_audio.py --text "RAW TEXT" - + Specify `--cache_dir` to generate .far grammars once and re-used them for faster inference """ class NormalizerWithAudio(Normalizer): """ - Normalizer class that converts text from written to spoken form. - Useful for TTS preprocessing. + Normalizer class that converts text from written to spoken form. + Useful for TTS preprocessing. Args: input_case: expected input capitalization @@ -282,7 +282,7 @@ def parse_args(): "--input_case", help="input capitalization", choices=["lower_cased", "cased"], default="cased", type=str ) parser.add_argument( - "--language", help="Select target language", choices=["en", "ru", "de"], default="en", type=str + "--language", help="Select target language", choices=["en", "ru", "de", "es"], default="en", type=str ) parser.add_argument("--audio_data", default=None, help="path to an audio file or .json manifest") parser.add_argument( diff --git a/tools/text_processing_deployment/pynini_export.py b/tools/text_processing_deployment/pynini_export.py --- a/tools/text_processing_deployment/pynini_export.py +++ b/tools/text_processing_deployment/pynini_export.py @@ -67,7 +67,7 @@ def tn_grammars(**kwargs): def export_grammars(output_dir, grammars): """ - Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. + Exports tokenizer_and_classify and verbalize Fsts as OpenFst finite state archive (FAR) files. Args: output_dir: directory to export FAR files to. Subdirectories will be created for tagger and verbalizer respectively. @@ -109,7 +109,7 @@ def parse_args(): if __name__ == '__main__': args = parse_args() - if args.language in ['ru', 'fr', 'es', 'vi'] and args.grammars == 'tn_grammars': + if args.language in ['ru', 'fr', 'vi'] and args.grammars == 'tn_grammars': raise ValueError('Only ITN grammars could be deployed in Sparrowhawk for the selected languages.') if args.language == 'en': @@ -148,6 +148,10 @@ def parse_args(): from nemo_text_processing.inverse_text_normalization.es.verbalizers.verbalize import ( VerbalizeFst as ITNVerbalizeFst, ) + from nemo_text_processing.text_normalization.es.taggers.tokenize_and_classify import ( + ClassifyFst as TNClassifyFst, + ) + from nemo_text_processing.text_normalization.es.verbalizers.verbalize import VerbalizeFst as TNVerbalizeFst elif args.language == 'fr': from nemo_text_processing.inverse_text_normalization.fr.taggers.tokenize_and_classify import ( ClassifyFst as ITNClassifyFst, </patch> </s> </patch>
diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_cardinal.txt @@ -0,0 +1,86 @@ +1~un +2~dos +3~tres +4~cuatro +5~cinco +6~seis +7~siete +8~ocho +9~nueve +10~diez +11~once +12~doce +13~trece +14~catorce +15~quince +16~diecisรฉis +17~diecisiete +18~dieciocho +19~diecinueve +20~veinte +21~veintiรบn +22~veintidรณs +23~veintitrรฉs +24~veinticuatro +25~veinticinco +26~veintisรฉis +27~veintisiete +28~veintiocho +29~veintinueve +30~treinta +31~treinta y un +40~cuarenta +41~cuarenta y un +50~cincuenta +51~cincuenta y un +60~sesenta +70~setenta +80~ochenta +90~noventa +100~cien +101~ciento un +120~ciento veinte +121~ciento veintiรบn +130~ciento treinta +131~ciento treinta y un +200~doscientos +201~doscientos un +300~trescientos +301~trescientos un +1000~mil +1 000~mil +1.000~mil +1001~mil un +1010~mil diez +1020~mil veinte +1021~mil veintiรบn +1100~mil cien +1101~mil ciento un +1110~mil ciento diez +1111~mil ciento once +1234~mil doscientos treinta y cuatro +2000~dos mil +2001~dos mil un +2010~dos mil diez +2020~dos mil veinte +2100~dos mil cien +2101~dos mil ciento un +2110~dos mil ciento diez +2111~dos mil ciento once +2222~dos mil doscientos veintidรณs +10000~diez mil +10 000~diez mil +10.000~diez mil +100000~cien mil +100 000~cien mil +100.000~cien mil +1 000 000~un millรณn +1.000.000~un millรณn +1 234 568~un millรณn doscientos treinta y cuatro mil quinientos sesenta y ocho +2.000.000~dos millones +1.000.000.000~mil millones +2.000.000.000~dos mil millones +3 000 000 000 000~tres billones +3.000.000.000.000~tres billones +100 000 000 000 000 000 000 000~cien mil trillones +100 000 000 000 000 000 000 001~cien mil trillones un diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_date.txt @@ -0,0 +1,13 @@ +1 enero~primero de enero +5 febrero~cinco de febrero +20 de marzo~veinte de marzo +abril 30~treinta de abril +31 marzo~treinta y uno de marzo +10 mayo 1990~diez de mayo de mil novecientos noventa +junio 11 2000~once de junio de dos mil +30 julio del 2020~treinta de julio del dos mil veinte +30-2-1990~treinta de febrero de mil novecientos noventa +30/2/1990~treinta de febrero de mil novecientos noventa +30.2.1990~treinta de febrero de mil novecientos noventa +1990-2-30~treinta de febrero de mil novecientos noventa +1990-02-30~treinta de febrero de mil novecientos noventa \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_decimal.txt @@ -0,0 +1,27 @@ +0,1~cero coma un +0,01~cero coma cero un +0,010~cero coma cero uno cero +1,0101~uno coma cero uno cero un +0,0~cero coma cero +1,0~uno coma cero +1,00~uno coma cero cero +1,1~uno coma un +233,32~doscientos treinta y tres coma treinta y dos +32,22 millones~treinta y dos coma veintidรณs millones +320 320,22 millones~trescientos veinte mil trescientos veinte coma veintidรณs millones +5.002,232~cinco mil dos coma doscientos treinta y dos +3,2 trillones~tres coma dos trillones +3 millones~tres millones +3 000 millones~tres mil millones +3000 millones~tres mil millones +3.000 millones~tres mil millones +3.001 millones~tres mil un millones +1 millรณn~un millรณn +1 000 millones~mil millones +1000 millones~mil millones +1.000 millones~mil millones +2,33302 millones~dos coma tres tres tres cero dos millones +1,5332 millรณn~uno coma cinco tres tres dos millรณn +1,53322 millรณn~uno coma cinco tres tres dos dos millรณn +1,53321 millรณn~uno coma cinco tres tres dos un millรณn +101,010101 millones~ciento uno coma cero uno cero uno cero un millones \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_electronic.txt @@ -0,0 +1,12 @@ +a.bc@gmail.com~a punto b c arroba gmail punto com +cdf@abc.edu~c d f arroba a b c punto e d u +abc@gmail.abc~a b c arroba gmail punto a b c +abc@abc.com~a b c arroba a b c punto com +asdf123@abc.com~a s d f uno dos tres arroba a b c punto com +a1b2@abc.com~a uno b dos arroba a b c punto com +ab3.sdd.3@gmail.com~a b tres punto s d d punto tres arroba gmail punto com +https://www.nvidia.com~h t t p s dos puntos barra barra w w w punto nvidia punto com +www.nvidia.com~w w w punto nvidia punto com +www.abc.es/efg~w w w punto a b c punto es barra e f g +www.abc.es~w w w punto a b c punto es +http://www.ourdailynews.com.sm~h t t p dos puntos barra barra w w w punto o u r d a i l y n e w s punto com punto s m \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_fraction.txt @@ -0,0 +1,76 @@ +1/2~medio +1 1/2~uno y medio +3/2~tres medios +1 3/2~uno y tres medios +1/3~un tercio +2/3~dos tercios +1/4~un cuarto +2/4~dos cuartos +1/5~un quinto +2/5~dos quintos +1/6~un sexto +2/6~dos sextos +1/7~un sรฉptimo +2/7~dos sรฉptimos +1/8~un octavo +2/8~dos octavos +1/9~un noveno +2/9~dos novenos +1/10~un dรฉcimo +2/10~dos dรฉcimos +1/11~un onceavo +1/12~un doceavo +1/13~un treceavo +1/14~un catorceavo +1/15~un quinceavo +1/16~un dieciseisavo +1/17~un diecisieteavo +1/18~un dieciochoavo +1/19~un diecinueveavo +1/20~un veinteavo +1/21~un veintiunavo +1/22~un veintidosavo +1/30~un treintavo +1/31~un treintaiunavo +1/40~un cuarentavo +1/41~un cuarentaiunavo +1/50~un cincuentavo +1/60~un sesentavo +1/70~un setentavo +1/80~un ochentavo +1/90~un noventavo +1/100~un centรฉsimo +2/100~dos centรฉsimos +1 2/100~uno y dos centรฉsimos +1/101~uno sobre ciento uno +1/110~uno sobre ciento diez +1/111~uno sobre ciento once +1/112~uno sobre ciento doce +1/123~uno sobre ciento veintitrรฉs +1/134~uno sobre ciento treinta y cuatro +1/200~un ducentรฉsimo +1/201~uno sobre doscientos uno +1/234~uno sobre doscientos treinta y cuatro +1/300~un tricentรฉsimo +1/345~uno sobre trescientos cuarenta y cinco +1/400~un cuadringentรฉsimo +1/456~uno sobre cuatrocientos cincuenta y seis +1/500~un quingentรฉsimo +1/600~un sexcentรฉsimo +1/700~un septingentรฉsimo +1/800~un octingentรฉsimo +1/900~un noningentรฉsimo +1/1000~un milรฉsimo +2/1000~dos milรฉsimos +1 2/1000~uno y dos milรฉsimos +1/1001~uno sobre mil uno +1/1100~uno sobre mil cien +1/1200~uno sobre mil doscientos +1/1234~uno sobre mil doscientos treinta y cuatro +1/2000~un dosmilรฉsimo +1/5000~un cincomilรฉsimo +1/10000~un diezmilรฉsimo +1/100.000~un cienmilรฉsimo +1/1.000.000~un millonรฉsimo +1/100.000.000~un cienmillonรฉsimo +1/1.200.000.000~un mildoscientosmillonรฉsimo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_measure.txt @@ -0,0 +1,17 @@ +1,2-a~uno coma dos a +a-5~a cinco +200 m~doscientos metros +3 h~tres horas +1 h~una hora +245 mph~doscientas cuarenta y cinco millas por hora +2 kg~dos kilogramos +60,2400 kg~sesenta coma dos cuatro cero cero kilogramos +-60,2400 kg~menos sesenta coma dos cuatro cero cero kilogramos +8,52 %~ocho coma cincuenta y dos por ciento +-8,52 %~menos ocho coma cincuenta y dos por ciento +1 %~uno por ciento +3 cm~tres centรญmetros +4 s~cuatro segundos +5 l~cinco litros +4,51/s~cuatro coma cincuenta y uno por segundo +0,0101 s~cero coma cero uno cero un segundos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_money.txt @@ -0,0 +1,24 @@ +$1~un dรณlar +1 $~un dรณlar +$1,50~un dรณlar cincuenta centavos +1,50 $~un dรณlar cincuenta centavos +ยฃ200.000.001~doscientos millones una libras +200.000.001 ยฃ~doscientos millones una libras +2 billones de euros~dos billones de euros +โ‚ฌ2 billones~dos billones de euros +โ‚ฌ 2 billones~dos billones de euros +โ‚ฌ 2,3 billones~dos coma tres billones de euros +2,3 billones de euros~dos coma tres billones de euros +โ‚ฌ5,50~cinco euros cincuenta cรฉntimos +5,50 โ‚ฌ~cinco euros cincuenta cรฉntimos +5,01 โ‚ฌ~cinco euros un cรฉntimo +5,01 ยฃ~cinco libras un penique +21 czk~veintiuna coronas checas +czk21~veintiuna coronas checas +czk21,1 millones~veintiuna coma una millones de coronas checas +czk 5,50 billones~cinco coma cincuenta billones de coronas checas +rs 5,50 billones~cinco coma cincuenta billones de rupias +czk5,50 billones~cinco coma cincuenta billones de coronas checas +0,55 $~cincuenta y cinco centavos +1,01 $~un dรณlar un centavo +ยฅ12,05~doce yenes cinco centavos \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_normalize_with_audio.txt @@ -0,0 +1,120 @@ +~121 +ciento veintiรบn +ciento veintiuno +ciento veintiuna +121 +~200 +doscientos +doscientas +200 +~201 +doscientos un +doscientos uno +doscientas una +201 +~1 +un +uno +una +1 +~550.000.001 +quinientos cincuenta millones un +quinientos cincuenta millones una +quinientos cincuenta millones uno +550.000.001 +~500.501 +quinientos mil quinientos un +quinientos mil quinientos uno +quinientas mil quinientas una +500.501 +~500.001.ยบ +quinientosmilรฉsimo primero +quingentรฉsimo milรฉsimo primero +quinientosmilรฉsimos primeros +quingentรฉsimos milรฉsimos primeros +500.001.ยบ +~500.001.ยช +quinientasmilรฉsima primera +quingentรฉsima milรฉsima primera +quinientasmilรฉsimas primeras +quingentรฉsimas milรฉsimas primeras +500.001.ยช +~11.ยช +dรฉcima primera +decimoprimera +dรฉcimas primeras +decimoprimeras +undรฉcima +undรฉcimas +11.ยช +~11.ยบ +dรฉcimo primero +decimoprimero +dรฉcimos primeros +decimoprimeros +undรฉcimo +undรฉcimos +11.ยบ +~12.ยบ +dรฉcimo segundo +decimosegundo +dรฉcimos segundos +decimosegundos +duodรฉcimo +duodรฉcimos +12.ยบ +~200,0101 +doscientos coma cero uno cero un +doscientos coma cero uno cero uno +doscientas coma cero una cero una +200,0101 +~1.000.200,21 +un millรณn doscientos coma veintiรบn +un millรณn doscientos coma veintiuno +un millรณn doscientas coma veintiuna +un millรณn doscientos coma dos un +un millรณn doscientos coma dos uno +un millรณn doscientas coma dos una +1.000.200,21 +~1/12 +un doceavo +una doceava parte +un duodรฉcimo +una duodรฉcima parte +uno sobre doce +1/12 +~5/200 +cinco ducentรฉsimos +cinco ducentรฉsimas partes +cinco sobre doscientos +5/200 +~1 5/3 +uno y cinco tercios +una y cinco terceras partes +uno y cinco sobre tres +una y cinco sobre tres +~1/5/2020 +primero de mayo de dos mil veinte +uno de mayo de dos mil veinte +cinco de enero de dos mil veinte +~$5,50 +cinco dรณlares con cincuenta +cinco dรณlares y cincuenta +cinco dรณlares cincuenta +cinco dรณlares con cincuenta centavos +cinco dรณlares y cincuenta centavos +cinco dรณlares cincuenta centavos +~2.30 h +dos y treinta +dos y media +tres menos treinta +tres menos media +treinta para las tres +~12.30 a.m. +doce y treinta de la medianoche +doce y treinta de la noche +doce y media de la medianoche +doce y media de la noche +una menos treinta de la maรฑana +una menos media de la maรฑana +treinta para la una de la maรฑana \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_ordinal.txt @@ -0,0 +1,137 @@ +1.แต‰สณ~primer +1.ยบ~primero +1.ยช~primera +2.ยบ~segundo +2.ยช~segunda +ii~segundo +II~segundo +3.แต‰สณ~tercer +3.ยบ~tercero +3.ยช~tercera +4.ยบ~cuarto +4.ยช~cuarta +5.ยบ~quinto +5.ยช~quinta +6.ยบ~sexto +6.ยช~sexta +7.ยบ~sรฉptimo +7.ยช~sรฉptima +8.ยบ~octavo +8.ยช~octava +9.ยบ~noveno +9.ยช~novena +10.ยบ~dรฉcimo +10.ยช~dรฉcima +11.แต‰สณ~decimoprimer +11.ยบ~undรฉcimo +11.ยช~undรฉcima +12.ยบ~duodรฉcimo +12.ยช~duodรฉcima +13.แต‰สณ~decimotercer +13.ยบ~decimotercero +13.ยช~decimotercera +14.ยบ~decimocuarto +14.ยช~decimocuarta +15.ยบ~decimoquinto +15.ยช~decimoquinta +16.ยบ~decimosexto +16.ยช~decimosexta +17.ยบ~decimosรฉptimo +17.ยช~decimosรฉptima +18.ยบ~decimoctavo +18.ยช~decimoctava +19.ยบ~decimonoveno +19.ยช~decimonovena +20.ยบ~vigรฉsimo +20.ยช~vigรฉsima +21.แต‰สณ~vigesimoprimer +21.ยบ~vigesimoprimero +21.ยช~vigesimoprimera +30.ยบ~trigรฉsimo +30.ยช~trigรฉsima +31.แต‰สณ~trigรฉsimo primer +31.ยบ~trigรฉsimo primero +31.ยช~trigรฉsima primera +40.ยบ~cuadragรฉsimo +40.ยช~cuadragรฉsima +41.แต‰สณ~cuadragรฉsimo primer +41.ยบ~cuadragรฉsimo primero +41.ยช~cuadragรฉsima primera +50.ยบ~quincuagรฉsimo +50.ยช~quincuagรฉsima +51.แต‰สณ~quincuagรฉsimo primer +51.ยบ~quincuagรฉsimo primero +51.ยช~quincuagรฉsima primera +60.ยบ~sexagรฉsimo +60.ยช~sexagรฉsima +70.ยบ~septuagรฉsimo +70.ยช~septuagรฉsima +80.ยบ~octogรฉsimo +80.ยช~octogรฉsima +90.ยบ~nonagรฉsimo +90.ยช~nonagรฉsima +100.ยบ~centรฉsimo +100.ยช~centรฉsima +101.แต‰สณ~centรฉsimo primer +101.ยบ~centรฉsimo primero +101.ยช~centรฉsima primera +134.ยบ~centรฉsimo trigรฉsimo cuarto +134.ยช~centรฉsima trigรฉsima cuarta +200.ยบ~ducentรฉsimo +200.ยช~ducentรฉsima +300.ยบ~tricentรฉsimo +300.ยช~tricentรฉsima +400.ยบ~cuadringentรฉsimo +400.ยช~cuadringentรฉsima +500.ยบ~quingentรฉsimo +500.ยช~quingentรฉsima +600.ยบ~sexcentรฉsimo +600.ยช~sexcentรฉsima +700.ยบ~septingentรฉsimo +700.ยช~septingentรฉsima +800.ยบ~octingentรฉsimo +800.ยช~octingentรฉsima +900.ยบ~noningentรฉsimo +900.ยช~noningentรฉsima +1000.ยบ~milรฉsimo +1000.ยช~milรฉsima +1001.แต‰สณ~milรฉsimo primer +1 000.ยบ~milรฉsimo +1 000.ยช~milรฉsima +1 001.แต‰สณ~milรฉsimo primer +1.000.ยบ~milรฉsimo +1.000.ยช~milรฉsima +1.001.แต‰สณ~milรฉsimo primer +1248.ยบ~milรฉsimo ducentรฉsimo cuadragรฉsimo octavo +1248.ยช~milรฉsima ducentรฉsima cuadragรฉsima octava +2000.ยบ~dosmilรฉsimo +100 000.ยบ~cienmilรฉsimo +i~primero +I~primero +ii~segundo +II~segundo +iii~tercero +III~tercero +iv~cuarto +IV~cuarto +V~quinto +VI~sexto +VII~sรฉptimo +VIII~octavo +IX~noveno +X~dรฉcimo +XI~undรฉcimo +XII~duodรฉcimo +XIII~decimotercero +XX~vigรฉsimo +XXI~vigesimoprimero +XXX~trigรฉsimo +XL~cuadragรฉsimo +L~quincuagรฉsimo +XC~nonagรฉsimo +C~centรฉsimo +CD~cuadringentรฉsimo +D~quingentรฉsimo +CM~noningentรฉsimo +999.ยบ~noningentรฉsimo nonagรฉsimo noveno +cmxcix~noningentรฉsimo nonagรฉsimo noveno \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_telephone.txt @@ -0,0 +1,3 @@ +123-123-5678~uno dos tres uno dos tres cinco seis siete ocho +123-456-789~uno dos tres cuatro cinco seis siete ocho nueve +1234-5678~uno dos tres cuatro cinco seis siete ocho \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_time.txt @@ -0,0 +1,26 @@ +1.00~una +1:00~una +01:00~una +01 h~una +3 h~tres horas +1 h~una hora +1.05 h~una y cinco +01.05 h~una y cinco +1.00 h~una +1.00 a.m.~una de la maรฑana +1.00 a.m~una de la maรฑana +1.00 p.m.~una de la tarde +1.00 p.m est~una de la tarde e s t +1.00 est~una e s t +5:02 est~cinco y dos e s t +5:02 p.m pst~cinco y dos de la noche p s t +5:02 p.m.~cinco y dos de la noche +12.15~doce y cuarto +12.15 a.m.~doce y cuarto de la noche +12.15 p.m.~doce y cuarto del mediodรญa +13.30~trece y media +14.05~catorce y cinco +24:50~veinticuatro y cincuenta +3:02:32 pst~tres horas dos minutos y treinta y dos segundos p s t +00:52~cero y cincuenta y dos +0:52~cero y cincuenta y dos diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_whitelist.txt @@ -0,0 +1,3 @@ +el dr.~el doctor +sr. rodriguez~seรฑor rodriguez +182 esq. toledo~ciento ochenta y dos esquina toledo \ No newline at end of file diff --git a/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/data_text_normalization/test_cases_word.txt @@ -0,0 +1,48 @@ +~ +yahoo!~yahoo! +veinte!~veinte! +โ€”~โ€” +aaa~aaa +aabach~aabach +aabenraa~aabenraa +aabye~aabye +aaccessed~aaccessed +aach~aach +aachen's~aachen's +aadri~aadri +aafia~aafia +aagaard~aagaard +aagadu~aagadu +aagard~aagard +aagathadi~aagathadi +aaghart's~aaghart's +aagnes~aagnes +aagomoni~aagomoni +aagon~aagon +aagoo~aagoo +aagot~aagot +aahar~aahar +aahh~aahh +aahperd~aahperd +aaibinterstate~aaibinterstate +aajab~aajab +aakasa~aakasa +aakervik~aakervik +aakirkeby~aakirkeby +aalam~aalam +aalbaek~aalbaek +aaldiu~aaldiu +aalem~aalem +a'ali~a'ali +aalilaassamthey~aalilaassamthey +aalin~aalin +aaliyan~aaliyan +aaliyan's~aaliyan's +aamadu~aamadu +aamara~aamara +aambala~aambala +aamera~aamera +aamer's~aamer's +aamina~aamina +aaminah~aaminah +aamjiwnaang~aamjiwnaang diff --git a/tests/nemo_text_processing/es/test_cardinal.py b/tests/nemo_text_processing/es/test_cardinal.py --- a/tests/nemo_text_processing/es/test_cardinal.py +++ b/tests/nemo_text_processing/es/test_cardinal.py @@ -22,7 +22,8 @@ class TestCardinal: - inverse_normalizer_es = ( + + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +33,34 @@ class TestCardinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_cardinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_date.py b/tests/nemo_text_processing/es/test_date.py --- a/tests/nemo_text_processing/es/test_date.py +++ b/tests/nemo_text_processing/es/test_date.py @@ -22,7 +22,7 @@ class TestDate: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDate: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_date.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_decimal.py b/tests/nemo_text_processing/es/test_decimal.py --- a/tests/nemo_text_processing/es/test_decimal.py +++ b/tests/nemo_text_processing/es/test_decimal.py @@ -22,7 +22,7 @@ class TestDecimal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -32,6 +32,34 @@ class TestDecimal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_decimal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_electronic.py b/tests/nemo_text_processing/es/test_electronic.py --- a/tests/nemo_text_processing/es/test_electronic.py +++ b/tests/nemo_text_processing/es/test_electronic.py @@ -35,3 +35,31 @@ class TestElectronic: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_electronic.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_fraction.py b/tests/nemo_text_processing/es/test_fraction.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_fraction.py @@ -0,0 +1,51 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import pytest +from nemo_text_processing.text_normalization.normalize import Normalizer +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, parse_test_case_file + + +class TestFraction: + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_fraction.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_measure.py b/tests/nemo_text_processing/es/test_measure.py --- a/tests/nemo_text_processing/es/test_measure.py +++ b/tests/nemo_text_processing/es/test_measure.py @@ -36,3 +36,31 @@ class TestMeasure: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_measure.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_money.py b/tests/nemo_text_processing/es/test_money.py --- a/tests/nemo_text_processing/es/test_money.py +++ b/tests/nemo_text_processing/es/test_money.py @@ -23,7 +23,7 @@ class TestMoney: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,34 @@ class TestMoney: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_money.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_normalization_with_audio.py b/tests/nemo_text_processing/es/test_normalization_with_audio.py new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_normalization_with_audio.py @@ -0,0 +1,43 @@ +# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import pytest +from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio +from parameterized import parameterized + +from ..utils import CACHE_DIR, PYNINI_AVAILABLE, get_test_cases_multiple + + +class TestNormalizeWithAudio: + + normalizer_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + @parameterized.expand(get_test_cases_multiple('es/data_text_normalization/test_cases_normalize_with_audio.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, n_tagged=1000, punct_post_process=False) + print(expected) + print("pred") + print(pred) + assert len(set(pred).intersection(set(expected))) == len( + expected + ), f'missing: {set(expected).difference(set(pred))}' diff --git a/tests/nemo_text_processing/es/test_ordinal.py b/tests/nemo_text_processing/es/test_ordinal.py --- a/tests/nemo_text_processing/es/test_ordinal.py +++ b/tests/nemo_text_processing/es/test_ordinal.py @@ -23,7 +23,7 @@ class TestOrdinal: - inverse_normalizer_es = ( + inverse_normalizer = ( InverseNormalizer(lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) if PYNINI_AVAILABLE else None ) @@ -33,6 +33,33 @@ class TestOrdinal: ) @pytest.mark.run_only_on('CPU') @pytest.mark.unit - def test_denorm_es(self, test_input, expected): - pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) + def test_denorm(self, test_input, expected): + pred = self.inverse_normalizer.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_ordinal.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=30, punct_post_process=False, + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh new file mode 100644 --- /dev/null +++ b/tests/nemo_text_processing/es/test_sparrowhawk_normalization.sh @@ -0,0 +1,84 @@ +#! /bin/sh + +PROJECT_DIR=/workspace/tests + +runtest () { + input=$1 + cd /workspace/sparrowhawk/documentation/grammars + + # read test file + while read testcase; do + IFS='~' read written spoken <<< $testcase + denorm_pred=$(echo $written | normalizer_main --config=sparrowhawk_configuration.ascii_proto 2>&1 | tail -n 1) + + # trim white space + spoken="$(echo -e "${spoken}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + denorm_pred="$(echo -e "${denorm_pred}" | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')" + + # input expected actual + assertEquals "$written" "$spoken" "$denorm_pred" + done < "$input" +} + +testTNCardinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_cardinal.txt + runtest $input +} + +testTNDate() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_date.txt + runtest $input +} + +testTNDecimal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_decimal.txt + runtest $input +} + +testTNElectronic() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_electronic.txt + runtest $input +} + +testTNFraction() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_fraction.txt + runtest $input +} + +testTNMoney() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_money.txt + runtest $input +} + +testTNOrdinal() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTelephone() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_ordinal.txt + runtest $input +} + +testTNTime() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_time.txt + runtest $input +} + +testTNMeasure() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_measure.txt + runtest $input +} + +testTNWhitelist() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_whitelist.txt + runtest $input +} + +testTNWord() { + input=$PROJECT_DIR/es/data_text_normalization/test_cases_word.txt + runtest $input +} + +# Load shUnit2 +. $PROJECT_DIR/../shunit2/shunit2 diff --git a/tests/nemo_text_processing/es/test_telephone.py b/tests/nemo_text_processing/es/test_telephone.py --- a/tests/nemo_text_processing/es/test_telephone.py +++ b/tests/nemo_text_processing/es/test_telephone.py @@ -36,3 +36,31 @@ class TestTelephone: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_telephone.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_time.py b/tests/nemo_text_processing/es/test_time.py --- a/tests/nemo_text_processing/es/test_time.py +++ b/tests/nemo_text_processing/es/test_time.py @@ -35,3 +35,31 @@ class TestTime: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_time.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=1000, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_whitelist.py b/tests/nemo_text_processing/es/test_whitelist.py --- a/tests/nemo_text_processing/es/test_whitelist.py +++ b/tests/nemo_text_processing/es/test_whitelist.py @@ -35,3 +35,30 @@ class TestWhitelist: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_whitelist.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer.normalize(test_input, verbose=False) + assert pred == expected + + if self.normalizer_with_audio: + pred_non_deterministic = self.normalizer_with_audio.normalize( + test_input, n_tagged=10, punct_post_process=False + ) + assert expected in pred_non_deterministic diff --git a/tests/nemo_text_processing/es/test_word.py b/tests/nemo_text_processing/es/test_word.py --- a/tests/nemo_text_processing/es/test_word.py +++ b/tests/nemo_text_processing/es/test_word.py @@ -35,3 +35,30 @@ class TestWord: def test_denorm_es(self, test_input, expected): pred = self.inverse_normalizer_es.inverse_normalize(test_input, verbose=False) assert pred == expected + + normalizer_es = ( + Normalizer(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE + else None + ) + normalizer_with_audio_es = ( + NormalizerWithAudio(input_case='cased', lang='es', cache_dir=CACHE_DIR, overwrite_cache=False) + if PYNINI_AVAILABLE and CACHE_DIR + else None + ) + + @parameterized.expand(parse_test_case_file('es/data_text_normalization/test_cases_word.txt')) + @pytest.mark.skipif( + not PYNINI_AVAILABLE, reason="`pynini` not installed, please install via nemo_text_processing/setup.sh" + ) + @pytest.mark.run_only_on('CPU') + @pytest.mark.unit + def test_norm(self, test_input, expected): + pred = self.normalizer_es.normalize(test_input, verbose=False) + assert pred == expected, f"input: {test_input}" + + if self.normalizer_with_audio_es: + pred_non_deterministic = self.normalizer_with_audio_es.normalize( + test_input, n_tagged=150, punct_post_process=False + ) + assert expected in pred_non_deterministic, f"input: {test_input}"
1.0
NVIDIA__NeMo-7582
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] 1 2 |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg 17 :target: https://badge.fury.io/py/nemo-toolkit 18 :alt: Release version 19 20 .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg 21 :target: https://badge.fury.io/py/nemo-toolkit 22 :alt: Python version 23 24 .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads 25 :target: https://pepy.tech/project/nemo-toolkit 26 :alt: PyPi total downloads 27 28 .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push 29 :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml 30 :alt: CodeQL 31 32 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 33 :target: https://github.com/psf/black 34 :alt: Code style: black 35 36 .. _main-readme: 37 38 **NVIDIA NeMo** 39 =============== 40 41 Introduction 42 ------------ 43 44 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), 45 text-to-speech synthesis (TTS), large language models (LLMs), and 46 natural language processing (NLP). 47 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) 48 and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 49 50 All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and 51 training is automatically scalable to 1000s of GPUs. 52 Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. 53 NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. 54 55 Getting started with NeMo is simple. 56 State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and 57 `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. 58 These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. 59 60 We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that 61 can all be run on `Google Colab <https://colab.research.google.com>`_. 62 63 For advanced users that want to train NeMo models from scratch or finetune existing NeMo models 64 we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. 65 66 For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. 67 The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ 68 which can be used to find the optimal model parallel configuration for training on a specific cluster. 69 70 Also see our `introductory video <https://www.youtube.com/embed/wBgpMf_KQVw>`_ for a high level overview of NeMo. 71 72 Key Features 73 ------------ 74 75 * Speech processing 76 * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ 77 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 78 * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ 79 * Jasper, QuartzNet, CitriNet, ContextNet 80 * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer 81 * Squeezeformer-CTC and Squeezeformer-Transducer 82 * LSTM-Transducer (RNNT) and LSTM-CTC 83 * Supports the following decoders/losses: 84 * CTC 85 * Transducer/RNNT 86 * Hybrid Transducer/CTC 87 * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ 88 * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ 89 * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ 90 * Beam Search decoding 91 * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 92 * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ 93 * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) 94 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 95 * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ 96 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 97 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ 98 * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet 99 * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) 100 * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer 101 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 102 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 103 * Natural Language Processing 104 * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ 105 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ 106 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 107 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 108 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 109 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 110 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 111 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 112 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 113 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 114 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 115 * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ 116 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 117 * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ 118 * Text-to-Speech Synthesis (TTS): 119 * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 120 * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 121 * Vocoders: HiFiGAN, UnivNet, WaveGlow 122 * End-to-End Models: VITS 123 * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 124 * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ 125 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ 126 * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ 127 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 128 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 129 * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ 130 131 132 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 133 134 Requirements 135 ------------ 136 137 1) Python 3.10 or above 138 2) Pytorch 1.13.1 or above 139 3) NVIDIA GPU for training 140 141 Documentation 142 ------------- 143 144 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 145 :alt: Documentation Status 146 :scale: 100% 147 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 148 149 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 150 :alt: Documentation Status 151 :scale: 100% 152 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 153 154 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 155 | Version | Status | Description | 156 +=========+=============+==========================================================================================================================================+ 157 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 158 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 159 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 160 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 161 162 Tutorials 163 --------- 164 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 165 166 Getting help with NeMo 167 ---------------------- 168 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 169 170 171 Installation 172 ------------ 173 174 Conda 175 ~~~~~ 176 177 We recommend installing NeMo in a fresh Conda environment. 178 179 .. code-block:: bash 180 181 conda create --name nemo python==3.10.12 182 conda activate nemo 183 184 Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. 185 186 .. code-block:: bash 187 188 conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia 189 190 The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. 191 192 Pip 193 ~~~ 194 Use this installation mode if you want the latest released version. 195 196 .. code-block:: bash 197 198 apt-get update && apt-get install -y libsndfile1 ffmpeg 199 pip install Cython 200 pip install nemo_toolkit['all'] 201 202 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 203 204 Pip from source 205 ~~~~~~~~~~~~~~~ 206 Use this installation mode if you want the version from a particular GitHub branch (e.g main). 207 208 .. code-block:: bash 209 210 apt-get update && apt-get install -y libsndfile1 ffmpeg 211 pip install Cython 212 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 213 214 215 From source 216 ~~~~~~~~~~~ 217 Use this installation mode if you are contributing to NeMo. 218 219 .. code-block:: bash 220 221 apt-get update && apt-get install -y libsndfile1 ffmpeg 222 git clone https://github.com/NVIDIA/NeMo 223 cd NeMo 224 ./reinstall.sh 225 226 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 227 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 228 229 RNNT 230 ~~~~ 231 Note that RNNT requires numba to be installed from conda. 232 233 .. code-block:: bash 234 235 conda remove numba 236 pip uninstall numba 237 conda install -c conda-forge numba 238 239 NeMo Megatron 240 ~~~~~~~~~~~~~ 241 NeMo Megatron training requires NVIDIA Apex to be installed. 242 Install it manually if not using the NVIDIA PyTorch container. 243 244 To install Apex, run 245 246 .. code-block:: bash 247 248 git clone https://github.com/NVIDIA/apex.git 249 cd apex 250 git checkout 52e18c894223800cb611682dce27d88050edf1de 251 pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ 252 253 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. 254 255 While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. 256 This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 257 258 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: 259 260 .. code-block:: bash 261 262 conda install -c nvidia cuda-nvprof=11.8 263 264 packaging is also needed: 265 266 .. code-block:: bash 267 268 pip install packaging 269 270 With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. 271 272 273 Transformer Engine 274 ~~~~~~~~~~~~~~~~~~ 275 NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ 276 Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. 277 `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. 278 279 .. code-block:: bash 280 281 pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable 282 283 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. 284 285 Transformer Engine requires PyTorch to be built with CUDA 11.8. 286 287 288 Flash Attention 289 ~~~~~~~~~~~~~~~~~~~~ 290 Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. 291 292 .. code-block:: bash 293 294 pip install flash-attn 295 pip install triton==2.0.0.dev20221202 296 297 NLP inference UI 298 ~~~~~~~~~~~~~~~~~~~~ 299 To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. 300 301 .. code-block:: bash 302 303 pip install gradio==3.34.0 304 305 NeMo Text Processing 306 ~~~~~~~~~~~~~~~~~~~~ 307 NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. 308 309 Docker containers: 310 ~~~~~~~~~~~~~~~~~~ 311 We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. 312 313 To use built container, please run 314 315 .. code-block:: bash 316 317 docker pull nvcr.io/nvidia/nemo:23.06 318 319 To build a nemo container with Dockerfile from a branch, please run 320 321 .. code-block:: bash 322 323 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 324 325 326 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. 327 328 .. code-block:: bash 329 330 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 331 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 332 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 333 334 Examples 335 -------- 336 337 Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 338 339 340 Contributing 341 ------------ 342 343 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 344 345 Publications 346 ------------ 347 348 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 349 350 License 351 ------- 352 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 353 [end of README.rst] [start of examples/asr/experimental/k2/align_speech_parallel.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 # Based on examples/asr/transcribe_speech_parallel.py 17 # ASR alignment with multi-GPU/multi-node support for large datasets 18 # It supports both tarred and non-tarred datasets 19 # Arguments 20 # model: path to a nemo/PTL checkpoint file or name of a pretrained model 21 # predict_ds: config of the dataset/dataloader 22 # aligner_args: aligner config 23 # output_path: path to store the predictions 24 # model_stride: model downsampling factor, 8 for Citrinet models and 4 for Conformer models 25 # 26 # Results of each GPU/worker is written into a file named 'predictions_{rank}.json, and aggregated results of all workers are written into 'predictions_all.json' 27 28 Example for non-tarred datasets: 29 30 python align_speech_parallel.py \ 31 model=stt_en_conformer_ctc_large \ 32 predict_ds.manifest_filepath=/dataset/manifest_file.json \ 33 predict_ds.batch_size=16 \ 34 output_path=/tmp/ 35 36 Example for tarred datasets: 37 38 python align_speech_parallel.py \ 39 predict_ds.is_tarred=true \ 40 predict_ds.manifest_filepath=/tarred_dataset/tarred_audio_manifest.json \ 41 predict_ds.tarred_audio_filepaths=/tarred_dataset/audio__OP_0..127_CL_.tar \ 42 ... 43 44 By default the trainer uses all the GPUs available and default precision is FP32. 45 By setting the trainer config you may control these configs. For example to do the predictions with AMP on just two GPUs: 46 47 python align_speech_parallel.py \ 48 trainer.precision=16 \ 49 trainer.gpus=2 \ 50 ... 51 52 You may control the dataloader's config by setting the predict_ds: 53 54 python align_speech_parallel.py \ 55 predict_ds.num_workers=8 \ 56 predict_ds.min_duration=2.0 \ 57 predict_ds.sample_rate=16000 \ 58 model=stt_en_conformer_ctc_small \ 59 ... 60 61 You may control the aligner's config by setting the aligner_args: 62 aligner_args.alignment_type=argmax \ 63 aligner_args.word_output=False \ 64 aligner_args.cpu_decoding=True \ 65 aligner_args.decode_batch_size=8 \ 66 aligner_args.ctc_cfg.prob_suppress_index=-1 \ 67 aligner_args.ctc_cfg.prob_suppress_value=0.5 \ 68 aligner_args.rnnt_cfg.predictor_window_size=10 \ 69 aligner_args.decoder_module_cfg.intersect_pruned=true \ 70 aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ 71 ... 72 73 """ 74 75 76 import os 77 from dataclasses import dataclass, is_dataclass 78 from typing import Optional 79 80 import pytorch_lightning as ptl 81 import torch 82 from omegaconf import MISSING, OmegaConf 83 84 from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter 85 from nemo.collections.asr.models import ASRModel 86 from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig 87 from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig 88 from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel 89 from nemo.core.config import TrainerConfig, hydra_runner 90 from nemo.utils import logging 91 from nemo.utils.get_rank import is_global_rank_zero 92 93 94 @dataclass 95 class ParallelAlignmentConfig: 96 model: Optional[str] = None # name 97 predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) 98 aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() 99 output_path: str = MISSING 100 model_stride: int = 8 101 102 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") 103 104 # there arguments will be ignored 105 return_predictions: bool = False 106 use_cer: bool = False 107 108 109 def match_train_config(predict_ds, train_ds): 110 # It copies the important configurations from the train dataset of the model 111 # into the predict_ds to be used for prediction. It is needed to match the training configurations. 112 if train_ds is None: 113 return 114 115 predict_ds.sample_rate = train_ds.get("sample_rate", 16000) 116 cfg_name_list = [ 117 "int_values", 118 "use_start_end_token", 119 "blank_index", 120 "unk_index", 121 "normalize", 122 "parser", 123 "eos_id", 124 "bos_id", 125 "pad_id", 126 ] 127 128 if is_dataclass(predict_ds): 129 predict_ds = OmegaConf.structured(predict_ds) 130 for cfg_name in cfg_name_list: 131 if hasattr(train_ds, cfg_name): 132 setattr(predict_ds, cfg_name, getattr(train_ds, cfg_name)) 133 134 return predict_ds 135 136 137 @hydra_runner(config_name="AlignmentConfig", schema=ParallelAlignmentConfig) 138 def main(cfg: ParallelAlignmentConfig): 139 if cfg.model.endswith(".nemo"): 140 logging.info("Attempting to initialize from .nemo file") 141 model = ASRModel.restore_from(restore_path=cfg.model, map_location="cpu") 142 elif cfg.model.endswith(".ckpt"): 143 logging.info("Attempting to initialize from .ckpt file") 144 model = ASRModel.load_from_checkpoint(checkpoint_path=cfg.model, map_location="cpu") 145 else: 146 logging.info( 147 "Attempting to initialize from a pretrained model as the model name does not have the extension of .nemo or .ckpt" 148 ) 149 model = ASRModel.from_pretrained(model_name=cfg.model, map_location="cpu") 150 151 trainer = ptl.Trainer(**cfg.trainer) 152 153 cfg.predict_ds.return_sample_id = True 154 cfg.return_predictions = False 155 cfg.use_cer = False 156 cfg.predict_ds = match_train_config(predict_ds=cfg.predict_ds, train_ds=model._cfg.train_ds) 157 data_loader = model._setup_dataloader_from_config(cfg.predict_ds) 158 159 os.makedirs(cfg.output_path, exist_ok=True) 160 # trainer.global_rank is not valid before predict() is called. Need this hack to find the correct global_rank. 161 global_rank = trainer.node_rank * trainer.num_devices + int(os.environ.get("LOCAL_RANK", 0)) 162 output_file = os.path.join(cfg.output_path, f"predictions_{global_rank}.json") 163 output_ctm_dir = os.path.join(cfg.output_path, "ctm") 164 predictor_writer = ASRCTMPredictionWriter( 165 dataset=data_loader.dataset, 166 output_file=output_file, 167 output_ctm_dir=output_ctm_dir, 168 time_per_frame=cfg.model_stride * model._cfg.preprocessor['window_stride'], 169 ) 170 trainer.callbacks.extend([predictor_writer]) 171 172 aligner_wrapper = AlignerWrapperModel(model=model, cfg=cfg.aligner_args) 173 trainer.predict(model=aligner_wrapper, dataloaders=data_loader, return_predictions=cfg.return_predictions) 174 samples_num = predictor_writer.close_output_file() 175 176 logging.info( 177 f"Prediction on rank {global_rank} is done for {samples_num} samples and results are stored in {output_file}." 178 ) 179 180 if torch.distributed.is_initialized(): 181 torch.distributed.barrier() 182 183 samples_num = 0 184 if is_global_rank_zero(): 185 output_file = os.path.join(cfg.output_path, f"predictions_all.json") 186 logging.info(f"Prediction files are being aggregated in {output_file}.") 187 with open(output_file, 'tw', encoding="utf-8") as outf: 188 for rank in range(trainer.world_size): 189 input_file = os.path.join(cfg.output_path, f"predictions_{rank}.json") 190 with open(input_file, 'r', encoding="utf-8") as inpf: 191 lines = inpf.readlines() 192 samples_num += len(lines) 193 outf.writelines(lines) 194 logging.info( 195 f"Prediction is done for {samples_num} samples in total on all workers and results are aggregated in {output_file}." 196 ) 197 198 199 if __name__ == '__main__': 200 main() 201 [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import re 17 from abc import abstractmethod 18 from dataclasses import dataclass, is_dataclass 19 from typing import Callable, Dict, List, Optional, Tuple, Union 20 21 import editdistance 22 import numpy as np 23 import torch 24 from omegaconf import OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.metrics.wer import move_dimension_to_the_front 28 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding as beam_decode 29 from nemo.collections.asr.parts.submodules import rnnt_greedy_decoding as greedy_decode 30 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 31 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 32 from nemo.utils import logging 33 34 __all__ = ['RNNTDecoding', 'RNNTWER'] 35 36 37 class AbstractRNNTDecoding(ConfidenceMixin): 38 """ 39 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 40 41 Args: 42 decoding_cfg: A dict-like object which contains the following key-value pairs. 43 strategy: str value which represents the type of decoding that can occur. 44 Possible values are : 45 - greedy, greedy_batch (for greedy decoding). 46 - beam, tsd, alsd (for beam search decoding). 47 48 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 49 tokens as well as the decoded string. Default is False in order to avoid double decoding 50 unless required. 51 52 preserve_alignments: Bool flag which preserves the history of logprobs generated during 53 decoding (sample / batched). When set to true, the Hypothesis will contain 54 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 55 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 56 57 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 58 with the `return_hypotheses` flag set to True. 59 60 The length of the list corresponds to the Acoustic Length (T). 61 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 62 U is the number of target tokens for the current timestep Ti. 63 64 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 65 word based timestamp mapping the output log-probabilities to discrete intervals of timestamps. 66 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 67 68 rnnt_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 69 Can take the following values - "char" for character/subword time stamps, "word" for word level 70 time stamps and "all" (default), for both character level and word level time stamps. 71 72 word_seperator: Str token representing the seperator between words. 73 74 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 75 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 76 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of ints. 77 78 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 79 scores. In order to obtain hypotheses with confidence scores, please utilize 80 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 81 82 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 83 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 84 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 85 86 The length of the list corresponds to the Acoustic Length (T). 87 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 88 U is the number of target tokens for the current timestep Ti. 89 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 90 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 91 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 92 93 The length of the list corresponds to the number of recognized tokens. 94 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 95 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 96 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 97 98 The length of the list corresponds to the number of recognized words. 99 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 100 from the `token_confidence`. 101 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 102 Valid options are `mean`, `min`, `max`, `prod`. 103 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 104 confidence scores. 105 106 name: The measure name (str). 107 Supported values: 108 - 'max_prob' for using the maximum token probability as a confidence. 109 - 'entropy' for using a normalized entropy of a log-likelihood vector. 110 111 entropy_type: Which type of entropy to use (str). 112 Used if confidence_measure_cfg.name is set to `entropy`. 113 Supported values: 114 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 115 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 116 Note that for this entropy, the alpha should comply the following inequality: 117 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 118 where V is the model vocabulary size. 119 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 120 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 121 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 122 More: https://en.wikipedia.org/wiki/Tsallis_entropy 123 - 'renyi' for the Rรฉnyi entropy. 124 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 125 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 126 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 127 128 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 129 When the alpha equals one, scaling is not applied to 'max_prob', 130 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 131 132 entropy_norm: A mapping of the entropy value to the interval [0,1]. 133 Supported values: 134 - 'lin' for using the linear mapping. 135 - 'exp' for using exponential mapping with linear shift. 136 137 The config may further contain the following sub-dictionaries: 138 "greedy": 139 max_symbols: int, describing the maximum number of target tokens to decode per 140 timestep during greedy decoding. Setting to larger values allows longer sentences 141 to be decoded, at the cost of increased execution time. 142 preserve_frame_confidence: Same as above, overrides above value. 143 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 144 145 "beam": 146 beam_size: int, defining the beam size for beam search. Must be >= 1. 147 If beam_size == 1, will perform cached greedy search. This might be slightly different 148 results compared to the greedy search above. 149 150 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 151 Set to True by default. 152 153 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 154 hypotheses after beam search has concluded. This flag is set by default. 155 156 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 157 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 158 at increased cost to execution time. 159 160 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 161 If an integer is provided, it can decode sequences of that particular maximum length. 162 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 163 where seq_len is the length of the acoustic model output (T). 164 165 NOTE: 166 If a float is provided, it can be greater than 1! 167 By default, a float of 2.0 is used so that a target sequence can be at most twice 168 as long as the acoustic model output length T. 169 170 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 171 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 172 173 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 174 in order to reduce expensive beam search cost later. int >= 0. 175 176 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 177 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 178 and affects the speed of inference since large values will perform large beam search in the next step. 179 180 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 181 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 182 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 183 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 184 expansion apart from the "most likely" candidate. 185 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 186 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 187 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 188 tuned on a validation set. 189 190 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 191 192 decoder: The Decoder/Prediction network module. 193 joint: The Joint network module. 194 blank_id: The id of the RNNT blank token. 195 """ 196 197 def __init__(self, decoding_cfg, decoder, joint, blank_id: int): 198 super(AbstractRNNTDecoding, self).__init__() 199 200 # Convert dataclass to config object 201 if is_dataclass(decoding_cfg): 202 decoding_cfg = OmegaConf.structured(decoding_cfg) 203 204 self.cfg = decoding_cfg 205 self.blank_id = blank_id 206 self.num_extra_outputs = joint.num_extra_outputs 207 self.big_blank_durations = self.cfg.get("big_blank_durations", None) 208 self.durations = self.cfg.get("durations", None) 209 self.compute_hypothesis_token_set = self.cfg.get("compute_hypothesis_token_set", False) 210 self.compute_langs = decoding_cfg.get('compute_langs', False) 211 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 212 self.joint_fused_batch_size = self.cfg.get('fused_batch_size', None) 213 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 214 self.word_seperator = self.cfg.get('word_seperator', ' ') 215 216 if self.durations is not None: # this means it's a TDT model. 217 if blank_id == 0: 218 raise ValueError("blank_id must equal len(non_blank_vocabs) for TDT models") 219 if self.big_blank_durations is not None: 220 raise ValueError("duration and big_blank_durations can't both be not None") 221 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 222 raise ValueError("currently only greedy and greedy_batch inference is supported for TDT models") 223 224 if self.big_blank_durations is not None: # this means it's a multi-blank model. 225 if blank_id == 0: 226 raise ValueError("blank_id must equal len(vocabs) for multi-blank RNN-T models") 227 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 228 raise ValueError( 229 "currently only greedy and greedy_batch inference is supported for multi-blank models" 230 ) 231 232 possible_strategies = ['greedy', 'greedy_batch', 'beam', 'tsd', 'alsd', 'maes'] 233 if self.cfg.strategy not in possible_strategies: 234 raise ValueError(f"Decoding strategy must be one of {possible_strategies}") 235 236 # Update preserve alignments 237 if self.preserve_alignments is None: 238 if self.cfg.strategy in ['greedy', 'greedy_batch']: 239 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 240 241 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 242 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 243 244 # Update compute timestamps 245 if self.compute_timestamps is None: 246 if self.cfg.strategy in ['greedy', 'greedy_batch']: 247 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 248 249 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 250 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 251 252 # Test if alignments are being preserved for RNNT 253 if self.compute_timestamps is True and self.preserve_alignments is False: 254 raise ValueError("If `compute_timesteps` flag is set, then `preserve_alignments` flag must also be set.") 255 256 # initialize confidence-related fields 257 self._init_confidence(self.cfg.get('confidence_cfg', None)) 258 259 # Confidence estimation is not implemented for these strategies 260 if ( 261 not self.preserve_frame_confidence 262 and self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes'] 263 and self.cfg.beam.get('preserve_frame_confidence', False) 264 ): 265 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 266 267 if self.cfg.strategy == 'greedy': 268 if self.big_blank_durations is None: 269 if self.durations is None: 270 self.decoding = greedy_decode.GreedyRNNTInfer( 271 decoder_model=decoder, 272 joint_model=joint, 273 blank_index=self.blank_id, 274 max_symbols_per_step=( 275 self.cfg.greedy.get('max_symbols', None) 276 or self.cfg.greedy.get('max_symbols_per_step', None) 277 ), 278 preserve_alignments=self.preserve_alignments, 279 preserve_frame_confidence=self.preserve_frame_confidence, 280 confidence_measure_cfg=self.confidence_measure_cfg, 281 ) 282 else: 283 self.decoding = greedy_decode.GreedyTDTInfer( 284 decoder_model=decoder, 285 joint_model=joint, 286 blank_index=self.blank_id, 287 durations=self.durations, 288 max_symbols_per_step=( 289 self.cfg.greedy.get('max_symbols', None) 290 or self.cfg.greedy.get('max_symbols_per_step', None) 291 ), 292 preserve_alignments=self.preserve_alignments, 293 preserve_frame_confidence=self.preserve_frame_confidence, 294 confidence_measure_cfg=self.confidence_measure_cfg, 295 ) 296 else: 297 self.decoding = greedy_decode.GreedyMultiblankRNNTInfer( 298 decoder_model=decoder, 299 joint_model=joint, 300 blank_index=self.blank_id, 301 big_blank_durations=self.big_blank_durations, 302 max_symbols_per_step=( 303 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 304 ), 305 preserve_alignments=self.preserve_alignments, 306 preserve_frame_confidence=self.preserve_frame_confidence, 307 confidence_measure_cfg=self.confidence_measure_cfg, 308 ) 309 310 elif self.cfg.strategy == 'greedy_batch': 311 if self.big_blank_durations is None: 312 if self.durations is None: 313 self.decoding = greedy_decode.GreedyBatchedRNNTInfer( 314 decoder_model=decoder, 315 joint_model=joint, 316 blank_index=self.blank_id, 317 max_symbols_per_step=( 318 self.cfg.greedy.get('max_symbols', None) 319 or self.cfg.greedy.get('max_symbols_per_step', None) 320 ), 321 preserve_alignments=self.preserve_alignments, 322 preserve_frame_confidence=self.preserve_frame_confidence, 323 confidence_measure_cfg=self.confidence_measure_cfg, 324 ) 325 else: 326 self.decoding = greedy_decode.GreedyBatchedTDTInfer( 327 decoder_model=decoder, 328 joint_model=joint, 329 blank_index=self.blank_id, 330 durations=self.durations, 331 max_symbols_per_step=( 332 self.cfg.greedy.get('max_symbols', None) 333 or self.cfg.greedy.get('max_symbols_per_step', None) 334 ), 335 preserve_alignments=self.preserve_alignments, 336 preserve_frame_confidence=self.preserve_frame_confidence, 337 confidence_measure_cfg=self.confidence_measure_cfg, 338 ) 339 340 else: 341 self.decoding = greedy_decode.GreedyBatchedMultiblankRNNTInfer( 342 decoder_model=decoder, 343 joint_model=joint, 344 blank_index=self.blank_id, 345 big_blank_durations=self.big_blank_durations, 346 max_symbols_per_step=( 347 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 348 ), 349 preserve_alignments=self.preserve_alignments, 350 preserve_frame_confidence=self.preserve_frame_confidence, 351 confidence_measure_cfg=self.confidence_measure_cfg, 352 ) 353 354 elif self.cfg.strategy == 'beam': 355 356 self.decoding = beam_decode.BeamRNNTInfer( 357 decoder_model=decoder, 358 joint_model=joint, 359 beam_size=self.cfg.beam.beam_size, 360 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 361 search_type='default', 362 score_norm=self.cfg.beam.get('score_norm', True), 363 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 364 preserve_alignments=self.preserve_alignments, 365 ) 366 367 elif self.cfg.strategy == 'tsd': 368 369 self.decoding = beam_decode.BeamRNNTInfer( 370 decoder_model=decoder, 371 joint_model=joint, 372 beam_size=self.cfg.beam.beam_size, 373 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 374 search_type='tsd', 375 score_norm=self.cfg.beam.get('score_norm', True), 376 tsd_max_sym_exp_per_step=self.cfg.beam.get('tsd_max_sym_exp', 10), 377 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 378 preserve_alignments=self.preserve_alignments, 379 ) 380 381 elif self.cfg.strategy == 'alsd': 382 383 self.decoding = beam_decode.BeamRNNTInfer( 384 decoder_model=decoder, 385 joint_model=joint, 386 beam_size=self.cfg.beam.beam_size, 387 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 388 search_type='alsd', 389 score_norm=self.cfg.beam.get('score_norm', True), 390 alsd_max_target_len=self.cfg.beam.get('alsd_max_target_len', 2), 391 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 392 preserve_alignments=self.preserve_alignments, 393 ) 394 395 elif self.cfg.strategy == 'maes': 396 397 self.decoding = beam_decode.BeamRNNTInfer( 398 decoder_model=decoder, 399 joint_model=joint, 400 beam_size=self.cfg.beam.beam_size, 401 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 402 search_type='maes', 403 score_norm=self.cfg.beam.get('score_norm', True), 404 maes_num_steps=self.cfg.beam.get('maes_num_steps', 2), 405 maes_prefix_alpha=self.cfg.beam.get('maes_prefix_alpha', 1), 406 maes_expansion_gamma=self.cfg.beam.get('maes_expansion_gamma', 2.3), 407 maes_expansion_beta=self.cfg.beam.get('maes_expansion_beta', 2.0), 408 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 409 preserve_alignments=self.preserve_alignments, 410 ngram_lm_model=self.cfg.beam.get('ngram_lm_model', None), 411 ngram_lm_alpha=self.cfg.beam.get('ngram_lm_alpha', 0.0), 412 hat_subtract_ilm=self.cfg.beam.get('hat_subtract_ilm', False), 413 hat_ilm_weight=self.cfg.beam.get('hat_ilm_weight', 0.0), 414 ) 415 416 else: 417 418 raise ValueError( 419 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 420 f"but was provided {self.cfg.strategy}" 421 ) 422 423 # Update the joint fused batch size or disable it entirely if needed. 424 self.update_joint_fused_batch_size() 425 426 def rnnt_decoder_predictions_tensor( 427 self, 428 encoder_output: torch.Tensor, 429 encoded_lengths: torch.Tensor, 430 return_hypotheses: bool = False, 431 partial_hypotheses: Optional[List[Hypothesis]] = None, 432 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 433 """ 434 Decode an encoder output by autoregressive decoding of the Decoder+Joint networks. 435 436 Args: 437 encoder_output: torch.Tensor of shape [B, D, T]. 438 encoded_lengths: torch.Tensor containing lengths of the padded encoder outputs. Shape [B]. 439 return_hypotheses: bool. If set to True it will return list of Hypothesis or NBestHypotheses 440 441 Returns: 442 If `return_best_hypothesis` is set: 443 A tuple (hypotheses, None): 444 hypotheses - list of Hypothesis (best hypothesis per sample). 445 Look at rnnt_utils.Hypothesis for more information. 446 447 If `return_best_hypothesis` is not set: 448 A tuple(hypotheses, all_hypotheses) 449 hypotheses - list of Hypothesis (best hypothesis per sample). 450 Look at rnnt_utils.Hypothesis for more information. 451 all_hypotheses - list of NBestHypotheses. Each NBestHypotheses further contains a sorted 452 list of all the hypotheses of the model per sample. 453 Look at rnnt_utils.NBestHypotheses for more information. 454 """ 455 # Compute hypotheses 456 with torch.inference_mode(): 457 hypotheses_list = self.decoding( 458 encoder_output=encoder_output, encoded_lengths=encoded_lengths, partial_hypotheses=partial_hypotheses 459 ) # type: [List[Hypothesis]] 460 461 # extract the hypotheses 462 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 463 464 prediction_list = hypotheses_list 465 466 if isinstance(prediction_list[0], NBestHypotheses): 467 hypotheses = [] 468 all_hypotheses = [] 469 470 for nbest_hyp in prediction_list: # type: NBestHypotheses 471 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 472 decoded_hyps = self.decode_hypothesis(n_hyps) # type: List[str] 473 474 # If computing timestamps 475 if self.compute_timestamps is True: 476 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 477 for hyp_idx in range(len(decoded_hyps)): 478 decoded_hyps[hyp_idx] = self.compute_rnnt_timestamps(decoded_hyps[hyp_idx], timestamp_type) 479 480 hypotheses.append(decoded_hyps[0]) # best hypothesis 481 all_hypotheses.append(decoded_hyps) 482 483 if return_hypotheses: 484 return hypotheses, all_hypotheses 485 486 best_hyp_text = [h.text for h in hypotheses] 487 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 488 return best_hyp_text, all_hyp_text 489 490 else: 491 hypotheses = self.decode_hypothesis(prediction_list) # type: List[str] 492 493 # If computing timestamps 494 if self.compute_timestamps is True: 495 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 496 for hyp_idx in range(len(hypotheses)): 497 hypotheses[hyp_idx] = self.compute_rnnt_timestamps(hypotheses[hyp_idx], timestamp_type) 498 499 if return_hypotheses: 500 # greedy decoding, can get high-level confidence scores 501 if self.preserve_frame_confidence and ( 502 self.preserve_word_confidence or self.preserve_token_confidence 503 ): 504 hypotheses = self.compute_confidence(hypotheses) 505 return hypotheses, None 506 507 best_hyp_text = [h.text for h in hypotheses] 508 return best_hyp_text, None 509 510 def decode_hypothesis(self, hypotheses_list: List[Hypothesis]) -> List[Union[Hypothesis, NBestHypotheses]]: 511 """ 512 Decode a list of hypotheses into a list of strings. 513 514 Args: 515 hypotheses_list: List of Hypothesis. 516 517 Returns: 518 A list of strings. 519 """ 520 for ind in range(len(hypotheses_list)): 521 # Extract the integer encoded hypothesis 522 prediction = hypotheses_list[ind].y_sequence 523 524 if type(prediction) != list: 525 prediction = prediction.tolist() 526 527 # RNN-T sample level is already preprocessed by implicit RNNT decoding 528 # Simply remove any blank and possibly big blank tokens 529 if self.big_blank_durations is not None: # multi-blank RNNT 530 num_extra_outputs = len(self.big_blank_durations) 531 prediction = [p for p in prediction if p < self.blank_id - num_extra_outputs] 532 elif self.durations is not None: # TDT model. 533 prediction = [p for p in prediction if p < self.blank_id] 534 else: # standard RNN-T 535 prediction = [p for p in prediction if p != self.blank_id] 536 537 # De-tokenize the integer tokens; if not computing timestamps 538 if self.compute_timestamps is True: 539 # keep the original predictions, wrap with the number of repetitions per token and alignments 540 # this is done so that `rnnt_decoder_predictions_tensor()` can process this hypothesis 541 # in order to compute exact time stamps. 542 alignments = copy.deepcopy(hypotheses_list[ind].alignments) 543 token_repetitions = [1] * len(alignments) # preserve number of repetitions per token 544 hypothesis = (prediction, alignments, token_repetitions) 545 else: 546 hypothesis = self.decode_tokens_to_str(prediction) 547 548 # TODO: remove 549 # collapse leading spaces before . , ? for PC models 550 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 551 552 if self.compute_hypothesis_token_set: 553 hypotheses_list[ind].tokens = self.decode_ids_to_tokens(prediction) 554 555 # De-tokenize the integer tokens 556 hypotheses_list[ind].text = hypothesis 557 558 return hypotheses_list 559 560 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 561 """ 562 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 563 Assumes that `frame_confidence` is present in the hypotheses. 564 565 Args: 566 hypotheses_list: List of Hypothesis. 567 568 Returns: 569 A list of hypotheses with high-level confidence scores. 570 """ 571 if self.exclude_blank_from_confidence: 572 for hyp in hypotheses_list: 573 hyp.token_confidence = hyp.non_blank_frame_confidence 574 else: 575 for hyp in hypotheses_list: 576 offset = 0 577 token_confidence = [] 578 if len(hyp.timestep) > 0: 579 for ts, te in zip(hyp.timestep, hyp.timestep[1:] + [len(hyp.frame_confidence)]): 580 if ts != te: 581 # <blank> tokens are considered to belong to the last non-blank token, if any. 582 token_confidence.append( 583 self._aggregate_confidence( 584 [hyp.frame_confidence[ts][offset]] 585 + [fc[0] for fc in hyp.frame_confidence[ts + 1 : te]] 586 ) 587 ) 588 offset = 0 589 else: 590 token_confidence.append(hyp.frame_confidence[ts][offset]) 591 offset += 1 592 hyp.token_confidence = token_confidence 593 if self.preserve_word_confidence: 594 for hyp in hypotheses_list: 595 hyp.word_confidence = self._aggregate_token_confidence(hyp) 596 return hypotheses_list 597 598 @abstractmethod 599 def decode_tokens_to_str(self, tokens: List[int]) -> str: 600 """ 601 Implemented by subclass in order to decoder a token id list into a string. 602 603 Args: 604 tokens: List of int representing the token ids. 605 606 Returns: 607 A decoded string. 608 """ 609 raise NotImplementedError() 610 611 @abstractmethod 612 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 613 """ 614 Implemented by subclass in order to decode a token id list into a token list. 615 A token list is the string representation of each token id. 616 617 Args: 618 tokens: List of int representing the token ids. 619 620 Returns: 621 A list of decoded tokens. 622 """ 623 raise NotImplementedError() 624 625 @abstractmethod 626 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 627 """ 628 Implemented by subclass in order to 629 compute the most likely language ID (LID) string given the tokens. 630 631 Args: 632 tokens: List of int representing the token ids. 633 634 Returns: 635 A decoded LID string. 636 """ 637 raise NotImplementedError() 638 639 @abstractmethod 640 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 641 """ 642 Implemented by subclass in order to 643 decode a token id list into language ID (LID) list. 644 645 Args: 646 tokens: List of int representing the token ids. 647 648 Returns: 649 A list of decoded LIDS. 650 """ 651 raise NotImplementedError() 652 653 def update_joint_fused_batch_size(self): 654 if self.joint_fused_batch_size is None: 655 # do nothing and let the Joint itself handle setting up of the fused batch 656 return 657 658 if not hasattr(self.decoding.joint, 'set_fused_batch_size'): 659 logging.warning( 660 "The joint module does not have `set_fused_batch_size(int)` as a setter function.\n" 661 "Ignoring update of joint fused batch size." 662 ) 663 return 664 665 if not hasattr(self.decoding.joint, 'set_fuse_loss_wer'): 666 logging.warning( 667 "The joint module does not have `set_fuse_loss_wer(bool, RNNTLoss, RNNTWER)` " 668 "as a setter function.\n" 669 "Ignoring update of joint fused batch size." 670 ) 671 return 672 673 if self.joint_fused_batch_size > 0: 674 self.decoding.joint.set_fused_batch_size(self.joint_fused_batch_size) 675 else: 676 logging.info("Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.") 677 self.decoding.joint.set_fuse_loss_wer(False) 678 679 def compute_rnnt_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 680 assert timestamp_type in ['char', 'word', 'all'] 681 682 # Unpack the temporary storage 683 decoded_prediction, alignments, token_repetitions = hypothesis.text 684 685 # Retrieve offsets 686 char_offsets = word_offsets = None 687 char_offsets = self._compute_offsets(hypothesis, token_repetitions, self.blank_id) 688 689 # finally, set the flattened decoded predictions to text field for later text decoding 690 hypothesis.text = decoded_prediction 691 692 # Assert number of offsets and hypothesis tokens are 1:1 match. 693 num_flattened_tokens = 0 694 for t in range(len(char_offsets)): 695 # Subtract one here for the extra RNNT BLANK token emitted to designate "End of timestep" 696 num_flattened_tokens += len(char_offsets[t]['char']) - 1 697 698 if num_flattened_tokens != len(hypothesis.text): 699 raise ValueError( 700 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 701 " have to be of the same length, but are: " 702 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 703 f" {len(hypothesis.text)}" 704 ) 705 706 encoded_char_offsets = copy.deepcopy(char_offsets) 707 708 # Correctly process the token ids to chars/subwords. 709 for i, offsets in enumerate(char_offsets): 710 decoded_chars = [] 711 for char in offsets['char'][:-1]: # ignore the RNNT Blank token at end of every timestep with -1 subset 712 decoded_chars.append(self.decode_tokens_to_str([int(char)])) 713 char_offsets[i]["char"] = decoded_chars 714 715 # detect char vs subword models 716 lens = [] 717 for v in char_offsets: 718 tokens = v["char"] 719 # each token may be either 1 unicode token or multiple unicode token 720 # for character based models, only 1 token is used 721 # for subword, more than one token can be used. 722 # Computing max, then summing up total lens is a test to check for char vs subword 723 # For char models, len(lens) == sum(lens) 724 # but this is violated for subword models. 725 max_len = max(len(c) for c in tokens) 726 lens.append(max_len) 727 728 # array of one or more chars implies subword based model with multiple char emitted per TxU step (via subword) 729 if sum(lens) > len(lens): 730 text_type = 'subword' 731 else: 732 # full array of ones implies character based model with 1 char emitted per TxU step 733 text_type = 'char' 734 735 # retrieve word offsets from character offsets 736 word_offsets = None 737 if timestamp_type in ['word', 'all']: 738 if text_type == 'char': 739 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 740 else: 741 # utilize the copy of char offsets with the correct integer ids for tokens 742 # so as to avoid tokenize -> detokenize -> compare -> merge steps. 743 word_offsets = self._get_word_offsets_subwords_sentencepiece( 744 encoded_char_offsets, 745 hypothesis, 746 decode_ids_to_tokens=self.decode_ids_to_tokens, 747 decode_tokens_to_str=self.decode_tokens_to_str, 748 ) 749 750 # attach results 751 if len(hypothesis.timestep) > 0: 752 timestep_info = hypothesis.timestep 753 else: 754 timestep_info = [] 755 756 # Setup defaults 757 hypothesis.timestep = {"timestep": timestep_info} 758 759 # Add char / subword time stamps 760 if char_offsets is not None and timestamp_type in ['char', 'all']: 761 hypothesis.timestep['char'] = char_offsets 762 763 # Add word time stamps 764 if word_offsets is not None and timestamp_type in ['word', 'all']: 765 hypothesis.timestep['word'] = word_offsets 766 767 # Convert the flattened token indices to text 768 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 769 770 return hypothesis 771 772 @staticmethod 773 def _compute_offsets( 774 hypothesis: Hypothesis, token_repetitions: List[int], rnnt_token: int 775 ) -> List[Dict[str, Union[str, int]]]: 776 """ 777 Utility method that calculates the indidual time indices where a token starts and ends. 778 779 Args: 780 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 781 emitted at every time step after rnnt collapse. 782 token_repetitions: A list of ints representing the number of repetitions of each emitted token. 783 rnnt_token: The integer of the rnnt blank token used during rnnt collapse. 784 785 Returns: 786 787 """ 788 start_index = 0 789 790 # If the exact timestep information is available, utilize the 1st non-rnnt blank token timestep 791 # as the start index. 792 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 793 start_index = max(0, hypothesis.timestep[0] - 1) 794 795 # Construct the start and end indices brackets 796 end_indices = np.asarray(token_repetitions).cumsum() 797 start_indices = np.concatenate(([start_index], end_indices[:-1])) 798 799 # Process the TxU dangling alignment tensor, containing pairs of (logits, label) 800 alignment_labels = [al_logits_labels for al_logits_labels in hypothesis.text[1]] 801 for t in range(len(alignment_labels)): 802 for u in range(len(alignment_labels[t])): 803 alignment_labels[t][u] = alignment_labels[t][u][1] # pick label from (logit, label) tuple 804 805 # Merge the results per token into a list of dictionaries 806 offsets = [ 807 {"char": a, "start_offset": s, "end_offset": e} 808 for a, s, e in zip(alignment_labels, start_indices, end_indices) 809 ] 810 811 # Filter out RNNT token (blank at [t][0] position). This is because blank can only occur at end of a 812 # time step for RNNT, so if 0th token is blank, then that timestep is skipped. 813 offsets = list(filter(lambda offsets: offsets["char"][0] != rnnt_token, offsets)) 814 return offsets 815 816 @staticmethod 817 def _get_word_offsets_chars( 818 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 819 ) -> Dict[str, Union[str, float]]: 820 """ 821 Utility method which constructs word time stamps out of character time stamps. 822 823 References: 824 This code is a port of the Hugging Face code for word time stamp construction. 825 826 Args: 827 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 828 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 829 830 Returns: 831 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 832 "end_offset". 833 """ 834 word_offsets = [] 835 836 last_state = "SPACE" 837 word = "" 838 start_offset = 0 839 end_offset = 0 840 for i, offset in enumerate(offsets): 841 chars = offset["char"] 842 for char in chars: 843 state = "SPACE" if char == word_delimiter_char else "WORD" 844 845 if state == last_state: 846 # If we are in the same state as before, we simply repeat what we've done before 847 end_offset = offset["end_offset"] 848 word += char 849 else: 850 # Switching state 851 if state == "SPACE": 852 # Finishing a word 853 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 854 else: 855 # Starting a new word 856 start_offset = offset["start_offset"] 857 end_offset = offset["end_offset"] 858 word = char 859 860 last_state = state 861 862 if last_state == "WORD": 863 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 864 865 return word_offsets 866 867 @staticmethod 868 def _get_word_offsets_subwords_sentencepiece( 869 offsets: Dict[str, Union[str, float]], 870 hypothesis: Hypothesis, 871 decode_ids_to_tokens: Callable[[List[int]], str], 872 decode_tokens_to_str: Callable[[List[int]], str], 873 ) -> Dict[str, Union[str, float]]: 874 """ 875 Utility method which constructs word time stamps out of sub-word time stamps. 876 877 **Note**: Only supports Sentencepiece based tokenizers ! 878 879 Args: 880 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 881 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 882 after rnnt collapse. 883 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 884 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 885 886 Returns: 887 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 888 "end_offset". 889 """ 890 word_offsets = [] 891 built_token = [] 892 previous_token_index = 0 893 # For every offset token 894 for i, offset in enumerate(offsets): 895 # For every subword token in offset token list (ignoring the RNNT Blank token at the end) 896 for char in offset['char'][:-1]: 897 char = int(char) 898 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if built_token: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 # This should only be done when these arrays contain more than one element. 931 if offsets and word_offsets: 932 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 933 934 # If there are any remaining tokens left, inject them all into the final word offset. 935 # The start offset of this token is the start time of the next token to process. 936 # The end offset of this token is the end time of the last token from offsets. 937 # Note that built_token is a flat list; but offsets contains a nested list which 938 # may have different dimensionality. 939 # As such, we can't rely on the length of the list of built_token to index offsets. 940 if built_token: 941 # start from the previous token index as this hasn't been committed to word_offsets yet 942 # if we still have content in built_token 943 start_offset = offsets[previous_token_index]["start_offset"] 944 word_offsets.append( 945 { 946 "word": decode_tokens_to_str(built_token), 947 "start_offset": start_offset, 948 "end_offset": offsets[-1]["end_offset"], 949 } 950 ) 951 built_token.clear() 952 953 return word_offsets 954 955 956 class RNNTDecoding(AbstractRNNTDecoding): 957 """ 958 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 959 960 Args: 961 decoding_cfg: A dict-like object which contains the following key-value pairs. 962 strategy: str value which represents the type of decoding that can occur. 963 Possible values are : 964 - greedy, greedy_batch (for greedy decoding). 965 - beam, tsd, alsd (for beam search decoding). 966 967 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 968 tokens as well as the decoded string. Default is False in order to avoid double decoding 969 unless required. 970 971 preserve_alignments: Bool flag which preserves the history of logprobs generated during 972 decoding (sample / batched). When set to true, the Hypothesis will contain 973 the non-null value for `logprobs` in it. Here, `alignments` is a List of List of 974 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 975 976 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 977 with the `return_hypotheses` flag set to True. 978 979 The length of the list corresponds to the Acoustic Length (T). 980 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 981 U is the number of target tokens for the current timestep Ti. 982 983 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 984 scores. In order to obtain hypotheses with confidence scores, please utilize 985 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 986 987 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 988 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 989 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 990 991 The length of the list corresponds to the Acoustic Length (T). 992 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 993 U is the number of target tokens for the current timestep Ti. 994 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 995 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 996 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 997 998 The length of the list corresponds to the number of recognized tokens. 999 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1000 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1001 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1002 1003 The length of the list corresponds to the number of recognized words. 1004 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1005 from the `token_confidence`. 1006 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1007 Valid options are `mean`, `min`, `max`, `prod`. 1008 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1009 confidence scores. 1010 1011 name: The measure name (str). 1012 Supported values: 1013 - 'max_prob' for using the maximum token probability as a confidence. 1014 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1015 1016 entropy_type: Which type of entropy to use (str). 1017 Used if confidence_measure_cfg.name is set to `entropy`. 1018 Supported values: 1019 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1020 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1021 Note that for this entropy, the alpha should comply the following inequality: 1022 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1023 where V is the model vocabulary size. 1024 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1025 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1026 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1027 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1028 - 'renyi' for the Rรฉnyi entropy. 1029 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1030 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1031 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1032 1033 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1034 When the alpha equals one, scaling is not applied to 'max_prob', 1035 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1036 1037 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1038 Supported values: 1039 - 'lin' for using the linear mapping. 1040 - 'exp' for using exponential mapping with linear shift. 1041 1042 The config may further contain the following sub-dictionaries: 1043 "greedy": 1044 max_symbols: int, describing the maximum number of target tokens to decode per 1045 timestep during greedy decoding. Setting to larger values allows longer sentences 1046 to be decoded, at the cost of increased execution time. 1047 1048 preserve_frame_confidence: Same as above, overrides above value. 1049 1050 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 1051 1052 "beam": 1053 beam_size: int, defining the beam size for beam search. Must be >= 1. 1054 If beam_size == 1, will perform cached greedy search. This might be slightly different 1055 results compared to the greedy search above. 1056 1057 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 1058 Set to True by default. 1059 1060 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1061 hypotheses after beam search has concluded. This flag is set by default. 1062 1063 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 1064 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 1065 at increased cost to execution time. 1066 1067 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 1068 If an integer is provided, it can decode sequences of that particular maximum length. 1069 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 1070 where seq_len is the length of the acoustic model output (T). 1071 1072 NOTE: 1073 If a float is provided, it can be greater than 1! 1074 By default, a float of 2.0 is used so that a target sequence can be at most twice 1075 as long as the acoustic model output length T. 1076 1077 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 1078 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 1079 1080 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 1081 in order to reduce expensive beam search cost later. int >= 0. 1082 1083 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 1084 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 1085 and affects the speed of inference since large values will perform large beam search in the next step. 1086 1087 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 1088 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 1089 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 1090 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 1091 expansion apart from the "most likely" candidate. 1092 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 1093 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 1094 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 1095 tuned on a validation set. 1096 1097 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 1098 1099 decoder: The Decoder/Prediction network module. 1100 joint: The Joint network module. 1101 vocabulary: The vocabulary (excluding the RNNT blank token) which will be used for decoding. 1102 """ 1103 1104 def __init__( 1105 self, decoding_cfg, decoder, joint, vocabulary, 1106 ): 1107 # we need to ensure blank is the last token in the vocab for the case of RNNT and Multi-blank RNNT. 1108 blank_id = len(vocabulary) + joint.num_extra_outputs 1109 1110 if hasattr(decoding_cfg, 'model_type') and decoding_cfg.model_type == 'tdt': 1111 blank_id = len(vocabulary) 1112 1113 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1114 1115 super(RNNTDecoding, self).__init__( 1116 decoding_cfg=decoding_cfg, decoder=decoder, joint=joint, blank_id=blank_id, 1117 ) 1118 1119 if isinstance(self.decoding, beam_decode.BeamRNNTInfer): 1120 self.decoding.set_decoding_type('char') 1121 1122 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1123 """ 1124 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1125 1126 Args: 1127 hypothesis: Hypothesis 1128 1129 Returns: 1130 A list of word-level confidence scores. 1131 """ 1132 return self._aggregate_token_confidence_chars(hypothesis.words, hypothesis.token_confidence) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c < self.blank_id - self.num_extra_outputs] 1159 return token_list 1160 1161 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 1162 """ 1163 Compute the most likely language ID (LID) string given the tokens. 1164 1165 Args: 1166 tokens: List of int representing the token ids. 1167 1168 Returns: 1169 A decoded LID string. 1170 """ 1171 lang = self.tokenizer.ids_to_lang(tokens) 1172 return lang 1173 1174 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 1175 """ 1176 Decode a token id list into language ID (LID) list. 1177 1178 Args: 1179 tokens: List of int representing the token ids. 1180 1181 Returns: 1182 A list of decoded LIDS. 1183 """ 1184 lang_list = self.tokenizer.ids_to_text_and_langs(tokens) 1185 return lang_list 1186 1187 1188 class RNNTWER(Metric): 1189 """ 1190 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. 1191 When doing distributed training/evaluation the result of res=WER(predictions, targets, target_lengths) calls 1192 will be all-reduced between all workers using SUM operations. 1193 Here contains two numbers res=[wer_numerator, wer_denominator]. WER=wer_numerator/wer_denominator. 1194 1195 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. 1196 Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1197 1198 Example: 1199 def validation_step(self, batch, batch_idx): 1200 ... 1201 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1202 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1203 return self.val_outputs 1204 1205 def on_validation_epoch_end(self): 1206 ... 1207 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1208 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1209 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1210 self.val_outputs.clear() # free memory 1211 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1212 1213 Args: 1214 decoding: RNNTDecoding object that will perform autoregressive decoding of the RNNT model. 1215 batch_dim_index: Index of the batch dimension. 1216 use_cer: Whether to use Character Error Rate isntead of Word Error Rate. 1217 log_prediction: Whether to log a single decoded sample per call. 1218 1219 Returns: 1220 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenshtein's 1221 distances for all prediction - reference pairs, total number of words in all references. 1222 """ 1223 1224 full_state_update = True 1225 1226 def __init__( 1227 self, decoding: RNNTDecoding, batch_dim_index=0, use_cer=False, log_prediction=True, dist_sync_on_step=False 1228 ): 1229 super(RNNTWER, self).__init__(dist_sync_on_step=dist_sync_on_step) 1230 self.decoding = decoding 1231 self.batch_dim_index = batch_dim_index 1232 self.use_cer = use_cer 1233 self.log_prediction = log_prediction 1234 self.blank_id = self.decoding.blank_id 1235 self.labels_map = self.decoding.labels_map 1236 1237 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1238 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1239 1240 def update( 1241 self, 1242 encoder_output: torch.Tensor, 1243 encoded_lengths: torch.Tensor, 1244 targets: torch.Tensor, 1245 target_lengths: torch.Tensor, 1246 ) -> torch.Tensor: 1247 words = 0 1248 scores = 0 1249 references = [] 1250 with torch.no_grad(): 1251 # prediction_cpu_tensor = tensors[0].long().cpu() 1252 targets_cpu_tensor = targets.long().cpu() 1253 targets_cpu_tensor = move_dimension_to_the_front(targets_cpu_tensor, self.batch_dim_index) 1254 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1255 1256 # iterate over batch 1257 for ind in range(targets_cpu_tensor.shape[0]): 1258 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1259 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1260 1261 reference = self.decoding.decode_tokens_to_str(target) 1262 references.append(reference) 1263 1264 hypotheses, _ = self.decoding.rnnt_decoder_predictions_tensor(encoder_output, encoded_lengths) 1265 1266 if self.log_prediction: 1267 logging.info(f"\n") 1268 logging.info(f"reference :{references[0]}") 1269 logging.info(f"predicted :{hypotheses[0]}") 1270 1271 for h, r in zip(hypotheses, references): 1272 if self.use_cer: 1273 h_list = list(h) 1274 r_list = list(r) 1275 else: 1276 h_list = h.split() 1277 r_list = r.split() 1278 words += len(r_list) 1279 # Compute Levenshtein's distance 1280 scores += editdistance.eval(h_list, r_list) 1281 1282 self.scores += torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1283 self.words += torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1284 # return torch.tensor([scores, words]).to(predictions.device) 1285 1286 def compute(self): 1287 wer = self.scores.float() / self.words 1288 return wer, self.scores.detach(), self.words.detach() 1289 1290 1291 @dataclass 1292 class RNNTDecodingConfig: 1293 model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" 1294 strategy: str = "greedy_batch" 1295 1296 compute_hypothesis_token_set: bool = False 1297 1298 # preserve decoding alignments 1299 preserve_alignments: Optional[bool] = None 1300 1301 # confidence config 1302 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 1307 # compute RNNT time stamps 1308 compute_timestamps: Optional[bool] = None 1309 1310 # compute language IDs 1311 compute_langs: bool = False 1312 1313 # token representing word seperator 1314 word_seperator: str = " " 1315 1316 # type of timestamps to calculate 1317 rnnt_timestamp_type: str = "all" # can be char, word or all for both 1318 1319 # greedy decoding config 1320 greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() 1321 1322 # beam decoding config 1323 beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) 1324 1325 # can be used to change temperature for decoding 1326 temperature: float = 1.0 1327 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/metrics/wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import re 16 from abc import abstractmethod 17 from dataclasses import dataclass, is_dataclass 18 from typing import Callable, Dict, List, Optional, Tuple, Union 19 20 import editdistance 21 import jiwer 22 import numpy as np 23 import torch 24 from omegaconf import DictConfig, OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding 28 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 29 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 30 from nemo.utils import logging, logging_mode 31 32 __all__ = ['word_error_rate', 'word_error_rate_detail', 'WER', 'move_dimension_to_the_front'] 33 34 35 def word_error_rate(hypotheses: List[str], references: List[str], use_cer=False) -> float: 36 """ 37 Computes Average Word Error rate between two texts represented as 38 corresponding lists of string. 39 40 Hypotheses and references must have same length. 41 42 Args: 43 hypotheses (list): list of hypotheses 44 references(list) : list of references 45 use_cer (bool): set True to enable cer 46 47 Returns: 48 wer (float): average word error rate 49 """ 50 scores = 0 51 words = 0 52 if len(hypotheses) != len(references): 53 raise ValueError( 54 "In word error rate calculation, hypotheses and reference" 55 " lists must have the same number of elements. But I got:" 56 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 57 ) 58 for h, r in zip(hypotheses, references): 59 if use_cer: 60 h_list = list(h) 61 r_list = list(r) 62 else: 63 h_list = h.split() 64 r_list = r.split() 65 words += len(r_list) 66 # May deprecate using editdistance in future release for here and rest of codebase 67 # once we confirm jiwer is reliable. 68 scores += editdistance.eval(h_list, r_list) 69 if words != 0: 70 wer = 1.0 * scores / words 71 else: 72 wer = float('inf') 73 return wer 74 75 76 def word_error_rate_detail( 77 hypotheses: List[str], references: List[str], use_cer=False 78 ) -> Tuple[float, int, float, float, float]: 79 """ 80 Computes Average Word Error Rate with details (insertion rate, deletion rate, substitution rate) 81 between two texts represented as corresponding lists of string. 82 83 Hypotheses and references must have same length. 84 85 Args: 86 hypotheses (list): list of hypotheses 87 references(list) : list of references 88 use_cer (bool): set True to enable cer 89 90 Returns: 91 wer (float): average word error rate 92 words (int): Total number of words/charactors of given reference texts 93 ins_rate (float): average insertion error rate 94 del_rate (float): average deletion error rate 95 sub_rate (float): average substitution error rate 96 """ 97 scores = 0 98 words = 0 99 ops_count = {'substitutions': 0, 'insertions': 0, 'deletions': 0} 100 101 if len(hypotheses) != len(references): 102 raise ValueError( 103 "In word error rate calculation, hypotheses and reference" 104 " lists must have the same number of elements. But I got:" 105 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 106 ) 107 108 for h, r in zip(hypotheses, references): 109 if use_cer: 110 h_list = list(h) 111 r_list = list(r) 112 else: 113 h_list = h.split() 114 r_list = r.split() 115 116 # To get rid of the issue that jiwer does not allow empty string 117 if len(r_list) == 0: 118 if len(h_list) != 0: 119 errors = len(h_list) 120 ops_count['insertions'] += errors 121 else: 122 errors = 0 123 else: 124 if use_cer: 125 measures = jiwer.cer(r, h, return_dict=True) 126 else: 127 measures = jiwer.compute_measures(r, h) 128 129 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 130 ops_count['insertions'] += measures['insertions'] 131 ops_count['deletions'] += measures['deletions'] 132 ops_count['substitutions'] += measures['substitutions'] 133 134 scores += errors 135 words += len(r_list) 136 137 if words != 0: 138 wer = 1.0 * scores / words 139 ins_rate = 1.0 * ops_count['insertions'] / words 140 del_rate = 1.0 * ops_count['deletions'] / words 141 sub_rate = 1.0 * ops_count['substitutions'] / words 142 else: 143 wer, ins_rate, del_rate, sub_rate = float('inf'), float('inf'), float('inf'), float('inf') 144 145 return wer, words, ins_rate, del_rate, sub_rate 146 147 148 def word_error_rate_per_utt(hypotheses: List[str], references: List[str], use_cer=False) -> Tuple[List[float], float]: 149 """ 150 Computes Word Error Rate per utterance and the average WER 151 between two texts represented as corresponding lists of string. 152 153 Hypotheses and references must have same length. 154 155 Args: 156 hypotheses (list): list of hypotheses 157 references(list) : list of references 158 use_cer (bool): set True to enable cer 159 160 Returns: 161 wer_per_utt (List[float]): word error rate per utterance 162 avg_wer (float): average word error rate 163 """ 164 scores = 0 165 words = 0 166 wer_per_utt = [] 167 168 if len(hypotheses) != len(references): 169 raise ValueError( 170 "In word error rate calculation, hypotheses and reference" 171 " lists must have the same number of elements. But I got:" 172 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 173 ) 174 175 for h, r in zip(hypotheses, references): 176 if use_cer: 177 h_list = list(h) 178 r_list = list(r) 179 else: 180 h_list = h.split() 181 r_list = r.split() 182 183 # To get rid of the issue that jiwer does not allow empty string 184 if len(r_list) == 0: 185 if len(h_list) != 0: 186 errors = len(h_list) 187 wer_per_utt.append(float('inf')) 188 else: 189 if use_cer: 190 measures = jiwer.cer(r, h, return_dict=True) 191 er = measures['cer'] 192 else: 193 measures = jiwer.compute_measures(r, h) 194 er = measures['wer'] 195 196 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 197 wer_per_utt.append(er) 198 199 scores += errors 200 words += len(r_list) 201 202 if words != 0: 203 avg_wer = 1.0 * scores / words 204 else: 205 avg_wer = float('inf') 206 207 return wer_per_utt, avg_wer 208 209 210 def move_dimension_to_the_front(tensor, dim_index): 211 all_dims = list(range(tensor.ndim)) 212 return tensor.permute(*([dim_index] + all_dims[:dim_index] + all_dims[dim_index + 1 :])) 213 214 215 class AbstractCTCDecoding(ConfidenceMixin): 216 """ 217 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs. 218 219 Args: 220 decoding_cfg: A dict-like object which contains the following key-value pairs. 221 strategy: str value which represents the type of decoding that can occur. 222 Possible values are : 223 - greedy (for greedy decoding). 224 - beam (for DeepSpeed KenLM based decoding). 225 226 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 227 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 228 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 229 230 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 231 Can take the following values - "char" for character/subword time stamps, "word" for word level 232 time stamps and "all" (default), for both character level and word level time stamps. 233 234 word_seperator: Str token representing the seperator between words. 235 236 preserve_alignments: Bool flag which preserves the history of logprobs generated during 237 decoding (sample / batched). When set to true, the Hypothesis will contain 238 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 239 240 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 241 scores. In order to obtain hypotheses with confidence scores, please utilize 242 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 243 244 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 245 generated during decoding. When set to true, the Hypothesis will contain 246 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 247 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 248 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 249 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 250 251 The length of the list corresponds to the number of recognized tokens. 252 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 253 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 254 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 255 256 The length of the list corresponds to the number of recognized words. 257 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 258 from the `token_confidence`. 259 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 260 Valid options are `mean`, `min`, `max`, `prod`. 261 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 262 confidence scores. 263 264 name: The measure name (str). 265 Supported values: 266 - 'max_prob' for using the maximum token probability as a confidence. 267 - 'entropy' for using a normalized entropy of a log-likelihood vector. 268 269 entropy_type: Which type of entropy to use (str). 270 Used if confidence_measure_cfg.name is set to `entropy`. 271 Supported values: 272 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 273 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 274 Note that for this entropy, the alpha should comply the following inequality: 275 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 276 where V is the model vocabulary size. 277 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 278 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 279 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 280 More: https://en.wikipedia.org/wiki/Tsallis_entropy 281 - 'renyi' for the Rรฉnyi entropy. 282 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 285 286 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 287 When the alpha equals one, scaling is not applied to 'max_prob', 288 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 289 290 entropy_norm: A mapping of the entropy value to the interval [0,1]. 291 Supported values: 292 - 'lin' for using the linear mapping. 293 - 'exp' for using exponential mapping with linear shift. 294 295 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 296 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 297 298 The config may further contain the following sub-dictionaries: 299 "greedy": 300 preserve_alignments: Same as above, overrides above value. 301 compute_timestamps: Same as above, overrides above value. 302 preserve_frame_confidence: Same as above, overrides above value. 303 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 304 305 "beam": 306 beam_size: int, defining the beam size for beam search. Must be >= 1. 307 If beam_size == 1, will perform cached greedy search. This might be slightly different 308 results compared to the greedy search above. 309 310 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 311 hypotheses after beam search has concluded. This flag is set by default. 312 313 beam_alpha: float, the strength of the Language model on the final score of a token. 314 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 315 316 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 317 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 318 319 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 320 If the path is invalid (file is not found at path), will raise a deferred error at the moment 321 of calculation of beam search, so that users may update / change the decoding strategy 322 to point to the correct file. 323 324 blank_id: The id of the RNNT blank token. 325 """ 326 327 def __init__(self, decoding_cfg, blank_id: int): 328 super().__init__() 329 330 # Convert dataclas to config 331 if is_dataclass(decoding_cfg): 332 decoding_cfg = OmegaConf.structured(decoding_cfg) 333 334 if not isinstance(decoding_cfg, DictConfig): 335 decoding_cfg = OmegaConf.create(decoding_cfg) 336 337 OmegaConf.set_struct(decoding_cfg, False) 338 339 # update minimal config 340 minimal_cfg = ['greedy'] 341 for item in minimal_cfg: 342 if item not in decoding_cfg: 343 decoding_cfg[item] = OmegaConf.create({}) 344 345 self.cfg = decoding_cfg 346 self.blank_id = blank_id 347 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 348 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 349 self.batch_dim_index = self.cfg.get('batch_dim_index', 0) 350 self.word_seperator = self.cfg.get('word_seperator', ' ') 351 352 possible_strategies = ['greedy', 'beam', 'pyctcdecode', 'flashlight'] 353 if self.cfg.strategy not in possible_strategies: 354 raise ValueError(f"Decoding strategy must be one of {possible_strategies}. Given {self.cfg.strategy}") 355 356 # Update preserve alignments 357 if self.preserve_alignments is None: 358 if self.cfg.strategy in ['greedy']: 359 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 360 else: 361 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 362 363 # Update compute timestamps 364 if self.compute_timestamps is None: 365 if self.cfg.strategy in ['greedy']: 366 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 367 elif self.cfg.strategy in ['beam']: 368 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 369 370 # initialize confidence-related fields 371 self._init_confidence(self.cfg.get('confidence_cfg', None)) 372 373 # Confidence estimation is not implemented for strategies other than `greedy` 374 if ( 375 not self.preserve_frame_confidence 376 and self.cfg.strategy != 'greedy' 377 and self.cfg.beam.get('preserve_frame_confidence', False) 378 ): 379 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 380 381 # we need timestamps to extract non-blank per-frame confidence 382 if self.compute_timestamps is not None: 383 self.compute_timestamps |= self.preserve_frame_confidence 384 385 if self.cfg.strategy == 'greedy': 386 387 self.decoding = ctc_greedy_decoding.GreedyCTCInfer( 388 blank_id=self.blank_id, 389 preserve_alignments=self.preserve_alignments, 390 compute_timestamps=self.compute_timestamps, 391 preserve_frame_confidence=self.preserve_frame_confidence, 392 confidence_measure_cfg=self.confidence_measure_cfg, 393 ) 394 395 elif self.cfg.strategy == 'beam': 396 397 self.decoding = ctc_beam_decoding.BeamCTCInfer( 398 blank_id=blank_id, 399 beam_size=self.cfg.beam.get('beam_size', 1), 400 search_type='default', 401 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 402 preserve_alignments=self.preserve_alignments, 403 compute_timestamps=self.compute_timestamps, 404 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 405 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 406 kenlm_path=self.cfg.beam.get('kenlm_path', None), 407 ) 408 409 self.decoding.override_fold_consecutive_value = False 410 411 elif self.cfg.strategy == 'pyctcdecode': 412 413 self.decoding = ctc_beam_decoding.BeamCTCInfer( 414 blank_id=blank_id, 415 beam_size=self.cfg.beam.get('beam_size', 1), 416 search_type='pyctcdecode', 417 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 418 preserve_alignments=self.preserve_alignments, 419 compute_timestamps=self.compute_timestamps, 420 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 421 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 422 kenlm_path=self.cfg.beam.get('kenlm_path', None), 423 pyctcdecode_cfg=self.cfg.beam.get('pyctcdecode_cfg', None), 424 ) 425 426 self.decoding.override_fold_consecutive_value = False 427 428 elif self.cfg.strategy == 'flashlight': 429 430 self.decoding = ctc_beam_decoding.BeamCTCInfer( 431 blank_id=blank_id, 432 beam_size=self.cfg.beam.get('beam_size', 1), 433 search_type='flashlight', 434 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 435 preserve_alignments=self.preserve_alignments, 436 compute_timestamps=self.compute_timestamps, 437 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 438 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 439 kenlm_path=self.cfg.beam.get('kenlm_path', None), 440 flashlight_cfg=self.cfg.beam.get('flashlight_cfg', None), 441 ) 442 443 self.decoding.override_fold_consecutive_value = False 444 445 else: 446 raise ValueError( 447 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 448 f"but was provided {self.cfg.strategy}" 449 ) 450 451 def ctc_decoder_predictions_tensor( 452 self, 453 decoder_outputs: torch.Tensor, 454 decoder_lengths: torch.Tensor = None, 455 fold_consecutive: bool = True, 456 return_hypotheses: bool = False, 457 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 458 """ 459 Decodes a sequence of labels to words 460 461 Args: 462 decoder_outputs: An integer torch.Tensor of shape [Batch, Time, {Vocabulary}] (if ``batch_index_dim == 0``) or [Time, Batch] 463 (if ``batch_index_dim == 1``) of integer indices that correspond to the index of some character in the 464 label set. 465 decoder_lengths: Optional tensor of length `Batch` which contains the integer lengths 466 of the sequence in the padded `predictions` tensor. 467 fold_consecutive: Bool, determine whether to perform "ctc collapse", folding consecutive tokens 468 into a single token. 469 return_hypotheses: Bool flag whether to return just the decoding predictions of the model 470 or a Hypothesis object that holds information such as the decoded `text`, 471 the `alignment` of emited by the CTC Model, and the `length` of the sequence (if available). 472 May also contain the log-probabilities of the decoder (if this method is called via 473 transcribe()) 474 475 Returns: 476 Either a list of str which represent the CTC decoded strings per sample, 477 or a list of Hypothesis objects containing additional information. 478 """ 479 480 if isinstance(decoder_outputs, torch.Tensor): 481 decoder_outputs = move_dimension_to_the_front(decoder_outputs, self.batch_dim_index) 482 483 if ( 484 hasattr(self.decoding, 'override_fold_consecutive_value') 485 and self.decoding.override_fold_consecutive_value is not None 486 ): 487 logging.info( 488 f"Beam search requires that consecutive ctc tokens are not folded. \n" 489 f"Overriding provided value of `fold_consecutive` = {fold_consecutive} to " 490 f"{self.decoding.override_fold_consecutive_value}", 491 mode=logging_mode.ONCE, 492 ) 493 fold_consecutive = self.decoding.override_fold_consecutive_value 494 495 with torch.inference_mode(): 496 # Resolve the forward step of the decoding strategy 497 hypotheses_list = self.decoding( 498 decoder_output=decoder_outputs, decoder_lengths=decoder_lengths 499 ) # type: List[List[Hypothesis]] 500 501 # extract the hypotheses 502 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 503 504 if isinstance(hypotheses_list[0], NBestHypotheses): 505 hypotheses = [] 506 all_hypotheses = [] 507 508 for nbest_hyp in hypotheses_list: # type: NBestHypotheses 509 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 510 decoded_hyps = self.decode_hypothesis( 511 n_hyps, fold_consecutive 512 ) # type: List[Union[Hypothesis, NBestHypotheses]] 513 514 # If computing timestamps 515 if self.compute_timestamps is True: 516 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 517 for hyp_idx in range(len(decoded_hyps)): 518 decoded_hyps[hyp_idx] = self.compute_ctc_timestamps(decoded_hyps[hyp_idx], timestamp_type) 519 520 hypotheses.append(decoded_hyps[0]) # best hypothesis 521 all_hypotheses.append(decoded_hyps) 522 523 if return_hypotheses: 524 return hypotheses, all_hypotheses 525 526 best_hyp_text = [h.text for h in hypotheses] 527 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 528 return best_hyp_text, all_hyp_text 529 530 else: 531 hypotheses = self.decode_hypothesis( 532 hypotheses_list, fold_consecutive 533 ) # type: List[Union[Hypothesis, NBestHypotheses]] 534 535 # If computing timestamps 536 if self.compute_timestamps is True: 537 # greedy decoding, can get high-level confidence scores 538 if return_hypotheses and (self.preserve_word_confidence or self.preserve_token_confidence): 539 hypotheses = self.compute_confidence(hypotheses) 540 else: 541 # remove unused token_repetitions from Hypothesis.text 542 for hyp in hypotheses: 543 hyp.text = hyp.text[:2] 544 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 545 for hyp_idx in range(len(hypotheses)): 546 hypotheses[hyp_idx] = self.compute_ctc_timestamps(hypotheses[hyp_idx], timestamp_type) 547 548 if return_hypotheses: 549 return hypotheses, None 550 551 best_hyp_text = [h.text for h in hypotheses] 552 return best_hyp_text, None 553 554 def decode_hypothesis( 555 self, hypotheses_list: List[Hypothesis], fold_consecutive: bool 556 ) -> List[Union[Hypothesis, NBestHypotheses]]: 557 """ 558 Decode a list of hypotheses into a list of strings. 559 560 Args: 561 hypotheses_list: List of Hypothesis. 562 fold_consecutive: Whether to collapse the ctc blank tokens or not. 563 564 Returns: 565 A list of strings. 566 """ 567 for ind in range(len(hypotheses_list)): 568 # Extract the integer encoded hypothesis 569 hyp = hypotheses_list[ind] 570 prediction = hyp.y_sequence 571 predictions_len = hyp.length if hyp.length > 0 else None 572 573 if fold_consecutive: 574 if type(prediction) != list: 575 prediction = prediction.numpy().tolist() 576 577 if predictions_len is not None: 578 prediction = prediction[:predictions_len] 579 580 # CTC decoding procedure 581 decoded_prediction = [] 582 token_lengths = [] # preserve token lengths 583 token_repetitions = [] # preserve number of repetitions per token 584 585 previous = self.blank_id 586 last_length = 0 587 last_repetition = 1 588 589 for pidx, p in enumerate(prediction): 590 if (p != previous or previous == self.blank_id) and p != self.blank_id: 591 decoded_prediction.append(p) 592 593 token_lengths.append(pidx - last_length) 594 last_length = pidx 595 token_repetitions.append(last_repetition) 596 last_repetition = 1 597 598 if p == previous and previous != self.blank_id: 599 last_repetition += 1 600 601 previous = p 602 603 if len(token_repetitions) > 0: 604 token_repetitions = token_repetitions[1:] + [last_repetition] 605 606 else: 607 if predictions_len is not None: 608 prediction = prediction[:predictions_len] 609 decoded_prediction = prediction[prediction != self.blank_id].tolist() 610 token_lengths = [1] * len(decoded_prediction) # preserve number of repetitions per token 611 token_repetitions = [1] * len(decoded_prediction) # preserve number of repetitions per token 612 613 # De-tokenize the integer tokens; if not computing timestamps 614 if self.compute_timestamps is True: 615 # keep the original predictions, wrap with the number of repetitions per token 616 # this is done so that `ctc_decoder_predictions_tensor()` can process this hypothesis 617 # in order to compute exact time stamps. 618 hypothesis = (decoded_prediction, token_lengths, token_repetitions) 619 else: 620 hypothesis = self.decode_tokens_to_str(decoded_prediction) 621 622 # TODO: remove 623 # collapse leading spaces before . , ? for PC models 624 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 625 626 # Preserve this wrapped hypothesis or decoded text tokens. 627 hypotheses_list[ind].text = hypothesis 628 629 return hypotheses_list 630 631 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 632 """ 633 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 634 Assumes that `frame_confidence` is present in the hypotheses. 635 636 Args: 637 hypotheses_list: List of Hypothesis. 638 639 Returns: 640 A list of hypotheses with high-level confidence scores. 641 """ 642 for hyp in hypotheses_list: 643 if not isinstance(hyp.text, tuple) or len(hyp.text) != 3: 644 # the method must have been called in the wrong place 645 raise ValueError( 646 """Wrong format of the `text` attribute of a hypothesis.\n 647 Expected: (decoded_prediction, token_repetitions)\n 648 The method invocation is expected between .decode_hypothesis() and .compute_ctc_timestamps()""" 649 ) 650 token_repetitions = hyp.text[2] 651 hyp.text = hyp.text[:2] 652 token_confidence = [] 653 if self.exclude_blank_from_confidence: 654 non_blank_frame_confidence = hyp.non_blank_frame_confidence 655 i = 0 656 for tr in token_repetitions: 657 # token repetition can be zero 658 j = i + tr 659 token_confidence.append(self._aggregate_confidence(non_blank_frame_confidence[i:j])) 660 i = j 661 else: 662 # <blank> tokens are considered to belong to the last non-blank token, if any. 663 token_lengths = hyp.text[1] 664 if len(token_lengths) > 0: 665 ts = token_lengths[0] 666 for tl in token_lengths[1:] + [len(hyp.frame_confidence)]: 667 token_confidence.append(self._aggregate_confidence(hyp.frame_confidence[ts : ts + tl])) 668 ts += tl 669 hyp.token_confidence = token_confidence 670 if self.preserve_word_confidence: 671 for hyp in hypotheses_list: 672 hyp.word_confidence = self._aggregate_token_confidence(hyp) 673 return hypotheses_list 674 675 @abstractmethod 676 def decode_tokens_to_str(self, tokens: List[int]) -> str: 677 """ 678 Implemented by subclass in order to decoder a token id list into a string. 679 680 Args: 681 tokens: List of int representing the token ids. 682 683 Returns: 684 A decoded string. 685 """ 686 raise NotImplementedError() 687 688 @abstractmethod 689 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 690 """ 691 Implemented by subclass in order to decode a token id list into a token list. 692 A token list is the string representation of each token id. 693 694 Args: 695 tokens: List of int representing the token ids. 696 697 Returns: 698 A list of decoded tokens. 699 """ 700 raise NotImplementedError() 701 702 def compute_ctc_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 703 """ 704 Method to compute time stamps at char/subword, and word level given some hypothesis. 705 Requires the input hypothesis to contain a `text` field that is the tuple. The tuple contains - 706 the ctc collapsed integer ids, and the number of repetitions of each token. 707 708 Args: 709 hypothesis: A Hypothesis object, with a wrapped `text` field. 710 The `text` field must contain a tuple with two values - 711 The ctc collapsed integer ids 712 A list of integers that represents the number of repetitions per token. 713 timestamp_type: A str value that represents the type of time stamp calculated. 714 Can be one of "char", "word" or "all" 715 716 Returns: 717 A Hypothesis object with a modified `timestep` value, which is now a dictionary containing 718 the time stamp information. 719 """ 720 assert timestamp_type in ['char', 'word', 'all'] 721 722 # Unpack the temporary storage, and set the decoded predictions 723 decoded_prediction, token_lengths = hypothesis.text 724 hypothesis.text = decoded_prediction 725 726 # Retrieve offsets 727 char_offsets = word_offsets = None 728 char_offsets = self._compute_offsets(hypothesis, token_lengths, self.blank_id) 729 730 # Assert number of offsets and hypothesis tokens are 1:1 match. 731 if len(char_offsets) != len(hypothesis.text): 732 raise ValueError( 733 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 734 " have to be of the same length, but are: " 735 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 736 f" {len(hypothesis.text)}" 737 ) 738 739 # Correctly process the token ids to chars/subwords. 740 for i, char in enumerate(hypothesis.text): 741 char_offsets[i]["char"] = self.decode_tokens_to_str([char]) 742 743 # detect char vs subword models 744 lens = [len(list(v["char"])) > 1 for v in char_offsets] 745 if any(lens): 746 text_type = 'subword' 747 else: 748 text_type = 'char' 749 750 # retrieve word offsets from character offsets 751 word_offsets = None 752 if timestamp_type in ['word', 'all']: 753 if text_type == 'char': 754 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 755 else: 756 word_offsets = self._get_word_offsets_subwords_sentencepiece( 757 char_offsets, 758 hypothesis, 759 decode_ids_to_tokens=self.decode_ids_to_tokens, 760 decode_tokens_to_str=self.decode_tokens_to_str, 761 ) 762 763 # attach results 764 if len(hypothesis.timestep) > 0: 765 timestep_info = hypothesis.timestep 766 else: 767 timestep_info = [] 768 769 # Setup defaults 770 hypothesis.timestep = {"timestep": timestep_info} 771 772 # Add char / subword time stamps 773 if char_offsets is not None and timestamp_type in ['char', 'all']: 774 hypothesis.timestep['char'] = char_offsets 775 776 # Add word time stamps 777 if word_offsets is not None and timestamp_type in ['word', 'all']: 778 hypothesis.timestep['word'] = word_offsets 779 780 # Convert the token indices to text 781 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 782 783 return hypothesis 784 785 @staticmethod 786 def _compute_offsets( 787 hypothesis: Hypothesis, token_lengths: List[int], ctc_token: int 788 ) -> List[Dict[str, Union[str, int]]]: 789 """ 790 Utility method that calculates the indidual time indices where a token starts and ends. 791 792 Args: 793 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 794 emitted at every time step after ctc collapse. 795 token_lengths: A list of ints representing the lengths of each emitted token. 796 ctc_token: The integer of the ctc blank token used during ctc collapse. 797 798 Returns: 799 800 """ 801 start_index = 0 802 803 # If the exact timestep information is available, utilize the 1st non-ctc blank token timestep 804 # as the start index. 805 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 806 start_index = max(0, hypothesis.timestep[0] - 1) 807 808 # Construct the start and end indices brackets 809 end_indices = np.asarray(token_lengths).cumsum() 810 start_indices = np.concatenate(([start_index], end_indices[:-1])) 811 812 # Merge the results per token into a list of dictionaries 813 offsets = [ 814 {"char": t, "start_offset": s, "end_offset": e} 815 for t, s, e in zip(hypothesis.text, start_indices, end_indices) 816 ] 817 818 # Filter out CTC token 819 offsets = list(filter(lambda offsets: offsets["char"] != ctc_token, offsets)) 820 return offsets 821 822 @staticmethod 823 def _get_word_offsets_chars( 824 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 825 ) -> Dict[str, Union[str, float]]: 826 """ 827 Utility method which constructs word time stamps out of character time stamps. 828 829 References: 830 This code is a port of the Hugging Face code for word time stamp construction. 831 832 Args: 833 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 834 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 835 836 Returns: 837 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 838 "end_offset". 839 """ 840 word_offsets = [] 841 842 last_state = "SPACE" 843 word = "" 844 start_offset = 0 845 end_offset = 0 846 for i, offset in enumerate(offsets): 847 char = offset["char"] 848 state = "SPACE" if char == word_delimiter_char else "WORD" 849 850 if state == last_state: 851 # If we are in the same state as before, we simply repeat what we've done before 852 end_offset = offset["end_offset"] 853 word += char 854 else: 855 # Switching state 856 if state == "SPACE": 857 # Finishing a word 858 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 859 else: 860 # Starting a new word 861 start_offset = offset["start_offset"] 862 end_offset = offset["end_offset"] 863 word = char 864 865 last_state = state 866 if last_state == "WORD": 867 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 868 869 return word_offsets 870 871 @staticmethod 872 def _get_word_offsets_subwords_sentencepiece( 873 offsets: Dict[str, Union[str, float]], 874 hypothesis: Hypothesis, 875 decode_ids_to_tokens: Callable[[List[int]], str], 876 decode_tokens_to_str: Callable[[List[int]], str], 877 ) -> Dict[str, Union[str, float]]: 878 """ 879 Utility method which constructs word time stamps out of sub-word time stamps. 880 881 **Note**: Only supports Sentencepiece based tokenizers ! 882 883 Args: 884 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 885 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 886 after ctc collapse. 887 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 888 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 889 890 Returns: 891 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 892 "end_offset". 893 """ 894 word_offsets = [] 895 built_token = [] 896 previous_token_index = 0 897 # For every collapsed sub-word token 898 for i, char in enumerate(hypothesis.text): 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if len(built_token) > 0: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 if len(word_offsets) == 0: 931 # alaptev: sometimes word_offsets can be empty 932 if len(built_token) > 0: 933 word_offsets.append( 934 { 935 "word": decode_tokens_to_str(built_token), 936 "start_offset": offsets[0]["start_offset"], 937 "end_offset": offsets[-1]["end_offset"], 938 } 939 ) 940 built_token.clear() 941 else: 942 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 943 944 # If there are any remaining tokens left, inject them all into the final word offset. 945 # Note: The start offset of this token is the start time of the first token inside build_token. 946 # Note: The end offset of this token is the end time of the last token inside build_token 947 if len(built_token) > 0: 948 word_offsets.append( 949 { 950 "word": decode_tokens_to_str(built_token), 951 "start_offset": offsets[-(len(built_token))]["start_offset"], 952 "end_offset": offsets[-1]["end_offset"], 953 } 954 ) 955 built_token.clear() 956 957 return word_offsets 958 959 @property 960 def preserve_alignments(self): 961 return self._preserve_alignments 962 963 @preserve_alignments.setter 964 def preserve_alignments(self, value): 965 self._preserve_alignments = value 966 967 if hasattr(self, 'decoding'): 968 self.decoding.preserve_alignments = value 969 970 @property 971 def compute_timestamps(self): 972 return self._compute_timestamps 973 974 @compute_timestamps.setter 975 def compute_timestamps(self, value): 976 self._compute_timestamps = value 977 978 if hasattr(self, 'decoding'): 979 self.decoding.compute_timestamps = value 980 981 @property 982 def preserve_frame_confidence(self): 983 return self._preserve_frame_confidence 984 985 @preserve_frame_confidence.setter 986 def preserve_frame_confidence(self, value): 987 self._preserve_frame_confidence = value 988 989 if hasattr(self, 'decoding'): 990 self.decoding.preserve_frame_confidence = value 991 992 993 class CTCDecoding(AbstractCTCDecoding): 994 """ 995 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs for character 996 based models. 997 998 Args: 999 decoding_cfg: A dict-like object which contains the following key-value pairs. 1000 strategy: str value which represents the type of decoding that can occur. 1001 Possible values are : 1002 - greedy (for greedy decoding). 1003 - beam (for DeepSpeed KenLM based decoding). 1004 1005 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 1006 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 1007 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 1008 1009 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 1010 Can take the following values - "char" for character/subword time stamps, "word" for word level 1011 time stamps and "all" (default), for both character level and word level time stamps. 1012 1013 word_seperator: Str token representing the seperator between words. 1014 1015 preserve_alignments: Bool flag which preserves the history of logprobs generated during 1016 decoding (sample / batched). When set to true, the Hypothesis will contain 1017 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 1018 1019 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 1020 scores. In order to obtain hypotheses with confidence scores, please utilize 1021 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 1022 1023 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 1024 generated during decoding. When set to true, the Hypothesis will contain 1025 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 1026 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 1027 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1028 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 1029 1030 The length of the list corresponds to the number of recognized tokens. 1031 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1032 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1033 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1034 1035 The length of the list corresponds to the number of recognized words. 1036 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1037 from the `token_confidence`. 1038 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1039 Valid options are `mean`, `min`, `max`, `prod`. 1040 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1041 confidence scores. 1042 1043 name: The measure name (str). 1044 Supported values: 1045 - 'max_prob' for using the maximum token probability as a confidence. 1046 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1047 1048 entropy_type: Which type of entropy to use (str). 1049 Used if confidence_measure_cfg.name is set to `entropy`. 1050 Supported values: 1051 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1052 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1053 Note that for this entropy, the alpha should comply the following inequality: 1054 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1055 where V is the model vocabulary size. 1056 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1057 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1058 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1059 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1060 - 'renyi' for the Rรฉnyi entropy. 1061 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1062 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1063 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1064 1065 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1066 When the alpha equals one, scaling is not applied to 'max_prob', 1067 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1068 1069 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1070 Supported values: 1071 - 'lin' for using the linear mapping. 1072 - 'exp' for using exponential mapping with linear shift. 1073 1074 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 1075 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 1076 1077 The config may further contain the following sub-dictionaries: 1078 "greedy": 1079 preserve_alignments: Same as above, overrides above value. 1080 compute_timestamps: Same as above, overrides above value. 1081 preserve_frame_confidence: Same as above, overrides above value. 1082 confidence_measure_cfg: Same as above, overrides confidence_cfg.measure_cfg. 1083 1084 "beam": 1085 beam_size: int, defining the beam size for beam search. Must be >= 1. 1086 If beam_size == 1, will perform cached greedy search. This might be slightly different 1087 results compared to the greedy search above. 1088 1089 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1090 hypotheses after beam search has concluded. This flag is set by default. 1091 1092 beam_alpha: float, the strength of the Language model on the final score of a token. 1093 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1094 1095 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 1096 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1097 1098 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 1099 If the path is invalid (file is not found at path), will raise a deferred error at the moment 1100 of calculation of beam search, so that users may update / change the decoding strategy 1101 to point to the correct file. 1102 1103 blank_id: The id of the RNNT blank token. 1104 """ 1105 1106 def __init__( 1107 self, decoding_cfg, vocabulary, 1108 ): 1109 blank_id = len(vocabulary) 1110 self.vocabulary = vocabulary 1111 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1112 1113 super().__init__(decoding_cfg=decoding_cfg, blank_id=blank_id) 1114 1115 # Finalize Beam Search Decoding framework 1116 if isinstance(self.decoding, ctc_beam_decoding.AbstractBeamCTCInfer): 1117 self.decoding.set_vocabulary(self.vocabulary) 1118 self.decoding.set_decoding_type('char') 1119 1120 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1121 """ 1122 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1123 1124 Args: 1125 hypothesis: Hypothesis 1126 1127 Returns: 1128 A list of word-level confidence scores. 1129 """ 1130 return self._aggregate_token_confidence_chars( 1131 self.decode_tokens_to_str(hypothesis.text[0]).split(), hypothesis.token_confidence 1132 ) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c != self.blank_id] 1159 return token_list 1160 1161 1162 class WER(Metric): 1163 """ 1164 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference 1165 texts. When doing distributed training/evaluation the result of ``res=WER(predictions, targets, target_lengths)`` 1166 calls will be all-reduced between all workers using SUM operations. Here ``res`` contains three numbers 1167 ``res=[wer, total_levenstein_distance, total_number_of_words]``. 1168 1169 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step 1170 results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1171 1172 Example: 1173 def validation_step(self, batch, batch_idx): 1174 ... 1175 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1176 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1177 return self.val_outputs 1178 1179 def on_validation_epoch_end(self): 1180 ... 1181 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1182 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1183 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1184 self.val_outputs.clear() # free memory 1185 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1186 1187 Args: 1188 decoding: An instance of CTCDecoding. 1189 use_cer: Whether to use Character Error Rate instead of Word Error Rate. 1190 log_prediction: Whether to log a single decoded sample per call. 1191 fold_consecutive: Whether repeated consecutive characters should be folded into one when decoding. 1192 1193 Returns: 1194 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenstein's 1195 distances for all prediction - reference pairs, total number of words in all references. 1196 """ 1197 1198 full_state_update: bool = True 1199 1200 def __init__( 1201 self, 1202 decoding: CTCDecoding, 1203 use_cer=False, 1204 log_prediction=True, 1205 fold_consecutive=True, 1206 dist_sync_on_step=False, 1207 ): 1208 super().__init__(dist_sync_on_step=dist_sync_on_step) 1209 1210 self.decoding = decoding 1211 self.use_cer = use_cer 1212 self.log_prediction = log_prediction 1213 self.fold_consecutive = fold_consecutive 1214 1215 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1216 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1217 1218 def update( 1219 self, 1220 predictions: torch.Tensor, 1221 targets: torch.Tensor, 1222 target_lengths: torch.Tensor, 1223 predictions_lengths: torch.Tensor = None, 1224 ): 1225 """ 1226 Updates metric state. 1227 Args: 1228 predictions: an integer torch.Tensor of shape ``[Batch, Time, {Vocabulary}]`` (if ``batch_dim_index == 0``) or 1229 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1230 targets: an integer torch.Tensor of shape ``[Batch, Time]`` (if ``batch_dim_index == 0``) or 1231 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1232 target_lengths: an integer torch.Tensor of shape ``[Batch]`` 1233 predictions_lengths: an integer torch.Tensor of shape ``[Batch]`` 1234 """ 1235 words = 0 1236 scores = 0 1237 references = [] 1238 with torch.no_grad(): 1239 # prediction_cpu_tensor = tensors[0].long().cpu() 1240 targets_cpu_tensor = targets.long().cpu() 1241 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1242 1243 # iterate over batch 1244 for ind in range(targets_cpu_tensor.shape[0]): 1245 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1246 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1247 reference = self.decoding.decode_tokens_to_str(target) 1248 references.append(reference) 1249 1250 hypotheses, _ = self.decoding.ctc_decoder_predictions_tensor( 1251 predictions, predictions_lengths, fold_consecutive=self.fold_consecutive 1252 ) 1253 1254 if self.log_prediction: 1255 logging.info(f"\n") 1256 logging.info(f"reference:{references[0]}") 1257 logging.info(f"predicted:{hypotheses[0]}") 1258 1259 for h, r in zip(hypotheses, references): 1260 if self.use_cer: 1261 h_list = list(h) 1262 r_list = list(r) 1263 else: 1264 h_list = h.split() 1265 r_list = r.split() 1266 words += len(r_list) 1267 # Compute Levenstein's distance 1268 scores += editdistance.eval(h_list, r_list) 1269 1270 self.scores = torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1271 self.words = torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1272 # return torch.tensor([scores, words]).to(predictions.device) 1273 1274 def compute(self): 1275 scores = self.scores.detach().float() 1276 words = self.words.detach().float() 1277 return scores / words, scores, words 1278 1279 1280 @dataclass 1281 class CTCDecodingConfig: 1282 strategy: str = "greedy" 1283 1284 # preserve decoding alignments 1285 preserve_alignments: Optional[bool] = None 1286 1287 # compute ctc time stamps 1288 compute_timestamps: Optional[bool] = None 1289 1290 # token representing word seperator 1291 word_seperator: str = " " 1292 1293 # type of timestamps to calculate 1294 ctc_timestamp_type: str = "all" # can be char, word or all for both 1295 1296 # batch dimension 1297 batch_dim_index: int = 0 1298 1299 # greedy decoding config 1300 greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() 1301 1302 # beam decoding config 1303 beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) 1304 1305 # confidence config 1306 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1307 1308 # can be used to change temperature for decoding 1309 temperature: float = 1.0 1310 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/asr/models/configs/aligner_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig 18 19 20 @dataclass 21 class AlignerCTCConfig: 22 prob_suppress_index: int = -1 23 prob_suppress_value: float = 1.0 24 25 26 @dataclass 27 class AlignerRNNTConfig: 28 predictor_window_size: int = 0 29 predictor_step_size: int = 1 30 31 32 @dataclass 33 class AlignerWrapperModelConfig: 34 alignment_type: str = "forced" 35 word_output: bool = True 36 cpu_decoding: bool = False 37 decode_batch_size: int = 0 38 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() 39 rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() 40 41 42 @dataclass 43 class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): 44 decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() 45 [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 22 from nemo.collections.asr.modules.audio_preprocessing import ( 23 AudioToMelSpectrogramPreprocessorConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class ASRDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[Any] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[Any] = None 40 tarred_shard_strategy: str = "scatter" 41 shard_manifests: bool = False 42 shuffle_n: int = 0 43 44 # Optional 45 int_values: Optional[int] = None 46 augmentor: Optional[Dict[str, Any]] = None 47 max_duration: Optional[float] = None 48 min_duration: Optional[float] = None 49 max_utts: int = 0 50 blank_index: int = -1 51 unk_index: int = -1 52 normalize: bool = False 53 trim: bool = True 54 parser: Optional[str] = 'en' 55 eos_id: Optional[int] = None 56 bos_id: Optional[int] = None 57 pad_id: int = 0 58 use_start_end_token: bool = False 59 return_sample_id: Optional[bool] = False 60 61 # bucketing params 62 bucketing_strategy: str = "synced_randomized" 63 bucketing_batch_size: Optional[Any] = None 64 bucketing_weights: Optional[List[int]] = None 65 66 67 @dataclass 68 class EncDecCTCConfig(model_cfg.ModelConfig): 69 # Model global arguments 70 sample_rate: int = 16000 71 repeat: int = 1 72 dropout: float = 0.0 73 separable: bool = False 74 labels: List[str] = MISSING 75 76 # Dataset configs 77 train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) 78 validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 79 test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 80 81 # Optimizer / Scheduler config 82 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 83 84 # Model component configs 85 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 86 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 87 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 88 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 89 decoding: CTCDecodingConfig = CTCDecodingConfig() 90 91 92 @dataclass 93 class EncDecCTCModelConfig(model_cfg.NemoConfig): 94 model: EncDecCTCConfig = EncDecCTCConfig() 95 96 97 @dataclass 98 class CacheAwareStreamingConfig: 99 chunk_size: int = 0 # the size of each chunk at each step, it can be a list of two integers to specify different chunk sizes for the first step and others 100 shift_size: int = 0 # the size of the shift in each step, it can be a list of two integers to specify different shift sizes for the first step and others 101 102 cache_drop_size: int = 0 # the number of steps to drop from the cache 103 last_channel_cache_size: int = 0 # the size of the needed cache for last channel layers 104 105 valid_out_len: int = 0 # the number of the steps in the final output which are valid (have the same value as in the offline mode) 106 107 pre_encode_cache_size: int = 0 # the size of the needed cache for the pre-encoding part of the model to avoid caching inside the pre-encoding layers 108 drop_extra_pre_encoded: int = 0 # the number of steps to get dropped after the pre-encoding layer 109 110 last_channel_num: int = 0 # number of the last channel layers (like MHA layers) which need caching in the model 111 last_time_num: int = 0 # number of the last time layers (like convolutions) which need caching in the model 112 [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderClassificationConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class EncDecClassificationDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[str] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[str] = None 40 tarred_shard_strategy: str = "scatter" 41 shuffle_n: int = 0 42 43 # Optional 44 int_values: Optional[int] = None 45 augmentor: Optional[Dict[str, Any]] = None 46 max_duration: Optional[float] = None 47 min_duration: Optional[float] = None 48 cal_labels_occurrence: Optional[bool] = False 49 50 # VAD Optional 51 vad_stream: Optional[bool] = None 52 window_length_in_sec: float = 0.31 53 shift_length_in_sec: float = 0.01 54 normalize_audio: bool = False 55 is_regression_task: bool = False 56 57 # bucketing params 58 bucketing_strategy: str = "synced_randomized" 59 bucketing_batch_size: Optional[Any] = None 60 bucketing_weights: Optional[List[int]] = None 61 62 63 @dataclass 64 class EncDecClassificationConfig(model_cfg.ModelConfig): 65 # Model global arguments 66 sample_rate: int = 16000 67 repeat: int = 1 68 dropout: float = 0.0 69 separable: bool = True 70 kernel_size_factor: float = 1.0 71 labels: List[str] = MISSING 72 timesteps: int = MISSING 73 74 # Dataset configs 75 train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 76 manifest_filepath=None, shuffle=True, trim_silence=False 77 ) 78 validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 79 manifest_filepath=None, shuffle=False 80 ) 81 test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 82 manifest_filepath=None, shuffle=False 83 ) 84 85 # Optimizer / Scheduler config 86 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 87 88 # Model component configs 89 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() 90 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 91 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 92 audio_length=timesteps 93 ) 94 95 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 96 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 97 98 99 @dataclass 100 class EncDecClassificationModelConfig(model_cfg.NemoConfig): 101 model: EncDecClassificationConfig = EncDecClassificationConfig() 102 [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import asdict, dataclass 16 from typing import Any, Dict, Optional, Tuple, Union 17 18 19 @dataclass 20 class DiarizerComponentConfig: 21 """Dataclass to imitate HydraConfig dict when accessing parameters.""" 22 23 def get(self, name: str, default: Optional[Any] = None): 24 return getattr(self, name, default) 25 26 def __iter__(self): 27 for key in asdict(self): 28 yield key 29 30 def dict(self) -> Dict: 31 return asdict(self) 32 33 34 @dataclass 35 class ASRDiarizerCTCDecoderParams: 36 pretrained_language_model: Optional[str] = None # KenLM model file: .arpa model file or .bin binary file. 37 beam_width: int = 32 38 alpha: float = 0.5 39 beta: float = 2.5 40 41 42 @dataclass 43 class ASRRealigningLMParams: 44 # Provide a KenLM language model in .arpa format. 45 arpa_language_model: Optional[str] = None 46 # Min number of words for the left context. 47 min_number_of_words: int = 3 48 # Max number of words for the right context. 49 max_number_of_words: int = 10 50 # The threshold for the difference between two log probability values from two hypotheses. 51 logprob_diff_threshold: float = 1.2 52 53 54 @dataclass 55 class ASRDiarizerParams(DiarizerComponentConfig): 56 # if True, speech segmentation for diarization is based on word-timestamps from ASR inference. 57 asr_based_vad: bool = False 58 # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD. 59 asr_based_vad_threshold: float = 1.0 60 # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null. 61 asr_batch_size: Optional[int] = None 62 # Native decoder delay. null is recommended to use the default values for each ASR model. 63 decoder_delay_in_sec: Optional[float] = None 64 # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. 65 word_ts_anchor_offset: Optional[float] = None 66 # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. 67 word_ts_anchor_pos: str = "start" 68 # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. 69 fix_word_ts_with_VAD: bool = False 70 # If True, use colored text to distinguish speakers in the output transcript. 71 colored_text: bool = False 72 # If True, the start and end time of each speaker turn is printed in the output transcript. 73 print_time: bool = True 74 # If True, the output transcript breaks the line to fix the line width (default is 90 chars) 75 break_lines: bool = False 76 77 78 @dataclass 79 class ASRDiarizerConfig(DiarizerComponentConfig): 80 model_path: Optional[str] = "stt_en_conformer_ctc_large" 81 parameters: ASRDiarizerParams = ASRDiarizerParams() 82 ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() 83 realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() 84 85 86 @dataclass 87 class VADParams(DiarizerComponentConfig): 88 window_length_in_sec: float = 0.15 # Window length in sec for VAD context input 89 shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction 90 smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) 91 overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter 92 onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech 93 offset: float = 0.1 # Offset threshold for detecting the end of a speech 94 pad_onset: float = 0.1 # Adding durations before each speech segment 95 pad_offset: float = 0 # Adding durations after each speech segment 96 min_duration_on: float = 0 # Threshold for small non_speech deletion 97 min_duration_off: float = 0.2 # Threshold for short speech segment deletion 98 filter_speech_first: bool = True 99 100 101 @dataclass 102 class VADConfig(DiarizerComponentConfig): 103 model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name 104 external_vad_manifest: Optional[str] = None 105 parameters: VADParams = VADParams() 106 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): 110 # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] 111 window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) 112 # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] 113 shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) 114 # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] 115 multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) 116 # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. 117 save_embeddings: bool = True 118 119 120 @dataclass 121 class SpeakerEmbeddingsConfig(DiarizerComponentConfig): 122 # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) 123 model_path: Optional[str] = None 124 parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() 125 126 127 @dataclass 128 class ClusteringParams(DiarizerComponentConfig): 129 # If True, use num of speakers value provided in manifest file. 130 oracle_num_speakers: bool = False 131 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. 132 max_num_speakers: int = 8 133 # If the number of segments is lower than this number, enhanced speaker counting is activated. 134 enhanced_count_thres: int = 80 135 # Determines the range of p-value search: 0 < p <= max_rp_threshold. 136 max_rp_threshold: float = 0.25 137 # The higher the number, the more values will be examined with more time. 138 sparse_search_volume: int = 30 139 # If True, take a majority vote on multiple p-values to estimate the number of speakers. 140 maj_vote_spk_count: bool = False 141 142 143 @dataclass 144 class ClusteringConfig(DiarizerComponentConfig): 145 parameters: ClusteringParams = ClusteringParams() 146 147 148 @dataclass 149 class MSDDParams(DiarizerComponentConfig): 150 # If True, use speaker embedding model in checkpoint, else provided speaker embedding model in config will be used. 151 use_speaker_model_from_ckpt: bool = True 152 # Batch size for MSDD inference. 153 infer_batch_size: int = 25 154 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. 155 sigmoid_threshold: Tuple[float] = (0.7,) 156 # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. 157 seq_eval_mode: bool = False 158 # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. 159 split_infer: bool = True 160 # The length of split short sequence when split_infer is True. 161 diar_window_length: int = 50 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): 174 manifest_filepath: Optional[str] = None 175 out_dir: Optional[str] = None 176 oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps 177 collar: float = 0.25 # Collar value for scoring 178 ignore_overlap: bool = True # Consider or ignore overlap segments while scoring 179 vad: VADConfig = VADConfig() 180 speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() 181 clustering: ClusteringConfig = ClusteringConfig() 182 msdd_model: MSDDConfig = MSDDConfig() 183 asr: ASRDiarizerConfig = ASRDiarizerConfig() 184 185 186 @dataclass 187 class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): 188 diarizer: DiarizerConfig = DiarizerConfig() 189 device: str = "cpu" 190 verbose: bool = False 191 batch_size: int = 64 192 num_workers: int = 1 193 sample_rate: int = 16000 194 name: str = "" 195 196 @classmethod 197 def init_config(cls, diar_model_path: str, vad_model_path: str, map_location: str, verbose: bool): 198 return NeuralDiarizerInferenceConfig( 199 DiarizerConfig( 200 vad=VADConfig(model_path=vad_model_path), msdd_model=MSDDConfig(model_path=diar_model_path), 201 ), 202 device=map_location, 203 verbose=verbose, 204 ) 205 [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig 18 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig 19 from nemo.core.config.modelPT import NemoConfig 20 21 22 @dataclass 23 class GraphModuleConfig: 24 criterion_type: str = "ml" 25 loss_type: str = "ctc" 26 split_batch_size: int = 0 27 dec_type: str = "topo" 28 transcribe_training: bool = True 29 backend_cfg: BackendConfig = BackendConfig() 30 31 32 @dataclass 33 class EncDecK2SeqConfig(EncDecCTCConfig): 34 graph_module_cfg: GraphModuleConfig = GraphModuleConfig() 35 36 37 @dataclass 38 class EncDecK2SeqModelConfig(NemoConfig): 39 model: EncDecK2SeqConfig = EncDecK2SeqConfig() 40 [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import classification_models_config as clf_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ( 27 ConvASRDecoderClassificationConfig, 28 ConvASREncoderConfig, 29 JasperEncoderConfig, 30 ) 31 from nemo.core.config import modelPT as model_cfg 32 33 34 # fmt: off 35 def matchboxnet_3x1x64(): 36 config = [ 37 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 38 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 39 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 40 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 41 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 42 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 43 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 44 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 45 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 46 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 47 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 48 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 49 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 50 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 51 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 52 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 53 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 54 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 55 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 56 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 57 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 58 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 59 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 60 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 61 ] 62 return config 63 64 65 def matchboxnet_3x1x64_vad(): 66 config = [ 67 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 68 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 69 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 70 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 71 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 72 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 73 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 74 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 75 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 76 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 77 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 78 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 79 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 80 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 81 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 82 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 83 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 84 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 85 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 86 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 87 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 88 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 89 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 90 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 91 ] 92 return config 93 94 95 # fmt: on 96 97 98 @dataclass 99 class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): 100 # Model global arguments 101 sample_rate: int = 16000 102 repeat: int = 1 103 dropout: float = 0.0 104 separable: bool = True 105 kernel_size_factor: float = 1.0 106 timesteps: int = 128 107 labels: List[str] = MISSING 108 109 # Dataset configs 110 train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 111 manifest_filepath=None, shuffle=True, trim_silence=False 112 ) 113 validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 114 manifest_filepath=None, shuffle=False 115 ) 116 test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 117 manifest_filepath=None, shuffle=False 118 ) 119 120 # Optimizer / Scheduler config 121 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 122 123 # Model general component configs 124 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) 125 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( 126 freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 127 ) 128 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 129 audio_length=128 130 ) 131 132 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 133 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 134 135 136 @dataclass 137 class MatchboxNetVADModelConfig(MatchboxNetModelConfig): 138 timesteps: int = 64 139 labels: List[str] = field(default_factory=lambda: ['background', 'speech']) 140 141 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = None 142 143 144 class EncDecClassificationModelConfigBuilder(model_cfg.ModelConfigBuilder): 145 VALID_CONFIGS = ['matchboxnet_3x1x64', 'matchboxnet_3x1x64_vad'] 146 147 def __init__(self, name: str = 'matchboxnet_3x1x64', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 148 if name not in EncDecClassificationModelConfigBuilder.VALID_CONFIGS: 149 raise ValueError("`name` must be one of : \n" f"{EncDecClassificationModelConfigBuilder.VALID_CONFIGS}") 150 151 self.name = name 152 153 if 'matchboxnet_3x1x64_vad' in name: 154 if encoder_cfg_func is None: 155 encoder_cfg_func = matchboxnet_3x1x64_vad 156 157 model_cfg = MatchboxNetVADModelConfig( 158 repeat=1, 159 separable=True, 160 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 161 decoder=ConvASRDecoderClassificationConfig(), 162 ) 163 164 elif 'matchboxnet_3x1x64' in name: 165 if encoder_cfg_func is None: 166 encoder_cfg_func = matchboxnet_3x1x64 167 168 model_cfg = MatchboxNetModelConfig( 169 repeat=1, 170 separable=False, 171 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 172 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 173 decoder=ConvASRDecoderClassificationConfig(), 174 ) 175 176 else: 177 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 178 179 super(EncDecClassificationModelConfigBuilder, self).__init__(model_cfg) 180 self.model_cfg: clf_cfg.EncDecClassificationConfig = model_cfg # enable type hinting 181 182 def set_labels(self, labels: List[str]): 183 self.model_cfg.labels = labels 184 185 def set_separable(self, separable: bool): 186 self.model_cfg.separable = separable 187 188 def set_repeat(self, repeat: int): 189 self.model_cfg.repeat = repeat 190 191 def set_sample_rate(self, sample_rate: int): 192 self.model_cfg.sample_rate = sample_rate 193 194 def set_dropout(self, dropout: float = 0.0): 195 self.model_cfg.dropout = dropout 196 197 def set_timesteps(self, timesteps: int): 198 self.model_cfg.timesteps = timesteps 199 200 def set_is_regression_task(self, is_regression_task: bool): 201 self.model_cfg.is_regression_task = is_regression_task 202 203 # Note: Autocomplete for users wont work without these overrides 204 # But practically it is not needed since python will infer at runtime 205 206 # def set_train_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 207 # super().set_train_ds(cfg) 208 # 209 # def set_validation_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 210 # super().set_validation_ds(cfg) 211 # 212 # def set_test_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 213 # super().set_test_ds(cfg) 214 215 def _finalize_cfg(self): 216 # propagate labels 217 self.model_cfg.train_ds.labels = self.model_cfg.labels 218 self.model_cfg.validation_ds.labels = self.model_cfg.labels 219 self.model_cfg.test_ds.labels = self.model_cfg.labels 220 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 221 222 # propagate num classes 223 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 224 225 # propagate sample rate 226 self.model_cfg.sample_rate = self.model_cfg.sample_rate 227 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 228 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 229 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 230 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 231 232 # propagate filters 233 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 234 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 235 236 # propagate timeteps 237 if self.model_cfg.crop_or_pad_augment is not None: 238 self.model_cfg.crop_or_pad_augment.audio_length = self.model_cfg.timesteps 239 240 # propagate separable 241 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 242 layer.separable = self.model_cfg.separable 243 244 # propagate repeat 245 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 246 layer.repeat = self.model_cfg.repeat 247 248 # propagate dropout 249 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 250 layer.dropout = self.model_cfg.dropout 251 252 def build(self) -> clf_cfg.EncDecClassificationConfig: 253 return super().build() 254 [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import asr_models_config as ctc_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMelSpectrogramPreprocessorConfig, 23 SpectrogramAugmentationConfig, 24 ) 25 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig, JasperEncoderConfig 26 from nemo.core.config import modelPT as model_cfg 27 28 29 # fmt: off 30 def qn_15x5(): 31 config = [ 32 JasperEncoderConfig(filters=256, repeat=1, kernel=[33], stride=[2], dilation=[1], dropout=0.0, 33 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 34 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 35 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 36 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 37 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 38 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 39 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 40 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 41 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 42 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 43 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 44 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 45 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 46 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 47 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 48 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 49 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 50 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 51 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 52 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 53 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 54 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 55 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 56 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 57 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 58 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 59 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 60 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 61 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 62 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 63 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 64 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 65 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 66 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 67 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 68 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 69 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 70 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 71 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 72 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 73 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 74 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 75 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 76 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 77 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 78 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 79 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 80 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 81 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 82 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 83 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 84 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 85 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 86 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 87 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 88 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 89 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 90 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 91 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 92 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 93 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 94 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 95 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 96 JasperEncoderConfig(filters=512, repeat=1, kernel=[87], stride=[1], dilation=[2], dropout=0.0, 97 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 98 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 99 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 100 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 101 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 102 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 103 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 104 ] 105 return config 106 107 108 def jasper_10x5_dr(): 109 config = [ 110 JasperEncoderConfig(filters=256, repeat=1, kernel=[11], stride=[2], dilation=[1], dropout=0.2, 111 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 112 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 113 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 114 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 115 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 116 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 117 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 118 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 119 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 120 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 121 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 122 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 123 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 124 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 125 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 126 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 127 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 128 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 129 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 130 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 131 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 132 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 133 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 134 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 135 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 136 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 137 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 138 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 139 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 140 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 141 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 142 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 143 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 144 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 145 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 146 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 147 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 148 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 149 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 150 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 151 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 152 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 153 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 154 JasperEncoderConfig(filters=896, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.4, 155 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 156 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 157 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 158 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.4, 159 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 160 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 161 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 162 ] 163 return config 164 # fmt: on 165 166 167 @dataclass 168 class JasperModelConfig(ctc_cfg.EncDecCTCConfig): 169 # Model global arguments 170 sample_rate: int = 16000 171 repeat: int = 1 172 dropout: float = 0.0 173 separable: bool = False 174 labels: List[str] = MISSING 175 176 # Dataset configs 177 train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( 178 manifest_filepath=None, shuffle=True, trim_silence=True 179 ) 180 validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 181 test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 182 183 # Optimizer / Scheduler config 184 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 185 186 # Model general component configs 187 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 188 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 189 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 190 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 191 192 193 @dataclass 194 class QuartzNetModelConfig(JasperModelConfig): 195 separable: bool = True 196 197 198 class EncDecCTCModelConfigBuilder(model_cfg.ModelConfigBuilder): 199 VALID_CONFIGS = ['quartznet_15x5', 'quartznet_15x5_zh', 'jasper_10x5dr'] 200 201 def __init__(self, name: str = 'quartznet_15x5', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 202 if name not in EncDecCTCModelConfigBuilder.VALID_CONFIGS: 203 raise ValueError("`name` must be one of : \n" f"{EncDecCTCModelConfigBuilder.VALID_CONFIGS}") 204 205 self.name = name 206 207 if 'quartznet_15x5' in name: 208 if encoder_cfg_func is None: 209 encoder_cfg_func = qn_15x5 210 211 model_cfg = QuartzNetModelConfig( 212 repeat=5, 213 separable=True, 214 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 215 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 216 decoder=ConvASRDecoderConfig(), 217 ) 218 219 elif 'jasper_10x5' in name: 220 if encoder_cfg_func is None: 221 encoder_cfg_func = jasper_10x5_dr 222 223 model_cfg = JasperModelConfig( 224 repeat=5, 225 separable=False, 226 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 227 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 228 decoder=ConvASRDecoderConfig(), 229 ) 230 231 else: 232 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 233 234 super(EncDecCTCModelConfigBuilder, self).__init__(model_cfg) 235 self.model_cfg: ctc_cfg.EncDecCTCConfig = model_cfg # enable type hinting 236 237 if 'zh' in name: 238 self.set_dataset_normalize(normalize=False) 239 240 def set_labels(self, labels: List[str]): 241 self.model_cfg.labels = labels 242 243 def set_separable(self, separable: bool): 244 self.model_cfg.separable = separable 245 246 def set_repeat(self, repeat: int): 247 self.model_cfg.repeat = repeat 248 249 def set_sample_rate(self, sample_rate: int): 250 self.model_cfg.sample_rate = sample_rate 251 252 def set_dropout(self, dropout: float = 0.0): 253 self.model_cfg.dropout = dropout 254 255 def set_dataset_normalize(self, normalize: bool): 256 self.model_cfg.train_ds.normalize = normalize 257 self.model_cfg.validation_ds.normalize = normalize 258 self.model_cfg.test_ds.normalize = normalize 259 260 # Note: Autocomplete for users wont work without these overrides 261 # But practically it is not needed since python will infer at runtime 262 263 # def set_train_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 264 # super().set_train_ds(cfg) 265 # 266 # def set_validation_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 267 # super().set_validation_ds(cfg) 268 # 269 # def set_test_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 270 # super().set_test_ds(cfg) 271 272 def _finalize_cfg(self): 273 # propagate labels 274 self.model_cfg.train_ds.labels = self.model_cfg.labels 275 self.model_cfg.validation_ds.labels = self.model_cfg.labels 276 self.model_cfg.test_ds.labels = self.model_cfg.labels 277 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 278 279 # propagate num classes 280 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 281 282 # propagate sample rate 283 self.model_cfg.sample_rate = self.model_cfg.sample_rate 284 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 285 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 286 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 287 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 288 289 # propagate filters 290 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 291 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 292 293 # propagate separable 294 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 295 layer.separable = self.model_cfg.separable 296 297 # propagate repeat 298 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 299 layer.repeat = self.model_cfg.repeat 300 301 # propagate dropout 302 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 303 layer.dropout = self.model_cfg.dropout 304 305 def build(self) -> ctc_cfg.EncDecCTCConfig: 306 return super().build() 307 [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import random 17 from abc import ABC, abstractmethod 18 from dataclasses import dataclass 19 from typing import Any, Dict, Optional, Tuple 20 21 import torch 22 from packaging import version 23 24 from nemo.collections.asr.parts.numba.spec_augment import SpecAugmentNumba, spec_augment_launch_heuristics 25 from nemo.collections.asr.parts.preprocessing.features import ( 26 FilterbankFeatures, 27 FilterbankFeaturesTA, 28 make_seq_mask_like, 29 ) 30 from nemo.collections.asr.parts.submodules.spectr_augment import SpecAugment, SpecCutout 31 from nemo.core.classes import Exportable, NeuralModule, typecheck 32 from nemo.core.neural_types import ( 33 AudioSignal, 34 LengthsType, 35 MelSpectrogramType, 36 MFCCSpectrogramType, 37 NeuralType, 38 SpectrogramType, 39 ) 40 from nemo.core.utils import numba_utils 41 from nemo.core.utils.numba_utils import __NUMBA_MINIMUM_VERSION__ 42 from nemo.utils import logging 43 44 try: 45 import torchaudio 46 import torchaudio.functional 47 import torchaudio.transforms 48 49 TORCHAUDIO_VERSION = version.parse(torchaudio.__version__) 50 TORCHAUDIO_VERSION_MIN = version.parse('0.5') 51 52 HAVE_TORCHAUDIO = True 53 except ModuleNotFoundError: 54 HAVE_TORCHAUDIO = False 55 56 __all__ = [ 57 'AudioToMelSpectrogramPreprocessor', 58 'AudioToSpectrogram', 59 'SpectrogramToAudio', 60 'AudioToMFCCPreprocessor', 61 'SpectrogramAugmentation', 62 'MaskedPatchAugmentation', 63 'CropOrPadSpectrogramAugmentation', 64 ] 65 66 67 class AudioPreprocessor(NeuralModule, ABC): 68 """ 69 An interface for Neural Modules that performs audio pre-processing, 70 transforming the wav files to features. 71 """ 72 73 def __init__(self, win_length, hop_length): 74 super().__init__() 75 76 self.win_length = win_length 77 self.hop_length = hop_length 78 79 self.torch_windows = { 80 'hann': torch.hann_window, 81 'hamming': torch.hamming_window, 82 'blackman': torch.blackman_window, 83 'bartlett': torch.bartlett_window, 84 'ones': torch.ones, 85 None: torch.ones, 86 } 87 88 @typecheck() 89 @torch.no_grad() 90 def forward(self, input_signal, length): 91 processed_signal, processed_length = self.get_features(input_signal, length) 92 93 return processed_signal, processed_length 94 95 @abstractmethod 96 def get_features(self, input_signal, length): 97 # Called by forward(). Subclasses should implement this. 98 pass 99 100 101 class AudioToMelSpectrogramPreprocessor(AudioPreprocessor, Exportable): 102 """Featurizer module that converts wavs to mel spectrograms. 103 104 Args: 105 sample_rate (int): Sample rate of the input audio data. 106 Defaults to 16000 107 window_size (float): Size of window for fft in seconds 108 Defaults to 0.02 109 window_stride (float): Stride of window for fft in seconds 110 Defaults to 0.01 111 n_window_size (int): Size of window for fft in samples 112 Defaults to None. Use one of window_size or n_window_size. 113 n_window_stride (int): Stride of window for fft in samples 114 Defaults to None. Use one of window_stride or n_window_stride. 115 window (str): Windowing function for fft. can be one of ['hann', 116 'hamming', 'blackman', 'bartlett'] 117 Defaults to "hann" 118 normalize (str): Can be one of ['per_feature', 'all_features']; all 119 other options disable feature normalization. 'all_features' 120 normalizes the entire spectrogram to be mean 0 with std 1. 121 'pre_features' normalizes per channel / freq instead. 122 Defaults to "per_feature" 123 n_fft (int): Length of FT window. If None, it uses the smallest power 124 of 2 that is larger than n_window_size. 125 Defaults to None 126 preemph (float): Amount of pre emphasis to add to audio. Can be 127 disabled by passing None. 128 Defaults to 0.97 129 features (int): Number of mel spectrogram freq bins to output. 130 Defaults to 64 131 lowfreq (int): Lower bound on mel basis in Hz. 132 Defaults to 0 133 highfreq (int): Lower bound on mel basis in Hz. 134 Defaults to None 135 log (bool): Log features. 136 Defaults to True 137 log_zero_guard_type(str): Need to avoid taking the log of zero. There 138 are two options: "add" or "clamp". 139 Defaults to "add". 140 log_zero_guard_value(float, or str): Add or clamp requires the number 141 to add with or clamp to. log_zero_guard_value can either be a float 142 or "tiny" or "eps". torch.finfo is used if "tiny" or "eps" is 143 passed. 144 Defaults to 2**-24. 145 dither (float): Amount of white-noise dithering. 146 Defaults to 1e-5 147 pad_to (int): Ensures that the output size of the time dimension is 148 a multiple of pad_to. 149 Defaults to 16 150 frame_splicing (int): Defaults to 1 151 exact_pad (bool): If True, sets stft center to False and adds padding, such that num_frames = audio_length 152 // hop_length. Defaults to False. 153 pad_value (float): The value that shorter mels are padded with. 154 Defaults to 0 155 mag_power (float): The power that the linear spectrogram is raised to 156 prior to multiplication with mel basis. 157 Defaults to 2 for a power spec 158 rng : Random number generator 159 nb_augmentation_prob (float) : Probability with which narrowband augmentation would be applied to 160 samples in the batch. 161 Defaults to 0.0 162 nb_max_freq (int) : Frequency above which all frequencies will be masked for narrowband augmentation. 163 Defaults to 4000 164 use_torchaudio: Whether to use the `torchaudio` implementation. 165 mel_norm: Normalization used for mel filterbank weights. 166 Defaults to 'slaney' (area normalization) 167 stft_exact_pad: Deprecated argument, kept for compatibility with older checkpoints. 168 stft_conv: Deprecated argument, kept for compatibility with older checkpoints. 169 """ 170 171 def save_to(self, save_path: str): 172 pass 173 174 @classmethod 175 def restore_from(cls, restore_path: str): 176 pass 177 178 @property 179 def input_types(self): 180 """Returns definitions of module input ports. 181 """ 182 return { 183 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 184 "length": NeuralType( 185 tuple('B'), LengthsType() 186 ), # Please note that length should be in samples not seconds. 187 } 188 189 @property 190 def output_types(self): 191 """Returns definitions of module output ports. 192 193 processed_signal: 194 0: AxisType(BatchTag) 195 1: AxisType(MelSpectrogramSignalTag) 196 2: AxisType(ProcessedTimeTag) 197 processed_length: 198 0: AxisType(BatchTag) 199 """ 200 return { 201 "processed_signal": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 202 "processed_length": NeuralType(tuple('B'), LengthsType()), 203 } 204 205 def __init__( 206 self, 207 sample_rate=16000, 208 window_size=0.02, 209 window_stride=0.01, 210 n_window_size=None, 211 n_window_stride=None, 212 window="hann", 213 normalize="per_feature", 214 n_fft=None, 215 preemph=0.97, 216 features=64, 217 lowfreq=0, 218 highfreq=None, 219 log=True, 220 log_zero_guard_type="add", 221 log_zero_guard_value=2 ** -24, 222 dither=1e-5, 223 pad_to=16, 224 frame_splicing=1, 225 exact_pad=False, 226 pad_value=0, 227 mag_power=2.0, 228 rng=None, 229 nb_augmentation_prob=0.0, 230 nb_max_freq=4000, 231 use_torchaudio: bool = False, 232 mel_norm="slaney", 233 stft_exact_pad=False, # Deprecated arguments; kept for config compatibility 234 stft_conv=False, # Deprecated arguments; kept for config compatibility 235 ): 236 super().__init__(n_window_size, n_window_stride) 237 238 self._sample_rate = sample_rate 239 if window_size and n_window_size: 240 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 241 if window_stride and n_window_stride: 242 raise ValueError( 243 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 244 ) 245 if window_size: 246 n_window_size = int(window_size * self._sample_rate) 247 if window_stride: 248 n_window_stride = int(window_stride * self._sample_rate) 249 250 # Given the long and similar argument list, point to the class and instantiate it by reference 251 if not use_torchaudio: 252 featurizer_class = FilterbankFeatures 253 else: 254 featurizer_class = FilterbankFeaturesTA 255 self.featurizer = featurizer_class( 256 sample_rate=self._sample_rate, 257 n_window_size=n_window_size, 258 n_window_stride=n_window_stride, 259 window=window, 260 normalize=normalize, 261 n_fft=n_fft, 262 preemph=preemph, 263 nfilt=features, 264 lowfreq=lowfreq, 265 highfreq=highfreq, 266 log=log, 267 log_zero_guard_type=log_zero_guard_type, 268 log_zero_guard_value=log_zero_guard_value, 269 dither=dither, 270 pad_to=pad_to, 271 frame_splicing=frame_splicing, 272 exact_pad=exact_pad, 273 pad_value=pad_value, 274 mag_power=mag_power, 275 rng=rng, 276 nb_augmentation_prob=nb_augmentation_prob, 277 nb_max_freq=nb_max_freq, 278 mel_norm=mel_norm, 279 stft_exact_pad=stft_exact_pad, # Deprecated arguments; kept for config compatibility 280 stft_conv=stft_conv, # Deprecated arguments; kept for config compatibility 281 ) 282 283 def input_example(self, max_batch: int = 8, max_dim: int = 32000, min_length: int = 200): 284 batch_size = torch.randint(low=1, high=max_batch, size=[1]).item() 285 max_length = torch.randint(low=min_length, high=max_dim, size=[1]).item() 286 signals = torch.rand(size=[batch_size, max_length]) * 2 - 1 287 lengths = torch.randint(low=min_length, high=max_dim, size=[batch_size]) 288 lengths[0] = max_length 289 return signals, lengths 290 291 def get_features(self, input_signal, length): 292 return self.featurizer(input_signal, length) 293 294 @property 295 def filter_banks(self): 296 return self.featurizer.filter_banks 297 298 299 class AudioToMFCCPreprocessor(AudioPreprocessor): 300 """Preprocessor that converts wavs to MFCCs. 301 Uses torchaudio.transforms.MFCC. 302 303 Args: 304 sample_rate: The sample rate of the audio. 305 Defaults to 16000. 306 window_size: Size of window for fft in seconds. Used to calculate the 307 win_length arg for mel spectrogram. 308 Defaults to 0.02 309 window_stride: Stride of window for fft in seconds. Used to caculate 310 the hop_length arg for mel spect. 311 Defaults to 0.01 312 n_window_size: Size of window for fft in samples 313 Defaults to None. Use one of window_size or n_window_size. 314 n_window_stride: Stride of window for fft in samples 315 Defaults to None. Use one of window_stride or n_window_stride. 316 window: Windowing function for fft. can be one of ['hann', 317 'hamming', 'blackman', 'bartlett', 'none', 'null']. 318 Defaults to 'hann' 319 n_fft: Length of FT window. If None, it uses the smallest power of 2 320 that is larger than n_window_size. 321 Defaults to None 322 lowfreq (int): Lower bound on mel basis in Hz. 323 Defaults to 0 324 highfreq (int): Lower bound on mel basis in Hz. 325 Defaults to None 326 n_mels: Number of mel filterbanks. 327 Defaults to 64 328 n_mfcc: Number of coefficients to retain 329 Defaults to 64 330 dct_type: Type of discrete cosine transform to use 331 norm: Type of norm to use 332 log: Whether to use log-mel spectrograms instead of db-scaled. 333 Defaults to True. 334 """ 335 336 @property 337 def input_types(self): 338 """Returns definitions of module input ports. 339 """ 340 return { 341 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 342 "length": NeuralType(tuple('B'), LengthsType()), 343 } 344 345 @property 346 def output_types(self): 347 """Returns definitions of module output ports. 348 """ 349 return { 350 "processed_signal": NeuralType(('B', 'D', 'T'), MFCCSpectrogramType()), 351 "processed_length": NeuralType(tuple('B'), LengthsType()), 352 } 353 354 def save_to(self, save_path: str): 355 pass 356 357 @classmethod 358 def restore_from(cls, restore_path: str): 359 pass 360 361 def __init__( 362 self, 363 sample_rate=16000, 364 window_size=0.02, 365 window_stride=0.01, 366 n_window_size=None, 367 n_window_stride=None, 368 window='hann', 369 n_fft=None, 370 lowfreq=0.0, 371 highfreq=None, 372 n_mels=64, 373 n_mfcc=64, 374 dct_type=2, 375 norm='ortho', 376 log=True, 377 ): 378 self._sample_rate = sample_rate 379 if not HAVE_TORCHAUDIO: 380 logging.error('Could not import torchaudio. Some features might not work.') 381 382 raise ModuleNotFoundError( 383 "torchaudio is not installed but is necessary for " 384 "AudioToMFCCPreprocessor. We recommend you try " 385 "building it from source for the PyTorch version you have." 386 ) 387 if window_size and n_window_size: 388 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 389 if window_stride and n_window_stride: 390 raise ValueError( 391 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 392 ) 393 # Get win_length (n_window_size) and hop_length (n_window_stride) 394 if window_size: 395 n_window_size = int(window_size * self._sample_rate) 396 if window_stride: 397 n_window_stride = int(window_stride * self._sample_rate) 398 399 super().__init__(n_window_size, n_window_stride) 400 401 mel_kwargs = {} 402 403 mel_kwargs['f_min'] = lowfreq 404 mel_kwargs['f_max'] = highfreq 405 mel_kwargs['n_mels'] = n_mels 406 407 mel_kwargs['n_fft'] = n_fft or 2 ** math.ceil(math.log2(n_window_size)) 408 409 mel_kwargs['win_length'] = n_window_size 410 mel_kwargs['hop_length'] = n_window_stride 411 412 # Set window_fn. None defaults to torch.ones. 413 window_fn = self.torch_windows.get(window, None) 414 if window_fn is None: 415 raise ValueError( 416 f"Window argument for AudioProcessor is invalid: {window}." 417 f"For no window function, use 'ones' or None." 418 ) 419 mel_kwargs['window_fn'] = window_fn 420 421 # Use torchaudio's implementation of MFCCs as featurizer 422 self.featurizer = torchaudio.transforms.MFCC( 423 sample_rate=self._sample_rate, 424 n_mfcc=n_mfcc, 425 dct_type=dct_type, 426 norm=norm, 427 log_mels=log, 428 melkwargs=mel_kwargs, 429 ) 430 431 def get_features(self, input_signal, length): 432 features = self.featurizer(input_signal) 433 seq_len = torch.ceil(length.to(torch.float32) / self.hop_length).to(dtype=torch.long) 434 return features, seq_len 435 436 437 class SpectrogramAugmentation(NeuralModule): 438 """ 439 Performs time and freq cuts in one of two ways. 440 SpecAugment zeroes out vertical and horizontal sections as described in 441 SpecAugment (https://arxiv.org/abs/1904.08779). Arguments for use with 442 SpecAugment are `freq_masks`, `time_masks`, `freq_width`, and `time_width`. 443 SpecCutout zeroes out rectangulars as described in Cutout 444 (https://arxiv.org/abs/1708.04552). Arguments for use with Cutout are 445 `rect_masks`, `rect_freq`, and `rect_time`. 446 447 Args: 448 freq_masks (int): how many frequency segments should be cut. 449 Defaults to 0. 450 time_masks (int): how many time segments should be cut 451 Defaults to 0. 452 freq_width (int): maximum number of frequencies to be cut in one 453 segment. 454 Defaults to 10. 455 time_width (int): maximum number of time steps to be cut in one 456 segment 457 Defaults to 10. 458 rect_masks (int): how many rectangular masks should be cut 459 Defaults to 0. 460 rect_freq (int): maximum size of cut rectangles along the frequency 461 dimension 462 Defaults to 5. 463 rect_time (int): maximum size of cut rectangles along the time 464 dimension 465 Defaults to 25. 466 """ 467 468 @property 469 def input_types(self): 470 """Returns definitions of module input types 471 """ 472 return { 473 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 474 "length": NeuralType(tuple('B'), LengthsType()), 475 } 476 477 @property 478 def output_types(self): 479 """Returns definitions of module output types 480 """ 481 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 482 483 def __init__( 484 self, 485 freq_masks=0, 486 time_masks=0, 487 freq_width=10, 488 time_width=10, 489 rect_masks=0, 490 rect_time=5, 491 rect_freq=20, 492 rng=None, 493 mask_value=0.0, 494 use_numba_spec_augment: bool = True, 495 ): 496 super().__init__() 497 498 if rect_masks > 0: 499 self.spec_cutout = SpecCutout(rect_masks=rect_masks, rect_time=rect_time, rect_freq=rect_freq, rng=rng,) 500 # self.spec_cutout.to(self._device) 501 else: 502 self.spec_cutout = lambda input_spec: input_spec 503 if freq_masks + time_masks > 0: 504 self.spec_augment = SpecAugment( 505 freq_masks=freq_masks, 506 time_masks=time_masks, 507 freq_width=freq_width, 508 time_width=time_width, 509 rng=rng, 510 mask_value=mask_value, 511 ) 512 else: 513 self.spec_augment = lambda input_spec, length: input_spec 514 515 # Check if numba is supported, and use a Numba kernel if it is 516 if use_numba_spec_augment and numba_utils.numba_cuda_is_supported(__NUMBA_MINIMUM_VERSION__): 517 logging.info('Numba CUDA SpecAugment kernel is being used') 518 self.spec_augment_numba = SpecAugmentNumba( 519 freq_masks=freq_masks, 520 time_masks=time_masks, 521 freq_width=freq_width, 522 time_width=time_width, 523 rng=rng, 524 mask_value=mask_value, 525 ) 526 else: 527 self.spec_augment_numba = None 528 529 @typecheck() 530 def forward(self, input_spec, length): 531 augmented_spec = self.spec_cutout(input_spec=input_spec) 532 533 # To run the Numba kernel, correct numba version is required as well as 534 # tensor must be on GPU and length must be provided 535 if self.spec_augment_numba is not None and spec_augment_launch_heuristics(augmented_spec, length): 536 augmented_spec = self.spec_augment_numba(input_spec=augmented_spec, length=length) 537 else: 538 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 539 return augmented_spec 540 541 542 class MaskedPatchAugmentation(NeuralModule): 543 """ 544 Zeroes out fixed size time patches of the spectrogram. 545 All samples in batch are guaranteed to have the same amount of masked time steps. 546 Optionally also performs frequency masking in the same way as SpecAugment. 547 Args: 548 patch_size (int): up to how many time steps does one patch consist of. 549 Defaults to 48. 550 mask_patches (float): how many patches should be masked in each sample. 551 if >= 1., interpreted as number of patches (after converting to int) 552 if <1., interpreted as fraction of total tokens to be masked (number of patches is rounded up) 553 Defaults to 10. 554 freq_masks (int): how many frequency segments should be cut. 555 Defaults to 0. 556 freq_width (int): maximum number of frequencies to be cut in a segment. 557 Defaults to 0. 558 """ 559 560 @property 561 def input_types(self): 562 """Returns definitions of module input types 563 """ 564 return { 565 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 566 "length": NeuralType(tuple('B'), LengthsType()), 567 } 568 569 @property 570 def output_types(self): 571 """Returns definitions of module output types 572 """ 573 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 574 575 def __init__( 576 self, patch_size: int = 48, mask_patches: float = 10.0, freq_masks: int = 0, freq_width: int = 0, 577 ): 578 super().__init__() 579 self.patch_size = patch_size 580 if mask_patches >= 1: 581 self.mask_patches = int(mask_patches) 582 elif mask_patches >= 0: 583 self._mask_fraction = mask_patches 584 self.mask_patches = None 585 else: 586 raise ValueError('mask_patches cannot be negative') 587 588 if freq_masks > 0: 589 self.spec_augment = SpecAugment(freq_masks=freq_masks, time_masks=0, freq_width=freq_width, time_width=0,) 590 else: 591 self.spec_augment = None 592 593 @typecheck() 594 def forward(self, input_spec, length): 595 augmented_spec = input_spec 596 597 min_len = torch.min(length) 598 599 if self.mask_patches is None: 600 # masking specified as fraction 601 len_fraction = int(min_len * self._mask_fraction) 602 mask_patches = len_fraction // self.patch_size + int(len_fraction % self.patch_size != 0) 603 else: 604 mask_patches = self.mask_patches 605 606 if min_len < self.patch_size * mask_patches: 607 mask_patches = min_len // self.patch_size 608 609 for idx in range(input_spec.shape[0]): 610 cur_len = length[idx] 611 patches = range(cur_len // self.patch_size) 612 masked_patches = random.sample(patches, mask_patches) 613 614 for mp in masked_patches: 615 augmented_spec[idx, :, mp * self.patch_size : (mp + 1) * self.patch_size] = 0.0 616 617 if self.spec_augment is not None: 618 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 619 620 return augmented_spec 621 622 623 class CropOrPadSpectrogramAugmentation(NeuralModule): 624 """ 625 Pad or Crop the incoming Spectrogram to a certain shape. 626 627 Args: 628 audio_length (int): the final number of timesteps that is required. 629 The signal will be either padded or cropped temporally to this 630 size. 631 """ 632 633 def __init__(self, audio_length): 634 super(CropOrPadSpectrogramAugmentation, self).__init__() 635 self.audio_length = audio_length 636 637 @typecheck() 638 @torch.no_grad() 639 def forward(self, input_signal, length): 640 image = input_signal 641 num_images = image.shape[0] 642 643 audio_length = self.audio_length 644 image_len = image.shape[-1] 645 646 # Crop long signal 647 if image_len > audio_length: # randomly slice 648 cutout_images = [] 649 offset = torch.randint(low=0, high=image_len - audio_length + 1, size=[num_images]) 650 651 for idx, offset in enumerate(offset): 652 cutout_images.append(image[idx : idx + 1, :, offset : offset + audio_length]) 653 654 image = torch.cat(cutout_images, dim=0) 655 del cutout_images 656 657 else: # symmetrically pad short signal with zeros 658 pad_left = (audio_length - image_len) // 2 659 pad_right = (audio_length - image_len) // 2 660 661 if (audio_length - image_len) % 2 == 1: 662 pad_right += 1 663 664 image = torch.nn.functional.pad(image, [pad_left, pad_right], mode="constant", value=0) 665 666 # Replace dynamic length sequences with static number of timesteps 667 length = (length * 0) + audio_length 668 669 return image, length 670 671 @property 672 def input_types(self): 673 """Returns definitions of module output ports. 674 """ 675 return { 676 "input_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 677 "length": NeuralType(tuple('B'), LengthsType()), 678 } 679 680 @property 681 def output_types(self): 682 """Returns definitions of module output ports. 683 """ 684 return { 685 "processed_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 686 "processed_length": NeuralType(tuple('B'), LengthsType()), 687 } 688 689 def save_to(self, save_path: str): 690 pass 691 692 @classmethod 693 def restore_from(cls, restore_path: str): 694 pass 695 696 697 class AudioToSpectrogram(NeuralModule): 698 """Transform a batch of input multi-channel signals into a batch of 699 STFT-based spectrograms. 700 701 Args: 702 fft_length: length of FFT 703 hop_length: length of hops/shifts of the sliding window 704 power: exponent for magnitude spectrogram. Default `None` will 705 return a complex-valued spectrogram 706 """ 707 708 def __init__(self, fft_length: int, hop_length: int, power: Optional[float] = None): 709 if not HAVE_TORCHAUDIO: 710 logging.error('Could not import torchaudio. Some features might not work.') 711 712 raise ModuleNotFoundError( 713 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 714 ) 715 716 super().__init__() 717 718 # For now, assume FFT length is divisible by two 719 if fft_length % 2 != 0: 720 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 721 722 self.stft = torchaudio.transforms.Spectrogram( 723 n_fft=fft_length, hop_length=hop_length, power=power, pad_mode='constant' 724 ) 725 726 # number of subbands 727 self.F = fft_length // 2 + 1 728 729 @property 730 def num_subbands(self) -> int: 731 return self.F 732 733 @property 734 def input_types(self) -> Dict[str, NeuralType]: 735 """Returns definitions of module output ports. 736 """ 737 return { 738 "input": NeuralType(('B', 'C', 'T'), AudioSignal()), 739 "input_length": NeuralType(('B',), LengthsType(), optional=True), 740 } 741 742 @property 743 def output_types(self) -> Dict[str, NeuralType]: 744 """Returns definitions of module output ports. 745 """ 746 return { 747 "output": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 748 "output_length": NeuralType(('B',), LengthsType()), 749 } 750 751 @typecheck() 752 def forward( 753 self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None 754 ) -> Tuple[torch.Tensor, torch.Tensor]: 755 """Convert a batch of C-channel input signals 756 into a batch of complex-valued spectrograms. 757 758 Args: 759 input: Time-domain input signal with C channels, shape (B, C, T) 760 input_length: Length of valid entries along the time dimension, shape (B,) 761 762 Returns: 763 Output spectrogram with F subbands and N time frames, shape (B, C, F, N) 764 and output length with shape (B,). 765 """ 766 B, T = input.size(0), input.size(-1) 767 input = input.view(B, -1, T) 768 769 # STFT output (B, C, F, N) 770 with torch.cuda.amp.autocast(enabled=False): 771 output = self.stft(input.float()) 772 773 if input_length is not None: 774 # Mask padded frames 775 output_length = self.get_output_length(input_length=input_length) 776 777 length_mask: torch.Tensor = make_seq_mask_like( 778 lengths=output_length, like=output, time_dim=-1, valid_ones=False 779 ) 780 output = output.masked_fill(length_mask, 0.0) 781 else: 782 # Assume all frames are valid for all examples in the batch 783 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 784 785 return output, output_length 786 787 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 788 """Get length of valid frames for the output. 789 790 Args: 791 input_length: number of valid samples, shape (B,) 792 793 Returns: 794 Number of valid frames, shape (B,) 795 """ 796 output_length = input_length.div(self.stft.hop_length, rounding_mode='floor').add(1).long() 797 return output_length 798 799 800 class SpectrogramToAudio(NeuralModule): 801 """Transform a batch of input multi-channel spectrograms into a batch of 802 time-domain multi-channel signals. 803 804 Args: 805 fft_length: length of FFT 806 hop_length: length of hops/shifts of the sliding window 807 power: exponent for magnitude spectrogram. Default `None` will 808 return a complex-valued spectrogram 809 """ 810 811 def __init__(self, fft_length: int, hop_length: int): 812 if not HAVE_TORCHAUDIO: 813 logging.error('Could not import torchaudio. Some features might not work.') 814 815 raise ModuleNotFoundError( 816 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 817 ) 818 819 super().__init__() 820 821 # For now, assume FFT length is divisible by two 822 if fft_length % 2 != 0: 823 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 824 825 self.istft = torchaudio.transforms.InverseSpectrogram( 826 n_fft=fft_length, hop_length=hop_length, pad_mode='constant' 827 ) 828 829 self.F = fft_length // 2 + 1 830 831 @property 832 def num_subbands(self) -> int: 833 return self.F 834 835 @property 836 def input_types(self) -> Dict[str, NeuralType]: 837 """Returns definitions of module output ports. 838 """ 839 return { 840 "input": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 841 "input_length": NeuralType(('B',), LengthsType(), optional=True), 842 } 843 844 @property 845 def output_types(self) -> Dict[str, NeuralType]: 846 """Returns definitions of module output ports. 847 """ 848 return { 849 "output": NeuralType(('B', 'C', 'T'), AudioSignal()), 850 "output_length": NeuralType(('B',), LengthsType()), 851 } 852 853 @typecheck() 854 def forward(self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None) -> torch.Tensor: 855 """Convert input complex-valued spectrogram to a time-domain 856 signal. Multi-channel IO is supported. 857 858 Args: 859 input: Input spectrogram for C channels, shape (B, C, F, N) 860 input_length: Length of valid entries along the time dimension, shape (B,) 861 862 Returns: 863 Time-domain signal with T time-domain samples and C channels, (B, C, T) 864 and output length with shape (B,). 865 """ 866 B, F, N = input.size(0), input.size(-2), input.size(-1) 867 assert F == self.F, f'Number of subbands F={F} not matching self.F={self.F}' 868 input = input.view(B, -1, F, N) 869 870 # iSTFT output (B, C, T) 871 with torch.cuda.amp.autocast(enabled=False): 872 output = self.istft(input.cfloat()) 873 874 if input_length is not None: 875 # Mask padded samples 876 output_length = self.get_output_length(input_length=input_length) 877 878 length_mask: torch.Tensor = make_seq_mask_like( 879 lengths=output_length, like=output, time_dim=-1, valid_ones=False 880 ) 881 output = output.masked_fill(length_mask, 0.0) 882 else: 883 # Assume all frames are valid for all examples in the batch 884 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 885 886 return output, output_length 887 888 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 889 """Get length of valid samples for the output. 890 891 Args: 892 input_length: number of valid frames, shape (B,) 893 894 Returns: 895 Number of valid samples, shape (B,) 896 """ 897 output_length = input_length.sub(1).mul(self.istft.hop_length).long() 898 return output_length 899 900 901 @dataclass 902 class AudioToMelSpectrogramPreprocessorConfig: 903 _target_: str = "nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor" 904 sample_rate: int = 16000 905 window_size: float = 0.02 906 window_stride: float = 0.01 907 n_window_size: Optional[int] = None 908 n_window_stride: Optional[int] = None 909 window: str = "hann" 910 normalize: str = "per_feature" 911 n_fft: Optional[int] = None 912 preemph: float = 0.97 913 features: int = 64 914 lowfreq: int = 0 915 highfreq: Optional[int] = None 916 log: bool = True 917 log_zero_guard_type: str = "add" 918 log_zero_guard_value: float = 2 ** -24 919 dither: float = 1e-5 920 pad_to: int = 16 921 frame_splicing: int = 1 922 exact_pad: bool = False 923 pad_value: int = 0 924 mag_power: float = 2.0 925 rng: Optional[str] = None 926 nb_augmentation_prob: float = 0.0 927 nb_max_freq: int = 4000 928 use_torchaudio: bool = False 929 mel_norm: str = "slaney" 930 stft_exact_pad: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 931 stft_conv: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 932 933 934 @dataclass 935 class AudioToMFCCPreprocessorConfig: 936 _target_: str = 'nemo.collections.asr.modules.AudioToMFCCPreprocessor' 937 sample_rate: int = 16000 938 window_size: float = 0.02 939 window_stride: float = 0.01 940 n_window_size: Optional[int] = None 941 n_window_stride: Optional[int] = None 942 window: str = 'hann' 943 n_fft: Optional[int] = None 944 lowfreq: Optional[float] = 0.0 945 highfreq: Optional[float] = None 946 n_mels: int = 64 947 n_mfcc: int = 64 948 dct_type: int = 2 949 norm: str = 'ortho' 950 log: bool = True 951 952 953 @dataclass 954 class SpectrogramAugmentationConfig: 955 _target_: str = "nemo.collections.asr.modules.SpectrogramAugmentation" 956 freq_masks: int = 0 957 time_masks: int = 0 958 freq_width: int = 0 959 time_width: Optional[Any] = 0 960 rect_masks: int = 0 961 rect_time: int = 0 962 rect_freq: int = 0 963 mask_value: float = 0 964 rng: Optional[Any] = None # random.Random() type 965 use_numba_spec_augment: bool = True 966 967 968 @dataclass 969 class CropOrPadSpectrogramAugmentationConfig: 970 audio_length: int 971 _target_: str = "nemo.collections.asr.modules.CropOrPadSpectrogramAugmentation" 972 973 974 @dataclass 975 class MaskedPatchAugmentationConfig: 976 patch_size: int = 48 977 mask_patches: float = 10.0 978 freq_masks: int = 0 979 freq_width: int = 0 980 _target_: str = "nemo.collections.asr.modules.MaskedPatchAugmentation" 981 [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/parts/k2/classes.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from abc import ABC 16 from dataclasses import dataclass 17 from typing import Any, Optional, Tuple 18 19 import torch 20 from omegaconf import DictConfig 21 22 from nemo.utils import logging 23 24 25 @dataclass 26 class GraphIntersectDenseConfig: 27 """Graph dense intersection config. 28 """ 29 30 search_beam: float = 20.0 31 output_beam: float = 10.0 32 min_active_states: int = 30 33 max_active_states: int = 10000 34 35 36 @dataclass 37 class GraphModuleConfig: 38 """Config for graph modules. 39 Typically used with graph losses and decoders. 40 """ 41 42 topo_type: str = "default" 43 topo_with_self_loops: bool = True 44 token_lm: Optional[Any] = None 45 intersect_pruned: bool = False 46 intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() 47 boost_coeff: float = 0.0 48 predictor_window_size: int = 0 49 predictor_step_size: int = 1 50 51 52 class ASRK2Mixin(ABC): 53 """k2 Mixin class that simplifies the construction of various models with k2-based losses. 54 55 It does the following: 56 - Sets up the graph loss and decoder (methods _init_k2 and update_k2_modules). 57 - Registers external graphs, if needed. 58 - Augments forward(...) with optional graph decoding to get accurate predictions. 59 """ 60 61 def _init_k2(self): 62 """ 63 k2-related initialization implementation. 64 65 This method is expected to run after the __init__ which sets self._cfg 66 self._cfg is expected to have the attribute graph_module_cfg 67 """ 68 if not hasattr(self, "_cfg"): 69 raise ValueError("self._cfg must be set before calling _init_k2().") 70 if not hasattr(self._cfg, "graph_module_cfg") or self._cfg.graph_module_cfg is None: 71 raise ValueError("self._cfg.graph_module_cfg must be set and cannot be None.") 72 self.graph_module_cfg = self._cfg.graph_module_cfg 73 74 # register token_lm for MAPLoss 75 criterion_type = self.graph_module_cfg.get("criterion_type", "ml") 76 self.use_graph_lm = criterion_type == "map" 77 if self.use_graph_lm: 78 token_lm_path = self.graph_module_cfg.backend_cfg.get("token_lm", None) 79 if token_lm_path is None: 80 raise ValueError( 81 f"graph_module_cfg.backend_cfg.token_lm is empty. It must be set for criterion_type == `{criterion_type}`" 82 ) 83 token_lm_path = self.register_artifact('graph_module_cfg.backend_cfg.token_lm', token_lm_path) 84 self.graph_module_cfg.backend_cfg["token_lm"] = token_lm_path 85 86 self.update_k2_modules(self.graph_module_cfg) 87 88 def update_k2_modules(self, input_cfg: DictConfig): 89 """ 90 Helper function to initialize or update k2 loss and transcribe_decoder. 91 92 Args: 93 input_cfg: DictConfig to take new parameters from. Schema is expected as in 94 nemo.collections.asr.models.configs.k2_sequence_models_config.GraphModuleConfig 95 """ 96 del self.loss 97 if hasattr(self, "transcribe_decoder"): 98 del self.transcribe_decoder 99 100 if hasattr(self, "joint"): 101 # RNNT 102 num_classes = self.joint.num_classes_with_blank - 1 103 else: 104 # CTC, MMI, ... 105 num_classes = self.decoder.num_classes_with_blank - 1 106 remove_consecutive = input_cfg.backend_cfg.get("topo_with_self_loops", True) and input_cfg.backend_cfg.get( 107 "topo_type", "default" 108 ) not in ["forced_blank", "identity",] 109 self._wer.remove_consecutive = remove_consecutive 110 111 from nemo.collections.asr.losses.lattice_losses import LatticeLoss 112 113 self.loss = LatticeLoss( 114 num_classes=num_classes, 115 reduction=self._cfg.get("ctc_reduction", "mean_batch"), 116 backend="k2", 117 criterion_type=input_cfg.get("criterion_type", "ml"), 118 loss_type=input_cfg.get("loss_type", "ctc"), 119 split_batch_size=input_cfg.get("split_batch_size", 0), 120 graph_module_cfg=input_cfg.backend_cfg, 121 ) 122 123 criterion_type = self.loss.criterion_type 124 self.use_graph_lm = criterion_type == "map" 125 transcribe_training = input_cfg.get("transcribe_training", False) 126 if transcribe_training and criterion_type == "ml": 127 logging.warning( 128 f"""You do not need to use transcribe_training=`{transcribe_training}` 129 with criterion_type=`{criterion_type}`. transcribe_training will be set to False.""" 130 ) 131 transcribe_training = False 132 self.transcribe_training = transcribe_training 133 if self.use_graph_lm: 134 from nemo.collections.asr.modules.graph_decoder import ViterbiDecoderWithGraph 135 136 self.transcribe_decoder = ViterbiDecoderWithGraph( 137 num_classes=num_classes, 138 backend="k2", 139 dec_type="token_lm", 140 return_type="1best", 141 return_ilabels=True, 142 output_aligned=True, 143 split_batch_size=input_cfg.get("split_batch_size", 0), 144 graph_module_cfg=input_cfg.backend_cfg, 145 ) 146 147 def _forward_k2_post_processing( 148 self, log_probs: torch.Tensor, encoded_length: torch.Tensor, greedy_predictions: torch.Tensor 149 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: 150 """ 151 k2-related post-processing parf of .forward() 152 153 Args: 154 log_probs: The log probabilities tensor of shape [B, T, D]. 155 encoded_length: The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 156 greedy_predictions: The greedy token predictions of the model of shape [B, T] 157 158 Returns: 159 A tuple of 3 elements - 160 1) The log probabilities tensor of shape [B, T, D]. 161 2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 162 3) The greedy token predictions of the model of shape [B, T] (via argmax) 163 """ 164 # greedy_predictions from .forward() are incorrect for criterion_type=`map` 165 # getting correct greedy_predictions, if needed 166 if self.use_graph_lm and (not self.training or self.transcribe_training): 167 greedy_predictions, encoded_length, _ = self.transcribe_decoder.forward( 168 log_probs=log_probs, log_probs_length=encoded_length 169 ) 170 return log_probs, encoded_length, greedy_predictions 171 [end of nemo/collections/asr/parts/k2/classes.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from dataclasses import dataclass 17 from typing import Any, Optional 18 19 import torch 20 from torch import nn as nn 21 22 from nemo.collections.asr.parts.submodules import multi_head_attention as mha 23 from nemo.collections.common.parts import adapter_modules 24 from nemo.core.classes.mixins import adapter_mixin_strategies 25 26 27 class MHAResidualAddAdapterStrategy(adapter_mixin_strategies.ResidualAddAdapterStrategy): 28 """ 29 An implementation of residual addition of an adapter module with its input for the MHA Adapters. 30 """ 31 32 def forward(self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin'): 33 """ 34 A basic strategy, comprising of a residual connection over the input, after forward pass by 35 the underlying adapter. Additional work is done to pack and unpack the dictionary of inputs and outputs. 36 37 Note: The `value` tensor is added to the output of the attention adapter as the residual connection. 38 39 Args: 40 input: A dictionary of multiple input arguments for the adapter module. 41 `query`, `key`, `value`: Original output tensor of the module, or the output of the 42 previous adapter (if more than one adapters are enabled). 43 `mask`: Attention mask. 44 `pos_emb`: Optional positional embedding for relative encoding. 45 adapter: The adapter module that is currently required to perform the forward pass. 46 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 47 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 48 49 Returns: 50 The result tensor, after one of the active adapters has finished its forward passes. 51 """ 52 out = self.compute_output(input, adapter, module=module) 53 54 # If not in training mode, or probability of stochastic depth is 0, skip step. 55 p = self.stochastic_depth 56 if not module.training or p == 0.0: 57 pass 58 else: 59 out = self.apply_stochastic_depth(out, input['value'], adapter, module=module) 60 61 # Return the residual connection output = input + adapter(input) 62 result = input['value'] + out 63 64 # If l2_lambda is activated, register the loss value 65 self.compute_auxiliary_losses(result, input['value'], adapter, module=module) 66 67 return result 68 69 def compute_output( 70 self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin' 71 ) -> torch.Tensor: 72 """ 73 Compute the output of a single adapter to some input. 74 75 Args: 76 input: Original output tensor of the module, or the output of the previous adapter (if more than 77 one adapters are enabled). 78 adapter: The adapter module that is currently required to perform the forward pass. 79 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 80 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 81 82 Returns: 83 The result tensor, after one of the active adapters has finished its forward passes. 84 """ 85 if isinstance(input, (list, tuple)): 86 out = adapter(*input) 87 elif isinstance(input, dict): 88 out = adapter(**input) 89 else: 90 out = adapter(input) 91 return out 92 93 94 @dataclass 95 class MHAResidualAddAdapterStrategyConfig(adapter_mixin_strategies.ResidualAddAdapterStrategyConfig): 96 _target_: str = "{0}.{1}".format( 97 MHAResidualAddAdapterStrategy.__module__, MHAResidualAddAdapterStrategy.__name__ 98 ) # mandatory field 99 100 101 class MultiHeadAttentionAdapter(mha.MultiHeadAttention, adapter_modules.AdapterModuleUtil): 102 """Multi-Head Attention layer of Transformer. 103 Args: 104 n_head (int): number of heads 105 n_feat (int): size of the features 106 dropout_rate (float): dropout rate 107 proj_dim (int, optional): Optional integer value for projection before computing attention. 108 If None, then there is no projection (equivalent to proj_dim = n_feat). 109 If > 0, then will project the n_feat to proj_dim before calculating attention. 110 If <0, then will equal n_head, so that each head has a projected dimension of 1. 111 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 112 """ 113 114 def __init__( 115 self, 116 n_head: int, 117 n_feat: int, 118 dropout_rate: float, 119 proj_dim: Optional[int] = None, 120 adapter_strategy: MHAResidualAddAdapterStrategy = None, 121 ): 122 super().__init__(n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, max_cache_len=0) 123 124 self.pre_norm = nn.LayerNorm(n_feat) 125 126 # Set the projection dim to number of heads automatically 127 if proj_dim is not None and proj_dim < 1: 128 proj_dim = n_head 129 130 self.proj_dim = proj_dim 131 132 # Recompute weights for projection dim 133 if self.proj_dim is not None: 134 if self.proj_dim % n_head != 0: 135 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 136 137 self.d_k = self.proj_dim // n_head 138 self.s_d_k = math.sqrt(self.d_k) 139 self.linear_q = nn.Linear(n_feat, self.proj_dim) 140 self.linear_k = nn.Linear(n_feat, self.proj_dim) 141 self.linear_v = nn.Linear(n_feat, self.proj_dim) 142 self.linear_out = nn.Linear(self.proj_dim, n_feat) 143 144 # Setup adapter strategy 145 self.setup_adapter_strategy(adapter_strategy) 146 147 # reset parameters for Q to be identity operation 148 self.reset_parameters() 149 150 def forward(self, query, key, value, mask, pos_emb=None, cache=None): 151 """Compute 'Scaled Dot Product Attention'. 152 Args: 153 query (torch.Tensor): (batch, time1, size) 154 key (torch.Tensor): (batch, time2, size) 155 value(torch.Tensor): (batch, time2, size) 156 mask (torch.Tensor): (batch, time1, time2) 157 cache (torch.Tensor) : (batch, time_cache, size) 158 159 returns: 160 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 161 cache (torch.Tensor) : (batch, time_cache_next, size) 162 """ 163 # Need to perform duplicate computations as at this point the tensors have been 164 # separated by the adapter forward 165 query = self.pre_norm(query) 166 key = self.pre_norm(key) 167 value = self.pre_norm(value) 168 169 return super().forward(query, key, value, mask, pos_emb, cache=cache) 170 171 def reset_parameters(self): 172 with torch.no_grad(): 173 nn.init.zeros_(self.linear_out.weight) 174 nn.init.zeros_(self.linear_out.bias) 175 176 def get_default_strategy_config(self) -> 'dataclass': 177 return MHAResidualAddAdapterStrategyConfig() 178 179 180 @dataclass 181 class MultiHeadAttentionAdapterConfig: 182 n_head: int 183 n_feat: int 184 dropout_rate: float = 0.0 185 proj_dim: Optional[int] = None 186 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 187 _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) 188 189 190 class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): 191 """Multi-Head Attention layer of Transformer-XL with support of relative positional encoding. 192 Paper: https://arxiv.org/abs/1901.02860 193 Args: 194 n_head (int): number of heads 195 n_feat (int): size of the features 196 dropout_rate (float): dropout rate 197 proj_dim (int, optional): Optional integer value for projection before computing attention. 198 If None, then there is no projection (equivalent to proj_dim = n_feat). 199 If > 0, then will project the n_feat to proj_dim before calculating attention. 200 If <0, then will equal n_head, so that each head has a projected dimension of 1. 201 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 202 """ 203 204 def __init__( 205 self, 206 n_head: int, 207 n_feat: int, 208 dropout_rate: float, 209 proj_dim: Optional[int] = None, 210 adapter_strategy: MHAResidualAddAdapterStrategyConfig = None, 211 ): 212 super().__init__( 213 n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, pos_bias_u=None, pos_bias_v=None, max_cache_len=0 214 ) 215 216 self.pre_norm = nn.LayerNorm(n_feat) 217 218 # Set the projection dim to number of heads automatically 219 if proj_dim is not None and proj_dim < 1: 220 proj_dim = n_head 221 222 self.proj_dim = proj_dim 223 224 # Recompute weights for projection dim 225 if self.proj_dim is not None: 226 if self.proj_dim % n_head != 0: 227 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 228 229 self.d_k = self.proj_dim // n_head 230 self.s_d_k = math.sqrt(self.d_k) 231 self.linear_q = nn.Linear(n_feat, self.proj_dim) 232 self.linear_k = nn.Linear(n_feat, self.proj_dim) 233 self.linear_v = nn.Linear(n_feat, self.proj_dim) 234 self.linear_out = nn.Linear(self.proj_dim, n_feat) 235 self.linear_pos = nn.Linear(n_feat, self.proj_dim, bias=False) 236 self.pos_bias_u = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 237 self.pos_bias_v = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 238 239 # Setup adapter strategy 240 self.setup_adapter_strategy(adapter_strategy) 241 242 # reset parameters for Q to be identity operation 243 self.reset_parameters() 244 245 def forward(self, query, key, value, mask, pos_emb, cache=None): 246 """Compute 'Scaled Dot Product Attention' with rel. positional encoding. 247 Args: 248 query (torch.Tensor): (batch, time1, size) 249 key (torch.Tensor): (batch, time2, size) 250 value(torch.Tensor): (batch, time2, size) 251 mask (torch.Tensor): (batch, time1, time2) 252 pos_emb (torch.Tensor) : (batch, time1, size) 253 cache (torch.Tensor) : (batch, time_cache, size) 254 Returns: 255 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 256 cache_next (torch.Tensor) : (batch, time_cache_next, size) 257 """ 258 # Need to perform duplicate computations as at this point the tensors have been 259 # separated by the adapter forward 260 query = self.pre_norm(query) 261 key = self.pre_norm(key) 262 value = self.pre_norm(value) 263 264 return super().forward(query, key, value, mask, pos_emb, cache=cache) 265 266 def reset_parameters(self): 267 with torch.no_grad(): 268 nn.init.zeros_(self.linear_out.weight) 269 nn.init.zeros_(self.linear_out.bias) 270 271 # NOTE: This exact procedure apparently highly important. 272 # Above operation is safe to do as self.linear_out.weight *= 0.0 (similar for bias) 273 # However: 274 # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 275 # For some reason at init sometimes it will cause the value of the tensor to become NaN 276 # All operations to compute matrix_ac and matrix_bd will then fail. 277 nn.init.zeros_(self.pos_bias_u) 278 nn.init.zeros_(self.pos_bias_v) 279 280 def get_default_strategy_config(self) -> 'dataclass': 281 return MHAResidualAddAdapterStrategyConfig() 282 283 284 @dataclass 285 class RelPositionMultiHeadAttentionAdapterConfig: 286 n_head: int 287 n_feat: int 288 dropout_rate: float = 0.0 289 proj_dim: Optional[int] = None 290 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 291 _target_: str = "{0}.{1}".format( 292 RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ 293 ) 294 295 296 class PositionalEncodingAdapter(mha.PositionalEncoding, adapter_modules.AdapterModuleUtil): 297 298 """ 299 Absolute positional embedding adapter. 300 301 .. note:: 302 303 Absolute positional embedding value is added to the input tensor *without residual connection* ! 304 Therefore, the input is changed, if you only require the positional embedding, drop the returned `x` ! 305 306 Args: 307 d_model (int): The input dimension of x. 308 max_len (int): The max sequence length. 309 xscale (float): The input scaling factor. Defaults to 1.0. 310 adapter_strategy (AbstractAdapterStrategy): By default, ReturnResultAdapterStrategyConfig. 311 An adapter composition function object. 312 NOTE: Since this is a positional encoding, it will not add a residual ! 313 """ 314 315 def __init__( 316 self, 317 d_model: int, 318 max_len: int = 5000, 319 xscale=1.0, 320 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 321 ): 322 323 super().__init__( 324 d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, 325 ) 326 327 # Setup adapter strategy 328 self.setup_adapter_strategy(adapter_strategy) 329 330 def get_default_strategy_config(self) -> 'dataclass': 331 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 332 333 334 @dataclass 335 class PositionalEncodingAdapterConfig: 336 d_model: int 337 max_len: int = 5000 338 xscale: float = 1.0 339 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 340 _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) 341 342 343 class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): 344 """ 345 Relative positional encoding for TransformerXL's layers 346 See : Appendix B in https://arxiv.org/abs/1901.02860 347 348 .. note:: 349 350 Relative positional embedding value is **not** added to the input tensor ! 351 Therefore, the input should be updated changed, if you only require the positional embedding, drop the returned `x` ! 352 353 Args: 354 d_model (int): embedding dim 355 max_len (int): maximum input length 356 xscale (bool): whether to scale the input by sqrt(d_model) 357 adapter_strategy: By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. 358 """ 359 360 def __init__( 361 self, 362 d_model: int, 363 max_len: int = 5000, 364 xscale=1.0, 365 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 366 ): 367 super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) 368 369 # Setup adapter strategy 370 self.setup_adapter_strategy(adapter_strategy) 371 372 def get_default_strategy_config(self) -> 'dataclass': 373 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 374 375 376 @dataclass 377 class RelPositionalEncodingAdapterConfig: 378 d_model: int 379 max_len: int = 5000 380 xscale: float = 1.0 381 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 382 _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) 383 [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import os 17 from dataclasses import dataclass 18 from typing import List, Optional, Tuple, Union 19 20 import torch 21 22 from nemo.collections.asr.parts.utils import rnnt_utils 23 from nemo.collections.common.tokenizers.tokenizer_spec import TokenizerSpec 24 from nemo.core.classes import Typing, typecheck 25 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 26 from nemo.utils import logging 27 28 DEFAULT_TOKEN_OFFSET = 100 29 30 31 def pack_hypotheses( 32 hypotheses: List[rnnt_utils.NBestHypotheses], logitlen: torch.Tensor, 33 ) -> List[rnnt_utils.NBestHypotheses]: 34 35 if logitlen is not None: 36 if hasattr(logitlen, 'cpu'): 37 logitlen_cpu = logitlen.to('cpu') 38 else: 39 logitlen_cpu = logitlen 40 41 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.NBestHypotheses 42 for candidate_idx, cand in enumerate(hyp.n_best_hypotheses): 43 cand.y_sequence = torch.tensor(cand.y_sequence, dtype=torch.long) 44 45 if logitlen is not None: 46 cand.length = logitlen_cpu[idx] 47 48 if cand.dec_state is not None: 49 cand.dec_state = _states_to_device(cand.dec_state) 50 51 return hypotheses 52 53 54 def _states_to_device(dec_state, device='cpu'): 55 if torch.is_tensor(dec_state): 56 dec_state = dec_state.to(device) 57 58 elif isinstance(dec_state, (list, tuple)): 59 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 60 61 return dec_state 62 63 64 class AbstractBeamCTCInfer(Typing): 65 """A beam CTC decoder. 66 67 Provides a common abstraction for sample level beam decoding. 68 69 Args: 70 blank_id: int, index of the blank token. Can be 0 or len(vocabulary). 71 beam_size: int, size of the beam used in the underlying beam search engine. 72 73 """ 74 75 @property 76 def input_types(self): 77 """Returns definitions of module input ports. 78 """ 79 return { 80 "decoder_output": NeuralType(('B', 'T', 'D'), LogprobsType()), 81 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 82 } 83 84 @property 85 def output_types(self): 86 """Returns definitions of module output ports. 87 """ 88 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 89 90 def __init__(self, blank_id: int, beam_size: int): 91 self.blank_id = blank_id 92 93 if beam_size < 1: 94 raise ValueError("Beam search size cannot be less than 1!") 95 96 self.beam_size = beam_size 97 98 # Variables set by corresponding setter methods 99 self.vocab = None 100 self.decoding_type = None 101 self.tokenizer = None 102 103 # Utility maps for vocabulary 104 self.vocab_index_map = None 105 self.index_vocab_map = None 106 107 # Internal variable, used to prevent double reduction of consecutive tokens (ctc collapse) 108 self.override_fold_consecutive_value = None 109 110 def set_vocabulary(self, vocab: List[str]): 111 """ 112 Set the vocabulary of the decoding framework. 113 114 Args: 115 vocab: List of str. Each token corresponds to its location in the vocabulary emitted by the model. 116 Note that this vocabulary must NOT contain the "BLANK" token. 117 """ 118 self.vocab = vocab 119 self.vocab_index_map = {v: i for i, v in enumerate(vocab)} 120 self.index_vocab_map = {i: v for i, v in enumerate(vocab)} 121 122 def set_decoding_type(self, decoding_type: str): 123 """ 124 Sets the decoding type of the framework. Can support either char or subword models. 125 126 Args: 127 decoding_type: Str corresponding to decoding type. Only supports "char" and "subword". 128 """ 129 decoding_type = decoding_type.lower() 130 supported_types = ['char', 'subword'] 131 132 if decoding_type not in supported_types: 133 raise ValueError( 134 f"Unsupported decoding type. Supported types = {supported_types}.\n" f"Given = {decoding_type}" 135 ) 136 137 self.decoding_type = decoding_type 138 139 def set_tokenizer(self, tokenizer: TokenizerSpec): 140 """ 141 Set the tokenizer of the decoding framework. 142 143 Args: 144 tokenizer: NeMo tokenizer object, which inherits from TokenizerSpec. 145 """ 146 self.tokenizer = tokenizer 147 148 @typecheck() 149 def forward( 150 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 151 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 152 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 153 Output token is generated auto-repressively. 154 155 Args: 156 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 157 decoder_lengths: list of int representing the length of each sequence 158 output sequence. 159 160 Returns: 161 packed list containing batch number of sentences (Hypotheses). 162 """ 163 raise NotImplementedError() 164 165 def __call__(self, *args, **kwargs): 166 return self.forward(*args, **kwargs) 167 168 169 class BeamCTCInfer(AbstractBeamCTCInfer): 170 """A greedy CTC decoder. 171 172 Provides a common abstraction for sample level and batch level greedy decoding. 173 174 Args: 175 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 176 preserve_alignments: Bool flag which preserves the history of logprobs generated during 177 decoding (sample / batched). When set to true, the Hypothesis will contain 178 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 179 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 180 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 181 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 182 183 """ 184 185 def __init__( 186 self, 187 blank_id: int, 188 beam_size: int, 189 search_type: str = "default", 190 return_best_hypothesis: bool = True, 191 preserve_alignments: bool = False, 192 compute_timestamps: bool = False, 193 beam_alpha: float = 1.0, 194 beam_beta: float = 0.0, 195 kenlm_path: str = None, 196 flashlight_cfg: Optional['FlashlightConfig'] = None, 197 pyctcdecode_cfg: Optional['PyCTCDecodeConfig'] = None, 198 ): 199 super().__init__(blank_id=blank_id, beam_size=beam_size) 200 201 self.search_type = search_type 202 self.return_best_hypothesis = return_best_hypothesis 203 self.preserve_alignments = preserve_alignments 204 self.compute_timestamps = compute_timestamps 205 206 if self.compute_timestamps: 207 raise ValueError(f"Currently this flag is not supported for beam search algorithms.") 208 209 self.vocab = None # This must be set by specific method by user before calling forward() ! 210 211 if search_type == "default" or search_type == "nemo": 212 self.search_algorithm = self.default_beam_search 213 elif search_type == "pyctcdecode": 214 self.search_algorithm = self._pyctcdecode_beam_search 215 elif search_type == "flashlight": 216 self.search_algorithm = self.flashlight_beam_search 217 else: 218 raise NotImplementedError( 219 f"The search type ({search_type}) supplied is not supported!\n" 220 f"Please use one of : (default, nemo, pyctcdecode)" 221 ) 222 223 # Log the beam search algorithm 224 logging.info(f"Beam search algorithm: {search_type}") 225 226 self.beam_alpha = beam_alpha 227 self.beam_beta = beam_beta 228 229 # Default beam search args 230 self.kenlm_path = kenlm_path 231 232 # PyCTCDecode params 233 if pyctcdecode_cfg is None: 234 pyctcdecode_cfg = PyCTCDecodeConfig() 235 self.pyctcdecode_cfg = pyctcdecode_cfg # type: PyCTCDecodeConfig 236 237 if flashlight_cfg is None: 238 flashlight_cfg = FlashlightConfig() 239 self.flashlight_cfg = flashlight_cfg 240 241 # Default beam search scorer functions 242 self.default_beam_scorer = None 243 self.pyctcdecode_beam_scorer = None 244 self.flashlight_beam_scorer = None 245 self.token_offset = 0 246 247 @typecheck() 248 def forward( 249 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 250 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 251 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 252 Output token is generated auto-repressively. 253 254 Args: 255 decoder_output: A tensor of size (batch, timesteps, features). 256 decoder_lengths: list of int representing the length of each sequence 257 output sequence. 258 259 Returns: 260 packed list containing batch number of sentences (Hypotheses). 261 """ 262 if self.vocab is None: 263 raise RuntimeError("Please set the vocabulary with `set_vocabulary()` before calling this function.") 264 265 if self.decoding_type is None: 266 raise ValueError("Please set the decoding type with `set_decoding_type()` before calling this function.") 267 268 with torch.no_grad(), torch.inference_mode(): 269 # Process each sequence independently 270 prediction_tensor = decoder_output 271 272 if prediction_tensor.ndim != 3: 273 raise ValueError( 274 f"`decoder_output` must be a tensor of shape [B, T, V] (log probs, float). " 275 f"Provided shape = {prediction_tensor.shape}" 276 ) 277 278 # determine type of input - logprobs or labels 279 out_len = decoder_lengths if decoder_lengths is not None else None 280 hypotheses = self.search_algorithm(prediction_tensor, out_len) 281 282 # Pack results into Hypotheses 283 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 284 285 # Pack the result 286 if self.return_best_hypothesis and isinstance(packed_result[0], rnnt_utils.NBestHypotheses): 287 packed_result = [res.n_best_hypotheses[0] for res in packed_result] # type: Hypothesis 288 289 return (packed_result,) 290 291 @torch.no_grad() 292 def default_beam_search( 293 self, x: torch.Tensor, out_len: torch.Tensor 294 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 295 """ 296 Open Seq2Seq Beam Search Algorithm (DeepSpeed) 297 298 Args: 299 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 300 and V is the vocabulary size. The tensor contains log-probabilities. 301 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 302 303 Returns: 304 A list of NBestHypotheses objects, one for each sequence in the batch. 305 """ 306 if self.compute_timestamps: 307 raise ValueError( 308 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 309 ) 310 311 if self.default_beam_scorer is None: 312 # Check for filepath 313 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 314 raise FileNotFoundError( 315 f"KenLM binary file not found at : {self.kenlm_path}. " 316 f"Please set a valid path in the decoding config." 317 ) 318 319 # perform token offset for subword models 320 if self.decoding_type == 'subword': 321 vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 322 else: 323 # char models 324 vocab = self.vocab 325 326 # Must import at runtime to avoid circular dependency due to module level import. 327 from nemo.collections.asr.modules.beam_search_decoder import BeamSearchDecoderWithLM 328 329 self.default_beam_scorer = BeamSearchDecoderWithLM( 330 vocab=vocab, 331 lm_path=self.kenlm_path, 332 beam_width=self.beam_size, 333 alpha=self.beam_alpha, 334 beta=self.beam_beta, 335 num_cpus=max(1, os.cpu_count()), 336 input_tensor=False, 337 ) 338 339 x = x.to('cpu') 340 341 with typecheck.disable_checks(): 342 data = [x[sample_id, : out_len[sample_id], :].softmax(dim=-1) for sample_id in range(len(x))] 343 beams_batch = self.default_beam_scorer.forward(log_probs=data, log_probs_length=None) 344 345 # For each sample in the batch 346 nbest_hypotheses = [] 347 for beams_idx, beams in enumerate(beams_batch): 348 # For each beam candidate / hypothesis in each sample 349 hypotheses = [] 350 for candidate_idx, candidate in enumerate(beams): 351 hypothesis = rnnt_utils.Hypothesis( 352 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 353 ) 354 355 # For subword encoding, NeMo will double encode the subword (multiple tokens) into a 356 # singular unicode id. In doing so, we preserve the semantic of the unicode token, and 357 # compress the size of the final KenLM ARPA / Binary file. 358 # In order to do double encoding, we shift the subword by some token offset. 359 # This step is ignored for character based models. 360 if self.decoding_type == 'subword': 361 pred_token_ids = [ord(c) - self.token_offset for c in candidate[1]] 362 else: 363 # Char models 364 pred_token_ids = [self.vocab_index_map[c] for c in candidate[1]] 365 366 # We preserve the token ids and the score for this hypothesis 367 hypothesis.y_sequence = pred_token_ids 368 hypothesis.score = candidate[0] 369 370 # If alignment must be preserved, we preserve a view of the output logprobs. 371 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 372 # require specific processing for each sample in the beam. 373 # This is done to preserve memory. 374 if self.preserve_alignments: 375 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 376 377 hypotheses.append(hypothesis) 378 379 # Wrap the result in NBestHypothesis. 380 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 381 nbest_hypotheses.append(hypotheses) 382 383 return nbest_hypotheses 384 385 @torch.no_grad() 386 def _pyctcdecode_beam_search( 387 self, x: torch.Tensor, out_len: torch.Tensor 388 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 389 """ 390 PyCTCDecode Beam Search Algorithm. Should support Char and Subword models. 391 392 Args: 393 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 394 and V is the vocabulary size. The tensor contains log-probabilities. 395 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 396 397 Returns: 398 A list of NBestHypotheses objects, one for each sequence in the batch. 399 """ 400 if self.compute_timestamps: 401 raise ValueError( 402 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 403 ) 404 405 try: 406 import pyctcdecode 407 except (ImportError, ModuleNotFoundError): 408 raise ImportError( 409 f"Could not load `pyctcdecode` library. Please install it from pip using :\n" 410 f"pip install --upgrade pyctcdecode" 411 ) 412 413 if self.pyctcdecode_beam_scorer is None: 414 self.pyctcdecode_beam_scorer = pyctcdecode.build_ctcdecoder( 415 labels=self.vocab, kenlm_model_path=self.kenlm_path, alpha=self.beam_alpha, beta=self.beam_beta 416 ) # type: pyctcdecode.BeamSearchDecoderCTC 417 418 x = x.to('cpu').numpy() 419 420 with typecheck.disable_checks(): 421 beams_batch = [] 422 for sample_id in range(len(x)): 423 logprobs = x[sample_id, : out_len[sample_id], :] 424 result = self.pyctcdecode_beam_scorer.decode_beams( 425 logprobs, 426 beam_width=self.beam_size, 427 beam_prune_logp=self.pyctcdecode_cfg.beam_prune_logp, 428 token_min_logp=self.pyctcdecode_cfg.token_min_logp, 429 prune_history=self.pyctcdecode_cfg.prune_history, 430 hotwords=self.pyctcdecode_cfg.hotwords, 431 hotword_weight=self.pyctcdecode_cfg.hotword_weight, 432 lm_start_state=None, 433 ) # Output format: text, last_lm_state, text_frames, logit_score, lm_score 434 beams_batch.append(result) 435 436 nbest_hypotheses = [] 437 for beams_idx, beams in enumerate(beams_batch): 438 hypotheses = [] 439 for candidate_idx, candidate in enumerate(beams): 440 # Candidate = (text, last_lm_state, text_frames, logit_score, lm_score) 441 hypothesis = rnnt_utils.Hypothesis( 442 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 443 ) 444 445 # TODO: Requires token ids to be returned rather than text. 446 if self.decoding_type == 'subword': 447 if self.tokenizer is None: 448 raise ValueError("Tokenizer must be provided for subword decoding. Use set_tokenizer().") 449 450 pred_token_ids = self.tokenizer.text_to_ids(candidate[0]) 451 else: 452 if self.vocab is None: 453 raise ValueError("Vocab must be provided for character decoding. Use set_vocab().") 454 455 chars = list(candidate[0]) 456 pred_token_ids = [self.vocab_index_map[c] for c in chars] 457 458 hypothesis.y_sequence = pred_token_ids 459 hypothesis.text = candidate[0] # text 460 hypothesis.score = candidate[4] # score 461 462 # Inject word level timestamps 463 hypothesis.timestep = candidate[2] # text_frames 464 465 if self.preserve_alignments: 466 hypothesis.alignments = torch.from_numpy(x[beams_idx][: out_len[beams_idx]]) 467 468 hypotheses.append(hypothesis) 469 470 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 471 nbest_hypotheses.append(hypotheses) 472 473 return nbest_hypotheses 474 475 @torch.no_grad() 476 def flashlight_beam_search( 477 self, x: torch.Tensor, out_len: torch.Tensor 478 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 479 """ 480 Flashlight Beam Search Algorithm. Should support Char and Subword models. 481 482 Args: 483 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 484 and V is the vocabulary size. The tensor contains log-probabilities. 485 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 486 487 Returns: 488 A list of NBestHypotheses objects, one for each sequence in the batch. 489 """ 490 if self.compute_timestamps: 491 raise ValueError( 492 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 493 ) 494 495 if self.flashlight_beam_scorer is None: 496 # Check for filepath 497 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 498 raise FileNotFoundError( 499 f"KenLM binary file not found at : {self.kenlm_path}. " 500 f"Please set a valid path in the decoding config." 501 ) 502 503 # perform token offset for subword models 504 # if self.decoding_type == 'subword': 505 # vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 506 # else: 507 # # char models 508 # vocab = self.vocab 509 510 # Must import at runtime to avoid circular dependency due to module level import. 511 from nemo.collections.asr.modules.flashlight_decoder import FlashLightKenLMBeamSearchDecoder 512 513 self.flashlight_beam_scorer = FlashLightKenLMBeamSearchDecoder( 514 lm_path=self.kenlm_path, 515 vocabulary=self.vocab, 516 tokenizer=self.tokenizer, 517 lexicon_path=self.flashlight_cfg.lexicon_path, 518 boost_path=self.flashlight_cfg.boost_path, 519 beam_size=self.beam_size, 520 beam_size_token=self.flashlight_cfg.beam_size_token, 521 beam_threshold=self.flashlight_cfg.beam_threshold, 522 lm_weight=self.beam_alpha, 523 word_score=self.beam_beta, 524 unk_weight=self.flashlight_cfg.unk_weight, 525 sil_weight=self.flashlight_cfg.sil_weight, 526 ) 527 528 x = x.to('cpu') 529 530 with typecheck.disable_checks(): 531 beams_batch = self.flashlight_beam_scorer.forward(log_probs=x) 532 533 # For each sample in the batch 534 nbest_hypotheses = [] 535 for beams_idx, beams in enumerate(beams_batch): 536 # For each beam candidate / hypothesis in each sample 537 hypotheses = [] 538 for candidate_idx, candidate in enumerate(beams): 539 hypothesis = rnnt_utils.Hypothesis( 540 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 541 ) 542 543 # We preserve the token ids and the score for this hypothesis 544 hypothesis.y_sequence = candidate['tokens'].tolist() 545 hypothesis.score = candidate['score'] 546 547 # If alignment must be preserved, we preserve a view of the output logprobs. 548 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 549 # require specific processing for each sample in the beam. 550 # This is done to preserve memory. 551 if self.preserve_alignments: 552 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 553 554 hypotheses.append(hypothesis) 555 556 # Wrap the result in NBestHypothesis. 557 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 558 nbest_hypotheses.append(hypotheses) 559 560 return nbest_hypotheses 561 562 def set_decoding_type(self, decoding_type: str): 563 super().set_decoding_type(decoding_type) 564 565 # Please check train_kenlm.py in scripts/asr_language_modeling/ to find out why we need 566 # TOKEN_OFFSET for BPE-based models 567 if self.decoding_type == 'subword': 568 self.token_offset = DEFAULT_TOKEN_OFFSET 569 570 571 @dataclass 572 class PyCTCDecodeConfig: 573 # These arguments cannot be imported from pyctcdecode (optional dependency) 574 # Therefore we copy the values explicitly 575 # Taken from pyctcdecode.constant 576 beam_prune_logp: float = -10.0 577 token_min_logp: float = -5.0 578 prune_history: bool = False 579 hotwords: Optional[List[str]] = None 580 hotword_weight: float = 10.0 581 582 583 @dataclass 584 class FlashlightConfig: 585 lexicon_path: Optional[str] = None 586 boost_path: Optional[str] = None 587 beam_size_token: int = 16 588 beam_threshold: float = 20.0 589 unk_weight: float = -math.inf 590 sil_weight: float = 0.0 591 592 593 @dataclass 594 class BeamCTCInferConfig: 595 beam_size: int 596 search_type: str = 'default' 597 preserve_alignments: bool = False 598 compute_timestamps: bool = False 599 return_best_hypothesis: bool = True 600 601 beam_alpha: float = 1.0 602 beam_beta: float = 0.0 603 kenlm_path: Optional[str] = None 604 605 flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() 606 pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() 607 [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import List, Optional 17 18 import torch 19 from omegaconf import DictConfig, OmegaConf 20 21 from nemo.collections.asr.parts.utils import rnnt_utils 22 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMeasureConfig, ConfidenceMeasureMixin 23 from nemo.core.classes import Typing, typecheck 24 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 25 from nemo.utils import logging 26 27 28 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 29 30 if logitlen is not None: 31 if hasattr(logitlen, 'cpu'): 32 logitlen_cpu = logitlen.to('cpu') 33 else: 34 logitlen_cpu = logitlen 35 36 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 37 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 38 39 if logitlen is not None: 40 hyp.length = logitlen_cpu[idx] 41 42 if hyp.dec_state is not None: 43 hyp.dec_state = _states_to_device(hyp.dec_state) 44 45 return hypotheses 46 47 48 def _states_to_device(dec_state, device='cpu'): 49 if torch.is_tensor(dec_state): 50 dec_state = dec_state.to(device) 51 52 elif isinstance(dec_state, (list, tuple)): 53 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 54 55 return dec_state 56 57 58 class GreedyCTCInfer(Typing, ConfidenceMeasureMixin): 59 """A greedy CTC decoder. 60 61 Provides a common abstraction for sample level and batch level greedy decoding. 62 63 Args: 64 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 65 preserve_alignments: Bool flag which preserves the history of logprobs generated during 66 decoding (sample / batched). When set to true, the Hypothesis will contain 67 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 68 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 69 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 70 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 71 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 72 generated during decoding. When set to true, the Hypothesis will contain 73 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 74 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 75 confidence scores. 76 77 name: The measure name (str). 78 Supported values: 79 - 'max_prob' for using the maximum token probability as a confidence. 80 - 'entropy' for using a normalized entropy of a log-likelihood vector. 81 82 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 83 Supported values: 84 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 85 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 86 Note that for this entropy, the alpha should comply the following inequality: 87 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 88 where V is the model vocabulary size. 89 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 90 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 91 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 92 More: https://en.wikipedia.org/wiki/Tsallis_entropy 93 - 'renyi' for the Rรฉnyi entropy. 94 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 95 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 96 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 97 98 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 99 When the alpha equals one, scaling is not applied to 'max_prob', 100 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 101 102 entropy_norm: A mapping of the entropy value to the interval [0,1]. 103 Supported values: 104 - 'lin' for using the linear mapping. 105 - 'exp' for using exponential mapping with linear shift. 106 107 """ 108 109 @property 110 def input_types(self): 111 """Returns definitions of module input ports. 112 """ 113 # Input can be of dimention - 114 # ('B', 'T', 'D') [Log probs] or ('B', 'T') [Labels] 115 116 return { 117 "decoder_output": NeuralType(None, LogprobsType()), 118 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 119 } 120 121 @property 122 def output_types(self): 123 """Returns definitions of module output ports. 124 """ 125 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 126 127 def __init__( 128 self, 129 blank_id: int, 130 preserve_alignments: bool = False, 131 compute_timestamps: bool = False, 132 preserve_frame_confidence: bool = False, 133 confidence_measure_cfg: Optional[DictConfig] = None, 134 ): 135 super().__init__() 136 137 self.blank_id = blank_id 138 self.preserve_alignments = preserve_alignments 139 # we need timestamps to extract non-blank per-frame confidence 140 self.compute_timestamps = compute_timestamps | preserve_frame_confidence 141 self.preserve_frame_confidence = preserve_frame_confidence 142 143 # set confidence calculation measure 144 self._init_confidence_measure(confidence_measure_cfg) 145 146 @typecheck() 147 def forward( 148 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 149 ): 150 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 151 Output token is generated auto-repressively. 152 153 Args: 154 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 155 decoder_lengths: list of int representing the length of each sequence 156 output sequence. 157 158 Returns: 159 packed list containing batch number of sentences (Hypotheses). 160 """ 161 with torch.inference_mode(): 162 hypotheses = [] 163 # Process each sequence independently 164 prediction_cpu_tensor = decoder_output.cpu() 165 166 if prediction_cpu_tensor.ndim < 2 or prediction_cpu_tensor.ndim > 3: 167 raise ValueError( 168 f"`decoder_output` must be a tensor of shape [B, T] (labels, int) or " 169 f"[B, T, V] (log probs, float). Provided shape = {prediction_cpu_tensor.shape}" 170 ) 171 172 # determine type of input - logprobs or labels 173 if prediction_cpu_tensor.ndim == 2: # labels 174 greedy_decode = self._greedy_decode_labels 175 else: 176 greedy_decode = self._greedy_decode_logprobs 177 178 for ind in range(prediction_cpu_tensor.shape[0]): 179 out_len = decoder_lengths[ind] if decoder_lengths is not None else None 180 hypothesis = greedy_decode(prediction_cpu_tensor[ind], out_len) 181 hypotheses.append(hypothesis) 182 183 # Pack results into Hypotheses 184 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 185 186 return (packed_result,) 187 188 @torch.no_grad() 189 def _greedy_decode_logprobs(self, x: torch.Tensor, out_len: torch.Tensor): 190 # x: [T, D] 191 # out_len: [seq_len] 192 193 # Initialize blank state and empty label set in Hypothesis 194 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 195 prediction = x.detach().cpu() 196 197 if out_len is not None: 198 prediction = prediction[:out_len] 199 200 prediction_logprobs, prediction_labels = prediction.max(dim=-1) 201 202 non_blank_ids = prediction_labels != self.blank_id 203 hypothesis.y_sequence = prediction_labels.numpy().tolist() 204 hypothesis.score = (prediction_logprobs[non_blank_ids]).sum() 205 206 if self.preserve_alignments: 207 # Preserve the logprobs, as well as labels after argmax 208 hypothesis.alignments = (prediction.clone(), prediction_labels.clone()) 209 210 if self.compute_timestamps: 211 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 212 213 if self.preserve_frame_confidence: 214 hypothesis.frame_confidence = self._get_confidence(prediction) 215 216 return hypothesis 217 218 @torch.no_grad() 219 def _greedy_decode_labels(self, x: torch.Tensor, out_len: torch.Tensor): 220 # x: [T] 221 # out_len: [seq_len] 222 223 # Initialize blank state and empty label set in Hypothesis 224 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 225 prediction_labels = x.detach().cpu() 226 227 if out_len is not None: 228 prediction_labels = prediction_labels[:out_len] 229 230 non_blank_ids = prediction_labels != self.blank_id 231 hypothesis.y_sequence = prediction_labels.numpy().tolist() 232 hypothesis.score = -1.0 233 234 if self.preserve_alignments: 235 raise ValueError("Requested for alignments, but predictions provided were labels, not log probabilities.") 236 237 if self.compute_timestamps: 238 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 239 240 if self.preserve_frame_confidence: 241 raise ValueError( 242 "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." 243 ) 244 245 return hypothesis 246 247 def __call__(self, *args, **kwargs): 248 return self.forward(*args, **kwargs) 249 250 251 @dataclass 252 class GreedyCTCInferConfig: 253 preserve_alignments: bool = False 254 compute_timestamps: bool = False 255 preserve_frame_confidence: bool = False 256 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 257 confidence_method_cfg: str = "DEPRECATED" 258 259 def __post_init__(self): 260 # OmegaConf.structured ensures that post_init check is always executed 261 self.confidence_measure_cfg = OmegaConf.structured( 262 self.confidence_measure_cfg 263 if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) 264 else ConfidenceMeasureConfig(**self.confidence_measure_cfg) 265 ) 266 if self.confidence_method_cfg != "DEPRECATED": 267 logging.warning( 268 "`confidence_method_cfg` is deprecated and will be removed in the future. " 269 "Please use `confidence_measure_cfg` instead." 270 ) 271 272 # TODO (alaptev): delete the following two lines sometime in the future 273 logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") 274 # OmegaConf.structured ensures that post_init check is always executed 275 self.confidence_measure_cfg = OmegaConf.structured( 276 self.confidence_method_cfg 277 if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) 278 else ConfidenceMeasureConfig(**self.confidence_method_cfg) 279 ) 280 [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 # Copyright 2017 Johns Hopkins University (Shinji Watanabe) 16 # 17 # Licensed under the Apache License, Version 2.0 (the "License"); 18 # you may not use this file except in compliance with the License. 19 # You may obtain a copy of the License at 20 # 21 # http://www.apache.org/licenses/LICENSE-2.0 22 # 23 # Unless required by applicable law or agreed to in writing, software 24 # distributed under the License is distributed on an "AS IS" BASIS, 25 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 26 # See the License for the specific language governing permissions and 27 # limitations under the License. 28 29 from dataclasses import dataclass 30 from typing import List, Optional, Tuple, Union 31 32 import numpy as np 33 import torch 34 from omegaconf import DictConfig, OmegaConf 35 36 from nemo.collections.asr.modules import rnnt_abstract 37 from nemo.collections.asr.parts.utils import rnnt_utils 38 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMeasureConfig, ConfidenceMeasureMixin 39 from nemo.collections.common.parts.rnn import label_collate 40 from nemo.core.classes import Typing, typecheck 41 from nemo.core.neural_types import AcousticEncodedRepresentation, ElementType, HypothesisType, LengthsType, NeuralType 42 from nemo.utils import logging 43 44 45 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 46 47 if hasattr(logitlen, 'cpu'): 48 logitlen_cpu = logitlen.to('cpu') 49 else: 50 logitlen_cpu = logitlen 51 52 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 53 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 54 hyp.length = logitlen_cpu[idx] 55 56 if hyp.dec_state is not None: 57 hyp.dec_state = _states_to_device(hyp.dec_state) 58 59 return hypotheses 60 61 62 def _states_to_device(dec_state, device='cpu'): 63 if torch.is_tensor(dec_state): 64 dec_state = dec_state.to(device) 65 66 elif isinstance(dec_state, (list, tuple)): 67 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 68 69 return dec_state 70 71 72 class _GreedyRNNTInfer(Typing, ConfidenceMeasureMixin): 73 """A greedy transducer decoder. 74 75 Provides a common abstraction for sample level and batch level greedy decoding. 76 77 Args: 78 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 79 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 80 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 81 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 82 to a sequence in a single time step; if set to None then there is 83 no limit. 84 preserve_alignments: Bool flag which preserves the history of alignments generated during 85 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 86 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 87 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 88 89 The length of the list corresponds to the Acoustic Length (T). 90 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 91 U is the number of target tokens for the current timestep Ti. 92 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 93 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 94 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 95 96 The length of the list corresponds to the Acoustic Length (T). 97 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 98 U is the number of target tokens for the current timestep Ti. 99 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 100 confidence scores. 101 102 name: The measure name (str). 103 Supported values: 104 - 'max_prob' for using the maximum token probability as a confidence. 105 - 'entropy' for using a normalized entropy of a log-likelihood vector. 106 107 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 108 Supported values: 109 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 110 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 111 Note that for this entropy, the alpha should comply the following inequality: 112 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 113 where V is the model vocabulary size. 114 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 115 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 116 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 117 More: https://en.wikipedia.org/wiki/Tsallis_entropy 118 - 'renyi' for the Rรฉnyi entropy. 119 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 120 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 121 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 122 123 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 124 When the alpha equals one, scaling is not applied to 'max_prob', 125 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 126 127 entropy_norm: A mapping of the entropy value to the interval [0,1]. 128 Supported values: 129 - 'lin' for using the linear mapping. 130 - 'exp' for using exponential mapping with linear shift. 131 """ 132 133 @property 134 def input_types(self): 135 """Returns definitions of module input ports. 136 """ 137 return { 138 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), 139 "encoded_lengths": NeuralType(tuple('B'), LengthsType()), 140 "partial_hypotheses": [NeuralType(elements_type=HypothesisType(), optional=True)], # must always be last 141 } 142 143 @property 144 def output_types(self): 145 """Returns definitions of module output ports. 146 """ 147 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 148 149 def __init__( 150 self, 151 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 152 joint_model: rnnt_abstract.AbstractRNNTJoint, 153 blank_index: int, 154 max_symbols_per_step: Optional[int] = None, 155 preserve_alignments: bool = False, 156 preserve_frame_confidence: bool = False, 157 confidence_measure_cfg: Optional[DictConfig] = None, 158 ): 159 super().__init__() 160 self.decoder = decoder_model 161 self.joint = joint_model 162 163 self._blank_index = blank_index 164 self._SOS = blank_index # Start of single index 165 self.max_symbols = max_symbols_per_step 166 self.preserve_alignments = preserve_alignments 167 self.preserve_frame_confidence = preserve_frame_confidence 168 169 # set confidence calculation measure 170 self._init_confidence_measure(confidence_measure_cfg) 171 172 def __call__(self, *args, **kwargs): 173 return self.forward(*args, **kwargs) 174 175 @torch.no_grad() 176 def _pred_step( 177 self, 178 label: Union[torch.Tensor, int], 179 hidden: Optional[torch.Tensor], 180 add_sos: bool = False, 181 batch_size: Optional[int] = None, 182 ) -> Tuple[torch.Tensor, torch.Tensor]: 183 """ 184 Common prediction step based on the AbstractRNNTDecoder implementation. 185 186 Args: 187 label: (int/torch.Tensor): Label or "Start-of-Signal" token. 188 hidden: (Optional torch.Tensor): RNN State vector 189 add_sos (bool): Whether to add a zero vector at the begging as "start of sentence" token. 190 batch_size: Batch size of the output tensor. 191 192 Returns: 193 g: (B, U, H) if add_sos is false, else (B, U + 1, H) 194 hid: (h, c) where h is the final sequence hidden state and c is 195 the final cell state: 196 h (tensor), shape (L, B, H) 197 c (tensor), shape (L, B, H) 198 """ 199 if isinstance(label, torch.Tensor): 200 # label: [batch, 1] 201 if label.dtype != torch.long: 202 label = label.long() 203 204 else: 205 # Label is an integer 206 if label == self._SOS: 207 return self.decoder.predict(None, hidden, add_sos=add_sos, batch_size=batch_size) 208 209 label = label_collate([[label]]) 210 211 # output: [B, 1, K] 212 return self.decoder.predict(label, hidden, add_sos=add_sos, batch_size=batch_size) 213 214 def _joint_step(self, enc, pred, log_normalize: Optional[bool] = None): 215 """ 216 Common joint step based on AbstractRNNTJoint implementation. 217 218 Args: 219 enc: Output of the Encoder model. A torch.Tensor of shape [B, 1, H1] 220 pred: Output of the Decoder model. A torch.Tensor of shape [B, 1, H2] 221 log_normalize: Whether to log normalize or not. None will log normalize only for CPU. 222 223 Returns: 224 logits of shape (B, T=1, U=1, V + 1) 225 """ 226 with torch.no_grad(): 227 logits = self.joint.joint(enc, pred) 228 229 if log_normalize is None: 230 if not logits.is_cuda: # Use log softmax only if on CPU 231 logits = logits.log_softmax(dim=len(logits.shape) - 1) 232 else: 233 if log_normalize: 234 logits = logits.log_softmax(dim=len(logits.shape) - 1) 235 236 return logits 237 238 239 class GreedyRNNTInfer(_GreedyRNNTInfer): 240 """A greedy transducer decoder. 241 242 Sequence level greedy decoding, performed auto-regressively. 243 244 Args: 245 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 246 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 247 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 248 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 249 to a sequence in a single time step; if set to None then there is 250 no limit. 251 preserve_alignments: Bool flag which preserves the history of alignments generated during 252 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 253 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 254 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 255 256 The length of the list corresponds to the Acoustic Length (T). 257 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 258 U is the number of target tokens for the current timestep Ti. 259 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 260 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 261 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 262 263 The length of the list corresponds to the Acoustic Length (T). 264 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 265 U is the number of target tokens for the current timestep Ti. 266 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 267 confidence scores. 268 269 name: The measure name (str). 270 Supported values: 271 - 'max_prob' for using the maximum token probability as a confidence. 272 - 'entropy' for using a normalized entropy of a log-likelihood vector. 273 274 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 275 Supported values: 276 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 277 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 278 Note that for this entropy, the alpha should comply the following inequality: 279 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 280 where V is the model vocabulary size. 281 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 282 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/Tsallis_entropy 285 - 'renyi' for the Rรฉnyi entropy. 286 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 287 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 288 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 289 290 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 291 When the alpha equals one, scaling is not applied to 'max_prob', 292 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 293 294 entropy_norm: A mapping of the entropy value to the interval [0,1]. 295 Supported values: 296 - 'lin' for using the linear mapping. 297 - 'exp' for using exponential mapping with linear shift. 298 """ 299 300 def __init__( 301 self, 302 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 303 joint_model: rnnt_abstract.AbstractRNNTJoint, 304 blank_index: int, 305 max_symbols_per_step: Optional[int] = None, 306 preserve_alignments: bool = False, 307 preserve_frame_confidence: bool = False, 308 confidence_measure_cfg: Optional[DictConfig] = None, 309 ): 310 super().__init__( 311 decoder_model=decoder_model, 312 joint_model=joint_model, 313 blank_index=blank_index, 314 max_symbols_per_step=max_symbols_per_step, 315 preserve_alignments=preserve_alignments, 316 preserve_frame_confidence=preserve_frame_confidence, 317 confidence_measure_cfg=confidence_measure_cfg, 318 ) 319 320 @typecheck() 321 def forward( 322 self, 323 encoder_output: torch.Tensor, 324 encoded_lengths: torch.Tensor, 325 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 326 ): 327 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 328 Output token is generated auto-regressively. 329 330 Args: 331 encoder_output: A tensor of size (batch, features, timesteps). 332 encoded_lengths: list of int representing the length of each sequence 333 output sequence. 334 335 Returns: 336 packed list containing batch number of sentences (Hypotheses). 337 """ 338 # Preserve decoder and joint training state 339 decoder_training_state = self.decoder.training 340 joint_training_state = self.joint.training 341 342 with torch.inference_mode(): 343 # Apply optional preprocessing 344 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 345 346 self.decoder.eval() 347 self.joint.eval() 348 349 hypotheses = [] 350 # Process each sequence independently 351 with self.decoder.as_frozen(), self.joint.as_frozen(): 352 for batch_idx in range(encoder_output.size(0)): 353 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 354 logitlen = encoded_lengths[batch_idx] 355 356 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 357 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 358 hypotheses.append(hypothesis) 359 360 # Pack results into Hypotheses 361 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 362 363 self.decoder.train(decoder_training_state) 364 self.joint.train(joint_training_state) 365 366 return (packed_result,) 367 368 @torch.no_grad() 369 def _greedy_decode( 370 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 371 ): 372 # x: [T, 1, D] 373 # out_len: [seq_len] 374 375 # Initialize blank state and empty label set in Hypothesis 376 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 377 378 if partial_hypotheses is not None: 379 hypothesis.last_token = partial_hypotheses.last_token 380 hypothesis.y_sequence = ( 381 partial_hypotheses.y_sequence.cpu().tolist() 382 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 383 else partial_hypotheses.y_sequence 384 ) 385 if partial_hypotheses.dec_state is not None: 386 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 387 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 388 389 if self.preserve_alignments: 390 # Alignments is a 2-dimensional dangling list representing T x U 391 hypothesis.alignments = [[]] 392 393 if self.preserve_frame_confidence: 394 hypothesis.frame_confidence = [[]] 395 396 # For timestep t in X_t 397 for time_idx in range(out_len): 398 # Extract encoder embedding at timestep t 399 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 400 f = x.narrow(dim=0, start=time_idx, length=1) 401 402 # Setup exit flags and counter 403 not_blank = True 404 symbols_added = 0 405 # While blank is not predicted, or we dont run out of max symbols per timestep 406 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 407 # In the first timestep, we initialize the network with RNNT Blank 408 # In later timesteps, we provide previous predicted label as input. 409 if hypothesis.last_token is None and hypothesis.dec_state is None: 410 last_label = self._SOS 411 else: 412 last_label = label_collate([[hypothesis.last_token]]) 413 414 # Perform prediction network and joint network steps. 415 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 416 # If preserving per-frame confidence, log_normalize must be true 417 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 418 0, 0, 0, : 419 ] 420 421 del g 422 423 # torch.max(0) op doesnt exist for FP 16. 424 if logp.dtype != torch.float32: 425 logp = logp.float() 426 427 # get index k, of max prob 428 v, k = logp.max(0) 429 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 430 431 if self.preserve_alignments: 432 # insert logprobs into last timestep 433 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 434 435 if self.preserve_frame_confidence: 436 # insert confidence into last timestep 437 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 438 439 del logp 440 441 # If blank token is predicted, exit inner loop, move onto next timestep t 442 if k == self._blank_index: 443 not_blank = False 444 else: 445 # Append token to label set, update RNN state. 446 hypothesis.y_sequence.append(k) 447 hypothesis.score += float(v) 448 hypothesis.timestep.append(time_idx) 449 hypothesis.dec_state = hidden_prime 450 hypothesis.last_token = k 451 452 # Increment token counter. 453 symbols_added += 1 454 455 if self.preserve_alignments: 456 # convert Ti-th logits into a torch array 457 hypothesis.alignments.append([]) # blank buffer for next timestep 458 459 if self.preserve_frame_confidence: 460 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 461 462 # Remove trailing empty list of Alignments 463 if self.preserve_alignments: 464 if len(hypothesis.alignments[-1]) == 0: 465 del hypothesis.alignments[-1] 466 467 # Remove trailing empty list of per-frame confidence 468 if self.preserve_frame_confidence: 469 if len(hypothesis.frame_confidence[-1]) == 0: 470 del hypothesis.frame_confidence[-1] 471 472 # Unpack the hidden states 473 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 474 475 return hypothesis 476 477 478 class GreedyBatchedRNNTInfer(_GreedyRNNTInfer): 479 """A batch level greedy transducer decoder. 480 481 Batch level greedy decoding, performed auto-regressively. 482 483 Args: 484 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 485 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 486 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 487 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 488 to a sequence in a single time step; if set to None then there is 489 no limit. 490 preserve_alignments: Bool flag which preserves the history of alignments generated during 491 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 492 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 493 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 494 495 The length of the list corresponds to the Acoustic Length (T). 496 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 497 U is the number of target tokens for the current timestep Ti. 498 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 499 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 500 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 501 502 The length of the list corresponds to the Acoustic Length (T). 503 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 504 U is the number of target tokens for the current timestep Ti. 505 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 506 confidence scores. 507 508 name: The measure name (str). 509 Supported values: 510 - 'max_prob' for using the maximum token probability as a confidence. 511 - 'entropy' for using a normalized entropy of a log-likelihood vector. 512 513 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 514 Supported values: 515 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 516 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 517 Note that for this entropy, the alpha should comply the following inequality: 518 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 519 where V is the model vocabulary size. 520 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 521 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 522 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 523 More: https://en.wikipedia.org/wiki/Tsallis_entropy 524 - 'renyi' for the Rรฉnyi entropy. 525 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 526 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 527 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 528 529 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 530 When the alpha equals one, scaling is not applied to 'max_prob', 531 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 532 533 entropy_norm: A mapping of the entropy value to the interval [0,1]. 534 Supported values: 535 - 'lin' for using the linear mapping. 536 - 'exp' for using exponential mapping with linear shift. 537 """ 538 539 def __init__( 540 self, 541 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 542 joint_model: rnnt_abstract.AbstractRNNTJoint, 543 blank_index: int, 544 max_symbols_per_step: Optional[int] = None, 545 preserve_alignments: bool = False, 546 preserve_frame_confidence: bool = False, 547 confidence_measure_cfg: Optional[DictConfig] = None, 548 ): 549 super().__init__( 550 decoder_model=decoder_model, 551 joint_model=joint_model, 552 blank_index=blank_index, 553 max_symbols_per_step=max_symbols_per_step, 554 preserve_alignments=preserve_alignments, 555 preserve_frame_confidence=preserve_frame_confidence, 556 confidence_measure_cfg=confidence_measure_cfg, 557 ) 558 559 # Depending on availability of `blank_as_pad` support 560 # switch between more efficient batch decoding technique 561 if self.decoder.blank_as_pad: 562 self._greedy_decode = self._greedy_decode_blank_as_pad 563 else: 564 self._greedy_decode = self._greedy_decode_masked 565 566 @typecheck() 567 def forward( 568 self, 569 encoder_output: torch.Tensor, 570 encoded_lengths: torch.Tensor, 571 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 572 ): 573 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 574 Output token is generated auto-regressively. 575 576 Args: 577 encoder_output: A tensor of size (batch, features, timesteps). 578 encoded_lengths: list of int representing the length of each sequence 579 output sequence. 580 581 Returns: 582 packed list containing batch number of sentences (Hypotheses). 583 """ 584 # Preserve decoder and joint training state 585 decoder_training_state = self.decoder.training 586 joint_training_state = self.joint.training 587 588 with torch.inference_mode(): 589 # Apply optional preprocessing 590 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 591 logitlen = encoded_lengths 592 593 self.decoder.eval() 594 self.joint.eval() 595 596 with self.decoder.as_frozen(), self.joint.as_frozen(): 597 inseq = encoder_output # [B, T, D] 598 hypotheses = self._greedy_decode( 599 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 600 ) 601 602 # Pack the hypotheses results 603 packed_result = pack_hypotheses(hypotheses, logitlen) 604 605 self.decoder.train(decoder_training_state) 606 self.joint.train(joint_training_state) 607 608 return (packed_result,) 609 610 def _greedy_decode_blank_as_pad( 611 self, 612 x: torch.Tensor, 613 out_len: torch.Tensor, 614 device: torch.device, 615 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 616 ): 617 if partial_hypotheses is not None: 618 raise NotImplementedError("`partial_hypotheses` support is not supported") 619 620 with torch.inference_mode(): 621 # x: [B, T, D] 622 # out_len: [B] 623 # device: torch.device 624 625 # Initialize list of Hypothesis 626 batchsize = x.shape[0] 627 hypotheses = [ 628 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 629 ] 630 631 # Initialize Hidden state matrix (shared by entire batch) 632 hidden = None 633 634 # If alignments need to be preserved, register a dangling list to hold the values 635 if self.preserve_alignments: 636 # alignments is a 3-dimensional dangling list representing B x T x U 637 for hyp in hypotheses: 638 hyp.alignments = [[]] 639 640 # If confidence scores need to be preserved, register a dangling list to hold the values 641 if self.preserve_frame_confidence: 642 # frame_confidence is a 3-dimensional dangling list representing B x T x U 643 for hyp in hypotheses: 644 hyp.frame_confidence = [[]] 645 646 # Last Label buffer + Last Label without blank buffer 647 # batch level equivalent of the last_label 648 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 649 650 # Mask buffers 651 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 652 653 # Get max sequence length 654 max_out_len = out_len.max() 655 for time_idx in range(max_out_len): 656 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 657 658 # Prepare t timestamp batch variables 659 not_blank = True 660 symbols_added = 0 661 662 # Reset blank mask 663 blank_mask.mul_(False) 664 665 # Update blank mask with time mask 666 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 667 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 668 blank_mask = time_idx >= out_len 669 # Start inner loop 670 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 671 # Batch prediction and joint network steps 672 # If very first prediction step, submit SOS tag (blank) to pred_step. 673 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 674 if time_idx == 0 and symbols_added == 0 and hidden is None: 675 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 676 else: 677 # Perform batch step prediction of decoder, getting new states and scores ("g") 678 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 679 680 # Batched joint step - Output = [B, V + 1] 681 # If preserving per-frame confidence, log_normalize must be true 682 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 683 :, 0, 0, : 684 ] 685 686 if logp.dtype != torch.float32: 687 logp = logp.float() 688 689 # Get index k, of max prob for batch 690 v, k = logp.max(1) 691 del g 692 693 # Update blank mask with current predicted blanks 694 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 695 k_is_blank = k == self._blank_index 696 blank_mask.bitwise_or_(k_is_blank) 697 all_blanks = torch.all(blank_mask) 698 699 del k_is_blank 700 701 # If preserving alignments, check if sequence length of sample has been reached 702 # before adding alignment 703 if self.preserve_alignments: 704 # Insert logprobs into last timestep per sample 705 logp_vals = logp.to('cpu') 706 logp_ids = logp_vals.max(1)[1] 707 for batch_idx, is_blank in enumerate(blank_mask): 708 # we only want to update non-blanks, unless we are at the last step in the loop where 709 # all elements produced blanks, otherwise there will be duplicate predictions 710 # saved in alignments 711 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 712 hypotheses[batch_idx].alignments[-1].append( 713 (logp_vals[batch_idx], logp_ids[batch_idx]) 714 ) 715 del logp_vals 716 717 # If preserving per-frame confidence, check if sequence length of sample has been reached 718 # before adding confidence scores 719 if self.preserve_frame_confidence: 720 # Insert probabilities into last timestep per sample 721 confidence = self._get_confidence(logp) 722 for batch_idx, is_blank in enumerate(blank_mask): 723 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 724 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 725 del logp 726 727 # If all samples predict / have predicted prior blanks, exit loop early 728 # This is equivalent to if single sample predicted k 729 if all_blanks: 730 not_blank = False 731 else: 732 # Collect batch indices where blanks occurred now/past 733 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 734 735 # Recover prior state for all samples which predicted blank now/past 736 if hidden is not None: 737 # LSTM has 2 states 738 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 739 740 elif len(blank_indices) > 0 and hidden is None: 741 # Reset state if there were some blank and other non-blank predictions in batch 742 # Original state is filled with zeros so we just multiply 743 # LSTM has 2 states 744 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 745 746 # Recover prior predicted label for all samples which predicted blank now/past 747 k[blank_indices] = last_label[blank_indices, 0] 748 749 # Update new label and hidden state for next iteration 750 last_label = k.clone().view(-1, 1) 751 hidden = hidden_prime 752 753 # Update predicted labels, accounting for time mask 754 # If blank was predicted even once, now or in the past, 755 # Force the current predicted label to also be blank 756 # This ensures that blanks propogate across all timesteps 757 # once they have occured (normally stopping condition of sample level loop). 758 for kidx, ki in enumerate(k): 759 if blank_mask[kidx] == 0: 760 hypotheses[kidx].y_sequence.append(ki) 761 hypotheses[kidx].timestep.append(time_idx) 762 hypotheses[kidx].score += float(v[kidx]) 763 symbols_added += 1 764 765 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 766 # Then preserve U at current timestep Ti 767 # Finally, forward the timestep history to Ti+1 for that sample 768 # All of this should only be done iff the current time index <= sample-level AM length. 769 # Otherwise ignore and move to next sample / next timestep. 770 if self.preserve_alignments: 771 772 # convert Ti-th logits into a torch array 773 for batch_idx in range(batchsize): 774 775 # this checks if current timestep <= sample-level AM length 776 # If current timestep > sample-level AM length, no alignments will be added 777 # Therefore the list of Uj alignments is empty here. 778 if len(hypotheses[batch_idx].alignments[-1]) > 0: 779 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 780 781 # Do the same if preserving per-frame confidence 782 if self.preserve_frame_confidence: 783 784 for batch_idx in range(batchsize): 785 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 786 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 787 788 # Remove trailing empty list of alignments at T_{am-len} x Uj 789 if self.preserve_alignments: 790 for batch_idx in range(batchsize): 791 if len(hypotheses[batch_idx].alignments[-1]) == 0: 792 del hypotheses[batch_idx].alignments[-1] 793 794 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 795 if self.preserve_frame_confidence: 796 for batch_idx in range(batchsize): 797 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 798 del hypotheses[batch_idx].frame_confidence[-1] 799 800 # Preserve states 801 for batch_idx in range(batchsize): 802 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 803 804 return hypotheses 805 806 def _greedy_decode_masked( 807 self, 808 x: torch.Tensor, 809 out_len: torch.Tensor, 810 device: torch.device, 811 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 812 ): 813 if partial_hypotheses is not None: 814 raise NotImplementedError("`partial_hypotheses` support is not supported") 815 816 # x: [B, T, D] 817 # out_len: [B] 818 # device: torch.device 819 820 # Initialize state 821 batchsize = x.shape[0] 822 hypotheses = [ 823 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 824 ] 825 826 # Initialize Hidden state matrix (shared by entire batch) 827 hidden = None 828 829 # If alignments need to be preserved, register a danling list to hold the values 830 if self.preserve_alignments: 831 # alignments is a 3-dimensional dangling list representing B x T x U 832 for hyp in hypotheses: 833 hyp.alignments = [[]] 834 else: 835 alignments = None 836 837 # If confidence scores need to be preserved, register a danling list to hold the values 838 if self.preserve_frame_confidence: 839 # frame_confidence is a 3-dimensional dangling list representing B x T x U 840 for hyp in hypotheses: 841 hyp.frame_confidence = [[]] 842 843 # Last Label buffer + Last Label without blank buffer 844 # batch level equivalent of the last_label 845 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 846 last_label_without_blank = last_label.clone() 847 848 # Mask buffers 849 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 850 851 # Get max sequence length 852 max_out_len = out_len.max() 853 854 with torch.inference_mode(): 855 for time_idx in range(max_out_len): 856 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 857 858 # Prepare t timestamp batch variables 859 not_blank = True 860 symbols_added = 0 861 862 # Reset blank mask 863 blank_mask.mul_(False) 864 865 # Update blank mask with time mask 866 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 867 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 868 blank_mask = time_idx >= out_len 869 870 # Start inner loop 871 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 872 # Batch prediction and joint network steps 873 # If very first prediction step, submit SOS tag (blank) to pred_step. 874 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 875 if time_idx == 0 and symbols_added == 0 and hidden is None: 876 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 877 else: 878 # Set a dummy label for the blank value 879 # This value will be overwritten by "blank" again the last label update below 880 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 881 last_label_without_blank_mask = last_label == self._blank_index 882 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 883 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 884 ~last_label_without_blank_mask 885 ] 886 887 # Perform batch step prediction of decoder, getting new states and scores ("g") 888 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 889 890 # Batched joint step - Output = [B, V + 1] 891 # If preserving per-frame confidence, log_normalize must be true 892 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 893 :, 0, 0, : 894 ] 895 896 if logp.dtype != torch.float32: 897 logp = logp.float() 898 899 # Get index k, of max prob for batch 900 v, k = logp.max(1) 901 del g 902 903 # Update blank mask with current predicted blanks 904 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 905 k_is_blank = k == self._blank_index 906 blank_mask.bitwise_or_(k_is_blank) 907 all_blanks = torch.all(blank_mask) 908 909 # If preserving alignments, check if sequence length of sample has been reached 910 # before adding alignment 911 if self.preserve_alignments: 912 # Insert logprobs into last timestep per sample 913 logp_vals = logp.to('cpu') 914 logp_ids = logp_vals.max(1)[1] 915 for batch_idx, is_blank in enumerate(blank_mask): 916 # we only want to update non-blanks, unless we are at the last step in the loop where 917 # all elements produced blanks, otherwise there will be duplicate predictions 918 # saved in alignments 919 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 920 hypotheses[batch_idx].alignments[-1].append( 921 (logp_vals[batch_idx], logp_ids[batch_idx]) 922 ) 923 924 del logp_vals 925 926 # If preserving per-frame confidence, check if sequence length of sample has been reached 927 # before adding confidence scores 928 if self.preserve_frame_confidence: 929 # Insert probabilities into last timestep per sample 930 confidence = self._get_confidence(logp) 931 for batch_idx, is_blank in enumerate(blank_mask): 932 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 933 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 934 del logp 935 936 # If all samples predict / have predicted prior blanks, exit loop early 937 # This is equivalent to if single sample predicted k 938 if blank_mask.all(): 939 not_blank = False 940 else: 941 # Collect batch indices where blanks occurred now/past 942 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 943 944 # Recover prior state for all samples which predicted blank now/past 945 if hidden is not None: 946 # LSTM has 2 states 947 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 948 949 elif len(blank_indices) > 0 and hidden is None: 950 # Reset state if there were some blank and other non-blank predictions in batch 951 # Original state is filled with zeros so we just multiply 952 # LSTM has 2 states 953 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 954 955 # Recover prior predicted label for all samples which predicted blank now/past 956 k[blank_indices] = last_label[blank_indices, 0] 957 958 # Update new label and hidden state for next iteration 959 last_label = k.view(-1, 1) 960 hidden = hidden_prime 961 962 # Update predicted labels, accounting for time mask 963 # If blank was predicted even once, now or in the past, 964 # Force the current predicted label to also be blank 965 # This ensures that blanks propogate across all timesteps 966 # once they have occured (normally stopping condition of sample level loop). 967 for kidx, ki in enumerate(k): 968 if blank_mask[kidx] == 0: 969 hypotheses[kidx].y_sequence.append(ki) 970 hypotheses[kidx].timestep.append(time_idx) 971 hypotheses[kidx].score += float(v[kidx]) 972 973 symbols_added += 1 974 975 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 976 # Then preserve U at current timestep Ti 977 # Finally, forward the timestep history to Ti+1 for that sample 978 # All of this should only be done iff the current time index <= sample-level AM length. 979 # Otherwise ignore and move to next sample / next timestep. 980 if self.preserve_alignments: 981 982 # convert Ti-th logits into a torch array 983 for batch_idx in range(batchsize): 984 985 # this checks if current timestep <= sample-level AM length 986 # If current timestep > sample-level AM length, no alignments will be added 987 # Therefore the list of Uj alignments is empty here. 988 if len(hypotheses[batch_idx].alignments[-1]) > 0: 989 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 990 991 # Do the same if preserving per-frame confidence 992 if self.preserve_frame_confidence: 993 994 for batch_idx in range(batchsize): 995 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 996 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 997 998 # Remove trailing empty list of alignments at T_{am-len} x Uj 999 if self.preserve_alignments: 1000 for batch_idx in range(batchsize): 1001 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1002 del hypotheses[batch_idx].alignments[-1] 1003 1004 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1005 if self.preserve_frame_confidence: 1006 for batch_idx in range(batchsize): 1007 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1008 del hypotheses[batch_idx].frame_confidence[-1] 1009 1010 # Preserve states 1011 for batch_idx in range(batchsize): 1012 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1013 1014 return hypotheses 1015 1016 1017 class ExportedModelGreedyBatchedRNNTInfer: 1018 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = None): 1019 self.encoder_model_path = encoder_model 1020 self.decoder_joint_model_path = decoder_joint_model 1021 self.max_symbols_per_step = max_symbols_per_step 1022 1023 # Will be populated at runtime 1024 self._blank_index = None 1025 1026 def __call__(self, audio_signal: torch.Tensor, length: torch.Tensor): 1027 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 1028 Output token is generated auto-regressively. 1029 1030 Args: 1031 encoder_output: A tensor of size (batch, features, timesteps). 1032 encoded_lengths: list of int representing the length of each sequence 1033 output sequence. 1034 1035 Returns: 1036 packed list containing batch number of sentences (Hypotheses). 1037 """ 1038 with torch.no_grad(): 1039 # Apply optional preprocessing 1040 encoder_output, encoded_lengths = self.run_encoder(audio_signal=audio_signal, length=length) 1041 1042 if torch.is_tensor(encoder_output): 1043 encoder_output = encoder_output.transpose(1, 2) 1044 else: 1045 encoder_output = encoder_output.transpose([0, 2, 1]) # (B, T, D) 1046 logitlen = encoded_lengths 1047 1048 inseq = encoder_output # [B, T, D] 1049 hypotheses, timestamps = self._greedy_decode(inseq, logitlen) 1050 1051 # Pack the hypotheses results 1052 packed_result = [rnnt_utils.Hypothesis(score=-1.0, y_sequence=[]) for _ in range(len(hypotheses))] 1053 for i in range(len(packed_result)): 1054 packed_result[i].y_sequence = torch.tensor(hypotheses[i], dtype=torch.long) 1055 packed_result[i].length = timestamps[i] 1056 1057 del hypotheses 1058 1059 return packed_result 1060 1061 def _greedy_decode(self, x, out_len): 1062 # x: [B, T, D] 1063 # out_len: [B] 1064 1065 # Initialize state 1066 batchsize = x.shape[0] 1067 hidden = self._get_initial_states(batchsize) 1068 target_lengths = torch.ones(batchsize, dtype=torch.int32) 1069 1070 # Output string buffer 1071 label = [[] for _ in range(batchsize)] 1072 timesteps = [[] for _ in range(batchsize)] 1073 1074 # Last Label buffer + Last Label without blank buffer 1075 # batch level equivalent of the last_label 1076 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long).numpy() 1077 if torch.is_tensor(x): 1078 last_label = torch.from_numpy(last_label).to(self.device) 1079 1080 # Mask buffers 1081 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool).numpy() 1082 1083 # Get max sequence length 1084 max_out_len = out_len.max() 1085 for time_idx in range(max_out_len): 1086 f = x[:, time_idx : time_idx + 1, :] # [B, 1, D] 1087 1088 if torch.is_tensor(f): 1089 f = f.transpose(1, 2) 1090 else: 1091 f = f.transpose([0, 2, 1]) 1092 1093 # Prepare t timestamp batch variables 1094 not_blank = True 1095 symbols_added = 0 1096 1097 # Reset blank mask 1098 blank_mask *= False 1099 1100 # Update blank mask with time mask 1101 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1102 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1103 blank_mask = time_idx >= out_len 1104 # Start inner loop 1105 while not_blank and (self.max_symbols_per_step is None or symbols_added < self.max_symbols_per_step): 1106 1107 # Batch prediction and joint network steps 1108 # If very first prediction step, submit SOS tag (blank) to pred_step. 1109 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1110 if time_idx == 0 and symbols_added == 0: 1111 g = torch.tensor([self._blank_index] * batchsize, dtype=torch.int32).view(-1, 1) 1112 else: 1113 if torch.is_tensor(last_label): 1114 g = last_label.type(torch.int32) 1115 else: 1116 g = last_label.astype(np.int32) 1117 1118 # Batched joint step - Output = [B, V + 1] 1119 joint_out, hidden_prime = self.run_decoder_joint(f, g, target_lengths, *hidden) 1120 logp, pred_lengths = joint_out 1121 logp = logp[:, 0, 0, :] 1122 1123 # Get index k, of max prob for batch 1124 if torch.is_tensor(logp): 1125 v, k = logp.max(1) 1126 else: 1127 k = np.argmax(logp, axis=1).astype(np.int32) 1128 1129 # Update blank mask with current predicted blanks 1130 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1131 k_is_blank = k == self._blank_index 1132 blank_mask |= k_is_blank 1133 1134 del k_is_blank 1135 del logp 1136 1137 # If all samples predict / have predicted prior blanks, exit loop early 1138 # This is equivalent to if single sample predicted k 1139 if blank_mask.all(): 1140 not_blank = False 1141 1142 else: 1143 # Collect batch indices where blanks occurred now/past 1144 if torch.is_tensor(blank_mask): 1145 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1146 else: 1147 blank_indices = blank_mask.astype(np.int32).nonzero() 1148 1149 if type(blank_indices) in (list, tuple): 1150 blank_indices = blank_indices[0] 1151 1152 # Recover prior state for all samples which predicted blank now/past 1153 if hidden is not None: 1154 # LSTM has 2 states 1155 for state_id in range(len(hidden)): 1156 hidden_prime[state_id][:, blank_indices, :] = hidden[state_id][:, blank_indices, :] 1157 1158 elif len(blank_indices) > 0 and hidden is None: 1159 # Reset state if there were some blank and other non-blank predictions in batch 1160 # Original state is filled with zeros so we just multiply 1161 # LSTM has 2 states 1162 for state_id in range(len(hidden_prime)): 1163 hidden_prime[state_id][:, blank_indices, :] *= 0.0 1164 1165 # Recover prior predicted label for all samples which predicted blank now/past 1166 k[blank_indices] = last_label[blank_indices, 0] 1167 1168 # Update new label and hidden state for next iteration 1169 if torch.is_tensor(k): 1170 last_label = k.clone().reshape(-1, 1) 1171 else: 1172 last_label = k.copy().reshape(-1, 1) 1173 hidden = hidden_prime 1174 1175 # Update predicted labels, accounting for time mask 1176 # If blank was predicted even once, now or in the past, 1177 # Force the current predicted label to also be blank 1178 # This ensures that blanks propogate across all timesteps 1179 # once they have occured (normally stopping condition of sample level loop). 1180 for kidx, ki in enumerate(k): 1181 if blank_mask[kidx] == 0: 1182 label[kidx].append(ki) 1183 timesteps[kidx].append(time_idx) 1184 1185 symbols_added += 1 1186 1187 return label, timesteps 1188 1189 def _setup_blank_index(self): 1190 raise NotImplementedError() 1191 1192 def run_encoder(self, audio_signal, length): 1193 raise NotImplementedError() 1194 1195 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1196 raise NotImplementedError() 1197 1198 def _get_initial_states(self, batchsize): 1199 raise NotImplementedError() 1200 1201 1202 class ONNXGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1203 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = 10): 1204 super().__init__( 1205 encoder_model=encoder_model, 1206 decoder_joint_model=decoder_joint_model, 1207 max_symbols_per_step=max_symbols_per_step, 1208 ) 1209 1210 try: 1211 import onnx 1212 import onnxruntime 1213 except (ModuleNotFoundError, ImportError): 1214 raise ImportError(f"`onnx` or `onnxruntime` could not be imported, please install the libraries.\n") 1215 1216 if torch.cuda.is_available(): 1217 # Try to use onnxruntime-gpu 1218 providers = ['TensorrtExecutionProvider', 'CUDAExecutionProvider'] 1219 else: 1220 # Fall back to CPU and onnxruntime-cpu 1221 providers = ['CPUExecutionProvider'] 1222 1223 onnx_session_opt = onnxruntime.SessionOptions() 1224 onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL 1225 1226 onnx_model = onnx.load(self.encoder_model_path) 1227 onnx.checker.check_model(onnx_model, full_check=True) 1228 self.encoder_model = onnx_model 1229 self.encoder = onnxruntime.InferenceSession( 1230 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1231 ) 1232 1233 onnx_model = onnx.load(self.decoder_joint_model_path) 1234 onnx.checker.check_model(onnx_model, full_check=True) 1235 self.decoder_joint_model = onnx_model 1236 self.decoder_joint = onnxruntime.InferenceSession( 1237 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1238 ) 1239 1240 logging.info("Successfully loaded encoder, decoder and joint onnx models !") 1241 1242 # Will be populated at runtime 1243 self._blank_index = None 1244 self.max_symbols_per_step = max_symbols_per_step 1245 1246 self._setup_encoder_input_output_keys() 1247 self._setup_decoder_joint_input_output_keys() 1248 self._setup_blank_index() 1249 1250 def _setup_encoder_input_output_keys(self): 1251 self.encoder_inputs = list(self.encoder_model.graph.input) 1252 self.encoder_outputs = list(self.encoder_model.graph.output) 1253 1254 def _setup_decoder_joint_input_output_keys(self): 1255 self.decoder_joint_inputs = list(self.decoder_joint_model.graph.input) 1256 self.decoder_joint_outputs = list(self.decoder_joint_model.graph.output) 1257 1258 def _setup_blank_index(self): 1259 # ASSUME: Single input with no time length information 1260 dynamic_dim = 257 1261 shapes = self.encoder_inputs[0].type.tensor_type.shape.dim 1262 ip_shape = [] 1263 for shape in shapes: 1264 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1265 ip_shape.append(dynamic_dim) # replace dynamic axes with constant 1266 else: 1267 ip_shape.append(int(shape.dim_value)) 1268 1269 enc_logits, encoded_length = self.run_encoder( 1270 audio_signal=torch.randn(*ip_shape), length=torch.randint(0, 1, size=(dynamic_dim,)) 1271 ) 1272 1273 # prepare states 1274 states = self._get_initial_states(batchsize=dynamic_dim) 1275 1276 # run decoder 1 step 1277 joint_out, states = self.run_decoder_joint(enc_logits, None, None, *states) 1278 log_probs, lengths = joint_out 1279 1280 self._blank_index = log_probs.shape[-1] - 1 # last token of vocab size is blank token 1281 logging.info( 1282 f"Enc-Dec-Joint step was evaluated, blank token id = {self._blank_index}; vocab size = {log_probs.shape[-1]}" 1283 ) 1284 1285 def run_encoder(self, audio_signal, length): 1286 if hasattr(audio_signal, 'cpu'): 1287 audio_signal = audio_signal.cpu().numpy() 1288 1289 if hasattr(length, 'cpu'): 1290 length = length.cpu().numpy() 1291 1292 ip = { 1293 self.encoder_inputs[0].name: audio_signal, 1294 self.encoder_inputs[1].name: length, 1295 } 1296 enc_out = self.encoder.run(None, ip) 1297 enc_out, encoded_length = enc_out # ASSUME: single output 1298 return enc_out, encoded_length 1299 1300 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1301 # ASSUME: Decoder is RNN Transducer 1302 if targets is None: 1303 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32) 1304 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32) 1305 1306 if hasattr(targets, 'cpu'): 1307 targets = targets.cpu().numpy() 1308 1309 if hasattr(target_length, 'cpu'): 1310 target_length = target_length.cpu().numpy() 1311 1312 ip = { 1313 self.decoder_joint_inputs[0].name: enc_logits, 1314 self.decoder_joint_inputs[1].name: targets, 1315 self.decoder_joint_inputs[2].name: target_length, 1316 } 1317 1318 num_states = 0 1319 if states is not None and len(states) > 0: 1320 num_states = len(states) 1321 for idx, state in enumerate(states): 1322 if hasattr(state, 'cpu'): 1323 state = state.cpu().numpy() 1324 1325 ip[self.decoder_joint_inputs[len(ip)].name] = state 1326 1327 dec_out = self.decoder_joint.run(None, ip) 1328 1329 # unpack dec output 1330 if num_states > 0: 1331 new_states = dec_out[-num_states:] 1332 dec_out = dec_out[:-num_states] 1333 else: 1334 new_states = None 1335 1336 return dec_out, new_states 1337 1338 def _get_initial_states(self, batchsize): 1339 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1340 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1341 num_states = len(input_state_nodes) 1342 if num_states == 0: 1343 return 1344 1345 input_states = [] 1346 for state_id in range(num_states): 1347 node = input_state_nodes[state_id] 1348 ip_shape = [] 1349 for shape_idx, shape in enumerate(node.type.tensor_type.shape.dim): 1350 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1351 ip_shape.append(batchsize) # replace dynamic axes with constant 1352 else: 1353 ip_shape.append(int(shape.dim_value)) 1354 1355 input_states.append(torch.zeros(*ip_shape)) 1356 1357 return input_states 1358 1359 1360 class TorchscriptGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1361 def __init__( 1362 self, 1363 encoder_model: str, 1364 decoder_joint_model: str, 1365 cfg: DictConfig, 1366 device: str, 1367 max_symbols_per_step: Optional[int] = 10, 1368 ): 1369 super().__init__( 1370 encoder_model=encoder_model, 1371 decoder_joint_model=decoder_joint_model, 1372 max_symbols_per_step=max_symbols_per_step, 1373 ) 1374 1375 self.cfg = cfg 1376 self.device = device 1377 1378 self.encoder = torch.jit.load(self.encoder_model_path, map_location=self.device) 1379 self.decoder_joint = torch.jit.load(self.decoder_joint_model_path, map_location=self.device) 1380 1381 logging.info("Successfully loaded encoder, decoder and joint torchscript models !") 1382 1383 # Will be populated at runtime 1384 self._blank_index = None 1385 self.max_symbols_per_step = max_symbols_per_step 1386 1387 self._setup_encoder_input_keys() 1388 self._setup_decoder_joint_input_keys() 1389 self._setup_blank_index() 1390 1391 def _setup_encoder_input_keys(self): 1392 arguments = self.encoder.forward.schema.arguments[1:] 1393 self.encoder_inputs = [arg for arg in arguments] 1394 1395 def _setup_decoder_joint_input_keys(self): 1396 arguments = self.decoder_joint.forward.schema.arguments[1:] 1397 self.decoder_joint_inputs = [arg for arg in arguments] 1398 1399 def _setup_blank_index(self): 1400 self._blank_index = len(self.cfg.joint.vocabulary) 1401 1402 logging.info(f"Blank token id = {self._blank_index}; vocab size = {len(self.cfg.joint.vocabulary) + 1}") 1403 1404 def run_encoder(self, audio_signal, length): 1405 enc_out = self.encoder(audio_signal, length) 1406 enc_out, encoded_length = enc_out # ASSUME: single output 1407 return enc_out, encoded_length 1408 1409 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1410 # ASSUME: Decoder is RNN Transducer 1411 if targets is None: 1412 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32, device=enc_logits.device) 1413 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32, device=enc_logits.device) 1414 1415 num_states = 0 1416 if states is not None and len(states) > 0: 1417 num_states = len(states) 1418 1419 dec_out = self.decoder_joint(enc_logits, targets, target_length, *states) 1420 1421 # unpack dec output 1422 if num_states > 0: 1423 new_states = dec_out[-num_states:] 1424 dec_out = dec_out[:-num_states] 1425 else: 1426 new_states = None 1427 1428 return dec_out, new_states 1429 1430 def _get_initial_states(self, batchsize): 1431 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1432 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1433 num_states = len(input_state_nodes) 1434 if num_states == 0: 1435 return 1436 1437 input_states = [] 1438 for state_id in range(num_states): 1439 # Hardcode shape size for LSTM (1 is for num layers in LSTM, which is flattened for export) 1440 ip_shape = [1, batchsize, self.cfg.model_defaults.pred_hidden] 1441 input_states.append(torch.zeros(*ip_shape, device=self.device)) 1442 1443 return input_states 1444 1445 1446 class GreedyMultiblankRNNTInfer(GreedyRNNTInfer): 1447 """A greedy transducer decoder for multi-blank RNN-T. 1448 1449 Sequence level greedy decoding, performed auto-regressively. 1450 1451 Args: 1452 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1453 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1454 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1455 big_blank_durations: a list containing durations for big blanks the model supports. 1456 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1457 to a sequence in a single time step; if set to None then there is 1458 no limit. 1459 preserve_alignments: Bool flag which preserves the history of alignments generated during 1460 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1461 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1462 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1463 The length of the list corresponds to the Acoustic Length (T). 1464 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1465 U is the number of target tokens for the current timestep Ti. 1466 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1467 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1468 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1469 The length of the list corresponds to the Acoustic Length (T). 1470 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1471 U is the number of target tokens for the current timestep Ti. 1472 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1473 confidence scores. 1474 1475 name: The measure name (str). 1476 Supported values: 1477 - 'max_prob' for using the maximum token probability as a confidence. 1478 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1479 1480 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 1481 Supported values: 1482 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1483 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1484 Note that for this entropy, the alpha should comply the following inequality: 1485 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1486 where V is the model vocabulary size. 1487 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1488 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1489 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1490 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1491 - 'renyi' for the Rรฉnyi entropy. 1492 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1493 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1494 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1495 1496 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1497 When the alpha equals one, scaling is not applied to 'max_prob', 1498 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1499 1500 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1501 Supported values: 1502 - 'lin' for using the linear mapping. 1503 - 'exp' for using exponential mapping with linear shift. 1504 """ 1505 1506 def __init__( 1507 self, 1508 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1509 joint_model: rnnt_abstract.AbstractRNNTJoint, 1510 blank_index: int, 1511 big_blank_durations: list, 1512 max_symbols_per_step: Optional[int] = None, 1513 preserve_alignments: bool = False, 1514 preserve_frame_confidence: bool = False, 1515 confidence_measure_cfg: Optional[DictConfig] = None, 1516 ): 1517 super().__init__( 1518 decoder_model=decoder_model, 1519 joint_model=joint_model, 1520 blank_index=blank_index, 1521 max_symbols_per_step=max_symbols_per_step, 1522 preserve_alignments=preserve_alignments, 1523 preserve_frame_confidence=preserve_frame_confidence, 1524 confidence_measure_cfg=confidence_measure_cfg, 1525 ) 1526 self.big_blank_durations = big_blank_durations 1527 self._SOS = blank_index - len(big_blank_durations) 1528 1529 @torch.no_grad() 1530 def _greedy_decode( 1531 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 1532 ): 1533 # x: [T, 1, D] 1534 # out_len: [seq_len] 1535 1536 # Initialize blank state and empty label set in Hypothesis 1537 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 1538 1539 if partial_hypotheses is not None: 1540 hypothesis.last_token = partial_hypotheses.last_token 1541 hypothesis.y_sequence = ( 1542 partial_hypotheses.y_sequence.cpu().tolist() 1543 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 1544 else partial_hypotheses.y_sequence 1545 ) 1546 if partial_hypotheses.dec_state is not None: 1547 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 1548 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 1549 1550 if self.preserve_alignments: 1551 # Alignments is a 2-dimensional dangling list representing T x U 1552 hypothesis.alignments = [[]] 1553 1554 if self.preserve_frame_confidence: 1555 hypothesis.frame_confidence = [[]] 1556 1557 # if this variable is > 1, it means the last emission was a big-blank and we need to skip frames. 1558 big_blank_duration = 1 1559 1560 # For timestep t in X_t 1561 for time_idx in range(out_len): 1562 if big_blank_duration > 1: 1563 # skip frames until big_blank_duration == 1. 1564 big_blank_duration -= 1 1565 continue 1566 # Extract encoder embedding at timestep t 1567 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 1568 f = x.narrow(dim=0, start=time_idx, length=1) 1569 1570 # Setup exit flags and counter 1571 not_blank = True 1572 symbols_added = 0 1573 1574 # While blank is not predicted, or we dont run out of max symbols per timestep 1575 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1576 # In the first timestep, we initialize the network with RNNT Blank 1577 # In later timesteps, we provide previous predicted label as input. 1578 if hypothesis.last_token is None and hypothesis.dec_state is None: 1579 last_label = self._SOS 1580 else: 1581 last_label = label_collate([[hypothesis.last_token]]) 1582 1583 # Perform prediction network and joint network steps. 1584 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 1585 # If preserving per-frame confidence, log_normalize must be true 1586 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1587 0, 0, 0, : 1588 ] 1589 1590 del g 1591 1592 # torch.max(0) op doesnt exist for FP 16. 1593 if logp.dtype != torch.float32: 1594 logp = logp.float() 1595 1596 # get index k, of max prob 1597 v, k = logp.max(0) 1598 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 1599 1600 # Note, we have non-blanks in the vocab first, followed by big blanks, and standard blank at last. 1601 # here we check if it's a big blank and if yes, set the duration variable. 1602 if k >= self._blank_index - len(self.big_blank_durations) and k < self._blank_index: 1603 big_blank_duration = self.big_blank_durations[self._blank_index - k - 1] 1604 1605 if self.preserve_alignments: 1606 # insert logprobs into last timestep 1607 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 1608 1609 if self.preserve_frame_confidence: 1610 # insert confidence into last timestep 1611 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 1612 1613 del logp 1614 1615 # If any type of blank token is predicted, exit inner loop, move onto next timestep t 1616 if k >= self._blank_index - len(self.big_blank_durations): 1617 not_blank = False 1618 else: 1619 # Append token to label set, update RNN state. 1620 hypothesis.y_sequence.append(k) 1621 hypothesis.score += float(v) 1622 hypothesis.timestep.append(time_idx) 1623 hypothesis.dec_state = hidden_prime 1624 hypothesis.last_token = k 1625 1626 # Increment token counter. 1627 symbols_added += 1 1628 1629 if self.preserve_alignments: 1630 # convert Ti-th logits into a torch array 1631 hypothesis.alignments.append([]) # blank buffer for next timestep 1632 1633 if self.preserve_frame_confidence: 1634 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 1635 1636 # Remove trailing empty list of Alignments 1637 if self.preserve_alignments: 1638 if len(hypothesis.alignments[-1]) == 0: 1639 del hypothesis.alignments[-1] 1640 1641 # Remove trailing empty list of per-frame confidence 1642 if self.preserve_frame_confidence: 1643 if len(hypothesis.frame_confidence[-1]) == 0: 1644 del hypothesis.frame_confidence[-1] 1645 1646 # Unpack the hidden states 1647 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 1648 1649 return hypothesis 1650 1651 1652 class GreedyBatchedMultiblankRNNTInfer(GreedyBatchedRNNTInfer): 1653 """A batch level greedy transducer decoder. 1654 Batch level greedy decoding, performed auto-regressively. 1655 Args: 1656 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1657 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1658 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1659 big_blank_durations: a list containing durations for big blanks the model supports. 1660 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1661 to a sequence in a single time step; if set to None then there is 1662 no limit. 1663 preserve_alignments: Bool flag which preserves the history of alignments generated during 1664 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1665 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1666 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1667 The length of the list corresponds to the Acoustic Length (T). 1668 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1669 U is the number of target tokens for the current timestep Ti. 1670 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1671 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1672 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1673 The length of the list corresponds to the Acoustic Length (T). 1674 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1675 U is the number of target tokens for the current timestep Ti. 1676 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 1677 confidence scores. 1678 1679 name: The measure name (str). 1680 Supported values: 1681 - 'max_prob' for using the maximum token probability as a confidence. 1682 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1683 1684 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 1685 Supported values: 1686 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1687 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1688 Note that for this entropy, the alpha should comply the following inequality: 1689 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1690 where V is the model vocabulary size. 1691 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1692 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1693 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1694 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1695 - 'renyi' for the Rรฉnyi entropy. 1696 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1697 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1698 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1699 1700 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1701 When the alpha equals one, scaling is not applied to 'max_prob', 1702 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1703 1704 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1705 Supported values: 1706 - 'lin' for using the linear mapping. 1707 - 'exp' for using exponential mapping with linear shift. 1708 """ 1709 1710 def __init__( 1711 self, 1712 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1713 joint_model: rnnt_abstract.AbstractRNNTJoint, 1714 blank_index: int, 1715 big_blank_durations: List[int], 1716 max_symbols_per_step: Optional[int] = None, 1717 preserve_alignments: bool = False, 1718 preserve_frame_confidence: bool = False, 1719 confidence_measure_cfg: Optional[DictConfig] = None, 1720 ): 1721 super().__init__( 1722 decoder_model=decoder_model, 1723 joint_model=joint_model, 1724 blank_index=blank_index, 1725 max_symbols_per_step=max_symbols_per_step, 1726 preserve_alignments=preserve_alignments, 1727 preserve_frame_confidence=preserve_frame_confidence, 1728 confidence_measure_cfg=confidence_measure_cfg, 1729 ) 1730 self.big_blank_durations = big_blank_durations 1731 1732 # Depending on availability of `blank_as_pad` support 1733 # switch between more efficient batch decoding technique 1734 if self.decoder.blank_as_pad: 1735 self._greedy_decode = self._greedy_decode_blank_as_pad 1736 else: 1737 self._greedy_decode = self._greedy_decode_masked 1738 self._SOS = blank_index - len(big_blank_durations) 1739 1740 def _greedy_decode_blank_as_pad( 1741 self, 1742 x: torch.Tensor, 1743 out_len: torch.Tensor, 1744 device: torch.device, 1745 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1746 ): 1747 if partial_hypotheses is not None: 1748 raise NotImplementedError("`partial_hypotheses` support is not supported") 1749 1750 with torch.inference_mode(): 1751 # x: [B, T, D] 1752 # out_len: [B] 1753 # device: torch.device 1754 1755 # Initialize list of Hypothesis 1756 batchsize = x.shape[0] 1757 hypotheses = [ 1758 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1759 ] 1760 1761 # Initialize Hidden state matrix (shared by entire batch) 1762 hidden = None 1763 1764 # If alignments need to be preserved, register a danling list to hold the values 1765 if self.preserve_alignments: 1766 # alignments is a 3-dimensional dangling list representing B x T x U 1767 for hyp in hypotheses: 1768 hyp.alignments = [[]] 1769 1770 # If confidence scores need to be preserved, register a danling list to hold the values 1771 if self.preserve_frame_confidence: 1772 # frame_confidence is a 3-dimensional dangling list representing B x T x U 1773 for hyp in hypotheses: 1774 hyp.frame_confidence = [[]] 1775 1776 # Last Label buffer + Last Label without blank buffer 1777 # batch level equivalent of the last_label 1778 last_label = torch.full([batchsize, 1], fill_value=self._SOS, dtype=torch.long, device=device) 1779 1780 # this mask is true for if the emission is *any type* of blank. 1781 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 1782 1783 # Get max sequence length 1784 max_out_len = out_len.max() 1785 1786 # We have a mask for each big blank. A mask is "true" means: the previous emission is exactly the big-blank 1787 # with the corresponding duration, or has larger duration. E.g., for big_blank_mask for duration 2, it will 1788 # be set true if the previous emission was a big blank with duration 4, or 3 or 2; but false if prevoius 1789 # emission was a standard blank (with duration = 1). 1790 big_blank_masks = [torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device)] * len( 1791 self.big_blank_durations 1792 ) 1793 1794 # if this variable > 1, it means the previous emission was big-blank and we need to skip frames. 1795 big_blank_duration = 1 1796 1797 for time_idx in range(max_out_len): 1798 if big_blank_duration > 1: 1799 # skip frames until big_blank_duration == 1 1800 big_blank_duration -= 1 1801 continue 1802 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 1803 1804 # Prepare t timestamp batch variables 1805 not_blank = True 1806 symbols_added = 0 1807 1808 # Reset all blank masks 1809 blank_mask.mul_(False) 1810 for i in range(len(big_blank_masks)): 1811 big_blank_masks[i].mul_(False) 1812 1813 # Update blank mask with time mask 1814 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1815 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1816 blank_mask = time_idx >= out_len 1817 for i in range(len(big_blank_masks)): 1818 big_blank_masks[i] = time_idx >= out_len 1819 1820 # Start inner loop 1821 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1822 # Batch prediction and joint network steps 1823 # If very first prediction step, submit SOS tag (blank) to pred_step. 1824 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1825 if time_idx == 0 and symbols_added == 0 and hidden is None: 1826 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 1827 else: 1828 # Perform batch step prediction of decoder, getting new states and scores ("g") 1829 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 1830 1831 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 1832 # If preserving per-frame confidence, log_normalize must be true 1833 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1834 :, 0, 0, : 1835 ] 1836 1837 if logp.dtype != torch.float32: 1838 logp = logp.float() 1839 1840 # Get index k, of max prob for batch 1841 v, k = logp.max(1) 1842 del g 1843 1844 # Update blank mask with current predicted blanks 1845 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1846 k_is_blank = k >= self._blank_index - len(self.big_blank_durations) 1847 blank_mask.bitwise_or_(k_is_blank) 1848 1849 for i in range(len(big_blank_masks)): 1850 # using <= since as we mentioned before, the mask doesn't store exact matches. 1851 # instead, it is True when the predicted blank's duration is >= the duration that the 1852 # mask corresponds to. 1853 k_is_big_blank = k <= self._blank_index - 1 - i 1854 1855 # need to do a bitwise_and since it could also be a non-blank. 1856 k_is_big_blank.bitwise_and_(k_is_blank) 1857 big_blank_masks[i].bitwise_or_(k_is_big_blank) 1858 1859 del k_is_blank 1860 1861 # If preserving alignments, check if sequence length of sample has been reached 1862 # before adding alignment 1863 if self.preserve_alignments: 1864 # Insert logprobs into last timestep per sample 1865 logp_vals = logp.to('cpu') 1866 logp_ids = logp_vals.max(1)[1] 1867 for batch_idx in range(batchsize): 1868 if time_idx < out_len[batch_idx]: 1869 hypotheses[batch_idx].alignments[-1].append( 1870 (logp_vals[batch_idx], logp_ids[batch_idx]) 1871 ) 1872 del logp_vals 1873 1874 # If preserving per-frame confidence, check if sequence length of sample has been reached 1875 # before adding confidence scores 1876 if self.preserve_frame_confidence: 1877 # Insert probabilities into last timestep per sample 1878 confidence = self._get_confidence(logp) 1879 for batch_idx in range(batchsize): 1880 if time_idx < out_len[batch_idx]: 1881 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 1882 del logp 1883 1884 # If all samples predict / have predicted prior blanks, exit loop early 1885 # This is equivalent to if single sample predicted k 1886 if blank_mask.all(): 1887 not_blank = False 1888 else: 1889 # Collect batch indices where blanks occurred now/past 1890 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1891 1892 # Recover prior state for all samples which predicted blank now/past 1893 if hidden is not None: 1894 # LSTM has 2 states 1895 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 1896 1897 elif len(blank_indices) > 0 and hidden is None: 1898 # Reset state if there were some blank and other non-blank predictions in batch 1899 # Original state is filled with zeros so we just multiply 1900 # LSTM has 2 states 1901 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 1902 1903 # Recover prior predicted label for all samples which predicted blank now/past 1904 k[blank_indices] = last_label[blank_indices, 0] 1905 1906 # Update new label and hidden state for next iteration 1907 last_label = k.clone().view(-1, 1) 1908 hidden = hidden_prime 1909 1910 # Update predicted labels, accounting for time mask 1911 # If blank was predicted even once, now or in the past, 1912 # Force the current predicted label to also be blank 1913 # This ensures that blanks propogate across all timesteps 1914 # once they have occured (normally stopping condition of sample level loop). 1915 for kidx, ki in enumerate(k): 1916 if blank_mask[kidx] == 0: 1917 hypotheses[kidx].y_sequence.append(ki) 1918 hypotheses[kidx].timestep.append(time_idx) 1919 hypotheses[kidx].score += float(v[kidx]) 1920 1921 symbols_added += 1 1922 1923 for i in range(len(big_blank_masks) + 1): 1924 # The task here is find the shortest blank duration of all batches. 1925 # so we start from the shortest blank duration and go up, 1926 # and stop once we found the duration whose corresponding mask isn't all True. 1927 if i == len(big_blank_masks) or not big_blank_masks[i].all(): 1928 big_blank_duration = self.big_blank_durations[i - 1] if i > 0 else 1 1929 break 1930 1931 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 1932 # Then preserve U at current timestep Ti 1933 # Finally, forward the timestep history to Ti+1 for that sample 1934 # All of this should only be done iff the current time index <= sample-level AM length. 1935 # Otherwise ignore and move to next sample / next timestep. 1936 if self.preserve_alignments: 1937 1938 # convert Ti-th logits into a torch array 1939 for batch_idx in range(batchsize): 1940 1941 # this checks if current timestep <= sample-level AM length 1942 # If current timestep > sample-level AM length, no alignments will be added 1943 # Therefore the list of Uj alignments is empty here. 1944 if len(hypotheses[batch_idx].alignments[-1]) > 0: 1945 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 1946 1947 # Do the same if preserving per-frame confidence 1948 if self.preserve_frame_confidence: 1949 1950 for batch_idx in range(batchsize): 1951 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 1952 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 1953 1954 # Remove trailing empty list of alignments at T_{am-len} x Uj 1955 if self.preserve_alignments: 1956 for batch_idx in range(batchsize): 1957 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1958 del hypotheses[batch_idx].alignments[-1] 1959 1960 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1961 if self.preserve_frame_confidence: 1962 for batch_idx in range(batchsize): 1963 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1964 del hypotheses[batch_idx].frame_confidence[-1] 1965 1966 # Preserve states 1967 for batch_idx in range(batchsize): 1968 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1969 1970 return hypotheses 1971 1972 def _greedy_decode_masked( 1973 self, 1974 x: torch.Tensor, 1975 out_len: torch.Tensor, 1976 device: torch.device, 1977 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1978 ): 1979 if partial_hypotheses is not None: 1980 raise NotImplementedError("`partial_hypotheses` support is not supported") 1981 1982 if self.big_blank_durations != [1] * len(self.big_blank_durations): 1983 raise NotImplementedError( 1984 "Efficient frame-skipping version for multi-blank masked decoding is not supported." 1985 ) 1986 1987 # x: [B, T, D] 1988 # out_len: [B] 1989 # device: torch.device 1990 1991 # Initialize state 1992 batchsize = x.shape[0] 1993 hypotheses = [ 1994 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1995 ] 1996 1997 # Initialize Hidden state matrix (shared by entire batch) 1998 hidden = None 1999 2000 # If alignments need to be preserved, register a danling list to hold the values 2001 if self.preserve_alignments: 2002 # alignments is a 3-dimensional dangling list representing B x T x U 2003 for hyp in hypotheses: 2004 hyp.alignments = [[]] 2005 else: 2006 hyp.alignments = None 2007 2008 # If confidence scores need to be preserved, register a danling list to hold the values 2009 if self.preserve_frame_confidence: 2010 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2011 for hyp in hypotheses: 2012 hyp.frame_confidence = [[]] 2013 2014 # Last Label buffer + Last Label without blank buffer 2015 # batch level equivalent of the last_label 2016 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2017 last_label_without_blank = last_label.clone() 2018 2019 # Mask buffers 2020 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2021 2022 # Get max sequence length 2023 max_out_len = out_len.max() 2024 2025 with torch.inference_mode(): 2026 for time_idx in range(max_out_len): 2027 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2028 2029 # Prepare t timestamp batch variables 2030 not_blank = True 2031 symbols_added = 0 2032 2033 # Reset blank mask 2034 blank_mask.mul_(False) 2035 2036 # Update blank mask with time mask 2037 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2038 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2039 blank_mask = time_idx >= out_len 2040 2041 # Start inner loop 2042 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 2043 # Batch prediction and joint network steps 2044 # If very first prediction step, submit SOS tag (blank) to pred_step. 2045 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2046 if time_idx == 0 and symbols_added == 0 and hidden is None: 2047 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2048 else: 2049 # Set a dummy label for the blank value 2050 # This value will be overwritten by "blank" again the last label update below 2051 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 2052 last_label_without_blank_mask = last_label >= self._blank_index 2053 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 2054 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 2055 ~last_label_without_blank_mask 2056 ] 2057 2058 # Perform batch step prediction of decoder, getting new states and scores ("g") 2059 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 2060 2061 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2062 # If preserving per-frame confidence, log_normalize must be true 2063 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 2064 :, 0, 0, : 2065 ] 2066 2067 if logp.dtype != torch.float32: 2068 logp = logp.float() 2069 2070 # Get index k, of max prob for batch 2071 v, k = logp.max(1) 2072 del g 2073 2074 # Update blank mask with current predicted blanks 2075 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2076 k_is_blank = k == self._blank_index 2077 blank_mask.bitwise_or_(k_is_blank) 2078 2079 # If preserving alignments, check if sequence length of sample has been reached 2080 # before adding alignment 2081 if self.preserve_alignments: 2082 # Insert logprobs into last timestep per sample 2083 logp_vals = logp.to('cpu') 2084 logp_ids = logp_vals.max(1)[1] 2085 for batch_idx in range(batchsize): 2086 if time_idx < out_len[batch_idx]: 2087 hypotheses[batch_idx].alignments[-1].append( 2088 (logp_vals[batch_idx], logp_ids[batch_idx]) 2089 ) 2090 del logp_vals 2091 2092 # If preserving per-frame confidence, check if sequence length of sample has been reached 2093 # before adding confidence scores 2094 if self.preserve_frame_confidence: 2095 # Insert probabilities into last timestep per sample 2096 confidence = self._get_confidence(logp) 2097 for batch_idx in range(batchsize): 2098 if time_idx < out_len[batch_idx]: 2099 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 2100 del logp 2101 2102 # If all samples predict / have predicted prior blanks, exit loop early 2103 # This is equivalent to if single sample predicted k 2104 if blank_mask.all(): 2105 not_blank = False 2106 else: 2107 # Collect batch indices where blanks occurred now/past 2108 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2109 2110 # Recover prior state for all samples which predicted blank now/past 2111 if hidden is not None: 2112 # LSTM has 2 states 2113 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2114 2115 elif len(blank_indices) > 0 and hidden is None: 2116 # Reset state if there were some blank and other non-blank predictions in batch 2117 # Original state is filled with zeros so we just multiply 2118 # LSTM has 2 states 2119 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2120 2121 # Recover prior predicted label for all samples which predicted blank now/past 2122 k[blank_indices] = last_label[blank_indices, 0] 2123 2124 # Update new label and hidden state for next iteration 2125 last_label = k.view(-1, 1) 2126 hidden = hidden_prime 2127 2128 # Update predicted labels, accounting for time mask 2129 # If blank was predicted even once, now or in the past, 2130 # Force the current predicted label to also be blank 2131 # This ensures that blanks propogate across all timesteps 2132 # once they have occured (normally stopping condition of sample level loop). 2133 for kidx, ki in enumerate(k): 2134 if blank_mask[kidx] == 0: 2135 hypotheses[kidx].y_sequence.append(ki) 2136 hypotheses[kidx].timestep.append(time_idx) 2137 hypotheses[kidx].score += float(v[kidx]) 2138 2139 symbols_added += 1 2140 2141 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 2142 # Then preserve U at current timestep Ti 2143 # Finally, forward the timestep history to Ti+1 for that sample 2144 # All of this should only be done iff the current time index <= sample-level AM length. 2145 # Otherwise ignore and move to next sample / next timestep. 2146 if self.preserve_alignments: 2147 2148 # convert Ti-th logits into a torch array 2149 for batch_idx in range(batchsize): 2150 2151 # this checks if current timestep <= sample-level AM length 2152 # If current timestep > sample-level AM length, no alignments will be added 2153 # Therefore the list of Uj alignments is empty here. 2154 if len(hypotheses[batch_idx].alignments[-1]) > 0: 2155 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 2156 2157 # Do the same if preserving per-frame confidence 2158 if self.preserve_frame_confidence: 2159 2160 for batch_idx in range(batchsize): 2161 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 2162 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 2163 2164 # Remove trailing empty list of alignments at T_{am-len} x Uj 2165 if self.preserve_alignments: 2166 for batch_idx in range(batchsize): 2167 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2168 del hypotheses[batch_idx].alignments[-1] 2169 2170 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2171 if self.preserve_frame_confidence: 2172 for batch_idx in range(batchsize): 2173 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2174 del hypotheses[batch_idx].frame_confidence[-1] 2175 2176 # Preserve states 2177 for batch_idx in range(batchsize): 2178 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2179 2180 return hypotheses 2181 2182 2183 @dataclass 2184 class GreedyRNNTInferConfig: 2185 max_symbols_per_step: Optional[int] = 10 2186 preserve_alignments: bool = False 2187 preserve_frame_confidence: bool = False 2188 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 2189 confidence_method_cfg: str = "DEPRECATED" 2190 2191 def __post_init__(self): 2192 # OmegaConf.structured ensures that post_init check is always executed 2193 self.confidence_measure_cfg = OmegaConf.structured( 2194 self.confidence_measure_cfg 2195 if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) 2196 else ConfidenceMeasureConfig(**self.confidence_measure_cfg) 2197 ) 2198 if self.confidence_method_cfg != "DEPRECATED": 2199 logging.warning( 2200 "`confidence_method_cfg` is deprecated and will be removed in the future. " 2201 "Please use `confidence_measure_cfg` instead." 2202 ) 2203 2204 # TODO (alaptev): delete the following two lines sometime in the future 2205 logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") 2206 # OmegaConf.structured ensures that post_init check is always executed 2207 self.confidence_measure_cfg = OmegaConf.structured( 2208 self.confidence_method_cfg 2209 if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) 2210 else ConfidenceMeasureConfig(**self.confidence_method_cfg) 2211 ) 2212 self.confidence_method_cfg = "DEPRECATED" 2213 2214 2215 @dataclass 2216 class GreedyBatchedRNNTInferConfig: 2217 max_symbols_per_step: Optional[int] = 10 2218 preserve_alignments: bool = False 2219 preserve_frame_confidence: bool = False 2220 confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() 2221 confidence_method_cfg: str = "DEPRECATED" 2222 2223 def __post_init__(self): 2224 # OmegaConf.structured ensures that post_init check is always executed 2225 self.confidence_measure_cfg = OmegaConf.structured( 2226 self.confidence_measure_cfg 2227 if isinstance(self.confidence_measure_cfg, ConfidenceMeasureConfig) 2228 else ConfidenceMeasureConfig(**self.confidence_measure_cfg) 2229 ) 2230 if self.confidence_method_cfg != "DEPRECATED": 2231 logging.warning( 2232 "`confidence_method_cfg` is deprecated and will be removed in the future. " 2233 "Please use `confidence_measure_cfg` instead." 2234 ) 2235 2236 # TODO (alaptev): delete the following two lines sometime in the future 2237 logging.warning("Re-writing `confidence_measure_cfg` with the value of `confidence_method_cfg`.") 2238 # OmegaConf.structured ensures that post_init check is always executed 2239 self.confidence_measure_cfg = OmegaConf.structured( 2240 self.confidence_method_cfg 2241 if isinstance(self.confidence_method_cfg, ConfidenceMeasureConfig) 2242 else ConfidenceMeasureConfig(**self.confidence_method_cfg) 2243 ) 2244 self.confidence_method_cfg = "DEPRECATED" 2245 2246 2247 class GreedyTDTInfer(_GreedyRNNTInfer): 2248 """A greedy TDT decoder. 2249 2250 Sequence level greedy decoding, performed auto-regressively. 2251 2252 Args: 2253 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2254 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2255 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2256 durations: a list containing durations for TDT. 2257 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2258 to a sequence in a single time step; if set to None then there is 2259 no limit. 2260 preserve_alignments: Bool flag which preserves the history of alignments generated during 2261 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2262 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2263 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2264 The length of the list corresponds to the Acoustic Length (T). 2265 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2266 U is the number of target tokens for the current timestep Ti. 2267 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2268 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2269 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2270 The length of the list corresponds to the Acoustic Length (T). 2271 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2272 U is the number of target tokens for the current timestep Ti. 2273 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 2274 confidence scores. 2275 2276 name: The measure name (str). 2277 Supported values: 2278 - 'max_prob' for using the maximum token probability as a confidence. 2279 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2280 2281 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 2282 Supported values: 2283 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2284 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2285 Note that for this entropy, the alpha should comply the following inequality: 2286 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2287 where V is the model vocabulary size. 2288 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2289 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2290 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2291 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2292 - 'renyi' for the Rรฉnyi entropy. 2293 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2294 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2295 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2296 2297 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2298 When the alpha equals one, scaling is not applied to 'max_prob', 2299 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2300 2301 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2302 Supported values: 2303 - 'lin' for using the linear mapping. 2304 - 'exp' for using exponential mapping with linear shift. 2305 """ 2306 2307 def __init__( 2308 self, 2309 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2310 joint_model: rnnt_abstract.AbstractRNNTJoint, 2311 blank_index: int, 2312 durations: list, 2313 max_symbols_per_step: Optional[int] = None, 2314 preserve_alignments: bool = False, 2315 preserve_frame_confidence: bool = False, 2316 confidence_measure_cfg: Optional[DictConfig] = None, 2317 ): 2318 super().__init__( 2319 decoder_model=decoder_model, 2320 joint_model=joint_model, 2321 blank_index=blank_index, 2322 max_symbols_per_step=max_symbols_per_step, 2323 preserve_alignments=preserve_alignments, 2324 preserve_frame_confidence=preserve_frame_confidence, 2325 confidence_measure_cfg=confidence_measure_cfg, 2326 ) 2327 self.durations = durations 2328 2329 @typecheck() 2330 def forward( 2331 self, 2332 encoder_output: torch.Tensor, 2333 encoded_lengths: torch.Tensor, 2334 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2335 ): 2336 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2337 Output token is generated auto-regressively. 2338 Args: 2339 encoder_output: A tensor of size (batch, features, timesteps). 2340 encoded_lengths: list of int representing the length of each sequence 2341 output sequence. 2342 Returns: 2343 packed list containing batch number of sentences (Hypotheses). 2344 """ 2345 # Preserve decoder and joint training state 2346 decoder_training_state = self.decoder.training 2347 joint_training_state = self.joint.training 2348 2349 with torch.inference_mode(): 2350 # Apply optional preprocessing 2351 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2352 2353 self.decoder.eval() 2354 self.joint.eval() 2355 2356 hypotheses = [] 2357 # Process each sequence independently 2358 with self.decoder.as_frozen(), self.joint.as_frozen(): 2359 for batch_idx in range(encoder_output.size(0)): 2360 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 2361 logitlen = encoded_lengths[batch_idx] 2362 2363 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 2364 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 2365 hypotheses.append(hypothesis) 2366 2367 # Pack results into Hypotheses 2368 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 2369 2370 self.decoder.train(decoder_training_state) 2371 self.joint.train(joint_training_state) 2372 2373 return (packed_result,) 2374 2375 @torch.no_grad() 2376 def _greedy_decode( 2377 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 2378 ): 2379 # x: [T, 1, D] 2380 # out_len: [seq_len] 2381 2382 # Initialize blank state and empty label set in Hypothesis 2383 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 2384 2385 if partial_hypotheses is not None: 2386 hypothesis.last_token = partial_hypotheses.last_token 2387 hypothesis.y_sequence = ( 2388 partial_hypotheses.y_sequence.cpu().tolist() 2389 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 2390 else partial_hypotheses.y_sequence 2391 ) 2392 if partial_hypotheses.dec_state is not None: 2393 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 2394 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 2395 2396 if self.preserve_alignments: 2397 # Alignments is a 2-dimensional dangling list representing T x U 2398 hypothesis.alignments = [[]] 2399 2400 if self.preserve_frame_confidence: 2401 hypothesis.frame_confidence = [[]] 2402 2403 time_idx = 0 2404 while time_idx < out_len: 2405 # Extract encoder embedding at timestep t 2406 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 2407 f = x.narrow(dim=0, start=time_idx, length=1) 2408 2409 # Setup exit flags and counter 2410 not_blank = True 2411 symbols_added = 0 2412 2413 need_loop = True 2414 # While blank is not predicted, or we dont run out of max symbols per timestep 2415 while need_loop and (self.max_symbols is None or symbols_added < self.max_symbols): 2416 # In the first timestep, we initialize the network with RNNT Blank 2417 # In later timesteps, we provide previous predicted label as input. 2418 if hypothesis.last_token is None and hypothesis.dec_state is None: 2419 last_label = self._SOS 2420 else: 2421 last_label = label_collate([[hypothesis.last_token]]) 2422 2423 # Perform prediction network and joint network steps. 2424 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 2425 # If preserving per-frame confidence, log_normalize must be true 2426 logits = self._joint_step(f, g, log_normalize=False) 2427 logp = logits[0, 0, 0, : -len(self.durations)] 2428 if self.preserve_frame_confidence: 2429 logp = torch.log_softmax(logp, -1) 2430 2431 duration_logp = torch.log_softmax(logits[0, 0, 0, -len(self.durations) :], dim=-1) 2432 del g 2433 2434 # torch.max(0) op doesnt exist for FP 16. 2435 if logp.dtype != torch.float32: 2436 logp = logp.float() 2437 2438 # get index k, of max prob 2439 v, k = logp.max(0) 2440 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 2441 2442 d_v, d_k = duration_logp.max(0) 2443 d_k = d_k.item() 2444 2445 skip = self.durations[d_k] 2446 2447 if self.preserve_alignments: 2448 # insert logprobs into last timestep 2449 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 2450 2451 if self.preserve_frame_confidence: 2452 # insert confidence into last timestep 2453 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 2454 2455 del logp 2456 2457 # If blank token is predicted, exit inner loop, move onto next timestep t 2458 if k == self._blank_index: 2459 not_blank = False 2460 else: 2461 # Append token to label set, update RNN state. 2462 hypothesis.y_sequence.append(k) 2463 hypothesis.score += float(v) 2464 hypothesis.timestep.append(time_idx) 2465 hypothesis.dec_state = hidden_prime 2466 hypothesis.last_token = k 2467 2468 # Increment token counter. 2469 symbols_added += 1 2470 time_idx += skip 2471 need_loop = skip == 0 2472 2473 # this rarely happens, but we manually increment the `skip` number 2474 # if blank is emitted and duration=0 is predicted. This prevents possible 2475 # infinite loops. 2476 if skip == 0: 2477 skip = 1 2478 2479 if self.preserve_alignments: 2480 # convert Ti-th logits into a torch array 2481 hypothesis.alignments.append([]) # blank buffer for next timestep 2482 2483 if self.preserve_frame_confidence: 2484 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 2485 2486 if symbols_added == self.max_symbols: 2487 time_idx += 1 2488 2489 # Remove trailing empty list of Alignments 2490 if self.preserve_alignments: 2491 if len(hypothesis.alignments[-1]) == 0: 2492 del hypothesis.alignments[-1] 2493 2494 # Remove trailing empty list of per-frame confidence 2495 if self.preserve_frame_confidence: 2496 if len(hypothesis.frame_confidence[-1]) == 0: 2497 del hypothesis.frame_confidence[-1] 2498 2499 # Unpack the hidden states 2500 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 2501 2502 return hypothesis 2503 2504 2505 class GreedyBatchedTDTInfer(_GreedyRNNTInfer): 2506 """A batch level greedy TDT decoder. 2507 Batch level greedy decoding, performed auto-regressively. 2508 Args: 2509 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2510 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2511 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2512 durations: a list containing durations. 2513 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2514 to a sequence in a single time step; if set to None then there is 2515 no limit. 2516 preserve_alignments: Bool flag which preserves the history of alignments generated during 2517 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2518 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2519 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2520 The length of the list corresponds to the Acoustic Length (T). 2521 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2522 U is the number of target tokens for the current timestep Ti. 2523 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2524 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2525 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2526 The length of the list corresponds to the Acoustic Length (T). 2527 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2528 U is the number of target tokens for the current timestep Ti. 2529 confidence_measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 2530 confidence scores. 2531 2532 name: The measure name (str). 2533 Supported values: 2534 - 'max_prob' for using the maximum token probability as a confidence. 2535 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2536 2537 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 2538 Supported values: 2539 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2540 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2541 Note that for this entropy, the alpha should comply the following inequality: 2542 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2543 where V is the model vocabulary size. 2544 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2545 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2546 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2547 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2548 - 'renyi' for the Rรฉnyi entropy. 2549 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2550 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2551 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2552 2553 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2554 When the alpha equals one, scaling is not applied to 'max_prob', 2555 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2556 2557 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2558 Supported values: 2559 - 'lin' for using the linear mapping. 2560 - 'exp' for using exponential mapping with linear shift. 2561 """ 2562 2563 def __init__( 2564 self, 2565 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2566 joint_model: rnnt_abstract.AbstractRNNTJoint, 2567 blank_index: int, 2568 durations: List[int], 2569 max_symbols_per_step: Optional[int] = None, 2570 preserve_alignments: bool = False, 2571 preserve_frame_confidence: bool = False, 2572 confidence_measure_cfg: Optional[DictConfig] = None, 2573 ): 2574 super().__init__( 2575 decoder_model=decoder_model, 2576 joint_model=joint_model, 2577 blank_index=blank_index, 2578 max_symbols_per_step=max_symbols_per_step, 2579 preserve_alignments=preserve_alignments, 2580 preserve_frame_confidence=preserve_frame_confidence, 2581 confidence_measure_cfg=confidence_measure_cfg, 2582 ) 2583 self.durations = durations 2584 2585 # Depending on availability of `blank_as_pad` support 2586 # switch between more efficient batch decoding technique 2587 if self.decoder.blank_as_pad: 2588 self._greedy_decode = self._greedy_decode_blank_as_pad 2589 else: 2590 self._greedy_decode = self._greedy_decode_masked 2591 2592 @typecheck() 2593 def forward( 2594 self, 2595 encoder_output: torch.Tensor, 2596 encoded_lengths: torch.Tensor, 2597 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2598 ): 2599 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2600 Output token is generated auto-regressively. 2601 Args: 2602 encoder_output: A tensor of size (batch, features, timesteps). 2603 encoded_lengths: list of int representing the length of each sequence 2604 output sequence. 2605 Returns: 2606 packed list containing batch number of sentences (Hypotheses). 2607 """ 2608 # Preserve decoder and joint training state 2609 decoder_training_state = self.decoder.training 2610 joint_training_state = self.joint.training 2611 2612 with torch.inference_mode(): 2613 # Apply optional preprocessing 2614 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2615 logitlen = encoded_lengths 2616 2617 self.decoder.eval() 2618 self.joint.eval() 2619 2620 with self.decoder.as_frozen(), self.joint.as_frozen(): 2621 inseq = encoder_output # [B, T, D] 2622 hypotheses = self._greedy_decode( 2623 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 2624 ) 2625 2626 # Pack the hypotheses results 2627 packed_result = pack_hypotheses(hypotheses, logitlen) 2628 2629 self.decoder.train(decoder_training_state) 2630 self.joint.train(joint_training_state) 2631 2632 return (packed_result,) 2633 2634 def _greedy_decode_blank_as_pad( 2635 self, 2636 x: torch.Tensor, 2637 out_len: torch.Tensor, 2638 device: torch.device, 2639 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2640 ): 2641 if partial_hypotheses is not None: 2642 raise NotImplementedError("`partial_hypotheses` support is not supported") 2643 2644 with torch.inference_mode(): 2645 # x: [B, T, D] 2646 # out_len: [B] 2647 # device: torch.device 2648 2649 # Initialize list of Hypothesis 2650 batchsize = x.shape[0] 2651 hypotheses = [ 2652 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 2653 ] 2654 2655 # Initialize Hidden state matrix (shared by entire batch) 2656 hidden = None 2657 2658 # If alignments need to be preserved, register a danling list to hold the values 2659 if self.preserve_alignments: 2660 # alignments is a 3-dimensional dangling list representing B x T x U 2661 for hyp in hypotheses: 2662 hyp.alignments = [[]] 2663 2664 # If confidence scores need to be preserved, register a danling list to hold the values 2665 if self.preserve_frame_confidence: 2666 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2667 for hyp in hypotheses: 2668 hyp.frame_confidence = [[]] 2669 2670 # Last Label buffer + Last Label without blank buffer 2671 # batch level equivalent of the last_label 2672 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2673 2674 # Mask buffers 2675 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2676 2677 # Get max sequence length 2678 max_out_len = out_len.max() 2679 2680 # skip means the number of frames the next decoding step should "jump" to. When skip == 1 2681 # it means the next decoding step will just use the next input frame. 2682 skip = 1 2683 for time_idx in range(max_out_len): 2684 if skip > 1: # if skip > 1 at the current step, we decrement it and skip the current frame. 2685 skip -= 1 2686 continue 2687 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2688 2689 # need_to_stay is a boolean indicates whether the next decoding step should remain in the same frame. 2690 need_to_stay = True 2691 symbols_added = 0 2692 2693 # Reset blank mask 2694 blank_mask.mul_(False) 2695 2696 # Update blank mask with time mask 2697 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2698 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2699 blank_mask = time_idx >= out_len 2700 2701 # Start inner loop 2702 while need_to_stay and (self.max_symbols is None or symbols_added < self.max_symbols): 2703 # Batch prediction and joint network steps 2704 # If very first prediction step, submit SOS tag (blank) to pred_step. 2705 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2706 if time_idx == 0 and symbols_added == 0 and hidden is None: 2707 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2708 else: 2709 # Perform batch step prediction of decoder, getting new states and scores ("g") 2710 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 2711 2712 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2713 # Note: log_normalize must not be True here since the joiner output is contanetation of both token logits and duration logits, 2714 # and they need to be normalized independently. 2715 joined = self._joint_step(f, g, log_normalize=None) 2716 logp = joined[:, 0, 0, : -len(self.durations)] 2717 duration_logp = joined[:, 0, 0, -len(self.durations) :] 2718 2719 if logp.dtype != torch.float32: 2720 logp = logp.float() 2721 duration_logp = duration_logp.float() 2722 2723 # get the max for both token and duration predictions. 2724 v, k = logp.max(1) 2725 dv, dk = duration_logp.max(1) 2726 2727 # here we set the skip value to be the minimum of all predicted durations, hense the "torch.min(dk)" call there. 2728 # Please refer to Section 5.2 of our paper https://arxiv.org/pdf/2304.06795.pdf for explanation of this. 2729 skip = self.durations[int(torch.min(dk))] 2730 2731 # this is a special case: if all batches emit blanks, we require that skip be at least 1 2732 # so we don't loop forever at the current frame. 2733 if blank_mask.all(): 2734 if skip == 0: 2735 skip = 1 2736 2737 need_to_stay = skip == 0 2738 del g 2739 2740 # Update blank mask with current predicted blanks 2741 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2742 k_is_blank = k == self._blank_index 2743 blank_mask.bitwise_or_(k_is_blank) 2744 2745 del k_is_blank 2746 del logp, duration_logp 2747 2748 # If all samples predict / have predicted prior blanks, exit loop early 2749 # This is equivalent to if single sample predicted k 2750 if not blank_mask.all(): 2751 # Collect batch indices where blanks occurred now/past 2752 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2753 2754 # Recover prior state for all samples which predicted blank now/past 2755 if hidden is not None: 2756 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2757 2758 elif len(blank_indices) > 0 and hidden is None: 2759 # Reset state if there were some blank and other non-blank predictions in batch 2760 # Original state is filled with zeros so we just multiply 2761 # LSTM has 2 states 2762 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2763 2764 # Recover prior predicted label for all samples which predicted blank now/past 2765 k[blank_indices] = last_label[blank_indices, 0] 2766 2767 # Update new label and hidden state for next iteration 2768 last_label = k.clone().view(-1, 1) 2769 hidden = hidden_prime 2770 2771 # Update predicted labels, accounting for time mask 2772 # If blank was predicted even once, now or in the past, 2773 # Force the current predicted label to also be blank 2774 # This ensures that blanks propogate across all timesteps 2775 # once they have occured (normally stopping condition of sample level loop). 2776 for kidx, ki in enumerate(k): 2777 if blank_mask[kidx] == 0: 2778 hypotheses[kidx].y_sequence.append(ki) 2779 hypotheses[kidx].timestep.append(time_idx) 2780 hypotheses[kidx].score += float(v[kidx]) 2781 2782 symbols_added += 1 2783 2784 # Remove trailing empty list of alignments at T_{am-len} x Uj 2785 if self.preserve_alignments: 2786 for batch_idx in range(batchsize): 2787 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2788 del hypotheses[batch_idx].alignments[-1] 2789 2790 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2791 if self.preserve_frame_confidence: 2792 for batch_idx in range(batchsize): 2793 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2794 del hypotheses[batch_idx].frame_confidence[-1] 2795 2796 # Preserve states 2797 for batch_idx in range(batchsize): 2798 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2799 2800 return hypotheses 2801 2802 def _greedy_decode_masked( 2803 self, 2804 x: torch.Tensor, 2805 out_len: torch.Tensor, 2806 device: torch.device, 2807 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2808 ): 2809 raise NotImplementedError("masked greedy-batched decode is not supported for TDT models.") 2810 [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from abc import ABC, abstractmethod 17 from dataclasses import dataclass 18 from functools import partial 19 from typing import List, Optional 20 21 import torch 22 from omegaconf import DictConfig, OmegaConf 23 24 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 25 from nemo.utils import logging 26 27 28 class ConfidenceMeasureConstants: 29 NAMES = ("max_prob", "entropy") 30 ENTROPY_TYPES = ("gibbs", "tsallis", "renyi") 31 ENTROPY_NORMS = ("lin", "exp") 32 33 @classmethod 34 def print(cls): 35 return ( 36 cls.__name__ 37 + ": " 38 + str({"NAMES": cls.NAMES, "ENTROPY_TYPES": cls.ENTROPY_TYPES, "ENTROPY_NORMS": cls.ENTROPY_NORMS}) 39 ) 40 41 42 class ConfidenceConstants: 43 AGGREGATIONS = ("mean", "min", "max", "prod") 44 45 @classmethod 46 def print(cls): 47 return cls.__name__ + ": " + str({"AGGREGATIONS": cls.AGGREGATIONS}) 48 49 50 @dataclass 51 class ConfidenceMeasureConfig: 52 """A Config which contains the measure name and settings to compute per-frame confidence scores. 53 54 Args: 55 name: The measure name (str). 56 Supported values: 57 - 'max_prob' for using the maximum token probability as a confidence. 58 - 'entropy' for using a normalized entropy of a log-likelihood vector. 59 60 entropy_type: Which type of entropy to use (str). 61 Used if confidence_measure_cfg.name is set to `entropy`. 62 Supported values: 63 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 64 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 65 Note that for this entropy, the alpha should comply the following inequality: 66 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 67 where V is the model vocabulary size. 68 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 69 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 70 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 71 More: https://en.wikipedia.org/wiki/Tsallis_entropy 72 - 'renyi' for the Rรฉnyi entropy. 73 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 74 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 75 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 76 77 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 78 When the alpha equals one, scaling is not applied to 'max_prob', 79 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 80 81 entropy_norm: A mapping of the entropy value to the interval [0,1]. 82 Supported values: 83 - 'lin' for using the linear mapping. 84 - 'exp' for using exponential mapping with linear shift. 85 """ 86 87 name: str = "entropy" 88 entropy_type: str = "tsallis" 89 alpha: float = 0.33 90 entropy_norm: str = "exp" 91 temperature: str = "DEPRECATED" 92 93 def __post_init__(self): 94 if self.temperature != "DEPRECATED": 95 logging.warning( 96 "`temperature` is deprecated and will be removed in the future. Please use `alpha` instead." 97 ) 98 99 # TODO (alaptev): delete the following two lines sometime in the future 100 logging.warning("Re-writing `alpha` with the value of `temperature`.") 101 # self.temperature has type str 102 self.alpha = float(self.temperature) 103 self.temperature = "DEPRECATED" 104 if self.name not in ConfidenceMeasureConstants.NAMES: 105 raise ValueError( 106 f"`name` must be one of the following: " 107 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.NAMES) + '`'}. Provided: `{self.name}`" 108 ) 109 if self.entropy_type not in ConfidenceMeasureConstants.ENTROPY_TYPES: 110 raise ValueError( 111 f"`entropy_type` must be one of the following: " 112 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.ENTROPY_TYPES) + '`'}. Provided: `{self.entropy_type}`" 113 ) 114 if self.alpha <= 0.0: 115 raise ValueError(f"`alpha` must be > 0. Provided: {self.alpha}") 116 if self.entropy_norm not in ConfidenceMeasureConstants.ENTROPY_NORMS: 117 raise ValueError( 118 f"`entropy_norm` must be one of the following: " 119 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.ENTROPY_NORMS) + '`'}. Provided: `{self.entropy_norm}`" 120 ) 121 122 123 @dataclass 124 class ConfidenceConfig: 125 """A config which contains the following key-value pairs related to confidence scores. 126 127 Args: 128 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 129 generated during decoding. When set to true, the Hypothesis will contain 130 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 131 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 132 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 133 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 134 135 The length of the list corresponds to the number of recognized tokens. 136 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 137 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 138 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 139 140 The length of the list corresponds to the number of recognized words. 141 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 142 from the `token_confidence`. 143 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 144 Valid options are `mean`, `min`, `max`, `prod`. 145 measure_cfg: A dict-like object which contains the measure name and settings to compute per-frame 146 confidence scores. 147 148 name: The measure name (str). 149 Supported values: 150 - 'max_prob' for using the maximum token probability as a confidence. 151 - 'entropy' for using a normalized entropy of a log-likelihood vector. 152 153 entropy_type: Which type of entropy to use (str). Used if confidence_measure_cfg.name is set to `entropy`. 154 Supported values: 155 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 156 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 157 Note that for this entropy, the alpha should comply the following inequality: 158 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 159 where V is the model vocabulary size. 160 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 161 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 162 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 163 More: https://en.wikipedia.org/wiki/Tsallis_entropy 164 - 'renyi' for the Rรฉnyi entropy. 165 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 166 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 167 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 168 169 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 170 When the alpha equals one, scaling is not applied to 'max_prob', 171 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 172 173 entropy_norm: A mapping of the entropy value to the interval [0,1]. 174 Supported values: 175 - 'lin' for using the linear mapping. 176 - 'exp' for using exponential mapping with linear shift. 177 """ 178 179 preserve_frame_confidence: bool = False 180 preserve_token_confidence: bool = False 181 preserve_word_confidence: bool = False 182 exclude_blank: bool = True 183 aggregation: str = "min" 184 measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() 185 method_cfg: str = "DEPRECATED" 186 187 def __post_init__(self): 188 # OmegaConf.structured ensures that post_init check is always executed 189 self.measure_cfg = OmegaConf.structured( 190 self.measure_cfg 191 if isinstance(self.measure_cfg, ConfidenceMeasureConfig) 192 else ConfidenceMeasureConfig(**self.measure_cfg) 193 ) 194 if self.method_cfg != "DEPRECATED": 195 logging.warning( 196 "`method_cfg` is deprecated and will be removed in the future. Please use `measure_cfg` instead." 197 ) 198 199 # TODO (alaptev): delete the following two lines sometime in the future 200 logging.warning("Re-writing `measure_cfg` with the value of `method_cfg`.") 201 # OmegaConf.structured ensures that post_init check is always executed 202 self.measure_cfg = OmegaConf.structured( 203 self.method_cfg 204 if isinstance(self.method_cfg, ConfidenceMeasureConfig) 205 else ConfidenceMeasureConfig(**self.method_cfg) 206 ) 207 self.method_cfg = "DEPRECATED" 208 if self.aggregation not in ConfidenceConstants.AGGREGATIONS: 209 raise ValueError( 210 f"`aggregation` has to be one of the following: " 211 f"{'`' + '`, `'.join(ConfidenceMeasureConstants.AGGREGATIONS) + '`'}. Provided: `{self.aggregation}`" 212 ) 213 214 215 def get_confidence_measure_bank(): 216 """Generate a dictionary with confidence measure functionals. 217 218 Supported confidence measures: 219 max_prob: normalized maximum probability 220 entropy_gibbs_lin: Gibbs entropy with linear normalization 221 entropy_gibbs_exp: Gibbs entropy with exponential normalization 222 entropy_tsallis_lin: Tsallis entropy with linear normalization 223 entropy_tsallis_exp: Tsallis entropy with exponential normalization 224 entropy_renyi_lin: Rรฉnyi entropy with linear normalization 225 entropy_renyi_exp: Rรฉnyi entropy with exponential normalization 226 227 Returns: 228 dictionary with lambda functions. 229 """ 230 # helper functions 231 # Gibbs entropy is implemented without alpha 232 neg_entropy_gibbs = lambda x: (x.exp() * x).sum(-1) 233 neg_entropy_alpha = lambda x, t: (x * t).exp().sum(-1) 234 neg_entropy_alpha_gibbs = lambda x, t: ((x * t).exp() * x).sum(-1) 235 # too big for a lambda 236 def entropy_tsallis_exp(x, v, t): 237 exp_neg_max_ent = math.exp((1 - math.pow(v, 1 - t)) / (1 - t)) 238 return (((1 - neg_entropy_alpha(x, t)) / (1 - t)).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 239 240 def entropy_gibbs_exp(x, v, t): 241 exp_neg_max_ent = math.pow(v, -t * math.pow(v, 1 - t)) 242 return ((neg_entropy_alpha_gibbs(x, t) * t).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 243 244 # use Gibbs entropies for Tsallis and Rรฉnyi with t == 1.0 245 entropy_gibbs_lin_baseline = lambda x, v: 1 + neg_entropy_gibbs(x) / math.log(v) 246 entropy_gibbs_exp_baseline = lambda x, v: (neg_entropy_gibbs(x).exp() * v - 1) / (v - 1) 247 # fill the measure bank 248 confidence_measure_bank = {} 249 # Maximum probability measure is implemented without alpha 250 confidence_measure_bank["max_prob"] = ( 251 lambda x, v, t: (x.max(dim=-1)[0].exp() * v - 1) / (v - 1) 252 if t == 1.0 253 else ((x.max(dim=-1)[0] * t).exp() * math.pow(v, t) - 1) / (math.pow(v, t) - 1) 254 ) 255 confidence_measure_bank["entropy_gibbs_lin"] = ( 256 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 257 if t == 1.0 258 else 1 + neg_entropy_alpha_gibbs(x, t) / math.log(v) / math.pow(v, 1 - t) 259 ) 260 confidence_measure_bank["entropy_gibbs_exp"] = ( 261 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_gibbs_exp(x, v, t) 262 ) 263 confidence_measure_bank["entropy_tsallis_lin"] = ( 264 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 265 if t == 1.0 266 else 1 + (1 - neg_entropy_alpha(x, t)) / (math.pow(v, 1 - t) - 1) 267 ) 268 confidence_measure_bank["entropy_tsallis_exp"] = ( 269 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_tsallis_exp(x, v, t) 270 ) 271 confidence_measure_bank["entropy_renyi_lin"] = ( 272 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 273 if t == 1.0 274 else 1 + neg_entropy_alpha(x, t).log2() / (t - 1) / math.log(v, 2) 275 ) 276 confidence_measure_bank["entropy_renyi_exp"] = ( 277 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) 278 if t == 1.0 279 else (neg_entropy_alpha(x, t).pow(1 / (t - 1)) * v - 1) / (v - 1) 280 ) 281 return confidence_measure_bank 282 283 284 def get_confidence_aggregation_bank(): 285 """Generate a dictionary with confidence aggregation functions. 286 287 Supported confidence measures: 288 min: minimum 289 max: maximum 290 mean: arithmetic mean 291 prod: product 292 293 Returns: 294 dictionary with functions. 295 """ 296 confidence_aggregation_bank = {"mean": lambda x: sum(x) / len(x), "min": min, "max": max} 297 # python 3.7 and earlier do not have math.prod 298 if hasattr(math, "prod"): 299 confidence_aggregation_bank["prod"] = math.prod 300 else: 301 import operator 302 from functools import reduce 303 304 confidence_aggregation_bank["prod"] = lambda x: reduce(operator.mul, x, 1) 305 return confidence_aggregation_bank 306 307 308 class ConfidenceMeasureMixin(ABC): 309 """Confidence Measure Mixin class. 310 311 It initializes per-frame confidence measure. 312 """ 313 314 def _init_confidence_measure(self, confidence_measure_cfg: Optional[DictConfig] = None): 315 """Initialize per-frame confidence measure from config. 316 """ 317 # OmegaConf.structured ensures that post_init check is always executed 318 confidence_measure_cfg = OmegaConf.structured( 319 ConfidenceMeasureConfig() 320 if confidence_measure_cfg is None 321 else ConfidenceMeasureConfig(**confidence_measure_cfg) 322 ) 323 324 # set confidence calculation measure 325 # we suppose that self.blank_id == len(vocabulary) 326 self.num_tokens = (self.blank_id if hasattr(self, "blank_id") else self._blank_index) + 1 327 self.alpha = confidence_measure_cfg.alpha 328 329 # init confidence measure bank 330 self.confidence_measure_bank = get_confidence_measure_bank() 331 332 measure = None 333 # construct measure_name 334 measure_name = "" 335 if confidence_measure_cfg.name == "max_prob": 336 measure_name = "max_prob" 337 elif confidence_measure_cfg.name == "entropy": 338 measure_name = '_'.join( 339 [confidence_measure_cfg.name, confidence_measure_cfg.entropy_type, confidence_measure_cfg.entropy_norm] 340 ) 341 else: 342 raise ValueError(f"Unsupported `confidence_measure_cfg.name`: `{confidence_measure_cfg.name}`") 343 if measure_name not in self.confidence_measure_bank: 344 raise ValueError(f"Unsupported measure setup: `{measure_name}`") 345 measure = partial(self.confidence_measure_bank[measure_name], v=self.num_tokens, t=self.alpha) 346 self._get_confidence = lambda x: measure(torch.nan_to_num(x)).tolist() 347 348 349 class ConfidenceMixin(ABC): 350 """Confidence Mixin class. 351 352 It is responsible for confidence estimation method initialization and high-level confidence score calculation. 353 """ 354 355 def _init_confidence(self, confidence_cfg: Optional[DictConfig] = None): 356 """Initialize confidence-related fields and confidence aggregation function from config. 357 """ 358 # OmegaConf.structured ensures that post_init check is always executed 359 confidence_cfg = OmegaConf.structured( 360 ConfidenceConfig() if confidence_cfg is None else ConfidenceConfig(**confidence_cfg) 361 ) 362 self.confidence_measure_cfg = confidence_cfg.measure_cfg 363 364 # extract the config 365 self.preserve_word_confidence = confidence_cfg.get('preserve_word_confidence', False) 366 # set preserve_frame_confidence and preserve_token_confidence to True 367 # if preserve_word_confidence is True 368 self.preserve_token_confidence = ( 369 confidence_cfg.get('preserve_token_confidence', False) | self.preserve_word_confidence 370 ) 371 # set preserve_frame_confidence to True if preserve_token_confidence is True 372 self.preserve_frame_confidence = ( 373 confidence_cfg.get('preserve_frame_confidence', False) | self.preserve_token_confidence 374 ) 375 self.exclude_blank_from_confidence = confidence_cfg.get('exclude_blank', True) 376 self.word_confidence_aggregation = confidence_cfg.get('aggregation', "min") 377 378 # define aggregation functions 379 self.confidence_aggregation_bank = get_confidence_aggregation_bank() 380 self._aggregate_confidence = self.confidence_aggregation_bank[self.word_confidence_aggregation] 381 382 # Update preserve frame confidence 383 if self.preserve_frame_confidence is False: 384 if self.cfg.strategy in ['greedy', 'greedy_batch']: 385 self.preserve_frame_confidence = self.cfg.greedy.get('preserve_frame_confidence', False) 386 # OmegaConf.structured ensures that post_init check is always executed 387 confidence_measure_cfg = OmegaConf.structured(self.cfg.greedy).get('confidence_measure_cfg', None) 388 self.confidence_measure_cfg = ( 389 OmegaConf.structured(ConfidenceMeasureConfig()) 390 if confidence_measure_cfg is None 391 else OmegaConf.structured(ConfidenceMeasureConfig(**confidence_measure_cfg)) 392 ) 393 394 @abstractmethod 395 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 396 """Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 397 Assumes that `frame_confidence` is present in the hypotheses. 398 399 Args: 400 hypotheses_list: List of Hypothesis. 401 402 Returns: 403 A list of hypotheses with high-level confidence scores. 404 """ 405 raise NotImplementedError() 406 407 @abstractmethod 408 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 409 """Implemented by subclass in order to aggregate token confidence to a word-level confidence. 410 411 Args: 412 hypothesis: Hypothesis 413 414 Returns: 415 A list of word-level confidence scores. 416 """ 417 raise NotImplementedError() 418 419 def _aggregate_token_confidence_chars(self, words: List[str], token_confidence: List[float]) -> List[float]: 420 """Implementation of token confidence aggregation for character-based models. 421 422 Args: 423 words: List of words of a hypothesis. 424 token_confidence: List of token-level confidence scores of a hypothesis. 425 426 Returns: 427 A list of word-level confidence scores. 428 """ 429 word_confidence = [] 430 i = 0 431 for word in words: 432 word_len = len(word) 433 word_confidence.append(self._aggregate_confidence(token_confidence[i : i + word_len])) 434 # we assume that there is exactly one space token between words and exclude it from word confidence 435 i += word_len + 1 436 return word_confidence 437 438 def _aggregate_token_confidence_subwords_sentencepiece( 439 self, words: List[str], token_confidence: List[float], token_ids: List[int] 440 ) -> List[float]: 441 """Implementation of token confidence aggregation for subword-based models. 442 443 **Note**: Only supports Sentencepiece based tokenizers ! 444 445 Args: 446 words: List of words of a hypothesis. 447 token_confidence: List of token-level confidence scores of a hypothesis. 448 token_ids: List of token ids of a hypothesis. 449 450 Returns: 451 A list of word-level confidence scores. 452 """ 453 word_confidence = [] 454 # run only if there are final words 455 if len(words) > 0: 456 j = 0 457 prev_unk = False 458 prev_underline = False 459 for i, token_id in enumerate(token_ids): 460 token = self.decode_ids_to_tokens([int(token_id)])[0] 461 token_text = self.decode_tokens_to_str([int(token_id)]) 462 # treat `<unk>` as a separate word regardless of the next token 463 # to match the result of `tokenizer.ids_to_text` 464 if (token != token_text or prev_unk) and i > j: 465 # do not add confidence for `โ–` if the current token starts with `โ–` 466 # to match the result of `tokenizer.ids_to_text` 467 if not prev_underline: 468 word_confidence.append(self._aggregate_confidence(token_confidence[j:i])) 469 j = i 470 prev_unk = token == '<unk>' 471 prev_underline = token == 'โ–' 472 if not prev_underline: 473 word_confidence.append(self._aggregate_confidence(token_confidence[j : len(token_ids)])) 474 if len(words) != len(word_confidence): 475 raise RuntimeError( 476 f"""Something went wrong with word-level confidence aggregation.\n 477 Please check these values for debugging:\n 478 len(words): {len(words)},\n 479 len(word_confidence): {len(word_confidence)},\n 480 recognized text: `{' '.join(words)}`""" 481 ) 482 return word_confidence 483 [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of nemo/collections/common/parts/adapter_modules.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, is_dataclass 16 from typing import Any, Optional 17 18 from hydra.utils import instantiate 19 from omegaconf import OmegaConf 20 from torch import nn as nn 21 22 from nemo.collections.common.parts.utils import activation_registry 23 from nemo.core.classes.mixins import access_mixins, adapter_mixin_strategies 24 25 26 class AdapterModuleUtil(access_mixins.AccessMixin): 27 """ 28 Base class of Adapter Modules, providing common functionality to all Adapter Modules. 29 """ 30 31 def setup_adapter_strategy(self, adapter_strategy: Optional[adapter_mixin_strategies.AbstractAdapterStrategy]): 32 """ 33 Setup adapter strategy of this class, enabling dynamic change in the way the adapter output is 34 merged with the input. 35 36 When called successfully, will assign the variable `adapter_strategy` to the module. 37 38 Args: 39 adapter_strategy: Can be a None or an implementation of AbstractAdapterStrategy. 40 """ 41 # set default adapter strategy 42 if adapter_strategy is None: 43 adapter_strategy = self.get_default_strategy_config() 44 45 if is_dataclass(adapter_strategy): 46 adapter_strategy = OmegaConf.structured(adapter_strategy) 47 OmegaConf.set_struct(adapter_strategy, False) 48 49 # The config must have the `_target_` field pointing to the actual adapter strategy class 50 # which will load that strategy dynamically to this module. 51 if isinstance(adapter_strategy, dict) or OmegaConf.is_config(adapter_strategy): 52 self.adapter_strategy = instantiate(adapter_strategy) 53 elif isinstance(adapter_strategy, adapter_mixin_strategies.AbstractAdapterStrategy): 54 self.adapter_strategy = adapter_strategy 55 else: 56 raise AttributeError(f'`adapter_strategy` provided is invalid : {adapter_strategy}') 57 58 def get_default_strategy_config(self) -> 'dataclass': 59 """ 60 Returns a default adapter module strategy. 61 """ 62 return adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 63 64 def adapter_unfreeze(self,): 65 """ 66 Sets the requires grad for all parameters in the adapter to True. 67 This method should be overridden for any custom unfreeze behavior that is required. 68 For example, if not all params of the adapter should be unfrozen. 69 """ 70 for param in self.parameters(): 71 param.requires_grad_(True) 72 73 74 class LinearAdapter(nn.Module, AdapterModuleUtil): 75 76 """ 77 Simple Linear Feedforward Adapter module with LayerNorm and singe hidden layer with activation function. 78 Note: The adapter explicitly initializes its final layer with all zeros in order to avoid affecting the 79 original model when all adapters are disabled. 80 81 Args: 82 in_features: Input dimension of the module. Note that for adapters, input_dim == output_dim. 83 dim: Hidden dimension of the feed forward network. 84 activation: Str name for an activation function. 85 norm_position: Str, can be `pre` or `post`. Defaults to `pre`. Determines whether the normalization 86 will occur in the first layer or the last layer. Certain architectures may prefer one over the other. 87 dropout: float value, whether to perform dropout on the output of the last layer of the adapter. 88 adapter_strategy: By default, ResidualAddAdapterStrategyConfig. An adapter composition function object. 89 """ 90 91 def __init__( 92 self, 93 in_features: int, 94 dim: int, 95 activation: str = 'swish', 96 norm_position: str = 'pre', 97 dropout: float = 0.0, 98 adapter_strategy: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig = None, 99 ): 100 super().__init__() 101 102 activation = activation_registry[activation]() 103 # If the activation can be executed in place, do so. 104 if hasattr(activation, 'inplace'): 105 activation.inplace = True 106 107 assert norm_position in ['pre', 'post'] 108 self.norm_position = norm_position 109 110 if norm_position == 'pre': 111 self.module = nn.Sequential( 112 nn.LayerNorm(in_features), 113 nn.Linear(in_features, dim, bias=False), 114 activation, 115 nn.Linear(dim, in_features, bias=False), 116 ) 117 118 elif norm_position == 'post': 119 self.module = nn.Sequential( 120 nn.Linear(in_features, dim, bias=False), 121 activation, 122 nn.Linear(dim, in_features, bias=False), 123 nn.LayerNorm(in_features), 124 ) 125 126 if dropout > 0.0: 127 self.dropout = nn.Dropout(dropout) 128 else: 129 self.dropout = None 130 131 # Setup adapter strategy 132 self.setup_adapter_strategy(adapter_strategy) 133 134 # reset parameters 135 self.reset_parameters() 136 137 def reset_parameters(self): 138 # Final layer initializations must be 0 139 if self.norm_position == 'pre': 140 self.module[-1].weight.data *= 0 141 142 elif self.norm_position == 'post': 143 self.module[-1].weight.data *= 0 144 self.module[-1].bias.data *= 0 145 146 def forward(self, x): 147 x = self.module(x) 148 149 # Add dropout if available 150 if self.dropout is not None: 151 x = self.dropout(x) 152 153 return x 154 155 156 @dataclass 157 class LinearAdapterConfig: 158 in_features: int 159 dim: int 160 activation: str = 'swish' 161 norm_position: str = 'pre' 162 dropout: float = 0.0 163 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 164 _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) 165 [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import re 15 from typing import List 16 17 import ipadic 18 import MeCab 19 from pangu import spacing 20 from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer 21 22 23 class EnJaProcessor: 24 """ 25 Tokenizer, Detokenizer and Normalizer utilities for Japanese & English 26 Args: 27 lang_id: One of ['en', 'ja']. 28 """ 29 30 def __init__(self, lang_id: str): 31 self.lang_id = lang_id 32 self.moses_tokenizer = MosesTokenizer(lang=lang_id) 33 self.moses_detokenizer = MosesDetokenizer(lang=lang_id) 34 self.normalizer = MosesPunctNormalizer( 35 lang=lang_id, pre_replace_unicode_punct=True, post_remove_control_chars=True 36 ) 37 38 def detokenize(self, tokens: List[str]) -> str: 39 """ 40 Detokenizes a list of tokens 41 Args: 42 tokens: list of strings as tokens 43 Returns: 44 detokenized Japanese or English string 45 """ 46 return self.moses_detokenizer.detokenize(tokens) 47 48 def tokenize(self, text) -> str: 49 """ 50 Tokenizes text using Moses. Returns a string of tokens. 51 """ 52 tokens = self.moses_tokenizer.tokenize(text) 53 return ' '.join(tokens) 54 55 def normalize(self, text) -> str: 56 # Normalization doesn't handle Japanese periods correctly; 57 # 'ใ€‚'becomes '.'. 58 if self.lang_id == 'en': 59 return self.normalizer.normalize(text) 60 else: 61 return text 62 63 64 class JaMecabProcessor: 65 """ 66 Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English 67 """ 68 69 def __init__(self): 70 self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") 71 72 def detokenize(self, text: List[str]) -> str: 73 RE_WS_IN_FW = re.compile( 74 r'([\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])\s+(?=[\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])' 75 ) 76 77 detokenize = lambda s: spacing(RE_WS_IN_FW.sub(r'\1', s)).strip() 78 return detokenize(' '.join(text)) 79 80 def tokenize(self, text) -> str: 81 """ 82 Tokenizes text using Moses. Returns a string of tokens. 83 """ 84 return self.mecab_tokenizer.parse(text).strip() 85 86 def normalize(self, text) -> str: 87 return text 88 [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Optional, Tuple 17 18 from omegaconf.omegaconf import MISSING 19 20 from nemo.collections.nlp.data.machine_translation.machine_translation_dataset import TranslationDataConfig 21 from nemo.collections.nlp.models.enc_dec_nlp_model import EncDecNLPModelConfig 22 from nemo.collections.nlp.modules.common.token_classifier import TokenClassifierConfig 23 from nemo.collections.nlp.modules.common.tokenizer_utils import TokenizerConfig 24 from nemo.collections.nlp.modules.common.transformer.transformer import ( 25 NeMoTransformerConfig, 26 NeMoTransformerEncoderConfig, 27 ) 28 from nemo.collections.nlp.modules.common.transformer.transformer_bottleneck import ( 29 NeMoTransformerBottleneckDecoderConfig, 30 NeMoTransformerBottleneckEncoderConfig, 31 ) 32 from nemo.core.config.modelPT import OptimConfig, SchedConfig 33 34 35 @dataclass 36 class MTSchedConfig(SchedConfig): 37 name: str = 'InverseSquareRootAnnealing' 38 warmup_ratio: Optional[float] = None 39 last_epoch: int = -1 40 41 42 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). 43 @dataclass 44 class MTOptimConfig(OptimConfig): 45 name: str = 'adam' 46 lr: float = 1e-3 47 betas: Tuple[float, float] = (0.9, 0.98) 48 weight_decay: float = 0.0 49 sched: Optional[MTSchedConfig] = MTSchedConfig() 50 51 52 @dataclass 53 class MTEncDecModelConfig(EncDecNLPModelConfig): 54 # machine translation configurations 55 num_val_examples: int = 3 56 num_test_examples: int = 3 57 max_generation_delta: int = 10 58 label_smoothing: Optional[float] = 0.0 59 beam_size: int = 4 60 len_pen: float = 0.0 61 src_language: Any = 'en' # Any = str or List[str] 62 tgt_language: Any = 'en' # Any = str or List[str] 63 find_unused_parameters: Optional[bool] = True 64 shared_tokenizer: Optional[bool] = True 65 multilingual: Optional[bool] = False 66 preproc_out_dir: Optional[str] = None 67 validate_input_ids: Optional[bool] = True 68 shared_embeddings: bool = False 69 70 # network architecture configuration 71 encoder_tokenizer: Any = MISSING 72 encoder: Any = MISSING 73 74 decoder_tokenizer: Any = MISSING 75 decoder: Any = MISSING 76 77 head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) 78 79 # dataset configurations 80 train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 81 src_file_name=MISSING, 82 tgt_file_name=MISSING, 83 tokens_in_batch=512, 84 clean=True, 85 shuffle=True, 86 cache_ids=False, 87 use_cache=False, 88 ) 89 validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 90 src_file_name=MISSING, 91 tgt_file_name=MISSING, 92 tokens_in_batch=512, 93 clean=False, 94 shuffle=False, 95 cache_ids=False, 96 use_cache=False, 97 ) 98 test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 99 src_file_name=MISSING, 100 tgt_file_name=MISSING, 101 tokens_in_batch=512, 102 clean=False, 103 shuffle=False, 104 cache_ids=False, 105 use_cache=False, 106 ) 107 optim: Optional[OptimConfig] = MTOptimConfig() 108 109 110 @dataclass 111 class AAYNBaseConfig(MTEncDecModelConfig): 112 113 # Attention is All You Need Base Configuration 114 encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 115 decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 116 117 encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( 118 library='nemo', 119 model_name=None, 120 pretrained=False, 121 hidden_size=512, 122 inner_size=2048, 123 num_layers=6, 124 num_attention_heads=8, 125 ffn_dropout=0.1, 126 attn_score_dropout=0.1, 127 attn_layer_dropout=0.1, 128 ) 129 130 decoder: NeMoTransformerConfig = NeMoTransformerConfig( 131 library='nemo', 132 model_name=None, 133 pretrained=False, 134 hidden_size=512, 135 inner_size=2048, 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass 145 class MTBottleneckModelConfig(AAYNBaseConfig): 146 model_type: str = 'nll' 147 min_logv: float = -6 148 latent_size: int = -1 # -1 will take value of encoder hidden 149 non_recon_warmup_batches: int = 200000 150 recon_per_token: bool = True 151 log_timing: bool = True 152 153 encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( 154 library='nemo', 155 model_name=None, 156 pretrained=False, 157 hidden_size=512, 158 inner_size=2048, 159 num_layers=6, 160 num_attention_heads=8, 161 ffn_dropout=0.1, 162 attn_score_dropout=0.1, 163 attn_layer_dropout=0.1, 164 arch='seq2seq', 165 hidden_steps=32, 166 hidden_blocks=1, 167 hidden_init_method='params', 168 ) 169 170 decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( 171 library='nemo', 172 model_name=None, 173 pretrained=False, 174 inner_size=2048, 175 num_layers=6, 176 num_attention_heads=8, 177 ffn_dropout=0.1, 178 attn_score_dropout=0.1, 179 attn_layer_dropout=0.1, 180 arch='seq2seq', 181 ) 182 [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, Optional 17 18 from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict 19 20 from nemo.collections.common.parts.adapter_modules import LinearAdapterConfig 21 from nemo.collections.nlp.data.token_classification.punctuation_capitalization_dataset import ( 22 PunctuationCapitalizationEvalDataConfig, 23 PunctuationCapitalizationTrainDataConfig, 24 legacy_data_config_to_new_data_config, 25 ) 26 from nemo.core.config import TrainerConfig 27 from nemo.core.config.modelPT import NemoConfig 28 from nemo.utils.exp_manager import ExpManagerConfig 29 30 31 @dataclass 32 class FreezeConfig: 33 is_enabled: bool = False 34 """Freeze audio encoder weight and add Conformer Layers on top of it""" 35 d_model: Optional[int] = 256 36 """`d_model` parameter of ``ConformerLayer``""" 37 d_ff: Optional[int] = 1024 38 """``d_ff`` parameter of ``ConformerLayer``""" 39 num_layers: Optional[int] = 8 40 """``num_layers`` number of ``ConformerLayer`` modules to add on top of audio encoder""" 41 42 43 @dataclass 44 class AdapterConfig: 45 config: Optional[LinearAdapterConfig] = None 46 """Linear adapter config see ``collections.common.parts.LinearAdapterConfig``""" 47 enable: bool = False 48 """Use adapters for audio encoder""" 49 50 51 @dataclass 52 class FusionConfig: 53 num_layers: Optional[int] = 4 54 """"Number of layers to use in fusion""" 55 num_attention_heads: Optional[int] = 4 56 """Number of attention heads to use in fusion""" 57 inner_size: Optional[int] = 2048 58 """Fusion inner size""" 59 60 61 @dataclass 62 class AudioEncoderConfig: 63 pretrained_model: str = MISSING 64 """A configuration for restoring pretrained audio encoder""" 65 freeze: Optional[FreezeConfig] = None 66 adapter: Optional[AdapterConfig] = None 67 fusion: Optional[FusionConfig] = None 68 69 70 @dataclass 71 class TokenizerConfig: 72 """A structure and default values of source text tokenizer.""" 73 74 vocab_file: Optional[str] = None 75 """A path to vocabulary file which is used in ``'word'``, ``'char'``, and HuggingFace tokenizers""" 76 77 tokenizer_name: str = MISSING 78 """A name of the tokenizer used for tokenization of source sequences. Possible options are ``'sentencepiece'``, 79 ``'word'``, ``'char'``, HuggingFace tokenizers (e.g. ``'bert-base-uncased'``). For more options see function 80 ``nemo.collections.nlp.modules.common.get_tokenizer``. The tokenizer must have properties ``cls_id``, ``pad_id``, 81 ``sep_id``, ``unk_id``.""" 82 83 special_tokens: Optional[Dict[str, str]] = None 84 """A dictionary with special tokens passed to constructors of ``'char'``, ``'word'``, ``'sentencepiece'``, and 85 various HuggingFace tokenizers.""" 86 87 tokenizer_model: Optional[str] = None 88 """A path to a tokenizer model required for ``'sentencepiece'`` tokenizer.""" 89 90 91 @dataclass 92 class LanguageModelConfig: 93 """ 94 A structure and default values of language model configuration of punctuation and capitalization model. BERT like 95 HuggingFace models are supported. Provide a valid ``pretrained_model_name`` and, optionally, you may 96 reinitialize model via ``config_file`` or ``config``. 97 98 Alternatively you can initialize the language model using ``lm_checkpoint``. 99 100 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 101 """ 102 103 pretrained_model_name: str = MISSING 104 """A mandatory parameter containing name of HuggingFace pretrained model. For example, ``'bert-base-uncased'``.""" 105 106 config_file: Optional[str] = None 107 """A path to a file with HuggingFace model config which is used to reinitialize language model.""" 108 109 config: Optional[Dict] = None 110 """A HuggingFace config which is used to reinitialize language model.""" 111 112 lm_checkpoint: Optional[str] = None 113 """A path to a ``torch`` checkpoint of a language model.""" 114 115 116 @dataclass 117 class HeadConfig: 118 """ 119 A structure and default values of configuration of capitalization or punctuation model head. This config defines a 120 multilayer perceptron which is applied to output of a language model. Number of units in the hidden layer is equal 121 to the dimension of the language model. 122 123 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 124 """ 125 126 num_fc_layers: int = 1 127 """A number of hidden layers in a multilayer perceptron.""" 128 129 fc_dropout: float = 0.1 130 """A dropout used in an MLP.""" 131 132 activation: str = 'relu' 133 """An activation used in hidden layers.""" 134 135 use_transformer_init: bool = True 136 """Whether to initialize the weights of the classifier head with the approach that was used for language model 137 initialization.""" 138 139 140 @dataclass 141 class ClassLabelsConfig: 142 """ 143 A structure and default values of a mandatory part of config which contains names of files which are saved in .nemo 144 checkpoint. These files can also be used for passing label vocabulary to the model. For using them as label 145 vocabularies you will need to provide path these files in parameter 146 ``model.common_dataset_parameters.label_vocab_dir``. Each line in labels files 147 contains 1 label. The values are sorted, ``<line number>==<label id>``, starting from ``0``. A label with ``0`` id 148 must contain neutral label which must be equal to ``model.common_dataset_parameters.pad_label``. 149 150 This config is a part of :class:`~CommonDatasetParametersConfig`. 151 """ 152 153 punct_labels_file: str = MISSING 154 """A name of punctuation labels file.""" 155 156 capit_labels_file: str = MISSING 157 """A name of capitalization labels file.""" 158 159 160 @dataclass 161 class CommonDatasetParametersConfig: 162 """ 163 A structure and default values of common dataset parameters config which includes label and loss mask information. 164 If you omit parameters ``punct_label_ids``, ``capit_label_ids``, ``label_vocab_dir``, then labels will be inferred 165 from a training dataset or loaded from a checkpoint. 166 167 Parameters ``ignore_extra_tokens`` and ``ignore_start_end`` are responsible for forming loss mask. A loss mask 168 defines on which tokens loss is computed. 169 170 This parameter is a part of config :class:`~PunctuationCapitalizationModelConfig`. 171 """ 172 173 pad_label: str = MISSING 174 """A mandatory parameter which should contain label used for punctuation and capitalization label padding. It 175 also serves as a neutral label for both punctuation and capitalization. If any of ``punct_label_ids``, 176 ``capit_label_ids`` parameters is provided, then ``pad_label`` must have ``0`` id in them. In addition, if ``label_vocab_dir`` 177 is provided, then ``pad_label`` must be on the first lines in files ``class_labels.punct_labels_file`` and 178 ``class_labels.capit_labels_file``.""" 179 180 ignore_extra_tokens: bool = False 181 """Whether to compute loss on not first tokens in words. If this parameter is ``True``, then loss mask is ``False`` 182 for all tokens in a word except the first.""" 183 184 ignore_start_end: bool = True 185 """If ``False``, then loss is computed on [CLS] and [SEP] tokens.""" 186 187 punct_label_ids: Optional[Dict[str, int]] = None 188 """A dictionary with punctuation label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this 189 parameter and pass label ids through ``class_labels.punct_labels_file`` or let the model to infer label ids from 190 dataset or load them from checkpoint.""" 191 192 capit_label_ids: Optional[Dict[str, int]] = None 193 """A dictionary with capitalization label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit 194 this parameter and pass label ids through ``class_labels.capit_labels_file`` or let model to infer label ids from 195 dataset or load them from checkpoint.""" 196 197 label_vocab_dir: Optional[str] = None 198 """A path to directory which contains class labels files. See :class:`ClassLabelsConfig`. If this parameter is 199 provided, then labels will be loaded from files which are located in ``label_vocab_dir`` and have names specified 200 in ``model.class_labels`` configuration section. A label specified in ``pad_label`` has to be on the first lines 201 of ``model.class_labels`` files.""" 202 203 204 @dataclass 205 class PunctuationCapitalizationModelConfig: 206 """ 207 A configuration of 208 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 209 model. 210 211 See an example of model config in 212 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 213 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 214 215 This config is a part of :class:`~PunctuationCapitalizationConfig`. 216 """ 217 218 class_labels: ClassLabelsConfig = ClassLabelsConfig() 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" 229 230 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 231 """A configuration for creating validation datasets and data loaders.""" 232 233 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 234 """A configuration for creating test datasets and data loaders.""" 235 236 punct_head: HeadConfig = HeadConfig() 237 """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" 238 239 capit_head: HeadConfig = HeadConfig() 240 """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" 241 242 tokenizer: Any = TokenizerConfig() 243 """A configuration for source text tokenizer.""" 244 245 language_model: LanguageModelConfig = LanguageModelConfig() 246 """A configuration of a BERT-like language model which serves as a model body.""" 247 248 optim: Optional[Any] = None 249 """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For 250 description see `Optimizers 251 <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/core/core.html#optimizers>`_ section in 252 documentation and `primer <https://github.com/NVIDIA/NeMo/blob/main/tutorials/00_NeMo_Primer.ipynb>_ tutorial.""" 253 254 255 @dataclass 256 class PunctuationCapitalizationLexicalAudioModelConfig(PunctuationCapitalizationModelConfig): 257 """ 258 A configuration of 259 :class:`~nemo.collections.nlp.models.token_classification.punctuation_lexical_audio_capitalization_model.PunctuationCapitalizationLexicalAudioModel` 260 model. 261 262 See an example of model config in 263 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 264 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_lexical_audio_config.yaml>`_ 265 266 Audio encoder can be frozen during training with ``freeze_audio_encoder`` parameter. 267 Adapter can be added to audio encoder with ``use_adapters`` and ``adapter_config`` parameters. 268 More conformer layers can be added on top of pretrained audio encoder with ``frozen_conf_d_model``, ``frozen_conf_d_ff`` and ``frozen_conf_num_layers`` parameters. 269 """ 270 271 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 272 """A configuration for creating training dataset and data loader.""" 273 274 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 275 """A configuration for creating validation datasets and data loaders.""" 276 277 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 278 """A configuration for creating test datasets and data loaders.""" 279 280 audio_encoder: Optional[AudioEncoderConfig] = None 281 282 restore_lexical_encoder_from: Optional[str] = None 283 """"Path to .nemo checkpoint to load weights from""" # add more comments 284 285 use_weighted_loss: Optional[bool] = False 286 """If set to ``True`` CrossEntropyLoss will be weighted""" 287 288 289 @dataclass 290 class PunctuationCapitalizationConfig(NemoConfig): 291 """ 292 A config for punctuation model training and testing. 293 294 See an example of full config in 295 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 296 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 297 """ 298 299 pretrained_model: Optional[str] = None 300 """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options 301 by calling method 302 :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. 303 """ 304 305 name: Optional[str] = 'Punctuation_and_Capitalization' 306 """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" 307 308 do_training: bool = True 309 """Whether to perform training of the model.""" 310 311 do_testing: bool = False 312 """Whether ot perform testing of the model.""" 313 314 model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() 315 """A configuration for the 316 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 317 model.""" 318 319 trainer: Optional[TrainerConfig] = TrainerConfig() 320 """Contains ``Trainer`` Lightning class constructor parameters.""" 321 322 exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) 323 """A configuration with various NeMo training options such as output directories, resuming from checkpoint, 324 tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" 325 326 327 @dataclass 328 class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): 329 model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() 330 331 332 def is_legacy_model_config(model_cfg: DictConfig) -> bool: 333 """ 334 Test if model config is old style config. Old style configs are configs which were used before 335 ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of 336 ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support 337 tarred datasets. 338 339 Args: 340 model_cfg: model configuration 341 342 Returns: 343 whether ``model_config`` is legacy 344 """ 345 return 'common_dataset_parameters' not in model_cfg 346 347 348 def legacy_model_config_to_new_model_config(model_cfg: DictConfig) -> DictConfig: 349 """ 350 Transform old style config into 351 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig`. 352 Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style 353 datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. 354 Old style configs do not support tarred datasets. 355 356 Args: 357 model_cfg: old style config 358 359 Returns: 360 model config which follows dataclass 361 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig` 362 """ 363 train_ds = model_cfg.get('train_ds') 364 validation_ds = model_cfg.get('validation_ds') 365 test_ds = model_cfg.get('test_ds') 366 dataset = model_cfg.dataset 367 punct_head_config = model_cfg.get('punct_head', {}) 368 capit_head_config = model_cfg.get('capit_head', {}) 369 omega_conf = OmegaConf.structured( 370 PunctuationCapitalizationModelConfig( 371 class_labels=model_cfg.class_labels, 372 common_dataset_parameters=CommonDatasetParametersConfig( 373 pad_label=dataset.pad_label, 374 ignore_extra_tokens=dataset.get( 375 'ignore_extra_tokens', CommonDatasetParametersConfig.ignore_extra_tokens 376 ), 377 ignore_start_end=dataset.get('ignore_start_end', CommonDatasetParametersConfig.ignore_start_end), 378 punct_label_ids=model_cfg.punct_label_ids, 379 capit_label_ids=model_cfg.capit_label_ids, 380 ), 381 train_ds=None 382 if train_ds is None 383 else legacy_data_config_to_new_data_config(train_ds, dataset, train=True), 384 validation_ds=None 385 if validation_ds is None 386 else legacy_data_config_to_new_data_config(validation_ds, dataset, train=False), 387 test_ds=None if test_ds is None else legacy_data_config_to_new_data_config(test_ds, dataset, train=False), 388 punct_head=HeadConfig( 389 num_fc_layers=punct_head_config.get('punct_num_fc_layers', HeadConfig.num_fc_layers), 390 fc_dropout=punct_head_config.get('fc_dropout', HeadConfig.fc_dropout), 391 activation=punct_head_config.get('activation', HeadConfig.activation), 392 use_transformer_init=punct_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 393 ), 394 capit_head=HeadConfig( 395 num_fc_layers=capit_head_config.get('capit_num_fc_layers', HeadConfig.num_fc_layers), 396 fc_dropout=capit_head_config.get('fc_dropout', HeadConfig.fc_dropout), 397 activation=capit_head_config.get('activation', HeadConfig.activation), 398 use_transformer_init=capit_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 399 ), 400 tokenizer=model_cfg.tokenizer, 401 language_model=model_cfg.language_model, 402 optim=model_cfg.optim, 403 ) 404 ) 405 with open_dict(omega_conf): 406 retain_during_legacy_conversion = model_cfg.get('retain_during_legacy_conversion', {}) 407 for key in retain_during_legacy_conversion.keys(): 408 omega_conf[key] = retain_during_legacy_conversion[key] 409 return omega_conf 410 [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """Transformer based language model.""" 16 from MeCab import Model 17 from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule 18 from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule 19 from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( 20 MegatronRetrievalTransformerEncoderModule, 21 ) 22 from nemo.collections.nlp.modules.common.megatron.utils import ( 23 ApexGuardDefaults, 24 init_method_normal, 25 scaled_init_method_normal, 26 ) 27 28 try: 29 from apex.transformer.enums import AttnMaskType, ModelType 30 31 HAVE_APEX = True 32 except (ImportError, ModuleNotFoundError): 33 HAVE_APEX = False 34 # fake missing classes with None attributes 35 AttnMaskType = ApexGuardDefaults() 36 ModelType = ApexGuardDefaults() 37 38 try: 39 from megatron.core import ModelParallelConfig 40 41 HAVE_MEGATRON_CORE = True 42 43 except (ImportError, ModuleNotFoundError): 44 45 ModelParallelConfig = ApexGuardDefaults 46 47 HAVE_MEGATRON_CORE = False 48 49 __all__ = [] 50 51 AVAILABLE_ENCODERS = ["transformer", "perceiver", "retro"] 52 53 54 def get_encoder_model( 55 config: ModelParallelConfig, 56 arch, 57 hidden_size, 58 ffn_hidden_size, 59 num_layers, 60 num_attention_heads, 61 apply_query_key_layer_scaling=False, 62 kv_channels=None, 63 init_method=None, 64 scaled_init_method=None, 65 encoder_attn_mask_type=AttnMaskType.padding, 66 pre_process=True, 67 post_process=True, 68 init_method_std=0.02, 69 megatron_amp_O2=False, 70 hidden_dropout=0.1, 71 attention_dropout=0.1, 72 ffn_dropout=0.0, 73 precision=16, 74 fp32_residual_connection=False, 75 activations_checkpoint_method=None, 76 activations_checkpoint_num_layers=1, 77 activations_checkpoint_granularity=None, 78 layernorm_epsilon=1e-5, 79 bias_activation_fusion=True, 80 bias_dropout_add_fusion=True, 81 masked_softmax_fusion=True, 82 persist_layer_norm=False, 83 openai_gelu=False, 84 activation="gelu", 85 onnx_safe=False, 86 bias=True, 87 normalization="layernorm", 88 headscale=False, 89 transformer_block_type="pre_ln", 90 hidden_steps=32, 91 parent_model_type=ModelType.encoder_or_decoder, 92 layer_type=None, 93 chunk_size=64, 94 num_self_attention_per_cross_attention=1, 95 layer_number_offset=0, # this is use only for attention norm_factor scaling 96 megatron_legacy=False, 97 normalize_attention_scores=True, 98 sequence_parallel=False, 99 num_moe_experts=1, 100 moe_frequency=1, 101 moe_dropout=0.0, 102 turn_off_rop=False, # turn off the RoP positional embedding 103 version=1, # model version 104 position_embedding_type='learned_absolute', 105 use_flash_attention=False, 106 ): 107 """Build language model and return along with the key to save.""" 108 109 if kv_channels is None: 110 assert ( 111 hidden_size % num_attention_heads == 0 112 ), 'hidden_size must be divisible by num_attention_heads if kv_channels is None' 113 kv_channels = hidden_size // num_attention_heads 114 115 if init_method is None: 116 init_method = init_method_normal(init_method_std) 117 118 if scaled_init_method is None: 119 scaled_init_method = scaled_init_method_normal(init_method_std, num_layers) 120 121 if arch == "transformer": 122 # Language encoder. 123 encoder = MegatronTransformerEncoderModule( 124 config=config, 125 init_method=init_method, 126 output_layer_init_method=scaled_init_method, 127 hidden_size=hidden_size, 128 num_layers=num_layers, 129 num_attention_heads=num_attention_heads, 130 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 131 kv_channels=kv_channels, 132 ffn_hidden_size=ffn_hidden_size, 133 encoder_attn_mask_type=encoder_attn_mask_type, 134 pre_process=pre_process, 135 post_process=post_process, 136 megatron_amp_O2=megatron_amp_O2, 137 hidden_dropout=hidden_dropout, 138 attention_dropout=attention_dropout, 139 ffn_dropout=ffn_dropout, 140 precision=precision, 141 fp32_residual_connection=fp32_residual_connection, 142 activations_checkpoint_method=activations_checkpoint_method, 143 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 144 activations_checkpoint_granularity=activations_checkpoint_granularity, 145 layernorm_epsilon=layernorm_epsilon, 146 bias_activation_fusion=bias_activation_fusion, 147 bias_dropout_add_fusion=bias_dropout_add_fusion, 148 masked_softmax_fusion=masked_softmax_fusion, 149 persist_layer_norm=persist_layer_norm, 150 openai_gelu=openai_gelu, 151 onnx_safe=onnx_safe, 152 activation=activation, 153 bias=bias, 154 normalization=normalization, 155 transformer_block_type=transformer_block_type, 156 headscale=headscale, 157 parent_model_type=parent_model_type, 158 megatron_legacy=megatron_legacy, 159 normalize_attention_scores=normalize_attention_scores, 160 num_moe_experts=num_moe_experts, 161 moe_frequency=moe_frequency, 162 moe_dropout=moe_dropout, 163 position_embedding_type=position_embedding_type, 164 use_flash_attention=use_flash_attention, 165 ) 166 elif arch == "retro": 167 encoder = MegatronRetrievalTransformerEncoderModule( 168 config=config, 169 init_method=init_method, 170 output_layer_init_method=scaled_init_method, 171 hidden_size=hidden_size, 172 num_layers=num_layers, 173 num_attention_heads=num_attention_heads, 174 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 175 kv_channels=kv_channels, 176 layer_type=layer_type, 177 ffn_hidden_size=ffn_hidden_size, 178 pre_process=pre_process, 179 post_process=post_process, 180 megatron_amp_O2=megatron_amp_O2, 181 hidden_dropout=hidden_dropout, 182 attention_dropout=attention_dropout, 183 precision=precision, 184 fp32_residual_connection=fp32_residual_connection, 185 activations_checkpoint_method=activations_checkpoint_method, 186 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 187 activations_checkpoint_granularity=activations_checkpoint_granularity, 188 layernorm_epsilon=layernorm_epsilon, 189 bias_activation_fusion=bias_activation_fusion, 190 bias_dropout_add_fusion=bias_dropout_add_fusion, 191 masked_softmax_fusion=masked_softmax_fusion, 192 persist_layer_norm=persist_layer_norm, 193 openai_gelu=openai_gelu, 194 onnx_safe=onnx_safe, 195 activation=activation, 196 bias=bias, 197 normalization=normalization, 198 transformer_block_type=transformer_block_type, 199 parent_model_type=parent_model_type, 200 chunk_size=chunk_size, 201 layer_number_offset=layer_number_offset, 202 megatron_legacy=megatron_legacy, 203 normalize_attention_scores=normalize_attention_scores, 204 turn_off_rop=turn_off_rop, 205 version=version, 206 ) 207 elif arch == "perceiver": 208 encoder = MegatronPerceiverEncoderModule( 209 config=config, 210 init_method=init_method, 211 output_layer_init_method=scaled_init_method, 212 hidden_size=hidden_size, 213 num_layers=num_layers, 214 num_attention_heads=num_attention_heads, 215 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 216 kv_channels=kv_channels, 217 ffn_hidden_size=ffn_hidden_size, 218 encoder_attn_mask_type=encoder_attn_mask_type, 219 pre_process=pre_process, 220 post_process=post_process, 221 megatron_amp_O2=megatron_amp_O2, 222 hidden_dropout=hidden_dropout, 223 attention_dropout=attention_dropout, 224 ffn_dropout=ffn_dropout, 225 precision=precision, 226 fp32_residual_connection=fp32_residual_connection, 227 activations_checkpoint_method=activations_checkpoint_method, 228 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 229 activations_checkpoint_granularity=activations_checkpoint_granularity, 230 layernorm_epsilon=layernorm_epsilon, 231 bias_activation_fusion=bias_activation_fusion, 232 bias_dropout_add_fusion=bias_dropout_add_fusion, 233 masked_softmax_fusion=masked_softmax_fusion, 234 persist_layer_norm=persist_layer_norm, 235 openai_gelu=openai_gelu, 236 onnx_safe=onnx_safe, 237 activation=activation, 238 bias=bias, 239 normalization=normalization, 240 transformer_block_type=transformer_block_type, 241 headscale=headscale, 242 parent_model_type=parent_model_type, 243 hidden_steps=hidden_steps, 244 num_self_attention_per_cross_attention=num_self_attention_per_cross_attention, 245 megatron_legacy=megatron_legacy, 246 normalize_attention_scores=normalize_attention_scores, 247 ) 248 else: 249 raise ValueError(f"Unknown encoder arch = {arch}. Available encoder arch = {AVAILABLE_ENCODERS}") 250 251 return encoder 252 [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/tts/models/fastpitch.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import contextlib 15 from dataclasses import dataclass 16 from pathlib import Path 17 from typing import List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import DictConfig, OmegaConf, open_dict 22 from pytorch_lightning import Trainer 23 from pytorch_lightning.loggers import TensorBoardLogger 24 25 from nemo.collections.common.parts.preprocessing import parsers 26 from nemo.collections.tts.losses.aligner_loss import BinLoss, ForwardSumLoss 27 from nemo.collections.tts.losses.fastpitchloss import DurationLoss, EnergyLoss, MelLoss, PitchLoss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.modules.fastpitch import FastPitchModule 30 from nemo.collections.tts.parts.mixins import FastPitchAdapterModelMixin 31 from nemo.collections.tts.parts.utils.callbacks import LoggingCallback 32 from nemo.collections.tts.parts.utils.helpers import ( 33 batch_from_ragged, 34 g2p_backward_compatible_support, 35 plot_alignment_to_numpy, 36 plot_spectrogram_to_numpy, 37 process_batch, 38 sample_tts_input, 39 ) 40 from nemo.core.classes import Exportable 41 from nemo.core.classes.common import PretrainedModelInfo, typecheck 42 from nemo.core.neural_types.elements import ( 43 Index, 44 LengthsType, 45 MelSpectrogramType, 46 ProbsType, 47 RegressionValuesType, 48 TokenDurationType, 49 TokenIndex, 50 TokenLogDurationType, 51 ) 52 from nemo.core.neural_types.neural_type import NeuralType 53 from nemo.utils import logging, model_utils 54 55 56 @dataclass 57 class G2PConfig: 58 _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" 59 phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" 60 heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" 61 phoneme_probability: float = 0.5 62 63 64 @dataclass 65 class TextTokenizer: 66 _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" 67 punct: bool = True 68 stresses: bool = True 69 chars: bool = True 70 apostrophe: bool = True 71 pad_with_space: bool = True 72 add_blank_at: bool = True 73 g2p: G2PConfig = G2PConfig() 74 75 76 @dataclass 77 class TextTokenizerConfig: 78 text_tokenizer: TextTokenizer = TextTokenizer() 79 80 81 class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): 82 """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" 83 84 def __init__(self, cfg: DictConfig, trainer: Trainer = None): 85 # Convert to Hydra 1.0 compatible DictConfig 86 cfg = model_utils.convert_model_config_to_dict_config(cfg) 87 cfg = model_utils.maybe_update_config_version(cfg) 88 89 # Setup normalizer 90 self.normalizer = None 91 self.text_normalizer_call = None 92 self.text_normalizer_call_kwargs = {} 93 self._setup_normalizer(cfg) 94 95 self.learn_alignment = cfg.get("learn_alignment", False) 96 97 # Setup vocabulary (=tokenizer) and input_fft_kwargs (supported only with self.learn_alignment=True) 98 input_fft_kwargs = {} 99 if self.learn_alignment: 100 self.vocab = None 101 102 self.ds_class = cfg.train_ds.dataset._target_ 103 self.ds_class_name = self.ds_class.split(".")[-1] 104 if not self.ds_class in [ 105 "nemo.collections.tts.data.dataset.TTSDataset", 106 "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset", 107 "nemo.collections.tts.torch.data.TTSDataset", 108 ]: 109 raise ValueError(f"Unknown dataset class: {self.ds_class}.") 110 111 self._setup_tokenizer(cfg) 112 assert self.vocab is not None 113 input_fft_kwargs["n_embed"] = len(self.vocab.tokens) 114 input_fft_kwargs["padding_idx"] = self.vocab.pad 115 116 self._parser = None 117 self._tb_logger = None 118 super().__init__(cfg=cfg, trainer=trainer) 119 120 self.bin_loss_warmup_epochs = cfg.get("bin_loss_warmup_epochs", 100) 121 self.log_images = cfg.get("log_images", False) 122 self.log_train_images = False 123 124 default_prosody_loss_scale = 0.1 if self.learn_alignment else 1.0 125 dur_loss_scale = cfg.get("dur_loss_scale", default_prosody_loss_scale) 126 pitch_loss_scale = cfg.get("pitch_loss_scale", default_prosody_loss_scale) 127 energy_loss_scale = cfg.get("energy_loss_scale", default_prosody_loss_scale) 128 129 self.mel_loss_fn = MelLoss() 130 self.pitch_loss_fn = PitchLoss(loss_scale=pitch_loss_scale) 131 self.duration_loss_fn = DurationLoss(loss_scale=dur_loss_scale) 132 self.energy_loss_fn = EnergyLoss(loss_scale=energy_loss_scale) 133 134 self.aligner = None 135 if self.learn_alignment: 136 aligner_loss_scale = cfg.get("aligner_loss_scale", 1.0) 137 self.aligner = instantiate(self._cfg.alignment_module) 138 self.forward_sum_loss_fn = ForwardSumLoss(loss_scale=aligner_loss_scale) 139 self.bin_loss_fn = BinLoss(loss_scale=aligner_loss_scale) 140 141 self.preprocessor = instantiate(self._cfg.preprocessor) 142 input_fft = instantiate(self._cfg.input_fft, **input_fft_kwargs) 143 output_fft = instantiate(self._cfg.output_fft) 144 duration_predictor = instantiate(self._cfg.duration_predictor) 145 pitch_predictor = instantiate(self._cfg.pitch_predictor) 146 speaker_encoder = instantiate(self._cfg.get("speaker_encoder", None)) 147 energy_embedding_kernel_size = cfg.get("energy_embedding_kernel_size", 0) 148 energy_predictor = instantiate(self._cfg.get("energy_predictor", None)) 149 150 # [TODO] may remove if we change the pre-trained config 151 # cfg: condition_types = [ "add" ] 152 n_speakers = cfg.get("n_speakers", 0) 153 speaker_emb_condition_prosody = cfg.get("speaker_emb_condition_prosody", False) 154 speaker_emb_condition_decoder = cfg.get("speaker_emb_condition_decoder", False) 155 speaker_emb_condition_aligner = cfg.get("speaker_emb_condition_aligner", False) 156 min_token_duration = cfg.get("min_token_duration", 0) 157 use_log_energy = cfg.get("use_log_energy", True) 158 if n_speakers > 1 and "add" not in input_fft.cond_input.condition_types: 159 input_fft.cond_input.condition_types.append("add") 160 if speaker_emb_condition_prosody: 161 duration_predictor.cond_input.condition_types.append("add") 162 pitch_predictor.cond_input.condition_types.append("add") 163 if speaker_emb_condition_decoder: 164 output_fft.cond_input.condition_types.append("add") 165 if speaker_emb_condition_aligner and self.aligner is not None: 166 self.aligner.cond_input.condition_types.append("add") 167 168 self.fastpitch = FastPitchModule( 169 input_fft, 170 output_fft, 171 duration_predictor, 172 pitch_predictor, 173 energy_predictor, 174 self.aligner, 175 speaker_encoder, 176 n_speakers, 177 cfg.symbols_embedding_dim, 178 cfg.pitch_embedding_kernel_size, 179 energy_embedding_kernel_size, 180 cfg.n_mel_channels, 181 min_token_duration, 182 cfg.max_token_duration, 183 use_log_energy, 184 ) 185 self._input_types = self._output_types = None 186 self.export_config = { 187 "emb_range": (0, self.fastpitch.encoder.word_emb.num_embeddings), 188 "enable_volume": False, 189 "enable_ragged_batches": False, 190 } 191 if self.fastpitch.speaker_emb is not None: 192 self.export_config["num_speakers"] = cfg.n_speakers 193 194 self.log_config = cfg.get("log_config", None) 195 196 # Adapter modules setup (from FastPitchAdapterModelMixin) 197 self.setup_adapters() 198 199 def _get_default_text_tokenizer_conf(self): 200 text_tokenizer: TextTokenizerConfig = TextTokenizerConfig() 201 return OmegaConf.create(OmegaConf.to_yaml(text_tokenizer)) 202 203 def _setup_normalizer(self, cfg): 204 if "text_normalizer" in cfg: 205 normalizer_kwargs = {} 206 207 if "whitelist" in cfg.text_normalizer: 208 normalizer_kwargs["whitelist"] = self.register_artifact( 209 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 210 ) 211 try: 212 import nemo_text_processing 213 214 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 215 except Exception as e: 216 logging.error(e) 217 raise ImportError( 218 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 219 ) 220 221 self.text_normalizer_call = self.normalizer.normalize 222 if "text_normalizer_call_kwargs" in cfg: 223 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 224 225 def _setup_tokenizer(self, cfg): 226 text_tokenizer_kwargs = {} 227 228 if "g2p" in cfg.text_tokenizer: 229 # for backward compatibility 230 if ( 231 self._is_model_being_restored() 232 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 233 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 234 ): 235 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 236 cfg.text_tokenizer.g2p["_target_"] 237 ) 238 239 g2p_kwargs = {} 240 241 if "phoneme_dict" in cfg.text_tokenizer.g2p: 242 g2p_kwargs["phoneme_dict"] = self.register_artifact( 243 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 244 ) 245 246 if "heteronyms" in cfg.text_tokenizer.g2p: 247 g2p_kwargs["heteronyms"] = self.register_artifact( 248 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 249 ) 250 251 # for backward compatability 252 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 253 254 # TODO @xueyang: rename the instance of tokenizer because vocab is misleading. 255 self.vocab = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 256 257 @property 258 def tb_logger(self): 259 if self._tb_logger is None: 260 if self.logger is None and self.logger.experiment is None: 261 return None 262 tb_logger = self.logger.experiment 263 for logger in self.trainer.loggers: 264 if isinstance(logger, TensorBoardLogger): 265 tb_logger = logger.experiment 266 break 267 self._tb_logger = tb_logger 268 return self._tb_logger 269 270 @property 271 def parser(self): 272 if self._parser is not None: 273 return self._parser 274 275 if self.learn_alignment: 276 self._parser = self.vocab.encode 277 else: 278 self._parser = parsers.make_parser( 279 labels=self._cfg.labels, 280 name='en', 281 unk_id=-1, 282 blank_id=-1, 283 do_normalize=True, 284 abbreviation_version="fastpitch", 285 make_table=False, 286 ) 287 return self._parser 288 289 def parse(self, str_input: str, normalize=True) -> torch.tensor: 290 if self.training: 291 logging.warning("parse() is meant to be called in eval mode.") 292 293 if normalize and self.text_normalizer_call is not None: 294 str_input = self.text_normalizer_call(str_input, **self.text_normalizer_call_kwargs) 295 296 if self.learn_alignment: 297 eval_phon_mode = contextlib.nullcontext() 298 if hasattr(self.vocab, "set_phone_prob"): 299 eval_phon_mode = self.vocab.set_phone_prob(prob=1.0) 300 301 # Disable mixed g2p representation if necessary 302 with eval_phon_mode: 303 tokens = self.parser(str_input) 304 else: 305 tokens = self.parser(str_input) 306 307 x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device) 308 return x 309 310 @typecheck( 311 input_types={ 312 "text": NeuralType(('B', 'T_text'), TokenIndex()), 313 "durs": NeuralType(('B', 'T_text'), TokenDurationType()), 314 "pitch": NeuralType(('B', 'T_audio'), RegressionValuesType()), 315 "energy": NeuralType(('B', 'T_audio'), RegressionValuesType(), optional=True), 316 "speaker": NeuralType(('B'), Index(), optional=True), 317 "pace": NeuralType(optional=True), 318 "spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 319 "attn_prior": NeuralType(('B', 'T_spec', 'T_text'), ProbsType(), optional=True), 320 "mel_lens": NeuralType(('B'), LengthsType(), optional=True), 321 "input_lens": NeuralType(('B'), LengthsType(), optional=True), 322 # reference_* data is used for multi-speaker FastPitch training 323 "reference_spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 324 "reference_spec_lens": NeuralType(('B'), LengthsType(), optional=True), 325 } 326 ) 327 def forward( 328 self, 329 *, 330 text, 331 durs=None, 332 pitch=None, 333 energy=None, 334 speaker=None, 335 pace=1.0, 336 spec=None, 337 attn_prior=None, 338 mel_lens=None, 339 input_lens=None, 340 reference_spec=None, 341 reference_spec_lens=None, 342 ): 343 return self.fastpitch( 344 text=text, 345 durs=durs, 346 pitch=pitch, 347 energy=energy, 348 speaker=speaker, 349 pace=pace, 350 spec=spec, 351 attn_prior=attn_prior, 352 mel_lens=mel_lens, 353 input_lens=input_lens, 354 reference_spec=reference_spec, 355 reference_spec_lens=reference_spec_lens, 356 ) 357 358 @typecheck(output_types={"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType())}) 359 def generate_spectrogram( 360 self, 361 tokens: 'torch.tensor', 362 speaker: Optional[int] = None, 363 pace: float = 1.0, 364 reference_spec: Optional['torch.tensor'] = None, 365 reference_spec_lens: Optional['torch.tensor'] = None, 366 ) -> torch.tensor: 367 if self.training: 368 logging.warning("generate_spectrogram() is meant to be called in eval mode.") 369 if isinstance(speaker, int): 370 speaker = torch.tensor([speaker]).to(self.device) 371 spect, *_ = self( 372 text=tokens, 373 durs=None, 374 pitch=None, 375 speaker=speaker, 376 pace=pace, 377 reference_spec=reference_spec, 378 reference_spec_lens=reference_spec_lens, 379 ) 380 return spect 381 382 def training_step(self, batch, batch_idx): 383 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 384 None, 385 None, 386 None, 387 None, 388 None, 389 None, 390 ) 391 if self.learn_alignment: 392 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 393 batch_dict = batch 394 else: 395 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 396 audio = batch_dict.get("audio") 397 audio_lens = batch_dict.get("audio_lens") 398 text = batch_dict.get("text") 399 text_lens = batch_dict.get("text_lens") 400 attn_prior = batch_dict.get("align_prior_matrix", None) 401 pitch = batch_dict.get("pitch", None) 402 energy = batch_dict.get("energy", None) 403 speaker = batch_dict.get("speaker_id", None) 404 reference_audio = batch_dict.get("reference_audio", None) 405 reference_audio_len = batch_dict.get("reference_audio_lens", None) 406 else: 407 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 408 409 mels, spec_len = self.preprocessor(input_signal=audio, length=audio_lens) 410 reference_spec, reference_spec_len = None, None 411 if reference_audio is not None: 412 reference_spec, reference_spec_len = self.preprocessor( 413 input_signal=reference_audio, length=reference_audio_len 414 ) 415 416 ( 417 mels_pred, 418 _, 419 _, 420 log_durs_pred, 421 pitch_pred, 422 attn_soft, 423 attn_logprob, 424 attn_hard, 425 attn_hard_dur, 426 pitch, 427 energy_pred, 428 energy_tgt, 429 ) = self( 430 text=text, 431 durs=durs, 432 pitch=pitch, 433 energy=energy, 434 speaker=speaker, 435 pace=1.0, 436 spec=mels if self.learn_alignment else None, 437 reference_spec=reference_spec, 438 reference_spec_lens=reference_spec_len, 439 attn_prior=attn_prior, 440 mel_lens=spec_len, 441 input_lens=text_lens, 442 ) 443 if durs is None: 444 durs = attn_hard_dur 445 446 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 447 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 448 loss = mel_loss + dur_loss 449 if self.learn_alignment: 450 ctc_loss = self.forward_sum_loss_fn(attn_logprob=attn_logprob, in_lens=text_lens, out_lens=spec_len) 451 bin_loss_weight = min(self.current_epoch / self.bin_loss_warmup_epochs, 1.0) * 1.0 452 bin_loss = self.bin_loss_fn(hard_attention=attn_hard, soft_attention=attn_soft) * bin_loss_weight 453 loss += ctc_loss + bin_loss 454 455 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 456 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 457 loss += pitch_loss + energy_loss 458 459 self.log("t_loss", loss) 460 self.log("t_mel_loss", mel_loss) 461 self.log("t_dur_loss", dur_loss) 462 self.log("t_pitch_loss", pitch_loss) 463 if energy_tgt is not None: 464 self.log("t_energy_loss", energy_loss) 465 if self.learn_alignment: 466 self.log("t_ctc_loss", ctc_loss) 467 self.log("t_bin_loss", bin_loss) 468 469 # Log images to tensorboard 470 if self.log_images and self.log_train_images and isinstance(self.logger, TensorBoardLogger): 471 self.log_train_images = False 472 473 self.tb_logger.add_image( 474 "train_mel_target", 475 plot_spectrogram_to_numpy(mels[0].data.cpu().float().numpy()), 476 self.global_step, 477 dataformats="HWC", 478 ) 479 spec_predict = mels_pred[0].data.cpu().float().numpy() 480 self.tb_logger.add_image( 481 "train_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 482 ) 483 if self.learn_alignment: 484 attn = attn_hard[0].data.cpu().float().numpy().squeeze() 485 self.tb_logger.add_image( 486 "train_attn", plot_alignment_to_numpy(attn.T), self.global_step, dataformats="HWC", 487 ) 488 soft_attn = attn_soft[0].data.cpu().float().numpy().squeeze() 489 self.tb_logger.add_image( 490 "train_soft_attn", plot_alignment_to_numpy(soft_attn.T), self.global_step, dataformats="HWC", 491 ) 492 493 return loss 494 495 def validation_step(self, batch, batch_idx): 496 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 497 None, 498 None, 499 None, 500 None, 501 None, 502 None, 503 ) 504 if self.learn_alignment: 505 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 506 batch_dict = batch 507 else: 508 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 509 audio = batch_dict.get("audio") 510 audio_lens = batch_dict.get("audio_lens") 511 text = batch_dict.get("text") 512 text_lens = batch_dict.get("text_lens") 513 attn_prior = batch_dict.get("align_prior_matrix", None) 514 pitch = batch_dict.get("pitch", None) 515 energy = batch_dict.get("energy", None) 516 speaker = batch_dict.get("speaker_id", None) 517 reference_audio = batch_dict.get("reference_audio", None) 518 reference_audio_len = batch_dict.get("reference_audio_lens", None) 519 else: 520 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 521 522 mels, mel_lens = self.preprocessor(input_signal=audio, length=audio_lens) 523 reference_spec, reference_spec_len = None, None 524 if reference_audio is not None: 525 reference_spec, reference_spec_len = self.preprocessor( 526 input_signal=reference_audio, length=reference_audio_len 527 ) 528 529 # Calculate val loss on ground truth durations to better align L2 loss in time 530 (mels_pred, _, _, log_durs_pred, pitch_pred, _, _, _, attn_hard_dur, pitch, energy_pred, energy_tgt,) = self( 531 text=text, 532 durs=durs, 533 pitch=pitch, 534 energy=energy, 535 speaker=speaker, 536 pace=1.0, 537 spec=mels if self.learn_alignment else None, 538 reference_spec=reference_spec, 539 reference_spec_lens=reference_spec_len, 540 attn_prior=attn_prior, 541 mel_lens=mel_lens, 542 input_lens=text_lens, 543 ) 544 if durs is None: 545 durs = attn_hard_dur 546 547 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 548 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 549 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 550 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 551 loss = mel_loss + dur_loss + pitch_loss + energy_loss 552 553 val_outputs = { 554 "val_loss": loss, 555 "mel_loss": mel_loss, 556 "dur_loss": dur_loss, 557 "pitch_loss": pitch_loss, 558 "energy_loss": energy_loss if energy_tgt is not None else None, 559 "mel_target": mels if batch_idx == 0 else None, 560 "mel_pred": mels_pred if batch_idx == 0 else None, 561 } 562 self.validation_step_outputs.append(val_outputs) 563 return val_outputs 564 565 def on_validation_epoch_end(self): 566 collect = lambda key: torch.stack([x[key] for x in self.validation_step_outputs]).mean() 567 val_loss = collect("val_loss") 568 mel_loss = collect("mel_loss") 569 dur_loss = collect("dur_loss") 570 pitch_loss = collect("pitch_loss") 571 self.log("val_loss", val_loss, sync_dist=True) 572 self.log("val_mel_loss", mel_loss, sync_dist=True) 573 self.log("val_dur_loss", dur_loss, sync_dist=True) 574 self.log("val_pitch_loss", pitch_loss, sync_dist=True) 575 if self.validation_step_outputs[0]["energy_loss"] is not None: 576 energy_loss = collect("energy_loss") 577 self.log("val_energy_loss", energy_loss, sync_dist=True) 578 579 _, _, _, _, _, spec_target, spec_predict = self.validation_step_outputs[0].values() 580 581 if self.log_images and isinstance(self.logger, TensorBoardLogger): 582 self.tb_logger.add_image( 583 "val_mel_target", 584 plot_spectrogram_to_numpy(spec_target[0].data.cpu().float().numpy()), 585 self.global_step, 586 dataformats="HWC", 587 ) 588 spec_predict = spec_predict[0].data.cpu().float().numpy() 589 self.tb_logger.add_image( 590 "val_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 591 ) 592 self.log_train_images = True 593 self.validation_step_outputs.clear() # free memory) 594 595 def _setup_train_dataloader(self, cfg): 596 phon_mode = contextlib.nullcontext() 597 if hasattr(self.vocab, "set_phone_prob"): 598 phon_mode = self.vocab.set_phone_prob(self.vocab.phoneme_probability) 599 600 with phon_mode: 601 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 602 603 sampler = dataset.get_sampler(cfg.dataloader_params.batch_size) 604 return torch.utils.data.DataLoader( 605 dataset, collate_fn=dataset.collate_fn, sampler=sampler, **cfg.dataloader_params 606 ) 607 608 def _setup_test_dataloader(self, cfg): 609 phon_mode = contextlib.nullcontext() 610 if hasattr(self.vocab, "set_phone_prob"): 611 phon_mode = self.vocab.set_phone_prob(0.0) 612 613 with phon_mode: 614 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 615 616 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 617 618 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 619 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 620 raise ValueError(f"No dataset for {name}") 621 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 622 raise ValueError(f"No dataloader_params for {name}") 623 if shuffle_should_be: 624 if 'shuffle' not in cfg.dataloader_params: 625 logging.warning( 626 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 627 "config. Manually setting to True" 628 ) 629 with open_dict(cfg.dataloader_params): 630 cfg.dataloader_params.shuffle = True 631 elif not cfg.dataloader_params.shuffle: 632 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 633 elif cfg.dataloader_params.shuffle: 634 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 635 636 if self.ds_class == "nemo.collections.tts.data.dataset.TTSDataset": 637 phon_mode = contextlib.nullcontext() 638 if hasattr(self.vocab, "set_phone_prob"): 639 phon_mode = self.vocab.set_phone_prob(prob=None if name == "val" else self.vocab.phoneme_probability) 640 641 with phon_mode: 642 dataset = instantiate( 643 cfg.dataset, 644 text_normalizer=self.normalizer, 645 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 646 text_tokenizer=self.vocab, 647 ) 648 else: 649 dataset = instantiate(cfg.dataset) 650 651 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 652 653 def setup_training_data(self, cfg): 654 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 655 self._train_dl = self._setup_train_dataloader(cfg) 656 else: 657 self._train_dl = self.__setup_dataloader_from_config(cfg) 658 659 def setup_validation_data(self, cfg): 660 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 661 self._validation_dl = self._setup_test_dataloader(cfg) 662 else: 663 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="val") 664 665 def setup_test_data(self, cfg): 666 """Omitted.""" 667 pass 668 669 def configure_callbacks(self): 670 if not self.log_config: 671 return [] 672 673 sample_ds_class = self.log_config.dataset._target_ 674 if sample_ds_class != "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 675 raise ValueError(f"Logging callback only supported for TextToSpeechDataset, got {sample_ds_class}") 676 677 data_loader = self._setup_test_dataloader(self.log_config) 678 679 generators = instantiate(self.log_config.generators) 680 log_dir = Path(self.log_config.log_dir) if self.log_config.log_dir else None 681 log_callback = LoggingCallback( 682 generators=generators, 683 data_loader=data_loader, 684 log_epochs=self.log_config.log_epochs, 685 epoch_frequency=self.log_config.epoch_frequency, 686 output_dir=log_dir, 687 loggers=self.trainer.loggers, 688 log_tensorboard=self.log_config.log_tensorboard, 689 log_wandb=self.log_config.log_wandb, 690 ) 691 692 return [log_callback] 693 694 @classmethod 695 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 696 """ 697 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 698 Returns: 699 List of available pre-trained models. 700 """ 701 list_of_models = [] 702 703 # en-US, single speaker, 22050Hz, LJSpeech (ARPABET). 704 model = PretrainedModelInfo( 705 pretrained_model_name="tts_en_fastpitch", 706 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/1.8.1/files/tts_en_fastpitch_align.nemo", 707 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is ARPABET-based.", 708 class_=cls, 709 ) 710 list_of_models.append(model) 711 712 # en-US, single speaker, 22050Hz, LJSpeech (IPA). 713 model = PretrainedModelInfo( 714 pretrained_model_name="tts_en_fastpitch_ipa", 715 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/IPA_1.13.0/files/tts_en_fastpitch_align_ipa.nemo", 716 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is IPA-based.", 717 class_=cls, 718 ) 719 list_of_models.append(model) 720 721 # en-US, multi-speaker, 44100Hz, HiFiTTS. 722 model = PretrainedModelInfo( 723 pretrained_model_name="tts_en_fastpitch_multispeaker", 724 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_multispeaker_fastpitchhifigan/versions/1.10.0/files/tts_en_fastpitch_multispeaker.nemo", 725 description="This model is trained on HiFITTS sampled at 44100Hz with and can be used to generate male and female English voices with an American accent.", 726 class_=cls, 727 ) 728 list_of_models.append(model) 729 730 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 21.02 731 model = PretrainedModelInfo( 732 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2102", 733 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2102.nemo", 734 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 21.02 Dataset sampled at 22050Hz and can be used to generate male German voices.", 735 class_=cls, 736 ) 737 list_of_models.append(model) 738 739 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 22.10 740 model = PretrainedModelInfo( 741 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2210", 742 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2210.nemo", 743 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 22.10 Dataset sampled at 22050Hz and can be used to generate male German voices.", 744 class_=cls, 745 ) 746 list_of_models.append(model) 747 748 # de-DE, multi-speaker, 5 speakers, 44100 Hz, HUI-Audio-Corpus-German Clean. 749 model = PretrainedModelInfo( 750 pretrained_model_name="tts_de_fastpitch_multispeaker_5", 751 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitch_multispeaker_5/versions/1.11.0/files/tts_de_fastpitch_multispeaker_5.nemo", 752 description="This model is trained on 5 speakers in HUI-Audio-Corpus-German clean subset sampled at 44100Hz with and can be used to generate male and female German voices.", 753 class_=cls, 754 ) 755 list_of_models.append(model) 756 757 # es, 174 speakers, 44100Hz, OpenSLR (IPA) 758 model = PretrainedModelInfo( 759 pretrained_model_name="tts_es_fastpitch_multispeaker", 760 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_es_multispeaker_fastpitchhifigan/versions/1.15.0/files/tts_es_fastpitch_multispeaker.nemo", 761 description="This model is trained on 174 speakers in 6 crowdsourced Latin American Spanish OpenSLR datasets sampled at 44100Hz and can be used to generate male and female Spanish voices with Latin American accents.", 762 class_=cls, 763 ) 764 list_of_models.append(model) 765 766 # zh, single female speaker, 22050Hz, SFSpeech Bilingual Chinese/English dataset, improved model using richer 767 # dict and jieba word segmenter for polyphone disambiguation. 768 model = PretrainedModelInfo( 769 pretrained_model_name="tts_zh_fastpitch_sfspeech", 770 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_zh_fastpitch_hifigan_sfspeech/versions/1.15.0/files/tts_zh_fastpitch_sfspeech.nemo", 771 description="This model is trained on a single female speaker in SFSpeech Bilingual Chinese/English dataset" 772 " sampled at 22050Hz and can be used to generate female Mandarin Chinese voices. It is improved" 773 " using richer dict and jieba word segmenter for polyphone disambiguation.", 774 class_=cls, 775 ) 776 list_of_models.append(model) 777 778 # en, multi speaker, LibriTTS, 16000 Hz 779 # stft 25ms 10ms matching ASR params 780 # for use during Enhlish ASR training/adaptation 781 model = PretrainedModelInfo( 782 pretrained_model_name="tts_en_fastpitch_for_asr_finetuning", 783 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch_spectrogram_enhancer_for_asr_finetuning/versions/1.20.0/files/tts_en_fastpitch_for_asr_finetuning.nemo", 784 description="This model is trained on LibriSpeech, train-960 subset." 785 " STFT parameters follow those commonly used in ASR: 25 ms window, 10 ms hop." 786 " This model is supposed to be used with its companion SpetrogramEnhancer for " 787 " ASR fine-tuning. Usage for regular TTS tasks is not advised.", 788 class_=cls, 789 ) 790 list_of_models.append(model) 791 792 return list_of_models 793 794 # Methods for model exportability 795 def _prepare_for_export(self, **kwargs): 796 super()._prepare_for_export(**kwargs) 797 798 tensor_shape = ('T') if self.export_config["enable_ragged_batches"] else ('B', 'T') 799 800 # Define input_types and output_types as required by export() 801 self._input_types = { 802 "text": NeuralType(tensor_shape, TokenIndex()), 803 "pitch": NeuralType(tensor_shape, RegressionValuesType()), 804 "pace": NeuralType(tensor_shape), 805 "volume": NeuralType(tensor_shape, optional=True), 806 "batch_lengths": NeuralType(('B'), optional=True), 807 "speaker": NeuralType(('B'), Index(), optional=True), 808 } 809 self._output_types = { 810 "spect": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 811 "num_frames": NeuralType(('B'), TokenDurationType()), 812 "durs_predicted": NeuralType(('B', 'T'), TokenDurationType()), 813 "log_durs_predicted": NeuralType(('B', 'T'), TokenLogDurationType()), 814 "pitch_predicted": NeuralType(('B', 'T'), RegressionValuesType()), 815 } 816 if self.export_config["enable_volume"]: 817 self._output_types["volume_aligned"] = NeuralType(('B', 'T'), RegressionValuesType()) 818 819 def _export_teardown(self): 820 self._input_types = self._output_types = None 821 822 @property 823 def disabled_deployment_input_names(self): 824 """Implement this method to return a set of input names disabled for export""" 825 disabled_inputs = set() 826 if self.fastpitch.speaker_emb is None: 827 disabled_inputs.add("speaker") 828 if not self.export_config["enable_ragged_batches"]: 829 disabled_inputs.add("batch_lengths") 830 if not self.export_config["enable_volume"]: 831 disabled_inputs.add("volume") 832 return disabled_inputs 833 834 @property 835 def input_types(self): 836 return self._input_types 837 838 @property 839 def output_types(self): 840 return self._output_types 841 842 def input_example(self, max_batch=1, max_dim=44): 843 """ 844 Generates input examples for tracing etc. 845 Returns: 846 A tuple of input examples. 847 """ 848 par = next(self.fastpitch.parameters()) 849 inputs = sample_tts_input(self.export_config, par.device, max_batch=max_batch, max_dim=max_dim) 850 if 'enable_ragged_batches' not in self.export_config: 851 inputs.pop('batch_lengths', None) 852 return (inputs,) 853 854 def forward_for_export(self, text, pitch, pace, volume=None, batch_lengths=None, speaker=None): 855 if self.export_config["enable_ragged_batches"]: 856 text, pitch, pace, volume_tensor, lens = batch_from_ragged( 857 text, pitch, pace, batch_lengths, padding_idx=self.fastpitch.encoder.padding_idx, volume=volume 858 ) 859 if volume is not None: 860 volume = volume_tensor 861 return self.fastpitch.infer(text=text, pitch=pitch, pace=pace, volume=volume, speaker=speaker) 862 863 def interpolate_speaker( 864 self, original_speaker_1, original_speaker_2, weight_speaker_1, weight_speaker_2, new_speaker_id 865 ): 866 """ 867 This method performs speaker interpolation between two original speakers the model is trained on. 868 869 Inputs: 870 original_speaker_1: Integer speaker ID of first existing speaker in the model 871 original_speaker_2: Integer speaker ID of second existing speaker in the model 872 weight_speaker_1: Floating point weight associated in to first speaker during weight combination 873 weight_speaker_2: Floating point weight associated in to second speaker during weight combination 874 new_speaker_id: Integer speaker ID of new interpolated speaker in the model 875 """ 876 if self.fastpitch.speaker_emb is None: 877 raise Exception( 878 "Current FastPitch model is not a multi-speaker FastPitch model. Speaker interpolation can only \ 879 be performed with a multi-speaker model" 880 ) 881 n_speakers = self.fastpitch.speaker_emb.weight.data.size()[0] 882 if original_speaker_1 >= n_speakers or original_speaker_2 >= n_speakers or new_speaker_id >= n_speakers: 883 raise Exception( 884 f"Parameters original_speaker_1, original_speaker_2, new_speaker_id should be less than the total \ 885 total number of speakers FastPitch was trained on (n_speakers = {n_speakers})." 886 ) 887 speaker_emb_1 = ( 888 self.fastpitch.speaker_emb(torch.tensor(original_speaker_1, dtype=torch.int32).cuda()).clone().detach() 889 ) 890 speaker_emb_2 = ( 891 self.fastpitch.speaker_emb(torch.tensor(original_speaker_2, dtype=torch.int32).cuda()).clone().detach() 892 ) 893 new_speaker_emb = weight_speaker_1 * speaker_emb_1 + weight_speaker_2 * speaker_emb_2 894 self.fastpitch.speaker_emb.weight.data[new_speaker_id] = new_speaker_emb 895 [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/tts/models/tacotron2.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import contextlib 16 from dataclasses import dataclass 17 from typing import Any, Dict, List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import MISSING, DictConfig, OmegaConf, open_dict 22 from omegaconf.errors import ConfigAttributeError 23 from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger 24 from torch import nn 25 26 from nemo.collections.common.parts.preprocessing import parsers 27 from nemo.collections.tts.losses.tacotron2loss import Tacotron2Loss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.parts.utils.helpers import ( 30 g2p_backward_compatible_support, 31 get_mask_from_lengths, 32 tacotron2_log_to_tb_func, 33 tacotron2_log_to_wandb_func, 34 ) 35 from nemo.core.classes.common import PretrainedModelInfo, typecheck 36 from nemo.core.neural_types.elements import ( 37 AudioSignal, 38 EmbeddedTextType, 39 LengthsType, 40 LogitsType, 41 MelSpectrogramType, 42 SequenceToSequenceAlignmentType, 43 ) 44 from nemo.core.neural_types.neural_type import NeuralType 45 from nemo.utils import logging, model_utils 46 47 48 @dataclass 49 class Preprocessor: 50 _target_: str = MISSING 51 pad_value: float = MISSING 52 53 54 @dataclass 55 class Tacotron2Config: 56 preprocessor: Preprocessor = Preprocessor() 57 encoder: Dict[Any, Any] = MISSING 58 decoder: Dict[Any, Any] = MISSING 59 postnet: Dict[Any, Any] = MISSING 60 labels: List = MISSING 61 train_ds: Optional[Dict[Any, Any]] = None 62 validation_ds: Optional[Dict[Any, Any]] = None 63 64 65 class Tacotron2Model(SpectrogramGenerator): 66 """Tacotron 2 Model that is used to generate mel spectrograms from text""" 67 68 def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None): 69 # Convert to Hydra 1.0 compatible DictConfig 70 cfg = model_utils.convert_model_config_to_dict_config(cfg) 71 cfg = model_utils.maybe_update_config_version(cfg) 72 73 # setup normalizer 74 self.normalizer = None 75 self.text_normalizer_call = None 76 self.text_normalizer_call_kwargs = {} 77 self._setup_normalizer(cfg) 78 79 # setup tokenizer 80 self.tokenizer = None 81 if hasattr(cfg, 'text_tokenizer'): 82 self._setup_tokenizer(cfg) 83 84 self.num_tokens = len(self.tokenizer.tokens) 85 self.tokenizer_pad = self.tokenizer.pad 86 self.tokenizer_unk = self.tokenizer.oov 87 # assert self.tokenizer is not None 88 else: 89 self.num_tokens = len(cfg.labels) + 3 90 91 super().__init__(cfg=cfg, trainer=trainer) 92 93 schema = OmegaConf.structured(Tacotron2Config) 94 # ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes 95 if isinstance(cfg, dict): 96 cfg = OmegaConf.create(cfg) 97 elif not isinstance(cfg, DictConfig): 98 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 99 # Ensure passed cfg is compliant with schema 100 try: 101 OmegaConf.merge(cfg, schema) 102 self.pad_value = cfg.preprocessor.pad_value 103 except ConfigAttributeError: 104 self.pad_value = cfg.preprocessor.params.pad_value 105 logging.warning( 106 "Your config is using an old NeMo yaml configuration. Please ensure that the yaml matches the " 107 "current version in the main branch for future compatibility." 108 ) 109 110 self._parser = None 111 self.audio_to_melspec_precessor = instantiate(cfg.preprocessor) 112 self.text_embedding = nn.Embedding(self.num_tokens, 512) 113 self.encoder = instantiate(self._cfg.encoder) 114 self.decoder = instantiate(self._cfg.decoder) 115 self.postnet = instantiate(self._cfg.postnet) 116 self.loss = Tacotron2Loss() 117 self.calculate_loss = True 118 119 @property 120 def parser(self): 121 if self._parser is not None: 122 return self._parser 123 124 ds_class_name = self._cfg.train_ds.dataset._target_.split(".")[-1] 125 if ds_class_name == "TTSDataset": 126 self._parser = None 127 elif hasattr(self._cfg, "labels"): 128 self._parser = parsers.make_parser( 129 labels=self._cfg.labels, 130 name='en', 131 unk_id=-1, 132 blank_id=-1, 133 do_normalize=True, 134 abbreviation_version="fastpitch", 135 make_table=False, 136 ) 137 else: 138 raise ValueError("Wanted to setup parser, but model does not have necessary paramaters") 139 140 return self._parser 141 142 def parse(self, text: str, normalize=True) -> torch.Tensor: 143 if self.training: 144 logging.warning("parse() is meant to be called in eval mode.") 145 if normalize and self.text_normalizer_call is not None: 146 text = self.text_normalizer_call(text, **self.text_normalizer_call_kwargs) 147 148 eval_phon_mode = contextlib.nullcontext() 149 if hasattr(self.tokenizer, "set_phone_prob"): 150 eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0) 151 152 with eval_phon_mode: 153 if self.tokenizer is not None: 154 tokens = self.tokenizer.encode(text) 155 else: 156 tokens = self.parser(text) 157 # Old parser doesn't add bos and eos ids, so maunally add it 158 tokens = [len(self._cfg.labels)] + tokens + [len(self._cfg.labels) + 1] 159 tokens_tensor = torch.tensor(tokens).unsqueeze_(0).to(self.device) 160 return tokens_tensor 161 162 @property 163 def input_types(self): 164 if self.training: 165 return { 166 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 167 "token_len": NeuralType(('B'), LengthsType()), 168 "audio": NeuralType(('B', 'T'), AudioSignal()), 169 "audio_len": NeuralType(('B'), LengthsType()), 170 } 171 else: 172 return { 173 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 174 "token_len": NeuralType(('B'), LengthsType()), 175 "audio": NeuralType(('B', 'T'), AudioSignal(), optional=True), 176 "audio_len": NeuralType(('B'), LengthsType(), optional=True), 177 } 178 179 @property 180 def output_types(self): 181 if not self.calculate_loss and not self.training: 182 return { 183 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 184 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 185 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 186 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 187 "pred_length": NeuralType(('B'), LengthsType()), 188 } 189 return { 190 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 191 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 192 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 193 "spec_target": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 194 "spec_target_len": NeuralType(('B'), LengthsType()), 195 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 196 } 197 198 @typecheck() 199 def forward(self, *, tokens, token_len, audio=None, audio_len=None): 200 if audio is not None and audio_len is not None: 201 spec_target, spec_target_len = self.audio_to_melspec_precessor(audio, audio_len) 202 else: 203 if self.training or self.calculate_loss: 204 raise ValueError( 205 f"'audio' and 'audio_len' can not be None when either 'self.training' or 'self.calculate_loss' is True." 206 ) 207 208 token_embedding = self.text_embedding(tokens).transpose(1, 2) 209 encoder_embedding = self.encoder(token_embedding=token_embedding, token_len=token_len) 210 211 if self.training: 212 spec_pred_dec, gate_pred, alignments = self.decoder( 213 memory=encoder_embedding, decoder_inputs=spec_target, memory_lengths=token_len 214 ) 215 else: 216 spec_pred_dec, gate_pred, alignments, pred_length = self.decoder( 217 memory=encoder_embedding, memory_lengths=token_len 218 ) 219 220 spec_pred_postnet = self.postnet(mel_spec=spec_pred_dec) 221 222 if not self.calculate_loss and not self.training: 223 return spec_pred_dec, spec_pred_postnet, gate_pred, alignments, pred_length 224 225 return spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments 226 227 @typecheck( 228 input_types={"tokens": NeuralType(('B', 'T'), EmbeddedTextType())}, 229 output_types={"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType())}, 230 ) 231 def generate_spectrogram(self, *, tokens): 232 self.eval() 233 self.calculate_loss = False 234 token_len = torch.tensor([len(i) for i in tokens]).to(self.device) 235 tensors = self(tokens=tokens, token_len=token_len) 236 spectrogram_pred = tensors[1] 237 238 if spectrogram_pred.shape[0] > 1: 239 # Silence all frames past the predicted end 240 mask = ~get_mask_from_lengths(tensors[-1]) 241 mask = mask.expand(spectrogram_pred.shape[1], mask.size(0), mask.size(1)) 242 mask = mask.permute(1, 0, 2) 243 spectrogram_pred.data.masked_fill_(mask, self.pad_value) 244 245 return spectrogram_pred 246 247 def training_step(self, batch, batch_idx): 248 audio, audio_len, tokens, token_len = batch 249 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, _ = self.forward( 250 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 251 ) 252 253 loss, _ = self.loss( 254 spec_pred_dec=spec_pred_dec, 255 spec_pred_postnet=spec_pred_postnet, 256 gate_pred=gate_pred, 257 spec_target=spec_target, 258 spec_target_len=spec_target_len, 259 pad_value=self.pad_value, 260 ) 261 262 output = { 263 'loss': loss, 264 'progress_bar': {'training_loss': loss}, 265 'log': {'loss': loss}, 266 } 267 return output 268 269 def validation_step(self, batch, batch_idx): 270 audio, audio_len, tokens, token_len = batch 271 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments = self.forward( 272 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 273 ) 274 275 loss, gate_target = self.loss( 276 spec_pred_dec=spec_pred_dec, 277 spec_pred_postnet=spec_pred_postnet, 278 gate_pred=gate_pred, 279 spec_target=spec_target, 280 spec_target_len=spec_target_len, 281 pad_value=self.pad_value, 282 ) 283 loss = { 284 "val_loss": loss, 285 "mel_target": spec_target, 286 "mel_postnet": spec_pred_postnet, 287 "gate": gate_pred, 288 "gate_target": gate_target, 289 "alignments": alignments, 290 } 291 self.validation_step_outputs.append(loss) 292 return loss 293 294 def on_validation_epoch_end(self): 295 if self.logger is not None and self.logger.experiment is not None: 296 logger = self.logger.experiment 297 for logger in self.trainer.loggers: 298 if isinstance(logger, TensorBoardLogger): 299 logger = logger.experiment 300 break 301 if isinstance(logger, TensorBoardLogger): 302 tacotron2_log_to_tb_func( 303 logger, 304 self.validation_step_outputs[0].values(), 305 self.global_step, 306 tag="val", 307 log_images=True, 308 add_audio=False, 309 ) 310 elif isinstance(logger, WandbLogger): 311 tacotron2_log_to_wandb_func( 312 logger, 313 self.validation_step_outputs[0].values(), 314 self.global_step, 315 tag="val", 316 log_images=True, 317 add_audio=False, 318 ) 319 avg_loss = torch.stack( 320 [x['val_loss'] for x in self.validation_step_outputs] 321 ).mean() # This reduces across batches, not workers! 322 self.log('val_loss', avg_loss) 323 self.validation_step_outputs.clear() # free memory 324 325 def _setup_normalizer(self, cfg): 326 if "text_normalizer" in cfg: 327 normalizer_kwargs = {} 328 329 if "whitelist" in cfg.text_normalizer: 330 normalizer_kwargs["whitelist"] = self.register_artifact( 331 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 332 ) 333 334 try: 335 import nemo_text_processing 336 337 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 338 except Exception as e: 339 logging.error(e) 340 raise ImportError( 341 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 342 ) 343 344 self.text_normalizer_call = self.normalizer.normalize 345 if "text_normalizer_call_kwargs" in cfg: 346 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 347 348 def _setup_tokenizer(self, cfg): 349 text_tokenizer_kwargs = {} 350 if "g2p" in cfg.text_tokenizer and cfg.text_tokenizer.g2p is not None: 351 # for backward compatibility 352 if ( 353 self._is_model_being_restored() 354 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 355 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 356 ): 357 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 358 cfg.text_tokenizer.g2p["_target_"] 359 ) 360 361 g2p_kwargs = {} 362 363 if "phoneme_dict" in cfg.text_tokenizer.g2p: 364 g2p_kwargs["phoneme_dict"] = self.register_artifact( 365 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 366 ) 367 368 if "heteronyms" in cfg.text_tokenizer.g2p: 369 g2p_kwargs["heteronyms"] = self.register_artifact( 370 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 371 ) 372 373 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 374 375 self.tokenizer = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 376 377 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 378 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 379 raise ValueError(f"No dataset for {name}") 380 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 381 raise ValueError(f"No dataloder_params for {name}") 382 if shuffle_should_be: 383 if 'shuffle' not in cfg.dataloader_params: 384 logging.warning( 385 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 386 "config. Manually setting to True" 387 ) 388 with open_dict(cfg.dataloader_params): 389 cfg.dataloader_params.shuffle = True 390 elif not cfg.dataloader_params.shuffle: 391 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 392 elif not shuffle_should_be and cfg.dataloader_params.shuffle: 393 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 394 395 dataset = instantiate( 396 cfg.dataset, 397 text_normalizer=self.normalizer, 398 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 399 text_tokenizer=self.tokenizer, 400 ) 401 402 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 403 404 def setup_training_data(self, cfg): 405 self._train_dl = self.__setup_dataloader_from_config(cfg) 406 407 def setup_validation_data(self, cfg): 408 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="validation") 409 410 @classmethod 411 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 412 """ 413 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 414 Returns: 415 List of available pre-trained models. 416 """ 417 list_of_models = [] 418 model = PretrainedModelInfo( 419 pretrained_model_name="tts_en_tacotron2", 420 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_tacotron2/versions/1.10.0/files/tts_en_tacotron2.nemo", 421 description="This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.", 422 class_=cls, 423 aliases=["Tacotron2-22050Hz"], 424 ) 425 list_of_models.append(model) 426 return list_of_models 427 [end of nemo/collections/tts/models/tacotron2.py] [start of nemo/core/config/modelPT.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Dict, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.core import config 21 from nemo.core.classes.dataset import DatasetConfig 22 from nemo.utils import exp_manager 23 24 25 @dataclass 26 class SchedConfig: 27 name: str = MISSING 28 min_lr: float = 0.0 29 last_epoch: int = -1 30 31 32 @dataclass 33 class OptimConfig: 34 name: str = MISSING 35 sched: Optional[SchedConfig] = None 36 37 38 @dataclass 39 class ModelConfig: 40 """ 41 Model component inside ModelPT 42 """ 43 44 # ... 45 train_ds: Optional[DatasetConfig] = None 46 validation_ds: Optional[DatasetConfig] = None 47 test_ds: Optional[DatasetConfig] = None 48 optim: Optional[OptimConfig] = None 49 50 51 @dataclass 52 class HydraConfig: 53 run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) 54 job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) 55 56 57 @dataclass 58 class NemoConfig: 59 name: str = MISSING 60 model: ModelConfig = MISSING 61 trainer: config.TrainerConfig = config.TrainerConfig( 62 strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' 63 ) 64 exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() 65 hydra: HydraConfig = HydraConfig() 66 67 68 class ModelConfigBuilder: 69 def __init__(self, model_cfg: ModelConfig): 70 """ 71 Base class for any Model Config Builder. 72 73 A Model Config Builder is a utility class that accepts a ModelConfig dataclass, 74 and via a set of utility methods (that are implemented by the subclassed ModelConfigBuilder), 75 builds a finalized ModelConfig that can be supplied to a NemoModel dataclass as 76 the `model` component. 77 78 Subclasses *must* implement the private method `_finalize_cfg`. 79 Inside this method, they must update `self.model_cfg` with all interdependent config 80 options that need to be set (either updated by user explicitly or with their default value). 81 82 The updated model config must then be preserved in `self.model_cfg`. 83 84 Example: 85 # Create the config builder 86 config_builder = <subclass>ModelConfigBuilder() 87 88 # Update the components of the config that are modifiable 89 config_builder.set_X(X) 90 config_builder.set_Y(Y) 91 92 # Create a "finalized" config dataclass that will contain all the updates 93 # that were specified by the builder 94 model_config = config_builder.build() 95 96 # Use model config as is (or further update values), then create a new Model 97 model = nemo.<domain>.models.<ModelName>Model(cfg=model_config, trainer=Trainer()) 98 99 Supported build methods: 100 - set_train_ds: All model configs can accept a subclass of `DatasetConfig` as their 101 training config. Subclasses can override this method to enable auto-complete 102 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 103 104 - set_validation_ds: All model configs can accept a subclass of `DatasetConfig` as their 105 validation config. Subclasses can override this method to enable auto-complete 106 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 107 108 - set_test_ds: All model configs can accept a subclass of `DatasetConfig` as their 109 test config. Subclasses can override this method to enable auto-complete 110 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 111 112 - set_optim: A build method that supports changes to the Optimizer (and optionally, 113 the Scheduler) used for training the model. The function accepts two inputs - 114 115 `cfg`: A subclass of `OptimizerParams` - any OptimizerParams subclass can be used, 116 in order to select an appropriate Optimizer. Examples: AdamParams. 117 118 `sched_cfg`: A subclass of `SchedulerParams` - any SchedulerParams subclass can be used, 119 in order to select an appropriate Scheduler. Examples: CosineAnnealingParams. 120 Note that this argument is optional. 121 122 - build(): The method which should return a "finalized" ModelConfig dataclass. 123 Subclasses *should* always override this method, and update the signature 124 of this method with the return type of the Dataclass, so that it enables 125 autocomplete for the user. 126 127 Example: 128 def build(self) -> EncDecCTCConfig: 129 return super().build() 130 131 Any additional build methods must be added by subclasses of ModelConfigBuilder. 132 133 Args: 134 model_cfg: 135 """ 136 self.model_cfg = model_cfg 137 self.train_ds_cfg = None 138 self.validation_ds_cfg = None 139 self.test_ds_cfg = None 140 self.optim_cfg = None 141 142 def set_train_ds(self, cfg: Optional[DatasetConfig] = None): 143 self.model_cfg.train_ds = cfg 144 145 def set_validation_ds(self, cfg: Optional[DatasetConfig] = None): 146 self.model_cfg.validation_ds = cfg 147 148 def set_test_ds(self, cfg: Optional[DatasetConfig] = None): 149 self.model_cfg.test_ds = cfg 150 151 def set_optim(self, cfg: config.OptimizerParams, sched_cfg: Optional[config.SchedulerParams] = None): 152 @dataclass 153 class WrappedOptimConfig(OptimConfig, cfg.__class__): 154 pass 155 156 # Setup optim 157 optim_name = cfg.__class__.__name__.replace("Params", "").lower() 158 wrapped_cfg = WrappedOptimConfig(name=optim_name, sched=None, **vars(cfg)) 159 160 if sched_cfg is not None: 161 162 @dataclass 163 class WrappedSchedConfig(SchedConfig, sched_cfg.__class__): 164 pass 165 166 # Setup scheduler 167 sched_name = sched_cfg.__class__.__name__.replace("Params", "") 168 wrapped_sched_cfg = WrappedSchedConfig(name=sched_name, **vars(sched_cfg)) 169 170 wrapped_cfg.sched = wrapped_sched_cfg 171 172 self.model_cfg.optim = wrapped_cfg 173 174 def _finalize_cfg(self): 175 raise NotImplementedError() 176 177 def build(self) -> ModelConfig: 178 # validate config 179 self._finalize_cfg() 180 181 return self.model_cfg 182 [end of nemo/core/config/modelPT.py] [start of nemo/utils/exp_manager.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import glob 16 import os 17 import subprocess 18 import sys 19 import time 20 import warnings 21 from dataclasses import dataclass 22 from datetime import timedelta 23 from pathlib import Path 24 from shutil import copy, move 25 from typing import Any, Dict, List, Optional, Tuple, Union 26 27 import pytorch_lightning 28 import torch 29 from hydra.core.hydra_config import HydraConfig 30 from hydra.utils import get_original_cwd 31 from omegaconf import DictConfig, OmegaConf, open_dict 32 from pytorch_lightning.callbacks import Callback, ModelCheckpoint 33 from pytorch_lightning.callbacks.early_stopping import EarlyStopping 34 from pytorch_lightning.callbacks.timer import Interval, Timer 35 from pytorch_lightning.loggers import MLFlowLogger, TensorBoardLogger, WandbLogger 36 from pytorch_lightning.loops import _TrainingEpochLoop 37 from pytorch_lightning.strategies.ddp import DDPStrategy 38 39 from nemo.collections.common.callbacks import EMA 40 from nemo.constants import NEMO_ENV_VARNAME_TESTING, NEMO_ENV_VARNAME_VERSION 41 from nemo.utils import logging, timers 42 from nemo.utils.app_state import AppState 43 from nemo.utils.callbacks import NeMoModelCheckpoint, PreemptionCallback 44 from nemo.utils.env_var_parsing import get_envbool 45 from nemo.utils.exceptions import NeMoBaseException 46 from nemo.utils.get_rank import is_global_rank_zero 47 from nemo.utils.lightning_logger_patch import add_filehandlers_to_pl_logger 48 from nemo.utils.loggers import ClearMLLogger, ClearMLParams, DLLogger, DLLoggerParams, MLFlowParams 49 from nemo.utils.model_utils import uninject_model_parallel_rank 50 51 52 class NotFoundError(NeMoBaseException): 53 """ Raised when a file or folder is not found""" 54 55 56 class LoggerMisconfigurationError(NeMoBaseException): 57 """ Raised when a mismatch between trainer.logger and exp_manager occurs""" 58 59 def __init__(self, message): 60 message = ( 61 message 62 + " You can disable lighning's trainer from creating a logger by passing logger=False to its constructor." 63 ) 64 super().__init__(message) 65 66 67 class CheckpointMisconfigurationError(NeMoBaseException): 68 """ Raised when a mismatch between trainer.callbacks and exp_manager occurs""" 69 70 71 @dataclass 72 class EarlyStoppingParams: 73 monitor: str = "val_loss" # The metric that early stopping should consider. 74 mode: str = "min" # inform early stopping whether to look for increase or decrease in monitor. 75 min_delta: float = 0.001 # smallest change to consider as improvement. 76 patience: int = 10 # how many (continuous) validation cycles to wait with no improvement and stopping training. 77 verbose: bool = True 78 strict: bool = True 79 check_finite: bool = True 80 stopping_threshold: Optional[float] = None 81 divergence_threshold: Optional[float] = None 82 check_on_train_epoch_end: Optional[bool] = None 83 log_rank_zero_only: bool = False 84 85 86 @dataclass 87 class CallbackParams: 88 filepath: Optional[str] = None # Deprecated 89 dirpath: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 90 filename: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 91 monitor: Optional[str] = "val_loss" 92 verbose: Optional[bool] = True 93 save_last: Optional[bool] = True 94 save_top_k: Optional[int] = 3 95 save_weights_only: Optional[bool] = False 96 mode: Optional[str] = "min" 97 auto_insert_metric_name: bool = True 98 every_n_epochs: Optional[int] = 1 99 every_n_train_steps: Optional[int] = None 100 train_time_interval: Optional[str] = None 101 prefix: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 102 postfix: str = ".nemo" 103 save_best_model: bool = False 104 always_save_nemo: bool = False 105 save_nemo_on_train_end: Optional[bool] = True # Whether to automatically save .nemo file durin on_train_end hook 106 model_parallel_size: Optional[int] = None # tensor parallel size * pipeline parallel size 107 save_on_train_epoch_end: Optional[bool] = False # Save after training, not after validation 108 109 110 @dataclass 111 class StepTimingParams: 112 reduction: Optional[str] = "mean" 113 # if True torch.cuda.synchronize() is called on start/stop 114 sync_cuda: Optional[bool] = False 115 # if positive, defines the size of a sliding window for computing mean 116 buffer_size: Optional[int] = 1 117 118 119 @dataclass 120 class EMAParams: 121 enable: Optional[bool] = False 122 decay: Optional[float] = 0.999 123 cpu_offload: Optional[bool] = False 124 validate_original_weights: Optional[bool] = False 125 every_n_steps: int = 1 126 127 128 @dataclass 129 class ExpManagerConfig: 130 """Experiment Manager config for validation of passed arguments. 131 """ 132 133 # Log dir creation parameters 134 explicit_log_dir: Optional[str] = None 135 exp_dir: Optional[str] = None 136 name: Optional[str] = None 137 version: Optional[str] = None 138 use_datetime_version: Optional[bool] = True 139 resume_if_exists: Optional[bool] = False 140 resume_past_end: Optional[bool] = False 141 resume_ignore_no_checkpoint: Optional[bool] = False 142 resume_from_checkpoint: Optional[str] = None 143 # Logging parameters 144 create_tensorboard_logger: Optional[bool] = True 145 summary_writer_kwargs: Optional[Dict[Any, Any]] = None 146 create_wandb_logger: Optional[bool] = False 147 wandb_logger_kwargs: Optional[Dict[Any, Any]] = None 148 create_mlflow_logger: Optional[bool] = False 149 mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() 150 create_dllogger_logger: Optional[bool] = False 151 dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() 152 create_clearml_logger: Optional[bool] = False 153 clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() 154 # Checkpointing parameters 155 create_checkpoint_callback: Optional[bool] = True 156 checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() 157 create_early_stopping_callback: Optional[bool] = False 158 early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() 159 create_preemption_callback: Optional[bool] = True 160 # Additional exp_manager arguments 161 files_to_copy: Optional[List[str]] = None 162 # logs timing of train/val/test steps 163 log_step_timing: Optional[bool] = True 164 step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() 165 # Configures creation of log files for different ranks 166 log_local_rank_0_only: Optional[bool] = False 167 log_global_rank_0_only: Optional[bool] = False 168 # disable initial validation when resuming from a checkpoint saved during validation 169 disable_validation_on_resume: Optional[bool] = True 170 ema: Optional[EMAParams] = EMAParams() 171 # Wall clock time limit 172 max_time_per_run: Optional[str] = None 173 # time to sleep non 0 ranks during initialization 174 seconds_to_sleep: float = 5 175 176 177 class TimingCallback(Callback): 178 """ 179 Logs execution time of train/val/test steps 180 """ 181 182 def __init__(self, timer_kwargs={}): 183 self.timer = timers.NamedTimer(**timer_kwargs) 184 185 def _on_batch_start(self, name): 186 # reset only if we do not return mean of a sliding window 187 if self.timer.buffer_size <= 0: 188 self.timer.reset(name) 189 190 self.timer.start(name) 191 192 def _on_batch_end(self, name, pl_module): 193 self.timer.stop(name) 194 # Set the `batch_size=1` as WAR for `dataloader_iter`, which is not used for any metric 195 pl_module.log( 196 name + ' in s', 197 self.timer[name], 198 on_step=True, 199 on_epoch=False, 200 batch_size=1, 201 prog_bar=(name == "train_step_timing"), 202 ) 203 204 def on_train_batch_start(self, trainer, pl_module, batch, batch_idx): 205 self._on_batch_start("train_step_timing") 206 207 def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): 208 self._on_batch_end("train_step_timing", pl_module) 209 210 def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 211 self._on_batch_start("validation_step_timing") 212 213 def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 214 self._on_batch_end("validation_step_timing", pl_module) 215 216 def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 217 self._on_batch_start("test_step_timing") 218 219 def on_test_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 220 self._on_batch_end("test_step_timing", pl_module) 221 222 def on_before_backward(self, trainer, pl_module, loss): 223 self._on_batch_start("train_backward_timing") 224 225 def on_after_backward(self, trainer, pl_module): 226 self._on_batch_end("train_backward_timing", pl_module) 227 228 229 def exp_manager(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None) -> Optional[Path]: 230 """ 231 exp_manager is a helper function used to manage folders for experiments. It follows the pytorch lightning paradigm 232 of exp_dir/model_or_experiment_name/version. If the lightning trainer has a logger, exp_manager will get exp_dir, 233 name, and version from the logger. Otherwise it will use the exp_dir and name arguments to create the logging 234 directory. exp_manager also allows for explicit folder creation via explicit_log_dir. 235 236 The version can be a datetime string or an integer. Datestime version can be disabled if use_datetime_version is set 237 to False. It optionally creates TensorBoardLogger, WandBLogger, DLLogger, MLFlowLogger, ClearMLLogger, 238 ModelCheckpoint objects from pytorch lightning. 239 It copies sys.argv, and git information if available to the logging directory. It creates a log file for each 240 process to log their output into. 241 242 exp_manager additionally has a resume feature (resume_if_exists) which can be used to continuing training from 243 the constructed log_dir. When you need to continue the training repeatedly (like on a cluster which you need 244 multiple consecutive jobs), you need to avoid creating the version folders. Therefore from v1.0.0, when 245 resume_if_exists is set to True, creating the version folders is ignored. 246 247 Args: 248 trainer (pytorch_lightning.Trainer): The lightning trainer. 249 cfg (DictConfig, dict): Can have the following keys: 250 251 - explicit_log_dir (str, Path): Can be used to override exp_dir/name/version folder creation. Defaults to 252 None, which will use exp_dir, name, and version to construct the logging directory. 253 - exp_dir (str, Path): The base directory to create the logging directory. Defaults to None, which logs to 254 ./nemo_experiments. 255 - name (str): The name of the experiment. Defaults to None which turns into "default" via name = name or 256 "default". 257 - version (str): The version of the experiment. Defaults to None which uses either a datetime string or 258 lightning's TensorboardLogger system of using version_{int}. 259 - use_datetime_version (bool): Whether to use a datetime string for version. Defaults to True. 260 - resume_if_exists (bool): Whether this experiment is resuming from a previous run. If True, it sets 261 trainer._checkpoint_connector._ckpt_path so that the trainer should auto-resume. exp_manager will move files 262 under log_dir to log_dir/run_{int}. Defaults to False. From v1.0.0, when resume_if_exists is True, 263 we would not create version folders to make it easier to find the log folder for next runs. 264 - resume_past_end (bool): exp_manager errors out if resume_if_exists is True and a checkpoint matching 265 ``*end.ckpt`` indicating a previous training run fully completed. This behaviour can be disabled, in which 266 case the ``*end.ckpt`` will be loaded by setting resume_past_end to True. Defaults to False. 267 - resume_ignore_no_checkpoint (bool): exp_manager errors out if resume_if_exists is True and no checkpoint 268 could be found. This behaviour can be disabled, in which case exp_manager will print a message and 269 continue without restoring, by setting resume_ignore_no_checkpoint to True. Defaults to False. 270 - resume_from_checkpoint (str): Can be used to specify a path to a specific checkpoint file to load from. This will 271 override any checkpoint found when resume_if_exists is True. Defaults to None. 272 - create_tensorboard_logger (bool): Whether to create a tensorboard logger and attach it to the pytorch 273 lightning trainer. Defaults to True. 274 - summary_writer_kwargs (dict): A dictionary of kwargs that can be passed to lightning's TensorboardLogger 275 class. Note that log_dir is passed by exp_manager and cannot exist in this dict. Defaults to None. 276 - create_wandb_logger (bool): Whether to create a Weights and Baises logger and attach it to the pytorch 277 lightning trainer. Defaults to False. 278 - wandb_logger_kwargs (dict): A dictionary of kwargs that can be passed to lightning's WandBLogger 279 class. Note that name and project are required parameters if create_wandb_logger is True. 280 Defaults to None. 281 - create_mlflow_logger (bool): Whether to create an MLFlow logger and attach it to the pytorch lightning 282 training. Defaults to False 283 - mlflow_logger_kwargs (dict): optional parameters for the MLFlow logger 284 - create_dllogger_logger (bool): Whether to create an DLLogger logger and attach it to the pytorch lightning 285 training. Defaults to False 286 - dllogger_logger_kwargs (dict): optional parameters for the DLLogger logger 287 - create_clearml_logger (bool): Whether to create an ClearML logger and attach it to the pytorch lightning 288 training. Defaults to False 289 - clearml_logger_kwargs (dict): optional parameters for the ClearML logger 290 - create_checkpoint_callback (bool): Whether to create a ModelCheckpoint callback and attach it to the 291 pytorch lightning trainer. The ModelCheckpoint saves the top 3 models with the best "val_loss", the most 292 recent checkpoint under ``*last.ckpt``, and the final checkpoint after training completes under ``*end.ckpt``. 293 Defaults to True. 294 - create_early_stopping_callback (bool): Flag to decide if early stopping should be used to stop training. Default is False. 295 See EarlyStoppingParams dataclass above. 296 - create_preemption_callback (bool): Flag to decide whether to enable preemption callback to save checkpoints and exit training 297 immediately upon preemption. Default is True. 298 - files_to_copy (list): A list of files to copy to the experiment logging directory. Defaults to None which 299 copies no files. 300 - log_local_rank_0_only (bool): Whether to only create log files for local rank 0. Defaults to False. 301 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 302 - log_global_rank_0_only (bool): Whether to only create log files for global rank 0. Defaults to False. 303 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 304 - max_time (str): The maximum wall clock time *per run*. This is intended to be used on clusters where you want 305 a checkpoint to be saved after this specified time and be able to resume from that checkpoint. Defaults to None. 306 - seconds_to_sleep (float): seconds to sleep non rank 0 processes for. Used to give enough time for rank 0 to initialize 307 308 returns: 309 log_dir (Path): The final logging directory where logging files are saved. Usually the concatenation of 310 exp_dir, name, and version. 311 """ 312 # Add rank information to logger 313 # Note: trainer.global_rank and trainer.is_global_zero are not set until trainer.fit, so have to hack around it 314 local_rank = int(os.environ.get("LOCAL_RANK", 0)) 315 global_rank = trainer.node_rank * trainer.num_devices + local_rank 316 logging.rank = global_rank 317 318 if cfg is None: 319 logging.error("exp_manager did not receive a cfg argument. It will be disabled.") 320 return 321 if trainer.fast_dev_run: 322 logging.info("Trainer was called with fast_dev_run. exp_manager will return without any functionality.") 323 return 324 325 # Ensure passed cfg is compliant with ExpManagerConfig 326 schema = OmegaConf.structured(ExpManagerConfig) 327 if isinstance(cfg, dict): 328 cfg = OmegaConf.create(cfg) 329 elif not isinstance(cfg, DictConfig): 330 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 331 cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True)) 332 cfg = OmegaConf.merge(schema, cfg) 333 334 error_checks(trainer, cfg) # Ensures that trainer options are compliant with NeMo and exp_manager arguments 335 336 log_dir, exp_dir, name, version = get_log_dir( 337 trainer=trainer, 338 exp_dir=cfg.exp_dir, 339 name=cfg.name, 340 version=cfg.version, 341 explicit_log_dir=cfg.explicit_log_dir, 342 use_datetime_version=cfg.use_datetime_version, 343 resume_if_exists=cfg.resume_if_exists, 344 ) 345 346 check_resume( 347 trainer, 348 log_dir, 349 cfg.resume_if_exists, 350 cfg.resume_past_end, 351 cfg.resume_ignore_no_checkpoint, 352 cfg.checkpoint_callback_params.dirpath, 353 cfg.resume_from_checkpoint, 354 ) 355 356 checkpoint_name = name 357 # If name returned from get_log_dir is "", use cfg.name for checkpointing 358 if checkpoint_name is None or checkpoint_name == '': 359 checkpoint_name = cfg.name or "default" 360 361 # Set mlflow name if it's not set, before the main name is erased 362 if cfg.create_mlflow_logger and (not cfg.mlflow_logger_kwargs.get("experiment_name", None)): 363 cfg.mlflow_logger_kwargs.experiment_name = cfg.name 364 logging.warning( 365 'mlflow logger specified but no experiment name set. Using the same as Tensorboard: %s', 366 cfg.mlflow_logger_kwargs.experiment_name, 367 ) 368 369 cfg.name = name # Used for configure_loggers so that the log_dir is properly set even if name is "" 370 cfg.version = version 371 372 # update app_state with log_dir, exp_dir, etc 373 app_state = AppState() 374 app_state.log_dir = log_dir 375 app_state.exp_dir = exp_dir 376 app_state.name = name 377 app_state.version = version 378 app_state.checkpoint_name = checkpoint_name 379 app_state.create_checkpoint_callback = cfg.create_checkpoint_callback 380 app_state.checkpoint_callback_params = cfg.checkpoint_callback_params 381 382 # Create the logging directory if it does not exist 383 os.makedirs(log_dir, exist_ok=True) # Cannot limit creation to global zero as all ranks write to own log file 384 logging.info(f'Experiments will be logged at {log_dir}') 385 trainer._default_root_dir = log_dir 386 387 if cfg.log_local_rank_0_only is True and cfg.log_global_rank_0_only is True: 388 raise ValueError( 389 f"Cannot set both log_local_rank_0_only and log_global_rank_0_only to True. Please set either one or neither." 390 ) 391 392 # This is set if the env var NEMO_TESTING is set to True. 393 nemo_testing = get_envbool(NEMO_ENV_VARNAME_TESTING, False) 394 395 # Handle logging to file 396 log_file = log_dir / f'nemo_log_globalrank-{global_rank}_localrank-{local_rank}.txt' 397 if cfg.log_local_rank_0_only is True and not nemo_testing: 398 if local_rank == 0: 399 logging.add_file_handler(log_file) 400 elif cfg.log_global_rank_0_only is True and not nemo_testing: 401 if global_rank == 0: 402 logging.add_file_handler(log_file) 403 else: 404 # Logs on all ranks. 405 logging.add_file_handler(log_file) 406 407 # For some reason, LearningRateLogger requires trainer to have a logger. Safer to create logger on all ranks 408 # not just global rank 0. 409 if ( 410 cfg.create_tensorboard_logger 411 or cfg.create_wandb_logger 412 or cfg.create_mlflow_logger 413 or cfg.create_dllogger_logger 414 or cfg.create_clearml_logger 415 ): 416 configure_loggers( 417 trainer, 418 exp_dir, 419 log_dir, 420 cfg.name, 421 cfg.version, 422 cfg.checkpoint_callback_params, 423 cfg.create_tensorboard_logger, 424 cfg.summary_writer_kwargs, 425 cfg.create_wandb_logger, 426 cfg.wandb_logger_kwargs, 427 cfg.create_mlflow_logger, 428 cfg.mlflow_logger_kwargs, 429 cfg.create_dllogger_logger, 430 cfg.dllogger_logger_kwargs, 431 cfg.create_clearml_logger, 432 cfg.clearml_logger_kwargs, 433 ) 434 435 # add loggers timing callbacks 436 if cfg.log_step_timing: 437 timing_callback = TimingCallback(timer_kwargs=cfg.step_timing_kwargs or {}) 438 trainer.callbacks.insert(0, timing_callback) 439 440 if cfg.ema.enable: 441 ema_callback = EMA( 442 decay=cfg.ema.decay, 443 validate_original_weights=cfg.ema.validate_original_weights, 444 cpu_offload=cfg.ema.cpu_offload, 445 every_n_steps=cfg.ema.every_n_steps, 446 ) 447 trainer.callbacks.append(ema_callback) 448 449 if cfg.create_early_stopping_callback: 450 early_stop_callback = EarlyStopping(**cfg.early_stopping_callback_params) 451 trainer.callbacks.append(early_stop_callback) 452 453 if cfg.create_checkpoint_callback: 454 configure_checkpointing( 455 trainer, 456 log_dir, 457 checkpoint_name, 458 cfg.resume_if_exists, 459 cfg.checkpoint_callback_params, 460 cfg.create_preemption_callback, 461 ) 462 463 if cfg.disable_validation_on_resume: 464 # extend training loop to skip initial validation when resuming from checkpoint 465 configure_no_restart_validation_training_loop(trainer) 466 # Setup a stateless timer for use on clusters. 467 if cfg.max_time_per_run is not None: 468 found_ptl_timer = False 469 for idx, callback in enumerate(trainer.callbacks): 470 if isinstance(callback, Timer): 471 # NOTE: PTL does not expose a `trainer.max_time`. By the time we are in this function, PTL has already setup a timer if the user specifies `trainer.max_time` so best we can do is replace that. 472 # Working: If only `trainer.max_time` is set - it behaves as a normal PTL timer. If only `exp_manager.max_time_per_run` is set - it behaves as a StateLessTimer. If both are set, it also behaves as a StateLessTimer. 473 logging.warning( 474 f'Found a PTL Timer callback, replacing with a StatelessTimer callback. This will happen if you set trainer.max_time as well as exp_manager.max_time_per_run.' 475 ) 476 trainer.callbacks[idx] = StatelessTimer(cfg.max_time_per_run) 477 found_ptl_timer = True 478 break 479 480 if not found_ptl_timer: 481 trainer.max_time = cfg.max_time_per_run 482 trainer.callbacks.append(StatelessTimer(cfg.max_time_per_run)) 483 484 if is_global_rank_zero(): 485 # Move files_to_copy to folder and add git information if present 486 if cfg.files_to_copy: 487 for _file in cfg.files_to_copy: 488 copy(Path(_file), log_dir) 489 490 # Create files for cmd args and git info 491 with open(log_dir / 'cmd-args.log', 'w', encoding='utf-8') as _file: 492 _file.write(" ".join(sys.argv)) 493 494 # Try to get git hash 495 git_repo, git_hash = get_git_hash() 496 if git_repo: 497 with open(log_dir / 'git-info.log', 'w', encoding='utf-8') as _file: 498 _file.write(f'commit hash: {git_hash}') 499 _file.write(get_git_diff()) 500 501 # Add err_file logging to global_rank zero 502 logging.add_err_file_handler(log_dir / 'nemo_error_log.txt') 503 504 # Add lightning file logging to global_rank zero 505 add_filehandlers_to_pl_logger(log_dir / 'lightning_logs.txt', log_dir / 'nemo_error_log.txt') 506 507 elif trainer.num_nodes * trainer.num_devices > 1: 508 # sleep other ranks so rank 0 can finish 509 # doing the initialization such as moving files 510 time.sleep(cfg.seconds_to_sleep) 511 512 return log_dir 513 514 515 def error_checks(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None): 516 """ 517 Checks that the passed trainer is compliant with NeMo and exp_manager's passed configuration. Checks that: 518 - Throws error when hydra has changed the working directory. This causes issues with lightning's DDP 519 - Throws error when trainer has loggers defined but create_tensorboard_logger or create_wandB_logger 520 or create_mlflow_logger or create_dllogger_logger is True 521 - Prints error messages when 1) run on multi-node and not Slurm, and 2) run on multi-gpu without DDP 522 """ 523 if HydraConfig.initialized() and get_original_cwd() != os.getcwd(): 524 raise ValueError( 525 "Hydra changed the working directory. This interferes with ExpManger's functionality. Please pass " 526 "hydra.run.dir=. to your python script." 527 ) 528 if trainer.logger is not None and ( 529 cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger 530 ): 531 raise LoggerMisconfigurationError( 532 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and either " 533 f"create_tensorboard_logger: {cfg.create_tensorboard_logger} or create_wandb_logger: " 534 f"{cfg.create_wandb_logger} or create_mlflow_logger: {cfg.create_mlflow_logger}" 535 f"or create_dllogger_logger: {cfg.create_mlflow_logger} was set to True. " 536 "These can only be used if trainer does not already have a logger." 537 ) 538 if trainer.num_nodes > 1 and not check_slurm(trainer): 539 logging.error( 540 "You are running multi-node training without SLURM handling the processes." 541 " Please note that this is not tested in NeMo and could result in errors." 542 ) 543 if trainer.num_devices > 1 and not isinstance(trainer.strategy, DDPStrategy): 544 logging.error( 545 "You are running multi-gpu without ddp.Please note that this is not tested in NeMo and could result in " 546 "errors." 547 ) 548 549 550 def check_resume( 551 trainer: 'pytorch_lightning.Trainer', 552 log_dir: str, 553 resume_if_exists: bool = False, 554 resume_past_end: bool = False, 555 resume_ignore_no_checkpoint: bool = False, 556 dirpath: str = None, 557 resume_from_checkpoint: str = None, 558 ): 559 """Checks that resume=True was used correctly with the arguments pass to exp_manager. Sets 560 trainer._checkpoint_connector._ckpt_path as necessary. 561 562 Returns: 563 log_dir (Path): The log_dir 564 exp_dir (str): The base exp_dir without name nor version 565 name (str): The name of the experiment 566 version (str): The version of the experiment 567 568 Raises: 569 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 570 ValueError: If resume is True, and there were more than 1 checkpoint could found. 571 """ 572 573 if not log_dir: 574 raise ValueError(f"Resuming requires the log_dir {log_dir} to be passed to exp_manager") 575 576 checkpoint = None 577 if resume_from_checkpoint: 578 checkpoint = resume_from_checkpoint 579 if resume_if_exists: 580 # Use <log_dir>/checkpoints/ unless `dirpath` is set 581 checkpoint_dir = Path(dirpath) if dirpath else Path(Path(log_dir) / "checkpoints") 582 583 # when using distributed checkpointing, checkpoint_dir is a directory of directories 584 # we check for this here 585 dist_checkpoints = [d for d in list(checkpoint_dir.glob("*")) if d.is_dir()] 586 end_dist_checkpoints = [d for d in dist_checkpoints if d.match("*end")] 587 last_dist_checkpoints = [d for d in dist_checkpoints if d.match("*last")] 588 589 end_checkpoints = end_dist_checkpoints if end_dist_checkpoints else list(checkpoint_dir.rglob("*end.ckpt")) 590 last_checkpoints = last_dist_checkpoints if last_dist_checkpoints else list(checkpoint_dir.rglob("*last.ckpt")) 591 592 if not checkpoint_dir.exists() or (not len(end_checkpoints) > 0 and not len(last_checkpoints) > 0): 593 if resume_ignore_no_checkpoint: 594 warn = f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. " 595 if checkpoint is None: 596 warn += "Training from scratch." 597 elif checkpoint == resume_from_checkpoint: 598 warn += f"Training from {resume_from_checkpoint}." 599 logging.warning(warn) 600 else: 601 raise NotFoundError( 602 f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. Cannot resume." 603 ) 604 elif len(end_checkpoints) > 0: 605 if resume_past_end: 606 if len(end_checkpoints) > 1: 607 if 'mp_rank' in str(end_checkpoints[0]): 608 checkpoint = end_checkpoints[0] 609 else: 610 raise ValueError(f"Multiple checkpoints {end_checkpoints} that matches *end.ckpt.") 611 else: 612 raise ValueError( 613 f"Found {end_checkpoints[0]} indicating that the last training run has already completed." 614 ) 615 elif len(last_checkpoints) > 1: 616 if 'mp_rank' in str(last_checkpoints[0]) or 'tp_rank' in str(last_checkpoints[0]): 617 checkpoint = last_checkpoints[0] 618 checkpoint = uninject_model_parallel_rank(checkpoint) 619 else: 620 raise ValueError(f"Multiple checkpoints {last_checkpoints} that matches *last.ckpt.") 621 else: 622 checkpoint = last_checkpoints[0] 623 624 # PTL 2.0 supports ckpt_path instead of resume_from_checkpoint as the trainer flag 625 if checkpoint is not None: 626 trainer.ckpt_path = str(checkpoint) 627 logging.info(f'Resuming training from checkpoint: {trainer.ckpt_path}') 628 629 if is_global_rank_zero(): 630 # Check to see if any files exist that need to be moved 631 files_to_move = [] 632 if Path(log_dir).exists(): 633 for child in Path(log_dir).iterdir(): 634 if child.is_file(): 635 files_to_move.append(child) 636 637 if len(files_to_move) > 0: 638 # Move old files to a new folder 639 other_run_dirs = Path(log_dir).glob("run_*") 640 run_count = 0 641 for fold in other_run_dirs: 642 if fold.is_dir(): 643 run_count += 1 644 new_run_dir = Path(Path(log_dir) / f"run_{run_count}") 645 new_run_dir.mkdir() 646 for _file in files_to_move: 647 move(str(_file), str(new_run_dir)) 648 649 650 def check_explicit_log_dir( 651 trainer: 'pytorch_lightning.Trainer', explicit_log_dir: Union[Path, str], exp_dir: str, name: str, version: str 652 ) -> Tuple[Path, str, str, str]: 653 """ Checks that the passed arguments are compatible with explicit_log_dir. 654 655 Returns: 656 log_dir (Path): the log_dir 657 exp_dir (str): the base exp_dir without name nor version 658 name (str): The name of the experiment 659 version (str): The version of the experiment 660 661 Raise: 662 LoggerMisconfigurationError 663 """ 664 if trainer.logger is not None: 665 raise LoggerMisconfigurationError( 666 "The pytorch lightning trainer that was passed to exp_manager contained a logger and explicit_log_dir: " 667 f"{explicit_log_dir} was pass to exp_manager. Please remove the logger from the lightning trainer." 668 ) 669 # Checking only (explicit_log_dir) vs (exp_dir and version). 670 # The `name` will be used as the actual name of checkpoint/archive. 671 if exp_dir or version: 672 logging.error( 673 f"exp_manager received explicit_log_dir: {explicit_log_dir} and at least one of exp_dir: {exp_dir}, " 674 f"or version: {version}. Please note that exp_dir, name, and version will be ignored." 675 ) 676 if is_global_rank_zero() and Path(explicit_log_dir).exists(): 677 logging.warning(f"Exp_manager is logging to {explicit_log_dir}, but it already exists.") 678 return Path(explicit_log_dir), str(explicit_log_dir), "", "" 679 680 681 def get_log_dir( 682 trainer: 'pytorch_lightning.Trainer', 683 exp_dir: str = None, 684 name: str = None, 685 version: str = None, 686 explicit_log_dir: str = None, 687 use_datetime_version: bool = True, 688 resume_if_exists: bool = False, 689 ) -> Tuple[Path, str, str, str]: 690 """ 691 Obtains the log_dir used for exp_manager. 692 693 Returns: 694 log_dir (Path): the log_dir 695 exp_dir (str): the base exp_dir without name nor version 696 name (str): The name of the experiment 697 version (str): The version of the experiment 698 explicit_log_dir (str): The explicit path to the log folder. Defaults to False. 699 use_datetime_version (bool): Uses date and time as the version of the log folder. Defaults to True. 700 resume_if_exists (bool): if resume_if_exists of the exp_manager's config is enabled or not. When enabled, the 701 version folders would not get created. 702 703 Raise: 704 LoggerMisconfigurationError: If trainer is incompatible with arguments 705 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 706 ValueError: If resume is True, and there were more than 1 checkpoint could found. 707 """ 708 if explicit_log_dir: # If explicit log_dir was passed, short circuit 709 return check_explicit_log_dir(trainer, explicit_log_dir, exp_dir, name, version) 710 711 # Default exp_dir to ./nemo_experiments if None was passed 712 _exp_dir = exp_dir 713 if exp_dir is None: 714 _exp_dir = str(Path.cwd() / 'nemo_experiments') 715 716 # If the user has already defined a logger for the trainer, use the logger defaults for logging directory 717 if trainer.logger is not None: 718 if trainer.logger.save_dir: 719 if exp_dir: 720 raise LoggerMisconfigurationError( 721 "The pytorch lightning trainer that was passed to exp_manager contained a logger, the logger's " 722 f"save_dir was not None, and exp_dir ({exp_dir}) was not None. If trainer.logger.save_dir " 723 "exists, exp_manager will use trainer.logger.save_dir as the logging directory and exp_dir " 724 "must be None." 725 ) 726 _exp_dir = trainer.logger.save_dir 727 if name: 728 raise LoggerMisconfigurationError( 729 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and name: " 730 f"{name} was also passed to exp_manager. If the trainer contains a " 731 "logger, exp_manager will use trainer.logger.name, and name passed to exp_manager must be None." 732 ) 733 name = trainer.logger.name 734 version = f"version_{trainer.logger.version}" 735 # Use user-defined exp_dir, project_name, exp_name, and versioning options 736 else: 737 name = name or "default" 738 version = version or os.environ.get(NEMO_ENV_VARNAME_VERSION, None) 739 740 if not version: 741 if resume_if_exists: 742 logging.warning( 743 "No version folders would be created under the log folder as 'resume_if_exists' is enabled." 744 ) 745 version = None 746 elif is_global_rank_zero(): 747 if use_datetime_version: 748 version = time.strftime('%Y-%m-%d_%H-%M-%S') 749 else: 750 tensorboard_logger = TensorBoardLogger(save_dir=Path(_exp_dir), name=name, version=version) 751 version = f"version_{tensorboard_logger.version}" 752 os.environ[NEMO_ENV_VARNAME_VERSION] = "" if version is None else version 753 754 log_dir = Path(_exp_dir) / Path(str(name)) / Path("" if version is None else str(version)) 755 return log_dir, str(_exp_dir), name, version 756 757 758 def get_git_hash(): 759 """ 760 Helper function that tries to get the commit hash if running inside a git folder 761 762 returns: 763 Bool: Whether the git subprocess ran without error 764 str: git subprocess output or error message 765 """ 766 try: 767 return ( 768 True, 769 subprocess.check_output(['git', 'rev-parse', 'HEAD'], stderr=subprocess.STDOUT).decode(), 770 ) 771 except subprocess.CalledProcessError as err: 772 return False, "{}\n".format(err.output.decode("utf-8")) 773 774 775 def get_git_diff(): 776 """ 777 Helper function that tries to get the git diff if running inside a git folder 778 779 returns: 780 Bool: Whether the git subprocess ran without error 781 str: git subprocess output or error message 782 """ 783 try: 784 return subprocess.check_output(['git', 'diff'], stderr=subprocess.STDOUT).decode() 785 except subprocess.CalledProcessError as err: 786 return "{}\n".format(err.output.decode("utf-8")) 787 788 789 def configure_loggers( 790 trainer: 'pytorch_lightning.Trainer', 791 exp_dir: [Path, str], 792 log_dir: [Path, str], 793 name: str, 794 version: str, 795 checkpoint_callback_params: dict, 796 create_tensorboard_logger: bool, 797 summary_writer_kwargs: dict, 798 create_wandb_logger: bool, 799 wandb_kwargs: dict, 800 create_mlflow_logger: bool, 801 mlflow_kwargs: dict, 802 create_dllogger_logger: bool, 803 dllogger_kwargs: dict, 804 create_clearml_logger: bool, 805 clearml_kwargs: dict, 806 ): 807 """ 808 Creates TensorboardLogger and/or WandBLogger / MLFlowLogger / DLlogger / ClearMLLogger and attach them to trainer. 809 Raises ValueError if summary_writer_kwargs or wandb_kwargs are misconfigured. 810 """ 811 # Potentially create tensorboard logger and/or WandBLogger / MLFlowLogger / DLLogger 812 logger_list = [] 813 if create_tensorboard_logger: 814 if summary_writer_kwargs is None: 815 summary_writer_kwargs = {} 816 elif "log_dir" in summary_writer_kwargs: 817 raise ValueError( 818 "You cannot pass `log_dir` as part of `summary_writer_kwargs`. `log_dir` is handled by lightning's " 819 "TensorBoardLogger logger." 820 ) 821 tensorboard_logger = TensorBoardLogger(save_dir=exp_dir, name=name, version=version, **summary_writer_kwargs) 822 logger_list.append(tensorboard_logger) 823 logging.info("TensorboardLogger has been set up") 824 825 if create_wandb_logger: 826 if wandb_kwargs is None: 827 wandb_kwargs = {} 828 if "name" not in wandb_kwargs and "project" not in wandb_kwargs: 829 raise ValueError("name and project are required for wandb_logger") 830 831 # Update the wandb save_dir 832 if wandb_kwargs.get('save_dir', None) is None: 833 wandb_kwargs['save_dir'] = exp_dir 834 os.makedirs(wandb_kwargs['save_dir'], exist_ok=True) 835 wandb_logger = WandbLogger(version=version, **wandb_kwargs) 836 837 logger_list.append(wandb_logger) 838 logging.info("WandBLogger has been set up") 839 840 if create_mlflow_logger: 841 mlflow_logger = MLFlowLogger(run_name=version, **mlflow_kwargs) 842 843 logger_list.append(mlflow_logger) 844 logging.info("MLFlowLogger has been set up") 845 846 if create_dllogger_logger: 847 dllogger_logger = DLLogger(**dllogger_kwargs) 848 849 logger_list.append(dllogger_logger) 850 logging.info("DLLogger has been set up") 851 852 if create_clearml_logger: 853 clearml_logger = ClearMLLogger( 854 clearml_cfg=clearml_kwargs, 855 log_dir=log_dir, 856 prefix=name, 857 save_best_model=checkpoint_callback_params.save_best_model, 858 ) 859 860 logger_list.append(clearml_logger) 861 logging.info("ClearMLLogger has been set up") 862 863 trainer._logger_connector.configure_logger(logger_list) 864 865 866 def configure_checkpointing( 867 trainer: 'pytorch_lightning.Trainer', 868 log_dir: Path, 869 name: str, 870 resume: bool, 871 params: 'DictConfig', 872 create_preemption_callback: bool, 873 ): 874 """ Adds ModelCheckpoint to trainer. Raises CheckpointMisconfigurationError if trainer already has a ModelCheckpoint 875 callback 876 """ 877 for callback in trainer.callbacks: 878 if isinstance(callback, ModelCheckpoint): 879 raise CheckpointMisconfigurationError( 880 "The pytorch lightning trainer that was passed to exp_manager contained a ModelCheckpoint " 881 "and create_checkpoint_callback was set to True. Please either set create_checkpoint_callback " 882 "to False, or remove ModelCheckpoint from the lightning trainer" 883 ) 884 # Create the callback and attach it to trainer 885 if "filepath" in params: 886 if params.filepath is not None: 887 logging.warning("filepath is deprecated. Please switch to dirpath and filename instead") 888 if params.dirpath is None: 889 params.dirpath = Path(params.filepath).parent 890 if params.filename is None: 891 params.filename = Path(params.filepath).name 892 with open_dict(params): 893 del params["filepath"] 894 if params.dirpath is None: 895 params.dirpath = Path(log_dir / 'checkpoints') 896 if params.filename is None: 897 params.filename = f'{name}--{{{params.monitor}:.4f}}-{{epoch}}' 898 if params.prefix is None: 899 params.prefix = name 900 NeMoModelCheckpoint.CHECKPOINT_NAME_LAST = params.filename + '-last' 901 902 logging.debug(params.dirpath) 903 logging.debug(params.filename) 904 logging.debug(params.prefix) 905 906 if "val" in params.monitor: 907 if ( 908 trainer.max_epochs is not None 909 and trainer.max_epochs != -1 910 and trainer.max_epochs < trainer.check_val_every_n_epoch 911 ): 912 logging.error( 913 "The checkpoint callback was told to monitor a validation value but trainer.max_epochs(" 914 f"{trainer.max_epochs}) was less than trainer.check_val_every_n_epoch({trainer.check_val_every_n_epoch}" 915 f"). It is very likely this run will fail with ModelCheckpoint(monitor='{params.monitor}') not found " 916 "in the returned metrics. Please ensure that validation is run within trainer.max_epochs." 917 ) 918 elif trainer.max_steps is not None and trainer.max_steps != -1: 919 logging.warning( 920 "The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to " 921 f"{trainer.max_steps}. Please ensure that max_steps will run for at least " 922 f"{trainer.check_val_every_n_epoch} epochs to ensure that checkpointing will not error out." 923 ) 924 925 checkpoint_callback = NeMoModelCheckpoint(n_resume=resume, **params) 926 checkpoint_callback.last_model_path = trainer.ckpt_path or "" 927 if 'mp_rank' in checkpoint_callback.last_model_path or 'tp_rank' in checkpoint_callback.last_model_path: 928 checkpoint_callback.last_model_path = uninject_model_parallel_rank(checkpoint_callback.last_model_path) 929 trainer.callbacks.append(checkpoint_callback) 930 if create_preemption_callback: 931 # Check if cuda is avialable as preemption is supported only on GPUs 932 if torch.cuda.is_available(): 933 ## By default PreemptionCallback handles SIGTERM. To handle other signals pass the signal in the call as below: 934 ## PreemptionCallback(checkpoint_callback, signal.SIGCHLD) 935 preemption_callback = PreemptionCallback(checkpoint_callback) 936 trainer.callbacks.append(preemption_callback) 937 else: 938 logging.info("Preemption is supported only on GPUs, disabling preemption") 939 940 941 def check_slurm(trainer): 942 try: 943 return trainer.accelerator_connector.is_slurm_managing_tasks 944 except AttributeError: 945 return False 946 947 948 class StatelessTimer(Timer): 949 """Extension of PTL timers to be per run.""" 950 951 def __init__(self, duration: timedelta = None, interval: str = Interval.step, verbose: bool = True,) -> None: 952 super().__init__(duration, interval, verbose) 953 954 # Override PTL Timer's state dict to not store elapsed time information so that we can restore and continue training. 955 def state_dict(self) -> Dict[str, Any]: 956 return {} 957 958 def load_state_dict(self, state_dict: Dict[str, Any]) -> None: 959 return 960 961 962 def configure_no_restart_validation_training_loop(trainer: pytorch_lightning.Trainer) -> None: 963 if type(trainer.fit_loop.epoch_loop) != _TrainingEpochLoop: 964 warnings.warn("Detected custom epoch loop. Skipping no validation on restart support.", UserWarning) 965 return 966 ## Pass trainer object to avoid trainer getting overwritten as None 967 loop = SkipResumeTrainingValidationLoop(trainer, trainer.min_steps, trainer.max_steps) 968 trainer.fit_loop.epoch_loop = loop 969 970 971 class SkipResumeTrainingValidationLoop(_TrainingEpochLoop): 972 """ 973 Extend the PTL Epoch loop to skip validating when resuming. 974 This happens when resuming a checkpoint that has already run validation, but loading restores 975 the training state before validation has run. 976 """ 977 978 def _should_check_val_fx(self) -> bool: 979 if self.restarting and self.global_step % self.trainer.val_check_batch == 0: 980 return False 981 return super()._should_check_val_fx() 982 983 984 def clean_exp_ckpt(exp_log_dir: Union[str, Path], remove_ckpt: bool = True, remove_nemo: bool = False): 985 """ 986 Helper method that removes Pytorch Lightning .ckpt files or NeMo .nemo files from the checkpoint directory 987 988 Args: 989 exp_log_dir: str path to the root directory of the current experiment. 990 remove_ckpt: bool, whether to remove all *.ckpt files in the checkpoints directory. 991 remove_nemo: bool, whether to remove all *.nemo files in the checkpoints directory. 992 """ 993 exp_log_dir = str(exp_log_dir) 994 995 if remove_ckpt: 996 logging.info("Deleting *.ckpt files ...") 997 ckpt_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.ckpt")) 998 for filepath in ckpt_files: 999 os.remove(filepath) 1000 logging.info(f"Deleted file : {filepath}") 1001 1002 if remove_nemo: 1003 logging.info("Deleting *.nemo files ...") 1004 nemo_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.nemo")) 1005 for filepath in nemo_files: 1006 os.remove(filepath) 1007 logging.info(f"Deleted file : {filepath}") 1008 [end of nemo/utils/exp_manager.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR model with CTC decoder. To evaluate a model with 19 # Transducer (RNN-T) decoder use another script 'scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py'. 20 # NeMo's beam search decoders are capable of using the KenLM's N-gram models 21 # to find the best candidates. This script supports both character level and BPE level 22 # encodings and models which is detected automatically from the type of the model. 23 # You may train the LM model with 'scripts/asr_language_modeling/ngram_lm/train_kenlm.py'. 24 25 # Config Help 26 27 To discover all arguments of the script, please run : 28 python eval_beamsearch_ngram.py --help 29 python eval_beamsearch_ngram.py --cfg job 30 31 # USAGE 32 33 python eval_beamsearch_ngram.py nemo_model_file=<path to the .nemo file of the model> \ 34 input_manifest=<path to the evaluation JSON manifest file> \ 35 kenlm_model_file=<path to the binary KenLM model> \ 36 beam_width=[<list of the beam widths, separated with commas>] \ 37 beam_alpha=[<list of the beam alphas, separated with commas>] \ 38 beam_beta=[<list of the beam betas, separated with commas>] \ 39 preds_output_folder=<optional folder to store the predictions> \ 40 probs_cache_file=null \ 41 decoding_mode=beamsearch_ngram 42 ... 43 44 45 # Grid Search for Hyper parameters 46 47 For grid search, you can provide a list of arguments as follows - 48 49 beam_width=[4,8,16,....] \ 50 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 51 beam_beta=[-1.0,-0.5,0.0,...,1.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 from dataclasses import dataclass, field, is_dataclass 64 from pathlib import Path 65 from typing import List, Optional 66 67 import editdistance 68 import numpy as np 69 import torch 70 from omegaconf import MISSING, OmegaConf 71 from sklearn.model_selection import ParameterGrid 72 from tqdm.auto import tqdm 73 74 import nemo.collections.asr as nemo_asr 75 from nemo.collections.asr.models import EncDecHybridRNNTCTCModel 76 from nemo.collections.asr.parts.submodules import ctc_beam_decoding 77 from nemo.collections.asr.parts.utils.transcribe_utils import PunctuationCapitalization, TextProcessingConfig 78 from nemo.core.config import hydra_runner 79 from nemo.utils import logging 80 81 # fmt: off 82 83 84 @dataclass 85 class EvalBeamSearchNGramConfig: 86 """ 87 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 88 """ 89 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 90 nemo_model_file: str = MISSING 91 92 # File paths 93 input_manifest: str = MISSING # The manifest file of the evaluation set 94 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 95 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 96 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 97 98 # Parameters for inference 99 acoustic_batch_size: int = 16 # The batch size to calculate log probabilities 100 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 101 device: str = "cuda" # The device to load the model onto to calculate log probabilities 102 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 103 104 # Beam Search hyperparameters 105 106 # The decoding scheme to be used for evaluation. 107 # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] 108 decoding_mode: str = "beamsearch_ngram" 109 110 beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding 111 beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding 112 beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding 113 114 decoding_strategy: str = "beam" 115 decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) 116 117 text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( 118 punctuation_marks = ".,?", 119 separate_punctuation = False, 120 do_lowercase = False, 121 rm_punctuation = False, 122 ) 123 # fmt: on 124 125 126 def beam_search_eval( 127 model: nemo_asr.models.ASRModel, 128 cfg: EvalBeamSearchNGramConfig, 129 all_probs: List[torch.Tensor], 130 target_transcripts: List[str], 131 preds_output_file: str = None, 132 lm_path: str = None, 133 beam_alpha: float = 1.0, 134 beam_beta: float = 0.0, 135 beam_width: int = 128, 136 beam_batch_size: int = 128, 137 progress_bar: bool = True, 138 punctuation_capitalization: PunctuationCapitalization = None, 139 ): 140 level = logging.getEffectiveLevel() 141 logging.setLevel(logging.CRITICAL) 142 # Reset config 143 model.change_decoding_strategy(None) 144 145 # Override the beam search config with current search candidate configuration 146 cfg.decoding.beam_size = beam_width 147 cfg.decoding.beam_alpha = beam_alpha 148 cfg.decoding.beam_beta = beam_beta 149 cfg.decoding.return_best_hypothesis = False 150 cfg.decoding.kenlm_path = cfg.kenlm_model_file 151 152 # Update model's decoding strategy config 153 model.cfg.decoding.strategy = cfg.decoding_strategy 154 model.cfg.decoding.beam = cfg.decoding 155 156 # Update model's decoding strategy 157 if isinstance(model, EncDecHybridRNNTCTCModel): 158 model.change_decoding_strategy(model.cfg.decoding, decoder_type='ctc') 159 decoding = model.ctc_decoding 160 else: 161 model.change_decoding_strategy(model.cfg.decoding) 162 decoding = model.decoding 163 logging.setLevel(level) 164 165 wer_dist_first = cer_dist_first = 0 166 wer_dist_best = cer_dist_best = 0 167 words_count = 0 168 chars_count = 0 169 sample_idx = 0 170 if preds_output_file: 171 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 172 173 if progress_bar: 174 it = tqdm( 175 range(int(np.ceil(len(all_probs) / beam_batch_size))), 176 desc=f"Beam search decoding with width={beam_width}, alpha={beam_alpha}, beta={beam_beta}", 177 ncols=120, 178 ) 179 else: 180 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 181 for batch_idx in it: 182 # disabling type checking 183 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 184 probs_lens = torch.tensor([prob.shape[0] for prob in probs_batch]) 185 with torch.no_grad(): 186 packed_batch = torch.zeros(len(probs_batch), max(probs_lens), probs_batch[0].shape[-1], device='cpu') 187 188 for prob_index in range(len(probs_batch)): 189 packed_batch[prob_index, : probs_lens[prob_index], :] = torch.tensor( 190 probs_batch[prob_index], device=packed_batch.device, dtype=packed_batch.dtype 191 ) 192 193 _, beams_batch = decoding.ctc_decoder_predictions_tensor( 194 packed_batch, decoder_lengths=probs_lens, return_hypotheses=True, 195 ) 196 197 for beams_idx, beams in enumerate(beams_batch): 198 target = target_transcripts[sample_idx + beams_idx] 199 target_split_w = target.split() 200 target_split_c = list(target) 201 words_count += len(target_split_w) 202 chars_count += len(target_split_c) 203 wer_dist_min = cer_dist_min = 10000 204 for candidate_idx, candidate in enumerate(beams): # type: (int, ctc_beam_decoding.rnnt_utils.Hypothesis) 205 pred_text = candidate.text 206 if cfg.text_processing.do_lowercase: 207 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 208 if cfg.text_processing.rm_punctuation: 209 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 210 if cfg.text_processing.separate_punctuation: 211 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 212 pred_split_w = pred_text.split() 213 wer_dist = editdistance.eval(target_split_w, pred_split_w) 214 pred_split_c = list(pred_text) 215 cer_dist = editdistance.eval(target_split_c, pred_split_c) 216 217 wer_dist_min = min(wer_dist_min, wer_dist) 218 cer_dist_min = min(cer_dist_min, cer_dist) 219 220 if candidate_idx == 0: 221 # first candidate 222 wer_dist_first += wer_dist 223 cer_dist_first += cer_dist 224 225 score = candidate.score 226 if preds_output_file: 227 out_file.write('{}\t{}\n'.format(pred_text, score)) 228 wer_dist_best += wer_dist_min 229 cer_dist_best += cer_dist_min 230 sample_idx += len(probs_batch) 231 232 if preds_output_file: 233 out_file.close() 234 logging.info(f"Stored the predictions of beam search decoding at '{preds_output_file}'.") 235 236 if lm_path: 237 logging.info( 238 'WER/CER with beam search decoding and N-gram model = {:.2%}/{:.2%}'.format( 239 wer_dist_first / words_count, cer_dist_first / chars_count 240 ) 241 ) 242 else: 243 logging.info( 244 'WER/CER with beam search decoding = {:.2%}/{:.2%}'.format( 245 wer_dist_first / words_count, cer_dist_first / chars_count 246 ) 247 ) 248 logging.info( 249 'Oracle WER/CER in candidates with perfect LM= {:.2%}/{:.2%}'.format( 250 wer_dist_best / words_count, cer_dist_best / chars_count 251 ) 252 ) 253 logging.info(f"=================================================================================") 254 255 return wer_dist_first / words_count, cer_dist_first / chars_count 256 257 258 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 259 def main(cfg: EvalBeamSearchNGramConfig): 260 logging.warning("This file will be renamed to eval_beamsearch_ngram_ctc.py in the future NeMo (1.21) release.") 261 if is_dataclass(cfg): 262 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 263 264 valid_decoding_modes = ["greedy", "beamsearch", "beamsearch_ngram"] 265 if cfg.decoding_mode not in valid_decoding_modes: 266 raise ValueError( 267 f"Given decoding_mode={cfg.decoding_mode} is invalid. Available options are :\n" f"{valid_decoding_modes}" 268 ) 269 270 if cfg.nemo_model_file.endswith('.nemo'): 271 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 272 else: 273 logging.warning( 274 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 275 ) 276 asr_model = nemo_asr.models.ASRModel.from_pretrained( 277 cfg.nemo_model_file, map_location=torch.device(cfg.device) 278 ) 279 280 target_transcripts = [] 281 manifest_dir = Path(cfg.input_manifest).parent 282 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 283 audio_file_paths = [] 284 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 285 data = json.loads(line) 286 audio_file = Path(data['audio_filepath']) 287 if not audio_file.is_file() and not audio_file.is_absolute(): 288 audio_file = manifest_dir / audio_file 289 target_transcripts.append(data['text']) 290 audio_file_paths.append(str(audio_file.absolute())) 291 292 punctuation_capitalization = PunctuationCapitalization(cfg.text_processing.punctuation_marks) 293 if cfg.text_processing.do_lowercase: 294 target_transcripts = punctuation_capitalization.do_lowercase(target_transcripts) 295 if cfg.text_processing.rm_punctuation: 296 target_transcripts = punctuation_capitalization.rm_punctuation(target_transcripts) 297 if cfg.text_processing.separate_punctuation: 298 target_transcripts = punctuation_capitalization.separate_punctuation(target_transcripts) 299 300 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 301 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 302 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 303 with open(cfg.probs_cache_file, 'rb') as probs_file: 304 all_probs = pickle.load(probs_file) 305 306 if len(all_probs) != len(audio_file_paths): 307 raise ValueError( 308 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 309 f"match the manifest file. You may need to delete the probabilities cached file." 310 ) 311 else: 312 313 @contextlib.contextmanager 314 def default_autocast(): 315 yield 316 317 if cfg.use_amp: 318 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 319 logging.info("AMP is enabled!\n") 320 autocast = torch.cuda.amp.autocast 321 322 else: 323 autocast = default_autocast 324 else: 325 326 autocast = default_autocast 327 328 with autocast(): 329 with torch.no_grad(): 330 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 331 asr_model.cur_decoder = 'ctc' 332 all_logits = asr_model.transcribe(audio_file_paths, batch_size=cfg.acoustic_batch_size, logprobs=True) 333 334 all_probs = all_logits 335 if cfg.probs_cache_file: 336 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 337 with open(cfg.probs_cache_file, 'wb') as f_dump: 338 pickle.dump(all_probs, f_dump) 339 340 wer_dist_greedy = 0 341 cer_dist_greedy = 0 342 words_count = 0 343 chars_count = 0 344 for batch_idx, probs in enumerate(all_probs): 345 preds = np.argmax(probs, axis=1) 346 preds_tensor = torch.tensor(preds, device='cpu').unsqueeze(0) 347 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 348 pred_text = asr_model.ctc_decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 349 else: 350 pred_text = asr_model._wer.decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 351 352 if cfg.text_processing.do_lowercase: 353 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 354 if cfg.text_processing.rm_punctuation: 355 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 356 if cfg.text_processing.separate_punctuation: 357 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 358 359 pred_split_w = pred_text.split() 360 target_split_w = target_transcripts[batch_idx].split() 361 pred_split_c = list(pred_text) 362 target_split_c = list(target_transcripts[batch_idx]) 363 364 wer_dist = editdistance.eval(target_split_w, pred_split_w) 365 cer_dist = editdistance.eval(target_split_c, pred_split_c) 366 367 wer_dist_greedy += wer_dist 368 cer_dist_greedy += cer_dist 369 words_count += len(target_split_w) 370 chars_count += len(target_split_c) 371 372 logging.info('Greedy WER/CER = {:.2%}/{:.2%}'.format(wer_dist_greedy / words_count, cer_dist_greedy / chars_count)) 373 374 asr_model = asr_model.to('cpu') 375 376 if cfg.decoding_mode == "beamsearch_ngram": 377 if not os.path.exists(cfg.kenlm_model_file): 378 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 379 lm_path = cfg.kenlm_model_file 380 else: 381 lm_path = None 382 383 # 'greedy' decoding_mode would skip the beam search decoding 384 if cfg.decoding_mode in ["beamsearch_ngram", "beamsearch"]: 385 if cfg.beam_width is None or cfg.beam_alpha is None or cfg.beam_beta is None: 386 raise ValueError("beam_width, beam_alpha and beam_beta are needed to perform beam search decoding.") 387 params = {'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'beam_beta': cfg.beam_beta} 388 hp_grid = ParameterGrid(params) 389 hp_grid = list(hp_grid) 390 391 best_wer_beam_size, best_cer_beam_size = None, None 392 best_wer_alpha, best_cer_alpha = None, None 393 best_wer_beta, best_cer_beta = None, None 394 best_wer, best_cer = 1e6, 1e6 395 396 logging.info(f"==============================Starting the beam search decoding===============================") 397 logging.info(f"Grid search size: {len(hp_grid)}") 398 logging.info(f"It may take some time...") 399 logging.info(f"==============================================================================================") 400 401 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 402 os.mkdir(cfg.preds_output_folder) 403 for hp in hp_grid: 404 if cfg.preds_output_folder: 405 preds_output_file = os.path.join( 406 cfg.preds_output_folder, 407 f"preds_out_width{hp['beam_width']}_alpha{hp['beam_alpha']}_beta{hp['beam_beta']}.tsv", 408 ) 409 else: 410 preds_output_file = None 411 412 candidate_wer, candidate_cer = beam_search_eval( 413 asr_model, 414 cfg, 415 all_probs=all_probs, 416 target_transcripts=target_transcripts, 417 preds_output_file=preds_output_file, 418 lm_path=lm_path, 419 beam_width=hp["beam_width"], 420 beam_alpha=hp["beam_alpha"], 421 beam_beta=hp["beam_beta"], 422 beam_batch_size=cfg.beam_batch_size, 423 progress_bar=True, 424 punctuation_capitalization=punctuation_capitalization, 425 ) 426 427 if candidate_cer < best_cer: 428 best_cer_beam_size = hp["beam_width"] 429 best_cer_alpha = hp["beam_alpha"] 430 best_cer_beta = hp["beam_beta"] 431 best_cer = candidate_cer 432 433 if candidate_wer < best_wer: 434 best_wer_beam_size = hp["beam_width"] 435 best_wer_alpha = hp["beam_alpha"] 436 best_wer_beta = hp["beam_beta"] 437 best_wer = candidate_wer 438 439 logging.info( 440 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 441 f'Beam alpha = {best_wer_alpha}, Beam beta = {best_wer_beta}' 442 ) 443 444 logging.info( 445 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 446 f'Beam alpha = {best_cer_alpha}, Beam beta = {best_cer_beta}' 447 ) 448 logging.info(f"=================================================================================") 449 450 451 if __name__ == '__main__': 452 main() 453 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR Transducer model. NeMo's beam search decoders are capable of using the 19 # KenLM's N-gram models to find the best candidates. This script supports both character level and BPE level 20 # encodings and models which is detected automatically from the type of the model. 21 # You may train the LM model with 'scripts/ngram_lm/train_kenlm.py'. 22 23 # Config Help 24 25 To discover all arguments of the script, please run : 26 python eval_beamsearch_ngram.py --help 27 python eval_beamsearch_ngram.py --cfg job 28 29 # USAGE 30 31 python eval_beamsearch_ngram_transducer.py nemo_model_file=<path to the .nemo file of the model> \ 32 input_manifest=<path to the evaluation JSON manifest file \ 33 kenlm_model_file=<path to the binary KenLM model> \ 34 beam_width=[<list of the beam widths, separated with commas>] \ 35 beam_alpha=[<list of the beam alphas, separated with commas>] \ 36 preds_output_folder=<optional folder to store the predictions> \ 37 probs_cache_file=null \ 38 decoding_strategy=<greedy_batch or maes decoding> 39 maes_prefix_alpha=[<list of the maes prefix alphas, separated with commas>] \ 40 maes_expansion_gamma=[<list of the maes expansion gammas, separated with commas>] \ 41 hat_subtract_ilm=<in case of HAT model: subtract internal LM or not> \ 42 hat_ilm_weight=[<in case of HAT model: list of the HAT internal LM weights, separated with commas>] \ 43 ... 44 45 46 # Grid Search for Hyper parameters 47 48 For grid search, you can provide a list of arguments as follows - 49 50 beam_width=[4,8,16,....] \ 51 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 import tempfile 64 from dataclasses import dataclass, field, is_dataclass 65 from pathlib import Path 66 from typing import List, Optional 67 68 import editdistance 69 import numpy as np 70 import torch 71 from omegaconf import MISSING, OmegaConf 72 from sklearn.model_selection import ParameterGrid 73 from tqdm.auto import tqdm 74 75 import nemo.collections.asr as nemo_asr 76 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding 77 from nemo.core.config import hydra_runner 78 from nemo.utils import logging 79 80 # fmt: off 81 82 83 @dataclass 84 class EvalBeamSearchNGramConfig: 85 """ 86 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 87 """ 88 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 89 nemo_model_file: str = MISSING 90 91 # File paths 92 input_manifest: str = MISSING # The manifest file of the evaluation set 93 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 94 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 95 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 96 97 # Parameters for inference 98 acoustic_batch_size: int = 128 # The batch size to calculate log probabilities 99 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 100 device: str = "cuda" # The device to load the model onto to calculate log probabilities 101 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 102 num_workers: int = 1 # Number of workers for DataLoader 103 104 # The decoding scheme to be used for evaluation 105 decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] 106 107 # Beam Search hyperparameters 108 beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding 109 beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding 110 111 maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding 112 maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding 113 114 # HAT related parameters (only for internal lm subtraction) 115 hat_subtract_ilm: bool = False 116 hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) 117 118 decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) 119 120 121 # fmt: on 122 123 124 def decoding_step( 125 model: nemo_asr.models.ASRModel, 126 cfg: EvalBeamSearchNGramConfig, 127 all_probs: List[torch.Tensor], 128 target_transcripts: List[str], 129 preds_output_file: str = None, 130 beam_batch_size: int = 128, 131 progress_bar: bool = True, 132 ): 133 level = logging.getEffectiveLevel() 134 logging.setLevel(logging.CRITICAL) 135 # Reset config 136 model.change_decoding_strategy(None) 137 138 cfg.decoding.hat_ilm_weight = cfg.decoding.hat_ilm_weight * cfg.hat_subtract_ilm 139 # Override the beam search config with current search candidate configuration 140 cfg.decoding.return_best_hypothesis = False 141 cfg.decoding.ngram_lm_model = cfg.kenlm_model_file 142 cfg.decoding.hat_subtract_ilm = cfg.hat_subtract_ilm 143 144 # Update model's decoding strategy config 145 model.cfg.decoding.strategy = cfg.decoding_strategy 146 model.cfg.decoding.beam = cfg.decoding 147 148 # Update model's decoding strategy 149 model.change_decoding_strategy(model.cfg.decoding) 150 logging.setLevel(level) 151 152 wer_dist_first = cer_dist_first = 0 153 wer_dist_best = cer_dist_best = 0 154 words_count = 0 155 chars_count = 0 156 sample_idx = 0 157 if preds_output_file: 158 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 159 160 if progress_bar: 161 if cfg.decoding_strategy == "greedy_batch": 162 description = "Greedy_batch decoding.." 163 else: 164 description = f"{cfg.decoding_strategy} decoding with bw={cfg.decoding.beam_size}, ba={cfg.decoding.ngram_lm_alpha}, ma={cfg.decoding.maes_prefix_alpha}, mg={cfg.decoding.maes_expansion_gamma}, hat_ilmw={cfg.decoding.hat_ilm_weight}" 165 it = tqdm(range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=description, ncols=120) 166 else: 167 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 168 for batch_idx in it: 169 # disabling type checking 170 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 171 probs_lens = torch.tensor([prob.shape[-1] for prob in probs_batch]) 172 with torch.no_grad(): 173 packed_batch = torch.zeros(len(probs_batch), probs_batch[0].shape[0], max(probs_lens), device='cpu') 174 175 for prob_index in range(len(probs_batch)): 176 packed_batch[prob_index, :, : probs_lens[prob_index]] = torch.tensor( 177 probs_batch[prob_index].unsqueeze(0), device=packed_batch.device, dtype=packed_batch.dtype 178 ) 179 best_hyp_batch, beams_batch = model.decoding.rnnt_decoder_predictions_tensor( 180 packed_batch, probs_lens, return_hypotheses=True, 181 ) 182 if cfg.decoding_strategy == "greedy_batch": 183 beams_batch = [[x] for x in best_hyp_batch] 184 185 for beams_idx, beams in enumerate(beams_batch): 186 target = target_transcripts[sample_idx + beams_idx] 187 target_split_w = target.split() 188 target_split_c = list(target) 189 words_count += len(target_split_w) 190 chars_count += len(target_split_c) 191 wer_dist_min = cer_dist_min = 10000 192 for candidate_idx, candidate in enumerate(beams): # type: (int, rnnt_beam_decoding.rnnt_utils.Hypothesis) 193 pred_text = candidate.text 194 pred_split_w = pred_text.split() 195 wer_dist = editdistance.eval(target_split_w, pred_split_w) 196 pred_split_c = list(pred_text) 197 cer_dist = editdistance.eval(target_split_c, pred_split_c) 198 199 wer_dist_min = min(wer_dist_min, wer_dist) 200 cer_dist_min = min(cer_dist_min, cer_dist) 201 202 if candidate_idx == 0: 203 # first candidate 204 wer_dist_first += wer_dist 205 cer_dist_first += cer_dist 206 207 score = candidate.score 208 if preds_output_file: 209 out_file.write('{}\t{}\n'.format(pred_text, score)) 210 wer_dist_best += wer_dist_min 211 cer_dist_best += cer_dist_min 212 sample_idx += len(probs_batch) 213 214 if cfg.decoding_strategy == "greedy_batch": 215 return wer_dist_first / words_count, cer_dist_first / chars_count 216 217 if preds_output_file: 218 out_file.close() 219 logging.info(f"Stored the predictions of {cfg.decoding_strategy} decoding at '{preds_output_file}'.") 220 221 if cfg.decoding.ngram_lm_model: 222 logging.info( 223 f"WER/CER with {cfg.decoding_strategy} decoding and N-gram model = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 224 ) 225 else: 226 logging.info( 227 f"WER/CER with {cfg.decoding_strategy} decoding = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 228 ) 229 logging.info( 230 f"Oracle WER/CER in candidates with perfect LM= {wer_dist_best / words_count:.2%}/{cer_dist_best / chars_count:.2%}" 231 ) 232 logging.info(f"=================================================================================") 233 234 return wer_dist_first / words_count, cer_dist_first / chars_count 235 236 237 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 238 def main(cfg: EvalBeamSearchNGramConfig): 239 if is_dataclass(cfg): 240 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 241 242 valid_decoding_strategis = ["greedy_batch", "beam", "tsd", "alsd", "maes"] 243 if cfg.decoding_strategy not in valid_decoding_strategis: 244 raise ValueError( 245 f"Given decoding_strategy={cfg.decoding_strategy} is invalid. Available options are :\n" 246 f"{valid_decoding_strategis}" 247 ) 248 249 if cfg.nemo_model_file.endswith('.nemo'): 250 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 251 else: 252 logging.warning( 253 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 254 ) 255 asr_model = nemo_asr.models.ASRModel.from_pretrained( 256 cfg.nemo_model_file, map_location=torch.device(cfg.device) 257 ) 258 259 if cfg.kenlm_model_file: 260 if not os.path.exists(cfg.kenlm_model_file): 261 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 262 if cfg.decoding_strategy != "maes": 263 raise ValueError(f"Decoding with kenlm model is supported only for maes decoding algorithm.") 264 lm_path = cfg.kenlm_model_file 265 else: 266 lm_path = None 267 cfg.beam_alpha = [0.0] 268 if cfg.hat_subtract_ilm: 269 assert lm_path, "kenlm must be set for hat internal lm subtraction" 270 271 if cfg.decoding_strategy != "maes": 272 cfg.maes_prefix_alpha, cfg.maes_expansion_gamma, cfg.hat_ilm_weight = [0], [0], [0] 273 274 target_transcripts = [] 275 manifest_dir = Path(cfg.input_manifest).parent 276 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 277 audio_file_paths = [] 278 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 279 data = json.loads(line) 280 audio_file = Path(data['audio_filepath']) 281 if not audio_file.is_file() and not audio_file.is_absolute(): 282 audio_file = manifest_dir / audio_file 283 target_transcripts.append(data['text']) 284 audio_file_paths.append(str(audio_file.absolute())) 285 286 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 287 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 288 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 289 with open(cfg.probs_cache_file, 'rb') as probs_file: 290 all_probs = pickle.load(probs_file) 291 292 if len(all_probs) != len(audio_file_paths): 293 raise ValueError( 294 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 295 f"match the manifest file. You may need to delete the probabilities cached file." 296 ) 297 else: 298 299 @contextlib.contextmanager 300 def default_autocast(): 301 yield 302 303 if cfg.use_amp: 304 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 305 logging.info("AMP is enabled!\n") 306 autocast = torch.cuda.amp.autocast 307 308 else: 309 autocast = default_autocast 310 else: 311 312 autocast = default_autocast 313 314 # manual calculation of encoder_embeddings 315 with autocast(): 316 with torch.no_grad(): 317 asr_model.eval() 318 asr_model.encoder.freeze() 319 device = next(asr_model.parameters()).device 320 all_probs = [] 321 with tempfile.TemporaryDirectory() as tmpdir: 322 with open(os.path.join(tmpdir, 'manifest.json'), 'w', encoding='utf-8') as fp: 323 for audio_file in audio_file_paths: 324 entry = {'audio_filepath': audio_file, 'duration': 100000, 'text': ''} 325 fp.write(json.dumps(entry) + '\n') 326 config = { 327 'paths2audio_files': audio_file_paths, 328 'batch_size': cfg.acoustic_batch_size, 329 'temp_dir': tmpdir, 330 'num_workers': cfg.num_workers, 331 'channel_selector': None, 332 'augmentor': None, 333 } 334 temporary_datalayer = asr_model._setup_transcribe_dataloader(config) 335 for test_batch in tqdm(temporary_datalayer, desc="Transcribing", disable=True): 336 encoded, encoded_len = asr_model.forward( 337 input_signal=test_batch[0].to(device), input_signal_length=test_batch[1].to(device) 338 ) 339 # dump encoder embeddings per file 340 for idx in range(encoded.shape[0]): 341 encoded_no_pad = encoded[idx, :, : encoded_len[idx]] 342 all_probs.append(encoded_no_pad) 343 344 if cfg.probs_cache_file: 345 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 346 with open(cfg.probs_cache_file, 'wb') as f_dump: 347 pickle.dump(all_probs, f_dump) 348 349 if cfg.decoding_strategy == "greedy_batch": 350 asr_model = asr_model.to('cpu') 351 candidate_wer, candidate_cer = decoding_step( 352 asr_model, 353 cfg, 354 all_probs=all_probs, 355 target_transcripts=target_transcripts, 356 beam_batch_size=cfg.beam_batch_size, 357 progress_bar=True, 358 ) 359 logging.info(f"Greedy batch WER/CER = {candidate_wer:.2%}/{candidate_cer:.2%}") 360 361 asr_model = asr_model.to('cpu') 362 363 # 'greedy_batch' decoding_strategy would skip the beam search decoding 364 if cfg.decoding_strategy in ["beam", "tsd", "alsd", "maes"]: 365 if cfg.beam_width is None or cfg.beam_alpha is None: 366 raise ValueError("beam_width and beam_alpha are needed to perform beam search decoding.") 367 params = { 368 'beam_width': cfg.beam_width, 369 'beam_alpha': cfg.beam_alpha, 370 'maes_prefix_alpha': cfg.maes_prefix_alpha, 371 'maes_expansion_gamma': cfg.maes_expansion_gamma, 372 'hat_ilm_weight': cfg.hat_ilm_weight, 373 } 374 hp_grid = ParameterGrid(params) 375 hp_grid = list(hp_grid) 376 377 best_wer_beam_size, best_cer_beam_size = None, None 378 best_wer_alpha, best_cer_alpha = None, None 379 best_wer, best_cer = 1e6, 1e6 380 381 logging.info( 382 f"==============================Starting the {cfg.decoding_strategy} decoding===============================" 383 ) 384 logging.info(f"Grid search size: {len(hp_grid)}") 385 logging.info(f"It may take some time...") 386 logging.info(f"==============================================================================================") 387 388 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 389 os.mkdir(cfg.preds_output_folder) 390 for hp in hp_grid: 391 if cfg.preds_output_folder: 392 results_file = f"preds_out_{cfg.decoding_strategy}_bw{hp['beam_width']}" 393 if cfg.decoding_strategy == "maes": 394 results_file = f"{results_file}_ma{hp['maes_prefix_alpha']}_mg{hp['maes_expansion_gamma']}" 395 if cfg.kenlm_model_file: 396 results_file = f"{results_file}_ba{hp['beam_alpha']}" 397 if cfg.hat_subtract_ilm: 398 results_file = f"{results_file}_hat_ilmw{hp['hat_ilm_weight']}" 399 preds_output_file = os.path.join(cfg.preds_output_folder, f"{results_file}.tsv") 400 else: 401 preds_output_file = None 402 403 cfg.decoding.beam_size = hp["beam_width"] 404 cfg.decoding.ngram_lm_alpha = hp["beam_alpha"] 405 cfg.decoding.maes_prefix_alpha = hp["maes_prefix_alpha"] 406 cfg.decoding.maes_expansion_gamma = hp["maes_expansion_gamma"] 407 cfg.decoding.hat_ilm_weight = hp["hat_ilm_weight"] 408 409 candidate_wer, candidate_cer = decoding_step( 410 asr_model, 411 cfg, 412 all_probs=all_probs, 413 target_transcripts=target_transcripts, 414 preds_output_file=preds_output_file, 415 beam_batch_size=cfg.beam_batch_size, 416 progress_bar=True, 417 ) 418 419 if candidate_cer < best_cer: 420 best_cer_beam_size = hp["beam_width"] 421 best_cer_alpha = hp["beam_alpha"] 422 best_cer_ma = hp["maes_prefix_alpha"] 423 best_cer_mg = hp["maes_expansion_gamma"] 424 best_cer_hat_ilm_weight = hp["hat_ilm_weight"] 425 best_cer = candidate_cer 426 427 if candidate_wer < best_wer: 428 best_wer_beam_size = hp["beam_width"] 429 best_wer_alpha = hp["beam_alpha"] 430 best_wer_ma = hp["maes_prefix_alpha"] 431 best_wer_ga = hp["maes_expansion_gamma"] 432 best_wer_hat_ilm_weight = hp["hat_ilm_weight"] 433 best_wer = candidate_wer 434 435 wer_hat_parameter = "" 436 if cfg.hat_subtract_ilm: 437 wer_hat_parameter = f"HAT ilm weight = {best_wer_hat_ilm_weight}, " 438 logging.info( 439 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 440 f'Beam alpha = {best_wer_alpha}, {wer_hat_parameter}' 441 f'maes_prefix_alpha = {best_wer_ma}, maes_expansion_gamma = {best_wer_ga} ' 442 ) 443 444 cer_hat_parameter = "" 445 if cfg.hat_subtract_ilm: 446 cer_hat_parameter = f"HAT ilm weight = {best_cer_hat_ilm_weight}" 447 logging.info( 448 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 449 f'Beam alpha = {best_cer_alpha}, {cer_hat_parameter} ' 450 f'maes_prefix_alpha = {best_cer_ma}, maes_expansion_gamma = {best_cer_mg}' 451 ) 452 logging.info(f"=================================================================================") 453 454 455 if __name__ == '__main__': 456 main() 457 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of scripts/confidence_ensembles/build_ensemble.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 This script provides a functionality to create confidence-based ensembles 17 from a collection of pretrained models. 18 19 For more details see the paper https://arxiv.org/abs/2306.15824 20 or tutorial in tutorials/asr/Confidence_Ensembles.ipynb 21 22 You would typically use this script by providing a yaml config file or overriding 23 default options from command line. 24 25 Usage examples: 26 27 1. Building an ensemble of two monolingual models with default settings (no confidence tuning). 28 29 python build_ensemble.py --config-path=. --config-name=ensemble_config.yaml 30 ensemble.0.model=stt_it_conformer_ctc_large 31 ensemble.0.training_manifest=<path to the Italian data of 100+ utterances (no transcription required)> 32 ensemble.1.model=stt_es_conformer_ctc_large 33 ensemble.1.training_manifest=<path to the Spanish data of 100+ utterances (no transcription required)> 34 output_path=<path to the desired location of the .nemo checkpoint> 35 36 You can have more than 2 models and can control transcription settings (e.g., batch size) 37 with ``transcription.<any argument of examples/asr/transcribe_speech.py>`` parameters. 38 39 2. If you want to get improved results, you can enable tuning of the confidence and logistic regression (LR) parameters. 40 E.g. 41 42 python build_ensemble.py 43 <all arguments like in the previous example> 44 ensemble.0.dev_manifest=<path to the dev data that's required for tuning> 45 ... 46 # IMPORTANT: see the note below if you use > 2 models! 47 ensemble.N.dev_manifest=<path to the dev data that's required for tuning> 48 tune_confidence=True # to allow confidence tuning. LR is tuned by default 49 50 As with any tuning, it is recommended to have reasonably large validation set for each model, 51 otherwise you might overfit to the validation data. 52 53 Note that if you add additional models (> 2) you will need to modify ensemble_config.yaml 54 or create a new one with added models in there. While it's theoretically possible to 55 fully override such parameters from commandline, hydra is very unfriendly for such 56 use-cases, so it's strongly recommended to be creating new configs. 57 58 3. If you want to precisely control tuning grid search, you can do that with 59 60 python build_ensemble.py 61 <all arguments as in the previous examples> 62 tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set 63 tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set 64 65 You can check the dataclasses in this file for the full list of supported 66 arguments and their default values. 67 """ 68 69 import atexit 70 71 # using default logging to be able to silence unnecessary messages from nemo 72 import logging 73 import os 74 import random 75 import sys 76 import tempfile 77 from copy import deepcopy 78 from dataclasses import dataclass 79 from pathlib import Path 80 from typing import Dict, List, Optional, Tuple 81 82 import joblib 83 import numpy as np 84 import pytorch_lightning as pl 85 from omegaconf import MISSING, DictConfig, OmegaConf 86 from sklearn.linear_model import LogisticRegression 87 from sklearn.metrics import confusion_matrix 88 from sklearn.pipeline import Pipeline, make_pipeline 89 from sklearn.preprocessing import StandardScaler 90 from tqdm import tqdm 91 92 from nemo.collections.asr.models.confidence_ensemble import ( 93 ConfidenceEnsembleModel, 94 ConfidenceSpec, 95 compute_confidence, 96 get_filtered_logprobs, 97 ) 98 from nemo.collections.asr.parts.utils.asr_confidence_utils import ( 99 ConfidenceConfig, 100 ConfidenceMeasureConfig, 101 get_confidence_aggregation_bank, 102 get_confidence_measure_bank, 103 ) 104 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 105 from nemo.core.config import hydra_runner 106 107 LOG = logging.getLogger(__file__) 108 109 # adding Python path. If not found, asking user to get the file 110 try: 111 sys.path.append(str(Path(__file__).parents[2] / "examples" / "asr")) 112 import transcribe_speech 113 except ImportError: 114 # if users run script normally from nemo repo, this shouldn't be triggered as 115 # we modify the path above. But if they downloaded the build_ensemble.py as 116 # an isolated script, we'd ask them to also download corresponding version 117 # of the transcribe_speech.py 118 print( 119 "Current script depends on 'examples/asr/transcribe_speech.py', but can't find it. " 120 "If it's not present, download it from the NeMo github manually and put inside this folder." 121 ) 122 123 124 @dataclass 125 class EnsembleConfig: 126 # .nemo path or pretrained name 127 model: str = MISSING 128 # path to the training data manifest (non-tarred) 129 training_manifest: str = MISSING 130 # specify to limit the number of training samples 131 # 100 is most likely enough, but setting higher default just in case 132 max_training_samples: int = 1000 133 # specify to provide dev data manifest for HP tuning 134 dev_manifest: Optional[str] = None 135 136 137 @dataclass 138 class TuneConfidenceConfig: 139 # important parameter, so should always be tuned 140 exclude_blank: Tuple[bool] = (True, False) 141 # prod is pretty much always worse, so not including by default 142 aggregation: Tuple[str] = ("mean", "min", "max") 143 # not including max prob, as there is always an entropy-based metric 144 # that's better but otherwise including everything 145 confidence_type: Tuple[str] = ( 146 "entropy_renyi_exp", 147 "entropy_renyi_lin", 148 "entropy_tsallis_exp", 149 "entropy_tsallis_lin", 150 "entropy_gibbs_lin", 151 "entropy_gibbs_exp", 152 ) 153 154 # TODO: currently it's not possible to efficiently tune temperature, as we always 155 # apply log-softmax in the decoder, so to try different values it will be required 156 # to rerun the decoding, which is very slow. To support this for one-off experiments 157 # it's possible to modify the code of CTC decoder / Transducer joint to 158 # remove log-softmax and then apply it directly in this script with the temperature 159 # 160 # Alternatively, one can run this script multiple times with different values of 161 # temperature and pick the best performing ensemble. Note that this will increase 162 # tuning time by the number of temperature values tried. On the other hand, 163 # the above approach is a lot more efficient and will only slightly increase 164 # the total tuning runtime. 165 166 # very important to tune for max prob, but for entropy metrics 1.0 is almost always best 167 # temperature: Tuple[float] = (1.0,) 168 169 # not that important, but can sometimes make a small difference 170 alpha: Tuple[float] = (0.25, 0.33, 0.5, 1.0) 171 172 def get_grid_size(self) -> int: 173 """Returns the total number of points in the search space.""" 174 if "max_prob" in self.confidence_type: 175 return ( 176 len(self.exclude_blank) 177 * len(self.aggregation) 178 * ((len(self.confidence_type) - 1) * len(self.alpha) + 1) 179 ) 180 return len(self.exclude_blank) * len(self.aggregation) * len(self.confidence_type) * len(self.alpha) 181 182 183 @dataclass 184 class TuneLogisticRegressionConfig: 185 # will have log-uniform grid over this range with that many points 186 # note that a value of 10000.0 (not regularization) is always added 187 C_num_points: int = 10 188 C_min: float = 0.0001 189 C_max: float = 10.0 190 191 # not too important 192 multi_class: Tuple[str] = ("ovr", "multinomial") 193 194 # should try to include weights directly if the data is too imbalanced 195 class_weight: Tuple = (None, "balanced") 196 197 # increase if getting many warnings that algorithm didn't converge 198 max_iter: int = 1000 199 200 201 @dataclass 202 class BuildEnsembleConfig: 203 # where to save the resulting ensemble model 204 output_path: str = MISSING 205 206 # each model specification 207 ensemble: List[EnsembleConfig] = MISSING 208 209 random_seed: int = 0 # for reproducibility 210 211 # default confidence, can override 212 confidence: ConfidenceConfig = ConfidenceConfig( 213 # we keep frame confidences and apply aggregation manually to get full-utterance confidence 214 preserve_frame_confidence=True, 215 exclude_blank=True, 216 aggregation="mean", 217 measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False 229 # used to specify what to tune over. By default runs tuning over some 230 # reasonalbe grid, so that it does not take forever. 231 # Can be changed as needed 232 tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() 233 234 # very fast to tune and can be important in case of imbalanced datasets 235 # will automatically set to False if dev data is not available 236 tune_logistic_regression: bool = True 237 tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() 238 239 def __post_init__(self): 240 """Checking that if any dev data is provided, all are provided. 241 242 Will also auto-set tune_logistic_regression to False if no dev data 243 is available. 244 245 If tune_confidence is set to True (user choice) and no dev data is 246 provided, will raise an error. 247 """ 248 num_dev_data = 0 249 for ensemble_cfg in self.ensemble: 250 num_dev_data += ensemble_cfg.dev_manifest is not None 251 if num_dev_data == 0: 252 if self.tune_confidence: 253 raise ValueError("tune_confidence is set to True, but no dev data is provided") 254 LOG.info("Setting tune_logistic_regression = False since no dev data is provided") 255 self.tune_logistic_regression = False 256 return 257 258 if num_dev_data < len(self.ensemble): 259 raise ValueError( 260 "Some ensemble configs specify dev data, but some don't. Either all have to specify it or none!" 261 ) 262 263 264 def calculate_score(features: np.ndarray, labels: np.ndarray, pipe: Pipeline) -> Tuple[float, np.ndarray]: 265 """Score is always calculated as mean of the per-class scores. 266 267 This is done to account for possible class imbalances. 268 269 Args: 270 features: numpy array of features of shape [N x D], where N is the 271 number of objects (typically a total number of utterances in 272 all datasets) and D is the total number of confidence scores 273 used to train the model (typically = number of models). 274 labels: numpy array of shape [N] contatining ground-truth model indices. 275 pipe: classification pipeline (currently, standardization + logistic 276 regression). 277 278 Returns: 279 tuple: score value in [0, 1] and full classification confusion matrix. 280 """ 281 predictions = pipe.predict(features) 282 conf_m = confusion_matrix(labels, predictions) 283 score = np.diag(conf_m).sum() / conf_m.sum() 284 return score, conf_m 285 286 287 def train_model_selection( 288 training_features: np.ndarray, 289 training_labels: np.ndarray, 290 dev_features: Optional[np.ndarray] = None, 291 dev_labels: Optional[np.ndarray] = None, 292 tune_lr: bool = False, 293 tune_lr_cfg: Optional[TuneLogisticRegressionConfig] = None, 294 verbose: bool = False, 295 ) -> Tuple[Pipeline, float]: 296 """Trains model selection block with an (optional) tuning of the parameters. 297 298 Returns a pipeline consisting of feature standardization and logistic 299 regression. If tune_lr is set to True, dev features/labels will be used 300 to tune the hyperparameters of the logistic regression with the grid 301 search that's defined via ``tune_lr_cfg``. 302 303 If no tuning is requested, uses the following parameters:: 304 305 best_pipe = make_pipeline( 306 StandardScaler(), 307 LogisticRegression( 308 multi_class="multinomial", 309 C=10000.0, 310 max_iter=1000, 311 class_weight="balanced", 312 ), 313 ) 314 315 Args: 316 training_features: numpy array of features of shape [N x D], where N is 317 the number of objects (typically a total number of utterances in 318 all training datasets) and D is the total number of confidence 319 scores used to train the model (typically = number of models). 320 training_labels: numpy array of shape [N] contatining ground-truth 321 model indices. 322 dev_features: same as training, but for the validation subset. 323 dev_labels: same as training, but for the validation subset. 324 tune_lr: controls whether tuning of LR hyperparameters is performed. 325 If set to True, it's required to also provide dev features/labels. 326 tune_lr_cfg: specifies what values of LR hyperparameters to try. 327 verbose: if True, will output final training/dev scores. 328 329 Returns: 330 tuple: trained model selection pipeline, best score (or -1 if no tuning 331 was done). 332 """ 333 if not tune_lr: 334 # default parameters: C=10000.0 disables regularization 335 best_pipe = make_pipeline( 336 StandardScaler(), 337 LogisticRegression(multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced"), 338 ) 339 max_score = -1 340 else: 341 C_pms = np.append( 342 np.exp(np.linspace(np.log(tune_lr_cfg.C_min), np.log(tune_lr_cfg.C_max), tune_lr_cfg.C_num_points)), 343 10000.0, 344 ) 345 max_score = 0 346 best_pipe = None 347 for class_weight in tune_lr_cfg.class_weight: 348 for multi_class in tune_lr_cfg.multi_class: 349 for C in C_pms: 350 pipe = make_pipeline( 351 StandardScaler(), 352 LogisticRegression( 353 multi_class=multi_class, C=C, max_iter=tune_lr_cfg.max_iter, class_weight=class_weight 354 ), 355 ) 356 pipe.fit(training_features, training_labels) 357 score, confusion = calculate_score(dev_features, dev_labels, pipe) 358 if score > max_score: 359 max_score = score 360 best_pipe = pipe 361 362 best_pipe.fit(training_features, training_labels) 363 if verbose: 364 accuracy, confusion = calculate_score(training_features, training_labels, best_pipe) 365 LOG.info("Training fit accuracy: %.4f", accuracy * 100.0) 366 LOG.info("Training confusion matrix:\n%s", str(confusion)) 367 if dev_features is not None and verbose: 368 accuracy, confusion = calculate_score(dev_features, dev_labels, best_pipe) 369 LOG.info("Dev fit accuracy: %.4f", accuracy * 100.0) 370 LOG.info("Dev confusion matrix:\n%s", str(confusion)) 371 372 return best_pipe, max_score 373 374 375 def subsample_manifest(manifest_file: str, max_samples: int) -> str: 376 """Will save a subsampled version of the manifest to the same folder. 377 378 Have to save to the same folder to support relative paths. 379 380 Args: 381 manifest_file: path to the manifest file that needs subsampling. 382 max_samples: how many samples to retain. Will randomly select that 383 many lines from the manifest. 384 385 Returns: 386 str: the path to the subsampled manifest file. 387 """ 388 with open(manifest_file, "rt", encoding="utf-8") as fin: 389 lines = fin.readlines() 390 if max_samples < len(lines): 391 lines = random.sample(lines, max_samples) 392 output_file = manifest_file + "-subsampled" 393 with open(output_file, "wt", encoding="utf-8") as fout: 394 fout.write("".join(lines)) 395 return output_file 396 397 398 def cleanup_subsampled_manifests(subsampled_manifests: List[str]): 399 """Removes all generated subsamples manifests.""" 400 for manifest in subsampled_manifests: 401 os.remove(manifest) 402 403 404 def compute_all_confidences( 405 hypothesis: Hypothesis, tune_confidence_cfg: TuneConfidenceConfig 406 ) -> Dict[ConfidenceSpec, float]: 407 """Computes a set of confidence scores from a given hypothesis. 408 409 Works with the output of both CTC and Transducer decoding. 410 411 Args: 412 hypothesis: generated hypothesis as returned from the transcribe 413 method of the ASR model. 414 tune_confidence_cfg: config specifying what confidence scores to 415 compute. 416 417 Returns: 418 dict: dictionary with confidenct spec -> confidence score mapping. 419 """ 420 conf_values = {} 421 422 for exclude_blank in tune_confidence_cfg.exclude_blank: 423 filtered_logprobs = get_filtered_logprobs(hypothesis, exclude_blank) 424 vocab_size = filtered_logprobs.shape[1] 425 for aggregation in tune_confidence_cfg.aggregation: 426 aggr_func = get_confidence_aggregation_bank()[aggregation] 427 for conf_type in tune_confidence_cfg.confidence_type: 428 conf_func = get_confidence_measure_bank()[conf_type] 429 if conf_type == "max_prob": # skipping alpha in this case 430 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=1.0)).cpu().item() 431 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, 1.0)] = conf_value 432 else: 433 for alpha in tune_confidence_cfg.alpha: 434 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=alpha)).cpu().item() 435 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, alpha)] = conf_value 436 437 return conf_values 438 439 440 def find_best_confidence( 441 train_confidences: List[List[Dict[ConfidenceSpec, float]]], 442 train_labels: List[int], 443 dev_confidences: List[List[Dict[ConfidenceSpec, float]]], 444 dev_labels: List[int], 445 tune_lr: bool, 446 tune_lr_config: TuneConfidenceConfig, 447 ) -> Tuple[ConfidenceConfig, Pipeline]: 448 """Finds the best confidence configuration for model selection. 449 450 Will loop over all values in the confidence dictionary and fit the LR 451 model (optionally tuning its HPs). The best performing confidence (on the 452 dev set) will be used for the final LR model. 453 454 Args: 455 train_confidences: this is an object of type 456 ``List[List[Dict[ConfidenceSpec, float]]]``. The shape of this 457 object is [M, N, S], where 458 M: number of models 459 N: number of utterances in all training sets 460 S: number of confidence scores to try 461 462 This argument will be used to construct np.array objects for each 463 of the confidence scores with the shape [M, N] 464 465 train_labels: ground-truth labels of the correct model for each data 466 points. This is a list of size [N] 467 dev_confidences: same as training, but for the validation subset. 468 dev_labels: same as training, but for the validation subset. 469 tune_lr: controls whether tuning of LR hyperparameters is performed. 470 tune_lr_cfg: specifies what values of LR hyperparameters to try. 471 472 Returns: 473 tuple: best confidence config, best model selection pipeline 474 """ 475 max_score = 0 476 best_pipe = None 477 best_conf_spec = None 478 LOG.info("Evaluation all confidences. Total grid size: %d", len(train_confidences[0][0].keys())) 479 for conf_spec in tqdm(train_confidences[0][0].keys()): 480 cur_train_confidences = [] 481 for model_confs in train_confidences: 482 cur_train_confidences.append([]) 483 for model_conf in model_confs: 484 cur_train_confidences[-1].append(model_conf[conf_spec]) 485 cur_dev_confidences = [] 486 for model_confs in dev_confidences: 487 cur_dev_confidences.append([]) 488 for model_conf in model_confs: 489 cur_dev_confidences[-1].append(model_conf[conf_spec]) 490 # transposing with zip(*list) 491 training_features = np.array(list(zip(*cur_train_confidences))) 492 training_labels = np.array(train_labels) 493 dev_features = np.array(list(zip(*cur_dev_confidences))) 494 dev_labels = np.array(dev_labels) 495 pipe, score = train_model_selection( 496 training_features, training_labels, dev_features, dev_labels, tune_lr, tune_lr_config, 497 ) 498 if max_score < score: 499 max_score = score 500 best_pipe = pipe 501 best_conf_spec = conf_spec 502 LOG.info("Found better parameters: %s. New score: %.4f", str(conf_spec), max_score) 503 504 return best_conf_spec.to_confidence_config(), best_pipe 505 506 507 @hydra_runner(config_name="BuildEnsembleConfig", schema=BuildEnsembleConfig) 508 def main(cfg: BuildEnsembleConfig): 509 # silencing all messages from nemo/ptl to avoid dumping tons of configs to the stdout 510 logging.getLogger('pytorch_lightning').setLevel(logging.CRITICAL) 511 logging.getLogger('nemo_logger').setLevel(logging.CRITICAL) 512 LOG.info(f'Build ensemble config:\n{OmegaConf.to_yaml(cfg)}') 513 514 # to ensure post init is called 515 cfg = BuildEnsembleConfig(**cfg) 516 517 pl.seed_everything(cfg.random_seed) 518 cfg.transcription.random_seed = None # seed is already applied 519 cfg.transcription.return_transcriptions = True 520 cfg.transcription.preserve_alignment = True 521 cfg.transcription.ctc_decoding.temperature = cfg.temperature 522 cfg.transcription.rnnt_decoding.temperature = cfg.temperature 523 # this ensures that generated output is after log-softmax for consistency with CTC 524 525 train_confidences = [] 526 dev_confidences = [] 527 train_labels = [] 528 dev_labels = [] 529 530 # registering clean-up function that will hold on to this list and 531 # should clean up even if there is partial error in some of the transcribe 532 # calls 533 subsampled_manifests = [] 534 atexit.register(cleanup_subsampled_manifests, subsampled_manifests) 535 536 # note that we loop over the same config. 537 # This is intentional, as we need to run all models on all datasets 538 # this loop will do the following things: 539 # 1. Goes through each model X each training dataset 540 # 2. Computes predictions by directly calling transcribe_speech.main 541 # 3. Converts transcription to the confidence score(s) as specified in the config 542 # 4. If dev sets are provided, computes the same for them 543 # 5. Creates a list of ground-truth model indices by mapping each model 544 # to its own training dataset as specified in the config. 545 # 6. After the loop, we either run tuning over all confidence scores or 546 # directly use a single score to fit logistic regression and save the 547 # final ensemble model. 548 for model_idx, model_cfg in enumerate(cfg.ensemble): 549 train_model_confidences = [] 550 dev_model_confidences = [] 551 for data_idx, data_cfg in enumerate(cfg.ensemble): 552 if model_idx == 0: # generating subsampled manifests only one time 553 subsampled_manifests.append( 554 subsample_manifest(data_cfg.training_manifest, data_cfg.max_training_samples) 555 ) 556 subsampled_manifest = subsampled_manifests[data_idx] 557 558 if model_cfg.model.endswith(".nemo"): 559 cfg.transcription.model_path = model_cfg.model 560 else: # assuming pretrained model 561 cfg.transcription.pretrained_name = model_cfg.model 562 563 cfg.transcription.dataset_manifest = subsampled_manifest 564 565 # training 566 with tempfile.NamedTemporaryFile() as output_file: 567 cfg.transcription.output_filename = output_file.name 568 LOG.info("Transcribing training dataset %d with model %d", data_idx, model_idx) 569 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 570 LOG.info("Generating confidence scores") 571 # TODO: parallelize this loop? 572 for transcription in tqdm(transcriptions): 573 if cfg.tune_confidence: 574 train_model_confidences.append( 575 compute_all_confidences(transcription, cfg.tune_confidence_config) 576 ) 577 else: 578 train_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 579 if model_idx == 0: # labels are the same for all models 580 train_labels.append(data_idx) 581 582 # optional dev 583 if data_cfg.dev_manifest is not None: 584 cfg.transcription.dataset_manifest = data_cfg.dev_manifest 585 with tempfile.NamedTemporaryFile() as output_file: 586 cfg.transcription.output_filename = output_file.name 587 LOG.info("Transcribing dev dataset %d with model %d", data_idx, model_idx) 588 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 589 LOG.info("Generating confidence scores") 590 for transcription in tqdm(transcriptions): 591 if cfg.tune_confidence: 592 dev_model_confidences.append( 593 compute_all_confidences(transcription, cfg.tune_confidence_config) 594 ) 595 else: 596 dev_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 597 if model_idx == 0: # labels are the same for all models 598 dev_labels.append(data_idx) 599 600 train_confidences.append(train_model_confidences) 601 if dev_model_confidences: 602 dev_confidences.append(dev_model_confidences) 603 604 if cfg.tune_confidence: 605 best_confidence, model_selection_block = find_best_confidence( 606 train_confidences, 607 train_labels, 608 dev_confidences, 609 dev_labels, 610 cfg.tune_logistic_regression, 611 cfg.tune_logistic_regression_config, 612 ) 613 else: 614 best_confidence = cfg.confidence 615 # transposing with zip(*list) 616 training_features = np.array(list(zip(*train_confidences))) 617 training_labels = np.array(train_labels) 618 if dev_confidences: 619 dev_features = np.array(list(zip(*dev_confidences))) 620 dev_labels = np.array(dev_labels) 621 else: 622 dev_features = None 623 dev_labels = None 624 model_selection_block, _ = train_model_selection( 625 training_features, 626 training_labels, 627 dev_features, 628 dev_labels, 629 cfg.tune_logistic_regression, 630 cfg.tune_logistic_regression_config, 631 verbose=True, 632 ) 633 634 with tempfile.TemporaryDirectory() as tmpdir: 635 model_selection_block_path = os.path.join(tmpdir, 'model_selection_block.pkl') 636 joblib.dump(model_selection_block, model_selection_block_path) 637 638 # creating ensemble checkpoint 639 ensemble_model = ConfidenceEnsembleModel( 640 cfg=DictConfig( 641 { 642 'model_selection_block': model_selection_block_path, 643 'confidence': best_confidence, 644 'temperature': cfg.temperature, 645 'load_models': [model_cfg.model for model_cfg in cfg.ensemble], 646 } 647 ), 648 trainer=None, 649 ) 650 ensemble_model.save_to(cfg.output_path) 651 652 653 if __name__ == '__main__': 654 main() 655 [end of scripts/confidence_ensembles/build_ensemble.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import json 16 import os 17 from dataclasses import dataclass, is_dataclass 18 from pathlib import Path 19 from typing import Optional 20 21 import pytorch_lightning as pl 22 import torch 23 from omegaconf import MISSING, OmegaConf 24 from sklearn.model_selection import ParameterGrid 25 26 from nemo.collections.asr.metrics.rnnt_wer import RNNTDecodingConfig 27 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 28 from nemo.collections.asr.models import ASRModel, EncDecRNNTModel 29 from nemo.collections.asr.parts.utils.asr_confidence_benchmarking_utils import ( 30 apply_confidence_parameters, 31 run_confidence_benchmark, 32 ) 33 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig 34 from nemo.core.config import hydra_runner 35 from nemo.utils import logging 36 37 """ 38 Get confidence metrics and curve plots for a given model, dataset, and confidence parameters. 39 40 # Arguments 41 model_path: Path to .nemo ASR checkpoint 42 pretrained_name: Name of pretrained ASR model (from NGC registry) 43 dataset_manifest: Path to dataset JSON manifest file (in NeMo format) 44 output_dir: Output directory to store a report and curve plot directories 45 46 batch_size: batch size during inference 47 num_workers: number of workers during inference 48 49 cuda: Optional int to enable or disable execution of model on certain CUDA device 50 amp: Bool to decide if Automatic Mixed Precision should be used during inference 51 audio_type: Str filetype of the audio. Supported = wav, flac, mp3 52 53 target_level: Word- or token-level confidence. Supported = word, token, auto (for computing both word and token) 54 confidence_cfg: Config with confidence parameters 55 grid_params: Dictionary with lists of parameters to iteratively benchmark on 56 57 # Usage 58 ASR model can be specified by either "model_path" or "pretrained_name". 59 Data for transcription are defined with "dataset_manifest". 60 Results are returned as a benchmark report and curve plots. 61 62 python benchmark_asr_confidence.py \ 63 model_path=null \ 64 pretrained_name=null \ 65 dataset_manifest="" \ 66 output_dir="" \ 67 batch_size=64 \ 68 num_workers=8 \ 69 cuda=0 \ 70 amp=True \ 71 target_level="word" \ 72 confidence_cfg.exclude_blank=False \ 73 'grid_params="{\"aggregation\": [\"min\", \"prod\"], \"alpha\": [0.33, 0.5]}"' 74 """ 75 76 77 def get_experiment_params(cfg): 78 """Get experiment parameters from a confidence config and generate the experiment name. 79 80 Returns: 81 List of experiment parameters. 82 String with the experiment name. 83 """ 84 blank = "no_blank" if cfg.exclude_blank else "blank" 85 aggregation = cfg.aggregation 86 method_name = cfg.measure_cfg.name 87 alpha = cfg.measure_cfg.alpha 88 if method_name == "entropy": 89 entropy_type = cfg.measure_cfg.entropy_type 90 entropy_norm = cfg.measure_cfg.entropy_norm 91 experiment_param_list = [ 92 aggregation, 93 str(cfg.exclude_blank), 94 method_name, 95 entropy_type, 96 entropy_norm, 97 str(alpha), 98 ] 99 experiment_str = "-".join([aggregation, blank, method_name, entropy_type, entropy_norm, str(alpha)]) 100 else: 101 experiment_param_list = [aggregation, str(cfg.exclude_blank), method_name, "-", "-", str(alpha)] 102 experiment_str = "-".join([aggregation, blank, method_name, str(alpha)]) 103 return experiment_param_list, experiment_str 104 105 106 @dataclass 107 class ConfidenceBenchmarkingConfig: 108 # Required configs 109 model_path: Optional[str] = None # Path to a .nemo file 110 pretrained_name: Optional[str] = None # Name of a pretrained model 111 dataset_manifest: str = MISSING 112 output_dir: str = MISSING 113 114 # General configs 115 batch_size: int = 32 116 num_workers: int = 4 117 118 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA 119 # device anyway, and do inference on CPU only if CUDA device is not found. 120 # If `cuda` is a negative number, inference will be on CPU only. 121 cuda: Optional[int] = None 122 amp: bool = False 123 audio_type: str = "wav" 124 125 # Confidence configs 126 target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) 127 confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) 128 grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on 129 130 131 @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) 132 def main(cfg: ConfidenceBenchmarkingConfig): 133 torch.set_grad_enabled(False) 134 135 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 136 137 if is_dataclass(cfg): 138 cfg = OmegaConf.structured(cfg) 139 140 if cfg.model_path is None and cfg.pretrained_name is None: 141 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None!") 142 143 # setup GPU 144 if cfg.cuda is None: 145 if torch.cuda.is_available(): 146 device = [0] # use 0th CUDA device 147 accelerator = 'gpu' 148 else: 149 device = 1 150 accelerator = 'cpu' 151 else: 152 device = [cfg.cuda] 153 accelerator = 'gpu' 154 155 map_location = torch.device('cuda:{}'.format(device[0]) if accelerator == 'gpu' else 'cpu') 156 157 # setup model 158 if cfg.model_path is not None: 159 # restore model from .nemo file path 160 model_cfg = ASRModel.restore_from(restore_path=cfg.model_path, return_config=True) 161 classpath = model_cfg.target # original class path 162 imported_class = model_utils.import_class_by_path(classpath) # type: ASRModel 163 logging.info(f"Restoring model : {imported_class.__name__}") 164 asr_model = imported_class.restore_from( 165 restore_path=cfg.model_path, map_location=map_location 166 ) # type: ASRModel 167 else: 168 # restore model by name 169 asr_model = ASRModel.from_pretrained( 170 model_name=cfg.pretrained_name, map_location=map_location 171 ) # type: ASRModel 172 173 trainer = pl.Trainer(devices=device, accelerator=accelerator) 174 asr_model.set_trainer(trainer) 175 asr_model = asr_model.eval() 176 177 # Check if ctc or rnnt model 178 is_rnnt = isinstance(asr_model, EncDecRNNTModel) 179 180 # Check that the model has the `change_decoding_strategy` method 181 if not hasattr(asr_model, 'change_decoding_strategy'): 182 raise RuntimeError("The asr_model you are using must have the `change_decoding_strategy` method.") 183 184 # get filenames and reference texts from manifest 185 filepaths = [] 186 reference_texts = [] 187 if os.stat(cfg.dataset_manifest).st_size == 0: 188 logging.error(f"The input dataset_manifest {cfg.dataset_manifest} is empty. Exiting!") 189 return None 190 manifest_dir = Path(cfg.dataset_manifest).parent 191 with open(cfg.dataset_manifest, 'r') as f: 192 for line in f: 193 item = json.loads(line) 194 audio_file = Path(item['audio_filepath']) 195 if not audio_file.is_file() and not audio_file.is_absolute(): 196 audio_file = manifest_dir / audio_file 197 filepaths.append(str(audio_file.absolute())) 198 reference_texts.append(item['text']) 199 200 # setup AMP (optional) 201 autocast = None 202 if cfg.amp and torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 203 logging.info("AMP enabled!\n") 204 autocast = torch.cuda.amp.autocast 205 206 # do grid-based benchmarking if grid_params is provided, otherwise a regular one 207 work_dir = Path(cfg.output_dir) 208 os.makedirs(work_dir, exist_ok=True) 209 report_legend = ( 210 ",".join( 211 [ 212 "model_type", 213 "aggregation", 214 "blank", 215 "method_name", 216 "entropy_type", 217 "entropy_norm", 218 "alpha", 219 "target_level", 220 "auc_roc", 221 "auc_pr", 222 "auc_nt", 223 "nce", 224 "ece", 225 "auc_yc", 226 "std_yc", 227 "max_yc", 228 ] 229 ) 230 + "\n" 231 ) 232 model_typename = "RNNT" if is_rnnt else "CTC" 233 report_file = work_dir / Path("report.csv") 234 if cfg.grid_params: 235 asr_model.change_decoding_strategy( 236 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 237 if is_rnnt 238 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 239 ) 240 params = json.loads(cfg.grid_params) 241 hp_grid = ParameterGrid(params) 242 hp_grid = list(hp_grid) 243 244 logging.info(f"==============================Running a benchmarking with grid search=========================") 245 logging.info(f"Grid search size: {len(hp_grid)}") 246 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directories near the file") 247 logging.info(f"==============================================================================================") 248 249 with open(report_file, "tw", encoding="utf-8") as f: 250 f.write(report_legend) 251 f.flush() 252 for i, hp in enumerate(hp_grid): 253 logging.info(f"Run # {i + 1}, grid: `{hp}`") 254 asr_model.change_decoding_strategy(apply_confidence_parameters(asr_model.cfg.decoding, hp)) 255 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 256 plot_dir = work_dir / Path(experiment_name) 257 results = run_confidence_benchmark( 258 asr_model, 259 cfg.target_level, 260 filepaths, 261 reference_texts, 262 cfg.batch_size, 263 cfg.num_workers, 264 plot_dir, 265 autocast, 266 ) 267 for level, result in results.items(): 268 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 269 f.flush() 270 else: 271 asr_model.change_decoding_strategy( 272 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 273 if is_rnnt 274 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 275 ) 276 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 277 plot_dir = work_dir / Path(experiment_name) 278 279 logging.info(f"==============================Running a single benchmarking===================================") 280 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directory `{plot_dir}`") 281 282 with open(report_file, "tw", encoding="utf-8") as f: 283 f.write(report_legend) 284 f.flush() 285 results = run_confidence_benchmark( 286 asr_model, 287 cfg.batch_size, 288 cfg.num_workers, 289 cfg.target_level, 290 filepaths, 291 reference_texts, 292 plot_dir, 293 autocast, 294 ) 295 for level, result in results.items(): 296 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 297 logging.info(f"===========================================Done===============================================") 298 299 300 if __name__ == '__main__': 301 main() 302 [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 """ 15 # This script converts an existing audio dataset with a manifest to 16 # a tarred and sharded audio dataset that can be read by the 17 # TarredAudioToTextDataLayer. 18 19 # Please make sure your audio_filepath DOES NOT CONTAIN '-sub'! 20 # Because we will use it to handle files which have duplicate filenames but with different offsets 21 # (see function create_shard for details) 22 23 24 # Bucketing can help to improve the training speed. You may use --buckets_num to specify the number of buckets. 25 # It creates multiple tarred datasets, one per bucket, based on the audio durations. 26 # The range of [min_duration, max_duration) is split into equal sized buckets. 27 # Recommend to use --sort_in_shards to speedup the training by reducing the paddings in the batches 28 # More info on how to use bucketing feature: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/datasets.html 29 30 # If valid NVIDIA DALI version is installed, will also generate the corresponding DALI index files that need to be 31 # supplied to the config in order to utilize webdataset for efficient large dataset handling. 32 # NOTE: DALI + Webdataset is NOT compatible with Bucketing support ! 33 34 # Usage: 35 1) Creating a new tarfile dataset 36 37 python convert_to_tarred_audio_dataset.py \ 38 --manifest_path=<path to the manifest file> \ 39 --target_dir=<path to output directory> \ 40 --num_shards=<number of tarfiles that will contain the audio> \ 41 --max_duration=<float representing maximum duration of audio samples> \ 42 --min_duration=<float representing minimum duration of audio samples> \ 43 --shuffle --shuffle_seed=1 \ 44 --sort_in_shards \ 45 --workers=-1 46 47 48 2) Concatenating more tarfiles to a pre-existing tarred dataset 49 50 python convert_to_tarred_audio_dataset.py \ 51 --manifest_path=<path to the tarred manifest file> \ 52 --metadata_path=<path to the metadata.yaml (or metadata_version_{X}.yaml) file> \ 53 --target_dir=<path to output directory where the original tarfiles are contained> \ 54 --max_duration=<float representing maximum duration of audio samples> \ 55 --min_duration=<float representing minimum duration of audio samples> \ 56 --shuffle --shuffle_seed=1 \ 57 --sort_in_shards \ 58 --workers=-1 \ 59 --concat_manifest_paths \ 60 <space separated paths to 1 or more manifest files to concatenate into the original tarred dataset> 61 62 3) Writing an empty metadata file 63 64 python convert_to_tarred_audio_dataset.py \ 65 --target_dir=<path to output directory> \ 66 # any other optional argument 67 --num_shards=8 \ 68 --max_duration=16.7 \ 69 --min_duration=0.01 \ 70 --shuffle \ 71 --workers=-1 \ 72 --sort_in_shards \ 73 --shuffle_seed=1 \ 74 --write_metadata 75 76 """ 77 import argparse 78 import copy 79 import json 80 import os 81 import random 82 import tarfile 83 from collections import defaultdict 84 from dataclasses import dataclass, field 85 from datetime import datetime 86 from typing import Any, List, Optional 87 88 from joblib import Parallel, delayed 89 from omegaconf import DictConfig, OmegaConf, open_dict 90 91 try: 92 import create_dali_tarred_dataset_index as dali_index 93 94 DALI_INDEX_SCRIPT_AVAILABLE = True 95 except (ImportError, ModuleNotFoundError, FileNotFoundError): 96 DALI_INDEX_SCRIPT_AVAILABLE = False 97 98 parser = argparse.ArgumentParser( 99 description="Convert an existing ASR dataset to tarballs compatible with TarredAudioToTextDataLayer." 100 ) 101 parser.add_argument( 102 "--manifest_path", default=None, type=str, required=False, help="Path to the existing dataset's manifest." 103 ) 104 105 parser.add_argument( 106 '--concat_manifest_paths', 107 nargs='+', 108 default=None, 109 type=str, 110 required=False, 111 help="Path to the additional dataset's manifests that will be concatenated with base dataset.", 112 ) 113 114 # Optional arguments 115 parser.add_argument( 116 "--target_dir", 117 default='./tarred', 118 type=str, 119 help="Target directory for resulting tarballs and manifest. Defaults to `./tarred`. Creates the path if necessary.", 120 ) 121 122 parser.add_argument( 123 "--metadata_path", required=False, default=None, type=str, help="Path to metadata file for the dataset.", 124 ) 125 126 parser.add_argument( 127 "--num_shards", 128 default=-1, 129 type=int, 130 help="Number of shards (tarballs) to create. Used for partitioning data among workers.", 131 ) 132 parser.add_argument( 133 '--max_duration', 134 default=None, 135 required=True, 136 type=float, 137 help='Maximum duration of audio clip in the dataset. By default, it is None and is required to be set.', 138 ) 139 parser.add_argument( 140 '--min_duration', 141 default=None, 142 type=float, 143 help='Minimum duration of audio clip in the dataset. By default, it is None and will not filter files.', 144 ) 145 parser.add_argument( 146 "--shuffle", 147 action='store_true', 148 help="Whether or not to randomly shuffle the samples in the manifest before tarring/sharding.", 149 ) 150 151 parser.add_argument( 152 "--keep_files_together", 153 action='store_true', 154 help="Whether or not to keep entries from the same file (but different offsets) together when sorting before tarring/sharding.", 155 ) 156 157 parser.add_argument( 158 "--sort_in_shards", 159 action='store_true', 160 help="Whether or not to sort samples inside the shards based on their duration.", 161 ) 162 163 parser.add_argument( 164 "--buckets_num", type=int, default=1, help="Number of buckets to create based on duration.", 165 ) 166 167 parser.add_argument("--shuffle_seed", type=int, default=None, help="Random seed for use if shuffling is enabled.") 168 parser.add_argument( 169 '--write_metadata', 170 action='store_true', 171 help=( 172 "Flag to write a blank metadata with the current call config. " 173 "Note that the metadata will not contain the number of shards, " 174 "and it must be filled out by the user." 175 ), 176 ) 177 parser.add_argument( 178 "--no_shard_manifests", 179 action='store_true', 180 help="Do not write sharded manifests along with the aggregated manifest.", 181 ) 182 parser.add_argument('--workers', type=int, default=1, help='Number of worker processes') 183 args = parser.parse_args() 184 185 186 @dataclass 187 class ASRTarredDatasetConfig: 188 num_shards: int = -1 189 shuffle: bool = False 190 max_duration: Optional[float] = None 191 min_duration: Optional[float] = None 192 shuffle_seed: Optional[int] = None 193 sort_in_shards: bool = True 194 shard_manifests: bool = True 195 keep_files_together: bool = False 196 197 198 @dataclass 199 class ASRTarredDatasetMetadata: 200 created_datetime: Optional[str] = None 201 version: int = 0 202 num_samples_per_shard: Optional[int] = None 203 is_concatenated_manifest: bool = False 204 205 dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() 206 history: Optional[List[Any]] = field(default_factory=lambda: []) 207 208 def __post_init__(self): 209 self.created_datetime = self.get_current_datetime() 210 211 def get_current_datetime(self): 212 return datetime.now().strftime("%m-%d-%Y %H-%M-%S") 213 214 @classmethod 215 def from_config(cls, config: DictConfig): 216 obj = cls() 217 obj.__dict__.update(**config) 218 return obj 219 220 @classmethod 221 def from_file(cls, filepath: str): 222 config = OmegaConf.load(filepath) 223 return ASRTarredDatasetMetadata.from_config(config=config) 224 225 226 class ASRTarredDatasetBuilder: 227 """ 228 Helper class that constructs a tarred dataset from scratch, or concatenates tarred datasets 229 together and constructs manifests for them. 230 """ 231 232 def __init__(self): 233 self.config = None 234 235 def configure(self, config: ASRTarredDatasetConfig): 236 """ 237 Sets the config generated from command line overrides. 238 239 Args: 240 config: ASRTarredDatasetConfig dataclass object. 241 """ 242 self.config = config # type: ASRTarredDatasetConfig 243 244 if self.config.num_shards < 0: 245 raise ValueError("`num_shards` must be > 0. Please fill in the metadata information correctly.") 246 247 def create_new_dataset(self, manifest_path: str, target_dir: str = "./tarred/", num_workers: int = 0): 248 """ 249 Creates a new tarred dataset from a given manifest file. 250 251 Args: 252 manifest_path: Path to the original ASR manifest. 253 target_dir: Output directory. 254 num_workers: Integer denoting number of parallel worker processes which will write tarfiles. 255 Defaults to 1 - which denotes sequential worker process. 256 257 Output: 258 Writes tarfiles, along with the tarred dataset compatible manifest file. 259 Also preserves a record of the metadata used to construct this tarred dataset. 260 """ 261 if self.config is None: 262 raise ValueError("Config has not been set. Please call `configure(config: ASRTarredDatasetConfig)`") 263 264 if manifest_path is None: 265 raise FileNotFoundError("Manifest filepath cannot be None !") 266 267 config = self.config # type: ASRTarredDatasetConfig 268 269 if not os.path.exists(target_dir): 270 os.makedirs(target_dir) 271 272 # Read the existing manifest 273 entries, total_duration, filtered_entries, filtered_duration = self._read_manifest(manifest_path, config) 274 275 if len(filtered_entries) > 0: 276 print(f"Filtered {len(filtered_entries)} files which amounts to {filtered_duration} seconds of audio.") 277 print( 278 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 279 ) 280 281 if len(entries) == 0: 282 print("No tarred dataset was created as there were 0 valid samples after filtering!") 283 return 284 if config.shuffle: 285 random.seed(config.shuffle_seed) 286 print("Shuffling...") 287 if config.keep_files_together: 288 filename_entries = defaultdict(list) 289 for ent in entries: 290 filename_entries[ent["audio_filepath"]].append(ent) 291 filenames = list(filename_entries.keys()) 292 random.shuffle(filenames) 293 shuffled_entries = [] 294 for filename in filenames: 295 shuffled_entries += filename_entries[filename] 296 entries = shuffled_entries 297 else: 298 random.shuffle(entries) 299 300 # Create shards and updated manifest entries 301 print(f"Number of samples added : {len(entries)}") 302 print(f"Remainder: {len(entries) % config.num_shards}") 303 304 start_indices = [] 305 end_indices = [] 306 # Build indices 307 for i in range(config.num_shards): 308 start_idx = (len(entries) // config.num_shards) * i 309 end_idx = start_idx + (len(entries) // config.num_shards) 310 print(f"Shard {i} has entries {start_idx} ~ {end_idx}") 311 files = set() 312 for ent_id in range(start_idx, end_idx): 313 files.add(entries[ent_id]["audio_filepath"]) 314 print(f"Shard {i} contains {len(files)} files") 315 if i == config.num_shards - 1: 316 # We discard in order to have the same number of entries per shard. 317 print(f"Have {len(entries) - end_idx} entries left over that will be discarded.") 318 319 start_indices.append(start_idx) 320 end_indices.append(end_idx) 321 322 manifest_folder, _ = os.path.split(manifest_path) 323 324 with Parallel(n_jobs=num_workers, verbose=config.num_shards) as parallel: 325 # Call parallel tarfile construction 326 new_entries_list = parallel( 327 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, i, manifest_folder) 328 for i, (start_idx, end_idx) in enumerate(zip(start_indices, end_indices)) 329 ) 330 331 if config.shard_manifests: 332 sharded_manifests_dir = target_dir + '/sharded_manifests' 333 if not os.path.exists(sharded_manifests_dir): 334 os.makedirs(sharded_manifests_dir) 335 336 for manifest in new_entries_list: 337 shard_id = manifest[0]['shard_id'] 338 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 339 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 340 for entry in manifest: 341 json.dump(entry, m2) 342 m2.write('\n') 343 344 # Flatten the list of list of entries to a list of entries 345 new_entries = [sample for manifest in new_entries_list for sample in manifest] 346 del new_entries_list 347 348 print("Total number of entries in manifest :", len(new_entries)) 349 350 # Write manifest 351 new_manifest_path = os.path.join(target_dir, 'tarred_audio_manifest.json') 352 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 353 for entry in new_entries: 354 json.dump(entry, m2) 355 m2.write('\n') 356 357 # Write metadata (default metadata for new datasets) 358 new_metadata_path = os.path.join(target_dir, 'metadata.yaml') 359 metadata = ASRTarredDatasetMetadata() 360 361 # Update metadata 362 metadata.dataset_config = config 363 metadata.num_samples_per_shard = len(new_entries) // config.num_shards 364 365 # Write metadata 366 metadata_yaml = OmegaConf.structured(metadata) 367 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 368 369 def create_concatenated_dataset( 370 self, 371 base_manifest_path: str, 372 manifest_paths: List[str], 373 metadata: ASRTarredDatasetMetadata, 374 target_dir: str = "./tarred_concatenated/", 375 num_workers: int = 1, 376 ): 377 """ 378 Creates new tarfiles in order to create a concatenated dataset, whose manifest contains the data for 379 both the original dataset as well as the new data submitted in manifest paths. 380 381 Args: 382 base_manifest_path: Path to the manifest file which contains the information for the original 383 tarred dataset (with flattened paths). 384 manifest_paths: List of one or more paths to manifest files that will be concatenated with above 385 base tarred dataset. 386 metadata: ASRTarredDatasetMetadata dataclass instance with overrides from command line. 387 target_dir: Output directory 388 389 Output: 390 Writes tarfiles which with indices mapping to a "concatenated" tarred dataset, 391 along with the tarred dataset compatible manifest file which includes information 392 about all the datasets that comprise the concatenated dataset. 393 394 Also preserves a record of the metadata used to construct this tarred dataset. 395 """ 396 if not os.path.exists(target_dir): 397 os.makedirs(target_dir) 398 399 if base_manifest_path is None: 400 raise FileNotFoundError("Base manifest filepath cannot be None !") 401 402 if manifest_paths is None or len(manifest_paths) == 0: 403 raise FileNotFoundError("List of additional manifest filepaths cannot be None !") 404 405 config = ASRTarredDatasetConfig(**(metadata.dataset_config)) 406 407 # Read the existing manifest (no filtering here) 408 base_entries, _, _, _ = self._read_manifest(base_manifest_path, config) 409 print(f"Read base manifest containing {len(base_entries)} samples.") 410 411 # Precompute number of samples per shard 412 if metadata.num_samples_per_shard is None: 413 num_samples_per_shard = len(base_entries) // config.num_shards 414 else: 415 num_samples_per_shard = metadata.num_samples_per_shard 416 417 print("Number of samples per shard :", num_samples_per_shard) 418 419 # Compute min and max duration and update config (if no metadata passed) 420 print(f"Selected max duration : {config.max_duration}") 421 print(f"Selected min duration : {config.min_duration}") 422 423 entries = [] 424 for new_manifest_idx in range(len(manifest_paths)): 425 new_entries, total_duration, filtered_new_entries, filtered_duration = self._read_manifest( 426 manifest_paths[new_manifest_idx], config 427 ) 428 429 if len(filtered_new_entries) > 0: 430 print( 431 f"Filtered {len(filtered_new_entries)} files which amounts to {filtered_duration:0.2f}" 432 f" seconds of audio from manifest {manifest_paths[new_manifest_idx]}." 433 ) 434 print( 435 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 436 ) 437 438 entries.extend(new_entries) 439 440 if len(entries) == 0: 441 print("No tarred dataset was created as there were 0 valid samples after filtering!") 442 return 443 444 if config.shuffle: 445 random.seed(config.shuffle_seed) 446 print("Shuffling...") 447 random.shuffle(entries) 448 449 # Drop last section of samples that cannot be added onto a chunk 450 drop_count = len(entries) % num_samples_per_shard 451 total_new_entries = len(entries) 452 entries = entries[:-drop_count] 453 454 print( 455 f"Dropping {drop_count} samples from total new samples {total_new_entries} since they cannot " 456 f"be added into a uniformly sized chunk." 457 ) 458 459 # Create shards and updated manifest entries 460 num_added_shards = len(entries) // num_samples_per_shard 461 462 print(f"Number of samples in base dataset : {len(base_entries)}") 463 print(f"Number of samples in additional datasets : {len(entries)}") 464 print(f"Number of added shards : {num_added_shards}") 465 print(f"Remainder: {len(entries) % num_samples_per_shard}") 466 467 start_indices = [] 468 end_indices = [] 469 shard_indices = [] 470 for i in range(num_added_shards): 471 start_idx = (len(entries) // num_added_shards) * i 472 end_idx = start_idx + (len(entries) // num_added_shards) 473 shard_idx = i + config.num_shards 474 print(f"Shard {shard_idx} has entries {start_idx + len(base_entries)} ~ {end_idx + len(base_entries)}") 475 476 start_indices.append(start_idx) 477 end_indices.append(end_idx) 478 shard_indices.append(shard_idx) 479 480 manifest_folder, _ = os.path.split(base_manifest_path) 481 482 with Parallel(n_jobs=num_workers, verbose=num_added_shards) as parallel: 483 # Call parallel tarfile construction 484 new_entries_list = parallel( 485 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, shard_idx, manifest_folder) 486 for i, (start_idx, end_idx, shard_idx) in enumerate(zip(start_indices, end_indices, shard_indices)) 487 ) 488 489 if config.shard_manifests: 490 sharded_manifests_dir = target_dir + '/sharded_manifests' 491 if not os.path.exists(sharded_manifests_dir): 492 os.makedirs(sharded_manifests_dir) 493 494 for manifest in new_entries_list: 495 shard_id = manifest[0]['shard_id'] 496 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 497 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 498 for entry in manifest: 499 json.dump(entry, m2) 500 m2.write('\n') 501 502 # Flatten the list of list of entries to a list of entries 503 new_entries = [sample for manifest in new_entries_list for sample in manifest] 504 del new_entries_list 505 506 # Write manifest 507 if metadata is None: 508 new_version = 1 # start with `1`, where `0` indicates the base manifest + dataset 509 else: 510 new_version = metadata.version + 1 511 512 print("Total number of entries in manifest :", len(base_entries) + len(new_entries)) 513 514 new_manifest_path = os.path.join(target_dir, f'tarred_audio_manifest_version_{new_version}.json') 515 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 516 # First write all the entries of base manifest 517 for entry in base_entries: 518 json.dump(entry, m2) 519 m2.write('\n') 520 521 # Finally write the new entries 522 for entry in new_entries: 523 json.dump(entry, m2) 524 m2.write('\n') 525 526 # Preserve historical metadata 527 base_metadata = metadata 528 529 # Write metadata (updated metadata for concatenated datasets) 530 new_metadata_path = os.path.join(target_dir, f'metadata_version_{new_version}.yaml') 531 metadata = ASRTarredDatasetMetadata() 532 533 # Update config 534 config.num_shards = config.num_shards + num_added_shards 535 536 # Update metadata 537 metadata.version = new_version 538 metadata.dataset_config = config 539 metadata.num_samples_per_shard = num_samples_per_shard 540 metadata.is_concatenated_manifest = True 541 metadata.created_datetime = metadata.get_current_datetime() 542 543 # Attach history 544 current_metadata = OmegaConf.structured(base_metadata.history) 545 metadata.history = current_metadata 546 547 # Write metadata 548 metadata_yaml = OmegaConf.structured(metadata) 549 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 550 551 def _read_manifest(self, manifest_path: str, config: ASRTarredDatasetConfig): 552 """Read and filters data from the manifest""" 553 # Read the existing manifest 554 entries = [] 555 total_duration = 0.0 556 filtered_entries = [] 557 filtered_duration = 0.0 558 with open(manifest_path, 'r', encoding='utf-8') as m: 559 for line in m: 560 entry = json.loads(line) 561 if (config.max_duration is None or entry['duration'] < config.max_duration) and ( 562 config.min_duration is None or entry['duration'] >= config.min_duration 563 ): 564 entries.append(entry) 565 total_duration += entry["duration"] 566 else: 567 filtered_entries.append(entry) 568 filtered_duration += entry['duration'] 569 570 return entries, total_duration, filtered_entries, filtered_duration 571 572 def _create_shard(self, entries, target_dir, shard_id, manifest_folder): 573 """Creates a tarball containing the audio files from `entries`. 574 """ 575 if self.config.sort_in_shards: 576 entries.sort(key=lambda x: x["duration"], reverse=False) 577 578 new_entries = [] 579 tar = tarfile.open(os.path.join(target_dir, f'audio_{shard_id}.tar'), mode='w', dereference=True) 580 581 count = dict() 582 for entry in entries: 583 # We squash the filename since we do not preserve directory structure of audio files in the tarball. 584 if os.path.exists(entry["audio_filepath"]): 585 audio_filepath = entry["audio_filepath"] 586 else: 587 audio_filepath = os.path.join(manifest_folder, entry["audio_filepath"]) 588 if not os.path.exists(audio_filepath): 589 raise FileNotFoundError(f"Could not find {entry['audio_filepath']}!") 590 591 base, ext = os.path.splitext(audio_filepath) 592 base = base.replace('/', '_') 593 # Need the following replacement as long as WebDataset splits on first period 594 base = base.replace('.', '_') 595 squashed_filename = f'{base}{ext}' 596 if squashed_filename not in count: 597 tar.add(audio_filepath, arcname=squashed_filename) 598 to_write = squashed_filename 599 count[squashed_filename] = 1 600 else: 601 to_write = base + "-sub" + str(count[squashed_filename]) + ext 602 count[squashed_filename] += 1 603 604 new_entry = { 605 'audio_filepath': to_write, 606 'duration': entry['duration'], 607 'shard_id': shard_id, # Keep shard ID for recordkeeping 608 } 609 610 if 'label' in entry: 611 new_entry['label'] = entry['label'] 612 613 if 'text' in entry: 614 new_entry['text'] = entry['text'] 615 616 if 'offset' in entry: 617 new_entry['offset'] = entry['offset'] 618 619 if 'lang' in entry: 620 new_entry['lang'] = entry['lang'] 621 622 new_entries.append(new_entry) 623 624 tar.close() 625 return new_entries 626 627 @classmethod 628 def setup_history(cls, base_metadata: ASRTarredDatasetMetadata, history: List[Any]): 629 if 'history' in base_metadata.keys(): 630 for history_val in base_metadata.history: 631 cls.setup_history(history_val, history) 632 633 if base_metadata is not None: 634 metadata_copy = copy.deepcopy(base_metadata) 635 with open_dict(metadata_copy): 636 metadata_copy.pop('history', None) 637 history.append(metadata_copy) 638 639 640 def main(): 641 if args.buckets_num > 1: 642 bucket_length = (args.max_duration - args.min_duration) / float(args.buckets_num) 643 for i in range(args.buckets_num): 644 min_duration = args.min_duration + i * bucket_length 645 max_duration = min_duration + bucket_length 646 if i == args.buckets_num - 1: 647 # add a small number to cover the samples with exactly duration of max_duration in the last bucket. 648 max_duration += 1e-5 649 target_dir = os.path.join(args.target_dir, f"bucket{i+1}") 650 print(f"Creating bucket {i+1} with min_duration={min_duration} and max_duration={max_duration} ...") 651 print(f"Results are being saved at: {target_dir}.") 652 create_tar_datasets(min_duration=min_duration, max_duration=max_duration, target_dir=target_dir) 653 print(f"Bucket {i+1} is created.") 654 else: 655 create_tar_datasets(min_duration=args.min_duration, max_duration=args.max_duration, target_dir=args.target_dir) 656 657 658 def create_tar_datasets(min_duration: float, max_duration: float, target_dir: str): 659 builder = ASRTarredDatasetBuilder() 660 661 shard_manifests = False if args.no_shard_manifests else True 662 663 if args.write_metadata: 664 metadata = ASRTarredDatasetMetadata() 665 dataset_cfg = ASRTarredDatasetConfig( 666 num_shards=args.num_shards, 667 shuffle=args.shuffle, 668 max_duration=max_duration, 669 min_duration=min_duration, 670 shuffle_seed=args.shuffle_seed, 671 sort_in_shards=args.sort_in_shards, 672 shard_manifests=shard_manifests, 673 keep_files_together=args.keep_files_together, 674 ) 675 metadata.dataset_config = dataset_cfg 676 677 output_path = os.path.join(target_dir, 'default_metadata.yaml') 678 OmegaConf.save(metadata, output_path, resolve=True) 679 print(f"Default metadata written to {output_path}") 680 exit(0) 681 682 if args.concat_manifest_paths is None or len(args.concat_manifest_paths) == 0: 683 print("Creating new tarred dataset ...") 684 685 # Create a tarred dataset from scratch 686 config = ASRTarredDatasetConfig( 687 num_shards=args.num_shards, 688 shuffle=args.shuffle, 689 max_duration=max_duration, 690 min_duration=min_duration, 691 shuffle_seed=args.shuffle_seed, 692 sort_in_shards=args.sort_in_shards, 693 shard_manifests=shard_manifests, 694 keep_files_together=args.keep_files_together, 695 ) 696 builder.configure(config) 697 builder.create_new_dataset(manifest_path=args.manifest_path, target_dir=target_dir, num_workers=args.workers) 698 699 else: 700 if args.buckets_num > 1: 701 raise ValueError("Concatenation feature does not support buckets_num > 1.") 702 print("Concatenating multiple tarred datasets ...") 703 704 # Implicitly update config from base details 705 if args.metadata_path is not None: 706 metadata = ASRTarredDatasetMetadata.from_file(args.metadata_path) 707 else: 708 raise ValueError("`metadata` yaml file path must be provided!") 709 710 # Preserve history 711 history = [] 712 builder.setup_history(OmegaConf.structured(metadata), history) 713 metadata.history = history 714 715 # Add command line overrides (everything other than num_shards) 716 metadata.dataset_config.max_duration = max_duration 717 metadata.dataset_config.min_duration = min_duration 718 metadata.dataset_config.shuffle = args.shuffle 719 metadata.dataset_config.shuffle_seed = args.shuffle_seed 720 metadata.dataset_config.sort_in_shards = args.sort_in_shards 721 metadata.dataset_config.shard_manifests = shard_manifests 722 723 builder.configure(metadata.dataset_config) 724 725 # Concatenate a tarred dataset onto a previous one 726 builder.create_concatenated_dataset( 727 base_manifest_path=args.manifest_path, 728 manifest_paths=args.concat_manifest_paths, 729 metadata=metadata, 730 target_dir=target_dir, 731 num_workers=args.workers, 732 ) 733 734 if DALI_INDEX_SCRIPT_AVAILABLE and dali_index.INDEX_CREATOR_AVAILABLE: 735 print("Constructing DALI Tarfile Index - ", target_dir) 736 index_config = dali_index.DALITarredIndexConfig(tar_dir=target_dir, workers=args.workers) 737 dali_index.main(index_config) 738 739 740 if __name__ == "__main__": 741 main() 742 [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of tools/nemo_forced_aligner/align.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import math 17 import os 18 from dataclasses import dataclass, field, is_dataclass 19 from pathlib import Path 20 from typing import List, Optional 21 22 import torch 23 from omegaconf import OmegaConf 24 from utils.data_prep import ( 25 add_t_start_end_to_utt_obj, 26 get_batch_starts_ends, 27 get_batch_variables, 28 get_manifest_lines_batch, 29 is_entry_in_all_lines, 30 is_entry_in_any_lines, 31 ) 32 from utils.make_ass_files import make_ass_files 33 from utils.make_ctm_files import make_ctm_files 34 from utils.make_output_manifest import write_manifest_out_line 35 from utils.viterbi_decoding import viterbi_decoding 36 37 from nemo.collections.asr.models.ctc_models import EncDecCTCModel 38 from nemo.collections.asr.models.hybrid_rnnt_ctc_models import EncDecHybridRNNTCTCModel 39 from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchASR 40 from nemo.collections.asr.parts.utils.transcribe_utils import setup_model 41 from nemo.core.config import hydra_runner 42 from nemo.utils import logging 43 44 """ 45 Align the utterances in manifest_filepath. 46 Results are saved in ctm files in output_dir. 47 48 Arguments: 49 pretrained_name: string specifying the name of a CTC NeMo ASR model which will be automatically downloaded 50 from NGC and used for generating the log-probs which we will use to do alignment. 51 Note: NFA can only use CTC models (not Transducer models) at the moment. 52 model_path: string specifying the local filepath to a CTC NeMo ASR model which will be used to generate the 53 log-probs which we will use to do alignment. 54 Note: NFA can only use CTC models (not Transducer models) at the moment. 55 Note: if a model_path is provided, it will override the pretrained_name. 56 manifest_filepath: filepath to the manifest of the data you want to align, 57 containing 'audio_filepath' and 'text' fields. 58 output_dir: the folder where output CTM files and new JSON manifest will be saved. 59 align_using_pred_text: if True, will transcribe the audio using the specified model and then use that transcription 60 as the reference text for the forced alignment. 61 transcribe_device: None, or a string specifying the device that will be used for generating log-probs (i.e. "transcribing"). 62 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 63 (otherwise will set it to 'cpu'). 64 viterbi_device: None, or string specifying the device that will be used for doing Viterbi decoding. 65 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 66 (otherwise will set it to 'cpu'). 67 batch_size: int specifying batch size that will be used for generating log-probs and doing Viterbi decoding. 68 use_local_attention: boolean flag specifying whether to try to use local attention for the ASR Model (will only 69 work if the ASR Model is a Conformer model). If local attention is used, we will set the local attention context 70 size to [64,64]. 71 additional_segment_grouping_separator: an optional string used to separate the text into smaller segments. 72 If this is not specified, then the whole text will be treated as a single segment. 73 remove_blank_tokens_from_ctm: a boolean denoting whether to remove <blank> tokens from token-level output CTMs. 74 audio_filepath_parts_in_utt_id: int specifying how many of the 'parts' of the audio_filepath 75 we will use (starting from the final part of the audio_filepath) to determine the 76 utt_id that will be used in the CTM files. Note also that any spaces that are present in the audio_filepath 77 will be replaced with dashes, so as not to change the number of space-separated elements in the 78 CTM files. 79 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 1 => utt_id will be "e1" 80 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 2 => utt_id will be "d_e1" 81 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 3 => utt_id will be "c_d_e1" 82 use_buffered_infer: False, if set True, using streaming to do get the logits for alignment 83 This flag is useful when aligning large audio file. 84 However, currently the chunk streaming inference does not support batch inference, 85 which means even you set batch_size > 1, it will only infer one by one instead of doing 86 the whole batch inference together. 87 chunk_len_in_secs: float chunk length in seconds 88 total_buffer_in_secs: float Length of buffer (chunk + left and right padding) in seconds 89 chunk_batch_size: int batch size for buffered chunk inference, 90 which will cut one audio into segments and do inference on chunk_batch_size segments at a time 91 92 simulate_cache_aware_streaming: False, if set True, using cache aware streaming to do get the logits for alignment 93 94 save_output_file_formats: List of strings specifying what type of output files to save (default: ["ctm", "ass"]) 95 ctm_file_config: CTMFileConfig to specify the configuration of the output CTM files 96 ass_file_config: ASSFileConfig to specify the configuration of the output ASS files 97 """ 98 99 100 @dataclass 101 class CTMFileConfig: 102 remove_blank_tokens: bool = False 103 # minimum duration (in seconds) for timestamps in the CTM.If any line in the CTM has a 104 # duration lower than this, it will be enlarged from the middle outwards until it 105 # meets the minimum_timestamp_duration, or reaches the beginning or end of the audio file. 106 # Note that this may cause timestamps to overlap. 107 minimum_timestamp_duration: float = 0 108 109 110 @dataclass 111 class ASSFileConfig: 112 fontsize: int = 20 113 vertical_alignment: str = "center" 114 # if resegment_text_to_fill_space is True, the ASS files will use new segments 115 # such that each segment will not take up more than (approximately) max_lines_per_segment 116 # when the ASS file is applied to a video 117 resegment_text_to_fill_space: bool = False 118 max_lines_per_segment: int = 2 119 text_already_spoken_rgb: List[int] = field(default_factory=lambda: [49, 46, 61]) # dark gray 120 text_being_spoken_rgb: List[int] = field(default_factory=lambda: [57, 171, 9]) # dark green 121 text_not_yet_spoken_rgb: List[int] = field(default_factory=lambda: [194, 193, 199]) # light gray 122 123 124 @dataclass 125 class AlignmentConfig: 126 # Required configs 127 pretrained_name: Optional[str] = None 128 model_path: Optional[str] = None 129 manifest_filepath: Optional[str] = None 130 output_dir: Optional[str] = None 131 132 # General configs 133 align_using_pred_text: bool = False 134 transcribe_device: Optional[str] = None 135 viterbi_device: Optional[str] = None 136 batch_size: int = 1 137 use_local_attention: bool = True 138 additional_segment_grouping_separator: Optional[str] = None 139 audio_filepath_parts_in_utt_id: int = 1 140 141 # Buffered chunked streaming configs 142 use_buffered_chunked_streaming: bool = False 143 chunk_len_in_secs: float = 1.6 144 total_buffer_in_secs: float = 4.0 145 chunk_batch_size: int = 32 146 147 # Cache aware streaming configs 148 simulate_cache_aware_streaming: Optional[bool] = False 149 150 # Output file configs 151 save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) 152 ctm_file_config: CTMFileConfig = CTMFileConfig() 153 ass_file_config: ASSFileConfig = ASSFileConfig() 154 155 156 @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) 157 def main(cfg: AlignmentConfig): 158 159 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 160 161 if is_dataclass(cfg): 162 cfg = OmegaConf.structured(cfg) 163 164 # Validate config 165 if cfg.model_path is None and cfg.pretrained_name is None: 166 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None") 167 168 if cfg.model_path is not None and cfg.pretrained_name is not None: 169 raise ValueError("One of cfg.model_path and cfg.pretrained_name must be None") 170 171 if cfg.manifest_filepath is None: 172 raise ValueError("cfg.manifest_filepath must be specified") 173 174 if cfg.output_dir is None: 175 raise ValueError("cfg.output_dir must be specified") 176 177 if cfg.batch_size < 1: 178 raise ValueError("cfg.batch_size cannot be zero or a negative number") 179 180 if cfg.additional_segment_grouping_separator == "" or cfg.additional_segment_grouping_separator == " ": 181 raise ValueError("cfg.additional_grouping_separator cannot be empty string or space character") 182 183 if cfg.ctm_file_config.minimum_timestamp_duration < 0: 184 raise ValueError("cfg.minimum_timestamp_duration cannot be a negative number") 185 186 if cfg.ass_file_config.vertical_alignment not in ["top", "center", "bottom"]: 187 raise ValueError("cfg.ass_file_config.vertical_alignment must be one of 'top', 'center' or 'bottom'") 188 189 for rgb_list in [ 190 cfg.ass_file_config.text_already_spoken_rgb, 191 cfg.ass_file_config.text_already_spoken_rgb, 192 cfg.ass_file_config.text_already_spoken_rgb, 193 ]: 194 if len(rgb_list) != 3: 195 raise ValueError( 196 "cfg.ass_file_config.text_already_spoken_rgb," 197 " cfg.ass_file_config.text_being_spoken_rgb," 198 " and cfg.ass_file_config.text_already_spoken_rgb all need to contain" 199 " exactly 3 elements." 200 ) 201 202 # Validate manifest contents 203 if not is_entry_in_all_lines(cfg.manifest_filepath, "audio_filepath"): 204 raise RuntimeError( 205 "At least one line in cfg.manifest_filepath does not contain an 'audio_filepath' entry. " 206 "All lines must contain an 'audio_filepath' entry." 207 ) 208 209 if cfg.align_using_pred_text: 210 if is_entry_in_any_lines(cfg.manifest_filepath, "pred_text"): 211 raise RuntimeError( 212 "Cannot specify cfg.align_using_pred_text=True when the manifest at cfg.manifest_filepath " 213 "contains 'pred_text' entries. This is because the audio will be transcribed and may produce " 214 "a different 'pred_text'. This may cause confusion." 215 ) 216 else: 217 if not is_entry_in_all_lines(cfg.manifest_filepath, "text"): 218 raise RuntimeError( 219 "At least one line in cfg.manifest_filepath does not contain a 'text' entry. " 220 "NFA requires all lines to contain a 'text' entry when cfg.align_using_pred_text=False." 221 ) 222 223 # init devices 224 if cfg.transcribe_device is None: 225 transcribe_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 226 else: 227 transcribe_device = torch.device(cfg.transcribe_device) 228 logging.info(f"Device to be used for transcription step (`transcribe_device`) is {transcribe_device}") 229 230 if cfg.viterbi_device is None: 231 viterbi_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 232 else: 233 viterbi_device = torch.device(cfg.viterbi_device) 234 logging.info(f"Device to be used for viterbi step (`viterbi_device`) is {viterbi_device}") 235 236 if transcribe_device.type == 'cuda' or viterbi_device.type == 'cuda': 237 logging.warning( 238 'One or both of transcribe_device and viterbi_device are GPUs. If you run into OOM errors ' 239 'it may help to change both devices to be the CPU.' 240 ) 241 242 # load model 243 model, _ = setup_model(cfg, transcribe_device) 244 model.eval() 245 246 if isinstance(model, EncDecHybridRNNTCTCModel): 247 model.change_decoding_strategy(decoder_type="ctc") 248 249 if cfg.use_local_attention: 250 logging.info( 251 "Flag use_local_attention is set to True => will try to use local attention for model if it allows it" 252 ) 253 model.change_attention_model(self_attention_model="rel_pos_local_attn", att_context_size=[64, 64]) 254 255 if not (isinstance(model, EncDecCTCModel) or isinstance(model, EncDecHybridRNNTCTCModel)): 256 raise NotImplementedError( 257 f"Model is not an instance of NeMo EncDecCTCModel or ENCDecHybridRNNTCTCModel." 258 " Currently only instances of these models are supported" 259 ) 260 261 if cfg.ctm_file_config.minimum_timestamp_duration > 0: 262 logging.warning( 263 f"cfg.ctm_file_config.minimum_timestamp_duration has been set to {cfg.ctm_file_config.minimum_timestamp_duration} seconds. " 264 "This may cause the alignments for some tokens/words/additional segments to be overlapping." 265 ) 266 267 buffered_chunk_params = {} 268 if cfg.use_buffered_chunked_streaming: 269 model_cfg = copy.deepcopy(model._cfg) 270 271 OmegaConf.set_struct(model_cfg.preprocessor, False) 272 # some changes for streaming scenario 273 model_cfg.preprocessor.dither = 0.0 274 model_cfg.preprocessor.pad_to = 0 275 276 if model_cfg.preprocessor.normalize != "per_feature": 277 logging.error( 278 "Only EncDecCTCModelBPE models trained with per_feature normalization are supported currently" 279 ) 280 # Disable config overwriting 281 OmegaConf.set_struct(model_cfg.preprocessor, True) 282 283 feature_stride = model_cfg.preprocessor['window_stride'] 284 model_stride_in_secs = feature_stride * cfg.model_downsample_factor 285 total_buffer = cfg.total_buffer_in_secs 286 chunk_len = float(cfg.chunk_len_in_secs) 287 tokens_per_chunk = math.ceil(chunk_len / model_stride_in_secs) 288 mid_delay = math.ceil((chunk_len + (total_buffer - chunk_len) / 2) / model_stride_in_secs) 289 logging.info(f"tokens_per_chunk is {tokens_per_chunk}, mid_delay is {mid_delay}") 290 291 model = FrameBatchASR( 292 asr_model=model, 293 frame_len=chunk_len, 294 total_buffer=cfg.total_buffer_in_secs, 295 batch_size=cfg.chunk_batch_size, 296 ) 297 buffered_chunk_params = { 298 "delay": mid_delay, 299 "model_stride_in_secs": model_stride_in_secs, 300 "tokens_per_chunk": tokens_per_chunk, 301 } 302 # get start and end line IDs of batches 303 starts, ends = get_batch_starts_ends(cfg.manifest_filepath, cfg.batch_size) 304 305 # init output_timestep_duration = None and we will calculate and update it during the first batch 306 output_timestep_duration = None 307 308 # init f_manifest_out 309 os.makedirs(cfg.output_dir, exist_ok=True) 310 tgt_manifest_name = str(Path(cfg.manifest_filepath).stem) + "_with_output_file_paths.json" 311 tgt_manifest_filepath = str(Path(cfg.output_dir) / tgt_manifest_name) 312 f_manifest_out = open(tgt_manifest_filepath, 'w') 313 314 # get alignment and save in CTM batch-by-batch 315 for start, end in zip(starts, ends): 316 manifest_lines_batch = get_manifest_lines_batch(cfg.manifest_filepath, start, end) 317 318 (log_probs_batch, y_batch, T_batch, U_batch, utt_obj_batch, output_timestep_duration,) = get_batch_variables( 319 manifest_lines_batch, 320 model, 321 cfg.additional_segment_grouping_separator, 322 cfg.align_using_pred_text, 323 cfg.audio_filepath_parts_in_utt_id, 324 output_timestep_duration, 325 cfg.simulate_cache_aware_streaming, 326 cfg.use_buffered_chunked_streaming, 327 buffered_chunk_params, 328 ) 329 330 alignments_batch = viterbi_decoding(log_probs_batch, y_batch, T_batch, U_batch, viterbi_device) 331 332 for utt_obj, alignment_utt in zip(utt_obj_batch, alignments_batch): 333 334 utt_obj = add_t_start_end_to_utt_obj(utt_obj, alignment_utt, output_timestep_duration) 335 336 if "ctm" in cfg.save_output_file_formats: 337 utt_obj = make_ctm_files(utt_obj, cfg.output_dir, cfg.ctm_file_config,) 338 339 if "ass" in cfg.save_output_file_formats: 340 utt_obj = make_ass_files(utt_obj, cfg.output_dir, cfg.ass_file_config) 341 342 write_manifest_out_line( 343 f_manifest_out, utt_obj, 344 ) 345 346 f_manifest_out.close() 347 348 return None 349 350 351 if __name__ == "__main__": 352 main() 353 [end of tools/nemo_forced_aligner/align.py] </code> Here is an example of a patch file. It consists of changes to the code base. It specifies the file names, the line numbers of each change, and the removed and added lines. A single patch file can contain changes to multiple files. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch> I need you to solve the provided issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the format shown above. Respond below:
NVIDIA/NeMo
8a892b86186dbdf61803d75570cb5c58471e9dda
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-09-30T01:26:50Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,9 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,9 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): @@ -2217,7 +2219,9 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = ConfidenceMeasureConfig() + confidence_measure_cfg: Optional[ConfidenceMeasureConfig] = field( + default_factory=lambda: ConfidenceMeasureConfig() + ) confidence_method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -181,7 +181,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - measure_cfg: ConfidenceMeasureConfig = ConfidenceMeasureConfig() + measure_cfg: ConfidenceMeasureConfig = field(default_factory=lambda: ConfidenceMeasureConfig()) method_cfg: str = "DEPRECATED" def __post_init__(self): diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMeasureConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -110,7 +110,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
NVIDIA__NeMo-7616
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4 </issue> <code> [start of README.rst] 1 2 |status| |documentation| |codeql| |license| |pypi| |pyversion| |downloads| |black| 3 4 .. |status| image:: http://www.repostatus.org/badges/latest/active.svg 5 :target: http://www.repostatus.org/#active 6 :alt: Project Status: Active โ€“ The project has reached a stable, usable state and is being actively developed. 7 8 .. |documentation| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 9 :alt: Documentation 10 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 11 12 .. |license| image:: https://img.shields.io/badge/License-Apache%202.0-brightgreen.svg 13 :target: https://github.com/NVIDIA/NeMo/blob/master/LICENSE 14 :alt: NeMo core license and license for collections in this repo 15 16 .. |pypi| image:: https://badge.fury.io/py/nemo-toolkit.svg 17 :target: https://badge.fury.io/py/nemo-toolkit 18 :alt: Release version 19 20 .. |pyversion| image:: https://img.shields.io/pypi/pyversions/nemo-toolkit.svg 21 :target: https://badge.fury.io/py/nemo-toolkit 22 :alt: Python version 23 24 .. |downloads| image:: https://static.pepy.tech/personalized-badge/nemo-toolkit?period=total&units=international_system&left_color=grey&right_color=brightgreen&left_text=downloads 25 :target: https://pepy.tech/project/nemo-toolkit 26 :alt: PyPi total downloads 27 28 .. |codeql| image:: https://github.com/nvidia/nemo/actions/workflows/codeql.yml/badge.svg?branch=main&event=push 29 :target: https://github.com/nvidia/nemo/actions/workflows/codeql.yml 30 :alt: CodeQL 31 32 .. |black| image:: https://img.shields.io/badge/code%20style-black-000000.svg 33 :target: https://github.com/psf/black 34 :alt: Code style: black 35 36 .. _main-readme: 37 38 **NVIDIA NeMo** 39 =============== 40 41 Introduction 42 ------------ 43 44 NVIDIA NeMo is a conversational AI toolkit built for researchers working on automatic speech recognition (ASR), 45 text-to-speech synthesis (TTS), large language models (LLMs), and 46 natural language processing (NLP). 47 The primary objective of NeMo is to help researchers from industry and academia to reuse prior work (code and pretrained models) 48 and make it easier to create new `conversational AI models <https://developer.nvidia.com/conversational-ai#started>`_. 49 50 All NeMo models are trained with `Lightning <https://github.com/Lightning-AI/lightning>`_ and 51 training is automatically scalable to 1000s of GPUs. 52 Additionally, NeMo Megatron LLM models can be trained up to 1 trillion parameters using tensor and pipeline model parallelism. 53 NeMo models can be optimized for inference and deployed for production use-cases with `NVIDIA Riva <https://developer.nvidia.com/riva>`_. 54 55 Getting started with NeMo is simple. 56 State of the Art pretrained NeMo models are freely available on `HuggingFace Hub <https://huggingface.co/models?library=nemo&sort=downloads&search=nvidia>`_ and 57 `NVIDIA NGC <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_. 58 These models can be used to transcribe audio, synthesize speech, or translate text in just a few lines of code. 59 60 We have extensive `tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_ that 61 can all be run on `Google Colab <https://colab.research.google.com>`_. 62 63 For advanced users that want to train NeMo models from scratch or finetune existing NeMo models 64 we have a full suite of `example scripts <https://github.com/NVIDIA/NeMo/tree/main/examples>`_ that support multi-GPU/multi-node training. 65 66 For scaling NeMo LLM training on Slurm clusters or public clouds, please see the `NVIDIA NeMo Megatron Launcher <https://github.com/NVIDIA/NeMo-Megatron-Launcher>`_. 67 The NM launcher has extensive recipes, scripts, utilities, and documentation for training NeMo LLMs and also has an `Autoconfigurator <https://github.com/NVIDIA/NeMo-Megatron-Launcher#53-using-autoconfigurator-to-find-the-optimal-configuration>`_ 68 which can be used to find the optimal model parallel configuration for training on a specific cluster. 69 70 Also see the two introductory videos below for a high level overview of NeMo. 71 72 * Developing State-Of-The-Art Conversational AI Models in Three Lines of Code. 73 * NVIDIA NeMo: Toolkit for Conversational AI at PyData Yerevan 2022. 74 75 |three_lines| |pydata| 76 77 .. |pydata| image:: https://img.youtube.com/vi/J-P6Sczmas8/maxres3.jpg 78 :target: https://www.youtube.com/embed/J-P6Sczmas8?mute=0&start=14&autoplay=0 79 :width: 600 80 :alt: Develop Conversational AI Models in 3 Lines 81 82 .. |three_lines| image:: https://img.youtube.com/vi/wBgpMf_KQVw/maxresdefault.jpg 83 :target: https://www.youtube.com/embed/wBgpMf_KQVw?mute=0&start=0&autoplay=0 84 :width: 600 85 :alt: Introduction at PyData@Yerevan 2022 86 87 Key Features 88 ------------ 89 90 * Speech processing 91 * `HuggingFace Space for Audio Transcription (File, Microphone and YouTube) <https://huggingface.co/spaces/smajumdar/nemo_multilingual_language_id>`_ 92 * `Automatic Speech Recognition (ASR) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html>`_ 93 * Supported ASR models: `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html>`_ 94 * Jasper, QuartzNet, CitriNet, ContextNet 95 * Conformer-CTC, Conformer-Transducer, FastConformer-CTC, FastConformer-Transducer 96 * Squeezeformer-CTC and Squeezeformer-Transducer 97 * LSTM-Transducer (RNNT) and LSTM-CTC 98 * Supports the following decoders/losses: 99 * CTC 100 * Transducer/RNNT 101 * Hybrid Transducer/CTC 102 * NeMo Original `Multi-blank Transducers <https://arxiv.org/abs/2211.03541>`_ and `Token-and-Duration Transducers (TDT) <https://arxiv.org/abs/2304.06795>`_ 103 * Streaming/Buffered ASR (CTC/Transducer) - `Chunked Inference Examples <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_chunked_inference>`_ 104 * Cache-aware Streaming Conformer with multiple lookaheads - `<https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#cache-aware-streaming-conformer>`_ 105 * Beam Search decoding 106 * `Language Modelling for ASR (CTC and RNNT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html>`_: N-gram LM in fusion with Beam Search decoding, Neural Rescoring with Transformer 107 * `Support of long audios for Conformer with memory efficient local attention <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/results.html#inference-on-long-audio>`_ 108 * `Speech Classification, Speech Command Recognition and Language Identification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/intro.html>`_: MatchboxNet (Command Recognition), AmberNet (LangID) 109 * `Voice activity Detection (VAD) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/speech_classification/models.html#marblenet-vad>`_: MarbleNet 110 * ASR with VAD Inference - `Example <https://github.com/NVIDIA/NeMo/tree/stable/examples/asr/asr_vad>`_ 111 * `Speaker Recognition <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/intro.html>`_: TitaNet, ECAPA_TDNN, SpeakerNet 112 * `Speaker Diarization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_diarization/intro.html>`_ 113 * Clustering Diarizer: TitaNet, ECAPA_TDNN, SpeakerNet 114 * Neural Diarizer: MSDD (Multi-scale Diarization Decoder) 115 * `Speech Intent Detection and Slot Filling <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_intent_slot/intro.html>`_: Conformer-Transformer 116 * `Pretrained models on different languages. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_: English, Spanish, German, Russian, Chinese, French, Italian, Polish, ... 117 * `NGC collection of pre-trained speech processing models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_asr>`_ 118 * Natural Language Processing 119 * `NeMo Megatron pre-training of Large Language Models <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/nlp/nemo_megatron/intro.html>`_ 120 * `Neural Machine Translation (NMT) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/machine_translation/machine_translation.html>`_ 121 * `Punctuation and Capitalization <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/punctuation_and_capitalization.html>`_ 122 * `Token classification (named entity recognition) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/token_classification.html>`_ 123 * `Text classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_classification.html>`_ 124 * `Joint Intent and Slot Classification <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/joint_intent_slot.html>`_ 125 * `Question answering <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/question_answering.html>`_ 126 * `GLUE benchmark <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/glue_benchmark.html>`_ 127 * `Information retrieval <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/information_retrieval.html>`_ 128 * `Entity Linking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/entity_linking.html>`_ 129 * `Dialogue State Tracking <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/sgd_qa.html>`_ 130 * `Prompt Learning <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/nemo_megatron/prompt_learning.html>`_ 131 * `NGC collection of pre-trained NLP models. <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_nlp>`_ 132 * `Synthetic Tabular Data Generation <https://developer.nvidia.com/blog/generating-synthetic-data-with-transformers-a-solution-for-enterprise-data-challenges/>`_ 133 * Text-to-Speech Synthesis (TTS): 134 * `Documentation <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tts/intro.html#>`_ 135 * Mel-Spectrogram generators: FastPitch, SSL FastPitch, Mixer-TTS/Mixer-TTS-X, RAD-TTS, Tacotron2 136 * Vocoders: HiFiGAN, UnivNet, WaveGlow 137 * End-to-End Models: VITS 138 * `Pre-trained Model Checkpoints in NVIDIA GPU Cloud (NGC) <https://ngc.nvidia.com/catalog/collections/nvidia:nemo_tts>`_ 139 * `Tools <https://github.com/NVIDIA/NeMo/tree/stable/tools>`_ 140 * `Text Processing (text normalization and inverse text normalization) <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/nlp/text_normalization/intro.html>`_ 141 * `NeMo Forced Aligner <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/nemo_forced_aligner.html>`_ 142 * `CTC-Segmentation tool <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/ctc_segmentation.html>`_ 143 * `Speech Data Explorer <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/tools/speech_data_explorer.html>`_: a dash-based tool for interactive exploration of ASR/TTS datasets 144 * `Speech Data Processor <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/tools/speech_data_processor.html>`_ 145 146 147 Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. 148 149 Requirements 150 ------------ 151 152 1) Python 3.10 or above 153 2) Pytorch 1.13.1 or above 154 3) NVIDIA GPU for training 155 156 Documentation 157 ------------- 158 159 .. |main| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=main 160 :alt: Documentation Status 161 :scale: 100% 162 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/ 163 164 .. |stable| image:: https://readthedocs.com/projects/nvidia-nemo/badge/?version=stable 165 :alt: Documentation Status 166 :scale: 100% 167 :target: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/ 168 169 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 170 | Version | Status | Description | 171 +=========+=============+==========================================================================================================================================+ 172 | Latest | |main| | `Documentation of the latest (i.e. main) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/>`_ | 173 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 174 | Stable | |stable| | `Documentation of the stable (i.e. most recent release) branch. <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/>`_ | 175 +---------+-------------+------------------------------------------------------------------------------------------------------------------------------------------+ 176 177 Tutorials 178 --------- 179 A great way to start with NeMo is by checking `one of our tutorials <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/tutorials.html>`_. 180 181 Getting help with NeMo 182 ---------------------- 183 FAQ can be found on NeMo's `Discussions board <https://github.com/NVIDIA/NeMo/discussions>`_. You are welcome to ask questions or start discussions there. 184 185 186 Installation 187 ------------ 188 189 Conda 190 ~~~~~ 191 192 We recommend installing NeMo in a fresh Conda environment. 193 194 .. code-block:: bash 195 196 conda create --name nemo python==3.10.12 197 conda activate nemo 198 199 Install PyTorch using their `configurator <https://pytorch.org/get-started/locally/>`_. 200 201 .. code-block:: bash 202 203 conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia 204 205 The command used to install PyTorch may depend on your system. Please use the configurator linked above to find the right command for your system. 206 207 Pip 208 ~~~ 209 Use this installation mode if you want the latest released version. 210 211 .. code-block:: bash 212 213 apt-get update && apt-get install -y libsndfile1 ffmpeg 214 pip install Cython 215 pip install nemo_toolkit['all'] 216 217 Depending on the shell used, you may need to use ``"nemo_toolkit[all]"`` instead in the above command. 218 219 Pip from source 220 ~~~~~~~~~~~~~~~ 221 Use this installation mode if you want the version from a particular GitHub branch (e.g main). 222 223 .. code-block:: bash 224 225 apt-get update && apt-get install -y libsndfile1 ffmpeg 226 pip install Cython 227 python -m pip install git+https://github.com/NVIDIA/NeMo.git@{BRANCH}#egg=nemo_toolkit[all] 228 229 230 From source 231 ~~~~~~~~~~~ 232 Use this installation mode if you are contributing to NeMo. 233 234 .. code-block:: bash 235 236 apt-get update && apt-get install -y libsndfile1 ffmpeg 237 git clone https://github.com/NVIDIA/NeMo 238 cd NeMo 239 ./reinstall.sh 240 241 If you only want the toolkit without additional conda-based dependencies, you may replace ``reinstall.sh`` 242 with ``pip install -e .`` when your PWD is the root of the NeMo repository. 243 244 RNNT 245 ~~~~ 246 Note that RNNT requires numba to be installed from conda. 247 248 .. code-block:: bash 249 250 conda remove numba 251 pip uninstall numba 252 conda install -c conda-forge numba 253 254 NeMo Megatron 255 ~~~~~~~~~~~~~ 256 NeMo Megatron training requires NVIDIA Apex to be installed. 257 Install it manually if not using the NVIDIA PyTorch container. 258 259 To install Apex, run 260 261 .. code-block:: bash 262 263 git clone https://github.com/NVIDIA/apex.git 264 cd apex 265 git checkout 52e18c894223800cb611682dce27d88050edf1de 266 pip install -v --no-build-isolation --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_layer_norm" --global-option="--distributed_adam" --global-option="--deprecated_fused_adam" ./ 267 268 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Apex or any other dependencies. 269 270 While installing Apex, it may raise an error if the CUDA version on your system does not match the CUDA version torch was compiled with. 271 This raise can be avoided by commenting it here: https://github.com/NVIDIA/apex/blob/master/setup.py#L32 272 273 cuda-nvprof is needed to install Apex. The version should match the CUDA version that you are using: 274 275 .. code-block:: bash 276 277 conda install -c nvidia cuda-nvprof=11.8 278 279 packaging is also needed: 280 281 .. code-block:: bash 282 283 pip install packaging 284 285 With the latest versions of Apex, the `pyproject.toml` file in Apex may need to be deleted in order to install locally. 286 287 288 Transformer Engine 289 ~~~~~~~~~~~~~~~~~~ 290 NeMo Megatron GPT has been integrated with `NVIDIA Transformer Engine <https://github.com/NVIDIA/TransformerEngine>`_ 291 Transformer Engine enables FP8 training on NVIDIA Hopper GPUs. 292 `Install <https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/installation.html>`_ it manually if not using the NVIDIA PyTorch container. 293 294 .. code-block:: bash 295 296 pip install --upgrade git+https://github.com/NVIDIA/TransformerEngine.git@stable 297 298 It is highly recommended to use the NVIDIA PyTorch or NeMo container if having issues installing Transformer Engine or any other dependencies. 299 300 Transformer Engine requires PyTorch to be built with CUDA 11.8. 301 302 303 Flash Attention 304 ~~~~~~~~~~~~~~~~~~~~ 305 Transformer Engine already supports Flash Attention for GPT models. If you want to use Flash Attention for non-causal models or use with attention bias (introduced from position encoding, e.g. Alibi), please install `flash-attn <https://github.com/HazyResearch/flash-attention>`_. 306 307 .. code-block:: bash 308 309 pip install flash-attn 310 pip install triton==2.0.0.dev20221202 311 312 NLP inference UI 313 ~~~~~~~~~~~~~~~~~~~~ 314 To launch the inference web UI server, please install the gradio `gradio <https://gradio.app/>`_. 315 316 .. code-block:: bash 317 318 pip install gradio==3.34.0 319 320 NeMo Text Processing 321 ~~~~~~~~~~~~~~~~~~~~ 322 NeMo Text Processing, specifically (Inverse) Text Normalization, is now a separate repository `https://github.com/NVIDIA/NeMo-text-processing <https://github.com/NVIDIA/NeMo-text-processing>`_. 323 324 Docker containers: 325 ~~~~~~~~~~~~~~~~~~ 326 We release NeMo containers alongside NeMo releases. For example, NeMo ``r1.20.0`` comes with container ``nemo:23.06``, you may find more details about released containers in `releases page <https://github.com/NVIDIA/NeMo/releases>`_. 327 328 To use built container, please run 329 330 .. code-block:: bash 331 332 docker pull nvcr.io/nvidia/nemo:23.06 333 334 To build a nemo container with Dockerfile from a branch, please run 335 336 .. code-block:: bash 337 338 DOCKER_BUILDKIT=1 docker build -f Dockerfile -t nemo:latest . 339 340 341 If you chose to work with main branch, we recommend using NVIDIA's PyTorch container version 23.06-py3 and then installing from GitHub. 342 343 .. code-block:: bash 344 345 docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g \ 346 -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit \ 347 stack=67108864 --device=/dev/snd nvcr.io/nvidia/pytorch:23.06-py3 348 349 Examples 350 -------- 351 352 Many examples can be found under the `"Examples" <https://github.com/NVIDIA/NeMo/tree/stable/examples>`_ folder. 353 354 355 Contributing 356 ------------ 357 358 We welcome community contributions! Please refer to the `CONTRIBUTING.md <https://github.com/NVIDIA/NeMo/blob/stable/CONTRIBUTING.md>`_ CONTRIBUTING.md for the process. 359 360 Publications 361 ------------ 362 363 We provide an ever growing list of publications that utilize the NeMo framework. Please refer to `PUBLICATIONS.md <https://github.com/NVIDIA/NeMo/tree/stable/PUBLICATIONS.md>`_. We welcome the addition of your own articles to this list ! 364 365 License 366 ------- 367 NeMo is under `Apache 2.0 license <https://github.com/NVIDIA/NeMo/blob/stable/LICENSE>`_. 368 [end of README.rst] [start of examples/asr/experimental/k2/align_speech_parallel.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 # Based on examples/asr/transcribe_speech_parallel.py 17 # ASR alignment with multi-GPU/multi-node support for large datasets 18 # It supports both tarred and non-tarred datasets 19 # Arguments 20 # model: path to a nemo/PTL checkpoint file or name of a pretrained model 21 # predict_ds: config of the dataset/dataloader 22 # aligner_args: aligner config 23 # output_path: path to store the predictions 24 # model_stride: model downsampling factor, 8 for Citrinet models and 4 for Conformer models 25 # 26 # Results of each GPU/worker is written into a file named 'predictions_{rank}.json, and aggregated results of all workers are written into 'predictions_all.json' 27 28 Example for non-tarred datasets: 29 30 python align_speech_parallel.py \ 31 model=stt_en_conformer_ctc_large \ 32 predict_ds.manifest_filepath=/dataset/manifest_file.json \ 33 predict_ds.batch_size=16 \ 34 output_path=/tmp/ 35 36 Example for tarred datasets: 37 38 python align_speech_parallel.py \ 39 predict_ds.is_tarred=true \ 40 predict_ds.manifest_filepath=/tarred_dataset/tarred_audio_manifest.json \ 41 predict_ds.tarred_audio_filepaths=/tarred_dataset/audio__OP_0..127_CL_.tar \ 42 ... 43 44 By default the trainer uses all the GPUs available and default precision is FP32. 45 By setting the trainer config you may control these configs. For example to do the predictions with AMP on just two GPUs: 46 47 python align_speech_parallel.py \ 48 trainer.precision=16 \ 49 trainer.gpus=2 \ 50 ... 51 52 You may control the dataloader's config by setting the predict_ds: 53 54 python align_speech_parallel.py \ 55 predict_ds.num_workers=8 \ 56 predict_ds.min_duration=2.0 \ 57 predict_ds.sample_rate=16000 \ 58 model=stt_en_conformer_ctc_small \ 59 ... 60 61 You may control the aligner's config by setting the aligner_args: 62 aligner_args.alignment_type=argmax \ 63 aligner_args.word_output=False \ 64 aligner_args.cpu_decoding=True \ 65 aligner_args.decode_batch_size=8 \ 66 aligner_args.ctc_cfg.prob_suppress_index=-1 \ 67 aligner_args.ctc_cfg.prob_suppress_value=0.5 \ 68 aligner_args.rnnt_cfg.predictor_window_size=10 \ 69 aligner_args.decoder_module_cfg.intersect_pruned=true \ 70 aligner_args.decoder_module_cfg.intersect_conf.search_beam=40 \ 71 ... 72 73 """ 74 75 76 import os 77 from dataclasses import dataclass, is_dataclass 78 from typing import Optional 79 80 import pytorch_lightning as ptl 81 import torch 82 from omegaconf import MISSING, OmegaConf 83 84 from nemo.collections.asr.data.audio_to_ctm_dataset import ASRCTMPredictionWriter 85 from nemo.collections.asr.models import ASRModel 86 from nemo.collections.asr.models.configs.aligner_config import K2AlignerWrapperModelConfig 87 from nemo.collections.asr.models.configs.asr_models_config import ASRDatasetConfig 88 from nemo.collections.asr.models.k2_aligner_model import AlignerWrapperModel 89 from nemo.core.config import TrainerConfig, hydra_runner 90 from nemo.utils import logging 91 from nemo.utils.get_rank import is_global_rank_zero 92 93 94 @dataclass 95 class ParallelAlignmentConfig: 96 model: Optional[str] = None # name 97 predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) 98 aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() 99 output_path: str = MISSING 100 model_stride: int = 8 101 102 trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") 103 104 # there arguments will be ignored 105 return_predictions: bool = False 106 use_cer: bool = False 107 108 109 def match_train_config(predict_ds, train_ds): 110 # It copies the important configurations from the train dataset of the model 111 # into the predict_ds to be used for prediction. It is needed to match the training configurations. 112 if train_ds is None: 113 return 114 115 predict_ds.sample_rate = train_ds.get("sample_rate", 16000) 116 cfg_name_list = [ 117 "int_values", 118 "use_start_end_token", 119 "blank_index", 120 "unk_index", 121 "normalize", 122 "parser", 123 "eos_id", 124 "bos_id", 125 "pad_id", 126 ] 127 128 if is_dataclass(predict_ds): 129 predict_ds = OmegaConf.structured(predict_ds) 130 for cfg_name in cfg_name_list: 131 if hasattr(train_ds, cfg_name): 132 setattr(predict_ds, cfg_name, getattr(train_ds, cfg_name)) 133 134 return predict_ds 135 136 137 @hydra_runner(config_name="AlignmentConfig", schema=ParallelAlignmentConfig) 138 def main(cfg: ParallelAlignmentConfig): 139 if cfg.model.endswith(".nemo"): 140 logging.info("Attempting to initialize from .nemo file") 141 model = ASRModel.restore_from(restore_path=cfg.model, map_location="cpu") 142 elif cfg.model.endswith(".ckpt"): 143 logging.info("Attempting to initialize from .ckpt file") 144 model = ASRModel.load_from_checkpoint(checkpoint_path=cfg.model, map_location="cpu") 145 else: 146 logging.info( 147 "Attempting to initialize from a pretrained model as the model name does not have the extension of .nemo or .ckpt" 148 ) 149 model = ASRModel.from_pretrained(model_name=cfg.model, map_location="cpu") 150 151 trainer = ptl.Trainer(**cfg.trainer) 152 153 cfg.predict_ds.return_sample_id = True 154 cfg.return_predictions = False 155 cfg.use_cer = False 156 cfg.predict_ds = match_train_config(predict_ds=cfg.predict_ds, train_ds=model._cfg.train_ds) 157 data_loader = model._setup_dataloader_from_config(cfg.predict_ds) 158 159 os.makedirs(cfg.output_path, exist_ok=True) 160 # trainer.global_rank is not valid before predict() is called. Need this hack to find the correct global_rank. 161 global_rank = trainer.node_rank * trainer.num_devices + int(os.environ.get("LOCAL_RANK", 0)) 162 output_file = os.path.join(cfg.output_path, f"predictions_{global_rank}.json") 163 output_ctm_dir = os.path.join(cfg.output_path, "ctm") 164 predictor_writer = ASRCTMPredictionWriter( 165 dataset=data_loader.dataset, 166 output_file=output_file, 167 output_ctm_dir=output_ctm_dir, 168 time_per_frame=cfg.model_stride * model._cfg.preprocessor['window_stride'], 169 ) 170 trainer.callbacks.extend([predictor_writer]) 171 172 aligner_wrapper = AlignerWrapperModel(model=model, cfg=cfg.aligner_args) 173 trainer.predict(model=aligner_wrapper, dataloaders=data_loader, return_predictions=cfg.return_predictions) 174 samples_num = predictor_writer.close_output_file() 175 176 logging.info( 177 f"Prediction on rank {global_rank} is done for {samples_num} samples and results are stored in {output_file}." 178 ) 179 180 if torch.distributed.is_initialized(): 181 torch.distributed.barrier() 182 183 samples_num = 0 184 if is_global_rank_zero(): 185 output_file = os.path.join(cfg.output_path, f"predictions_all.json") 186 logging.info(f"Prediction files are being aggregated in {output_file}.") 187 with open(output_file, 'tw', encoding="utf-8") as outf: 188 for rank in range(trainer.world_size): 189 input_file = os.path.join(cfg.output_path, f"predictions_{rank}.json") 190 with open(input_file, 'r', encoding="utf-8") as inpf: 191 lines = inpf.readlines() 192 samples_num += len(lines) 193 outf.writelines(lines) 194 logging.info( 195 f"Prediction is done for {samples_num} samples in total on all workers and results are aggregated in {output_file}." 196 ) 197 198 199 if __name__ == '__main__': 200 main() 201 [end of examples/asr/experimental/k2/align_speech_parallel.py] [start of nemo/collections/asr/metrics/rnnt_wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import re 17 from abc import abstractmethod 18 from dataclasses import dataclass, is_dataclass 19 from typing import Callable, Dict, List, Optional, Tuple, Union 20 21 import editdistance 22 import numpy as np 23 import torch 24 from omegaconf import OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.metrics.wer import move_dimension_to_the_front 28 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding as beam_decode 29 from nemo.collections.asr.parts.submodules import rnnt_greedy_decoding as greedy_decode 30 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 31 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 32 from nemo.utils import logging 33 34 __all__ = ['RNNTDecoding', 'RNNTWER'] 35 36 37 class AbstractRNNTDecoding(ConfidenceMixin): 38 """ 39 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 40 41 Args: 42 decoding_cfg: A dict-like object which contains the following key-value pairs. 43 strategy: str value which represents the type of decoding that can occur. 44 Possible values are : 45 - greedy, greedy_batch (for greedy decoding). 46 - beam, tsd, alsd (for beam search decoding). 47 48 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 49 tokens as well as the decoded string. Default is False in order to avoid double decoding 50 unless required. 51 52 preserve_alignments: Bool flag which preserves the history of logprobs generated during 53 decoding (sample / batched). When set to true, the Hypothesis will contain 54 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 55 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 56 57 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 58 with the `return_hypotheses` flag set to True. 59 60 The length of the list corresponds to the Acoustic Length (T). 61 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 62 U is the number of target tokens for the current timestep Ti. 63 64 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 65 word based timestamp mapping the output log-probabilities to discrete intervals of timestamps. 66 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 67 68 rnnt_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 69 Can take the following values - "char" for character/subword time stamps, "word" for word level 70 time stamps and "all" (default), for both character level and word level time stamps. 71 72 word_seperator: Str token representing the seperator between words. 73 74 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 75 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 76 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of ints. 77 78 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 79 scores. In order to obtain hypotheses with confidence scores, please utilize 80 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 81 82 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 83 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 84 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 85 86 The length of the list corresponds to the Acoustic Length (T). 87 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 88 U is the number of target tokens for the current timestep Ti. 89 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 90 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 91 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 92 93 The length of the list corresponds to the number of recognized tokens. 94 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 95 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 96 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 97 98 The length of the list corresponds to the number of recognized words. 99 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 100 from the `token_confidence`. 101 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 102 Valid options are `mean`, `min`, `max`, `prod`. 103 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 104 confidence scores. 105 106 name: The method name (str). 107 Supported values: 108 - 'max_prob' for using the maximum token probability as a confidence. 109 - 'entropy' for using a normalized entropy of a log-likelihood vector. 110 111 entropy_type: Which type of entropy to use (str). 112 Used if confidence_method_cfg.name is set to `entropy`. 113 Supported values: 114 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 115 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 116 Note that for this entropy, the alpha should comply the following inequality: 117 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 118 where V is the model vocabulary size. 119 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 120 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 121 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 122 More: https://en.wikipedia.org/wiki/Tsallis_entropy 123 - 'renyi' for the Rรฉnyi entropy. 124 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 125 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 126 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 127 128 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 129 When the alpha equals one, scaling is not applied to 'max_prob', 130 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 131 132 entropy_norm: A mapping of the entropy value to the interval [0,1]. 133 Supported values: 134 - 'lin' for using the linear mapping. 135 - 'exp' for using exponential mapping with linear shift. 136 137 The config may further contain the following sub-dictionaries: 138 "greedy": 139 max_symbols: int, describing the maximum number of target tokens to decode per 140 timestep during greedy decoding. Setting to larger values allows longer sentences 141 to be decoded, at the cost of increased execution time. 142 preserve_frame_confidence: Same as above, overrides above value. 143 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 144 145 "beam": 146 beam_size: int, defining the beam size for beam search. Must be >= 1. 147 If beam_size == 1, will perform cached greedy search. This might be slightly different 148 results compared to the greedy search above. 149 150 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 151 Set to True by default. 152 153 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 154 hypotheses after beam search has concluded. This flag is set by default. 155 156 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 157 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 158 at increased cost to execution time. 159 160 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 161 If an integer is provided, it can decode sequences of that particular maximum length. 162 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 163 where seq_len is the length of the acoustic model output (T). 164 165 NOTE: 166 If a float is provided, it can be greater than 1! 167 By default, a float of 2.0 is used so that a target sequence can be at most twice 168 as long as the acoustic model output length T. 169 170 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 171 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 172 173 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 174 in order to reduce expensive beam search cost later. int >= 0. 175 176 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 177 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 178 and affects the speed of inference since large values will perform large beam search in the next step. 179 180 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 181 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 182 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 183 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 184 expansion apart from the "most likely" candidate. 185 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 186 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 187 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 188 tuned on a validation set. 189 190 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 191 192 decoder: The Decoder/Prediction network module. 193 joint: The Joint network module. 194 blank_id: The id of the RNNT blank token. 195 """ 196 197 def __init__(self, decoding_cfg, decoder, joint, blank_id: int): 198 super(AbstractRNNTDecoding, self).__init__() 199 200 # Convert dataclass to config object 201 if is_dataclass(decoding_cfg): 202 decoding_cfg = OmegaConf.structured(decoding_cfg) 203 204 self.cfg = decoding_cfg 205 self.blank_id = blank_id 206 self.num_extra_outputs = joint.num_extra_outputs 207 self.big_blank_durations = self.cfg.get("big_blank_durations", None) 208 self.durations = self.cfg.get("durations", None) 209 self.compute_hypothesis_token_set = self.cfg.get("compute_hypothesis_token_set", False) 210 self.compute_langs = decoding_cfg.get('compute_langs', False) 211 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 212 self.joint_fused_batch_size = self.cfg.get('fused_batch_size', None) 213 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 214 self.word_seperator = self.cfg.get('word_seperator', ' ') 215 216 if self.durations is not None: # this means it's a TDT model. 217 if blank_id == 0: 218 raise ValueError("blank_id must equal len(non_blank_vocabs) for TDT models") 219 if self.big_blank_durations is not None: 220 raise ValueError("duration and big_blank_durations can't both be not None") 221 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 222 raise ValueError("currently only greedy and greedy_batch inference is supported for TDT models") 223 224 if self.big_blank_durations is not None: # this means it's a multi-blank model. 225 if blank_id == 0: 226 raise ValueError("blank_id must equal len(vocabs) for multi-blank RNN-T models") 227 if self.cfg.strategy not in ['greedy', 'greedy_batch']: 228 raise ValueError( 229 "currently only greedy and greedy_batch inference is supported for multi-blank models" 230 ) 231 232 possible_strategies = ['greedy', 'greedy_batch', 'beam', 'tsd', 'alsd', 'maes'] 233 if self.cfg.strategy not in possible_strategies: 234 raise ValueError(f"Decoding strategy must be one of {possible_strategies}") 235 236 # Update preserve alignments 237 if self.preserve_alignments is None: 238 if self.cfg.strategy in ['greedy', 'greedy_batch']: 239 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 240 241 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 242 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 243 244 # Update compute timestamps 245 if self.compute_timestamps is None: 246 if self.cfg.strategy in ['greedy', 'greedy_batch']: 247 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 248 249 elif self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes']: 250 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 251 252 # Test if alignments are being preserved for RNNT 253 if self.compute_timestamps is True and self.preserve_alignments is False: 254 raise ValueError("If `compute_timesteps` flag is set, then `preserve_alignments` flag must also be set.") 255 256 # initialize confidence-related fields 257 self._init_confidence(self.cfg.get('confidence_cfg', None)) 258 259 # Confidence estimation is not implemented for these strategies 260 if ( 261 not self.preserve_frame_confidence 262 and self.cfg.strategy in ['beam', 'tsd', 'alsd', 'maes'] 263 and self.cfg.beam.get('preserve_frame_confidence', False) 264 ): 265 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 266 267 if self.cfg.strategy == 'greedy': 268 if self.big_blank_durations is None: 269 if self.durations is None: 270 self.decoding = greedy_decode.GreedyRNNTInfer( 271 decoder_model=decoder, 272 joint_model=joint, 273 blank_index=self.blank_id, 274 max_symbols_per_step=( 275 self.cfg.greedy.get('max_symbols', None) 276 or self.cfg.greedy.get('max_symbols_per_step', None) 277 ), 278 preserve_alignments=self.preserve_alignments, 279 preserve_frame_confidence=self.preserve_frame_confidence, 280 confidence_method_cfg=self.confidence_method_cfg, 281 ) 282 else: 283 self.decoding = greedy_decode.GreedyTDTInfer( 284 decoder_model=decoder, 285 joint_model=joint, 286 blank_index=self.blank_id, 287 durations=self.durations, 288 max_symbols_per_step=( 289 self.cfg.greedy.get('max_symbols', None) 290 or self.cfg.greedy.get('max_symbols_per_step', None) 291 ), 292 preserve_alignments=self.preserve_alignments, 293 preserve_frame_confidence=self.preserve_frame_confidence, 294 confidence_method_cfg=self.confidence_method_cfg, 295 ) 296 else: 297 self.decoding = greedy_decode.GreedyMultiblankRNNTInfer( 298 decoder_model=decoder, 299 joint_model=joint, 300 blank_index=self.blank_id, 301 big_blank_durations=self.big_blank_durations, 302 max_symbols_per_step=( 303 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 304 ), 305 preserve_alignments=self.preserve_alignments, 306 preserve_frame_confidence=self.preserve_frame_confidence, 307 confidence_method_cfg=self.confidence_method_cfg, 308 ) 309 310 elif self.cfg.strategy == 'greedy_batch': 311 if self.big_blank_durations is None: 312 if self.durations is None: 313 self.decoding = greedy_decode.GreedyBatchedRNNTInfer( 314 decoder_model=decoder, 315 joint_model=joint, 316 blank_index=self.blank_id, 317 max_symbols_per_step=( 318 self.cfg.greedy.get('max_symbols', None) 319 or self.cfg.greedy.get('max_symbols_per_step', None) 320 ), 321 preserve_alignments=self.preserve_alignments, 322 preserve_frame_confidence=self.preserve_frame_confidence, 323 confidence_method_cfg=self.confidence_method_cfg, 324 ) 325 else: 326 self.decoding = greedy_decode.GreedyBatchedTDTInfer( 327 decoder_model=decoder, 328 joint_model=joint, 329 blank_index=self.blank_id, 330 durations=self.durations, 331 max_symbols_per_step=( 332 self.cfg.greedy.get('max_symbols', None) 333 or self.cfg.greedy.get('max_symbols_per_step', None) 334 ), 335 preserve_alignments=self.preserve_alignments, 336 preserve_frame_confidence=self.preserve_frame_confidence, 337 confidence_method_cfg=self.confidence_method_cfg, 338 ) 339 340 else: 341 self.decoding = greedy_decode.GreedyBatchedMultiblankRNNTInfer( 342 decoder_model=decoder, 343 joint_model=joint, 344 blank_index=self.blank_id, 345 big_blank_durations=self.big_blank_durations, 346 max_symbols_per_step=( 347 self.cfg.greedy.get('max_symbols', None) or self.cfg.greedy.get('max_symbols_per_step', None) 348 ), 349 preserve_alignments=self.preserve_alignments, 350 preserve_frame_confidence=self.preserve_frame_confidence, 351 confidence_method_cfg=self.confidence_method_cfg, 352 ) 353 354 elif self.cfg.strategy == 'beam': 355 356 self.decoding = beam_decode.BeamRNNTInfer( 357 decoder_model=decoder, 358 joint_model=joint, 359 beam_size=self.cfg.beam.beam_size, 360 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 361 search_type='default', 362 score_norm=self.cfg.beam.get('score_norm', True), 363 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 364 preserve_alignments=self.preserve_alignments, 365 ) 366 367 elif self.cfg.strategy == 'tsd': 368 369 self.decoding = beam_decode.BeamRNNTInfer( 370 decoder_model=decoder, 371 joint_model=joint, 372 beam_size=self.cfg.beam.beam_size, 373 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 374 search_type='tsd', 375 score_norm=self.cfg.beam.get('score_norm', True), 376 tsd_max_sym_exp_per_step=self.cfg.beam.get('tsd_max_sym_exp', 10), 377 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 378 preserve_alignments=self.preserve_alignments, 379 ) 380 381 elif self.cfg.strategy == 'alsd': 382 383 self.decoding = beam_decode.BeamRNNTInfer( 384 decoder_model=decoder, 385 joint_model=joint, 386 beam_size=self.cfg.beam.beam_size, 387 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 388 search_type='alsd', 389 score_norm=self.cfg.beam.get('score_norm', True), 390 alsd_max_target_len=self.cfg.beam.get('alsd_max_target_len', 2), 391 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 392 preserve_alignments=self.preserve_alignments, 393 ) 394 395 elif self.cfg.strategy == 'maes': 396 397 self.decoding = beam_decode.BeamRNNTInfer( 398 decoder_model=decoder, 399 joint_model=joint, 400 beam_size=self.cfg.beam.beam_size, 401 return_best_hypothesis=decoding_cfg.beam.get('return_best_hypothesis', True), 402 search_type='maes', 403 score_norm=self.cfg.beam.get('score_norm', True), 404 maes_num_steps=self.cfg.beam.get('maes_num_steps', 2), 405 maes_prefix_alpha=self.cfg.beam.get('maes_prefix_alpha', 1), 406 maes_expansion_gamma=self.cfg.beam.get('maes_expansion_gamma', 2.3), 407 maes_expansion_beta=self.cfg.beam.get('maes_expansion_beta', 2.0), 408 softmax_temperature=self.cfg.beam.get('softmax_temperature', 1.0), 409 preserve_alignments=self.preserve_alignments, 410 ngram_lm_model=self.cfg.beam.get('ngram_lm_model', None), 411 ngram_lm_alpha=self.cfg.beam.get('ngram_lm_alpha', 0.0), 412 hat_subtract_ilm=self.cfg.beam.get('hat_subtract_ilm', False), 413 hat_ilm_weight=self.cfg.beam.get('hat_ilm_weight', 0.0), 414 ) 415 416 else: 417 418 raise ValueError( 419 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 420 f"but was provided {self.cfg.strategy}" 421 ) 422 423 # Update the joint fused batch size or disable it entirely if needed. 424 self.update_joint_fused_batch_size() 425 426 def rnnt_decoder_predictions_tensor( 427 self, 428 encoder_output: torch.Tensor, 429 encoded_lengths: torch.Tensor, 430 return_hypotheses: bool = False, 431 partial_hypotheses: Optional[List[Hypothesis]] = None, 432 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 433 """ 434 Decode an encoder output by autoregressive decoding of the Decoder+Joint networks. 435 436 Args: 437 encoder_output: torch.Tensor of shape [B, D, T]. 438 encoded_lengths: torch.Tensor containing lengths of the padded encoder outputs. Shape [B]. 439 return_hypotheses: bool. If set to True it will return list of Hypothesis or NBestHypotheses 440 441 Returns: 442 If `return_best_hypothesis` is set: 443 A tuple (hypotheses, None): 444 hypotheses - list of Hypothesis (best hypothesis per sample). 445 Look at rnnt_utils.Hypothesis for more information. 446 447 If `return_best_hypothesis` is not set: 448 A tuple(hypotheses, all_hypotheses) 449 hypotheses - list of Hypothesis (best hypothesis per sample). 450 Look at rnnt_utils.Hypothesis for more information. 451 all_hypotheses - list of NBestHypotheses. Each NBestHypotheses further contains a sorted 452 list of all the hypotheses of the model per sample. 453 Look at rnnt_utils.NBestHypotheses for more information. 454 """ 455 # Compute hypotheses 456 with torch.inference_mode(): 457 hypotheses_list = self.decoding( 458 encoder_output=encoder_output, encoded_lengths=encoded_lengths, partial_hypotheses=partial_hypotheses 459 ) # type: [List[Hypothesis]] 460 461 # extract the hypotheses 462 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 463 464 prediction_list = hypotheses_list 465 466 if isinstance(prediction_list[0], NBestHypotheses): 467 hypotheses = [] 468 all_hypotheses = [] 469 470 for nbest_hyp in prediction_list: # type: NBestHypotheses 471 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 472 decoded_hyps = self.decode_hypothesis(n_hyps) # type: List[str] 473 474 # If computing timestamps 475 if self.compute_timestamps is True: 476 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 477 for hyp_idx in range(len(decoded_hyps)): 478 decoded_hyps[hyp_idx] = self.compute_rnnt_timestamps(decoded_hyps[hyp_idx], timestamp_type) 479 480 hypotheses.append(decoded_hyps[0]) # best hypothesis 481 all_hypotheses.append(decoded_hyps) 482 483 if return_hypotheses: 484 return hypotheses, all_hypotheses 485 486 best_hyp_text = [h.text for h in hypotheses] 487 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 488 return best_hyp_text, all_hyp_text 489 490 else: 491 hypotheses = self.decode_hypothesis(prediction_list) # type: List[str] 492 493 # If computing timestamps 494 if self.compute_timestamps is True: 495 timestamp_type = self.cfg.get('rnnt_timestamp_type', 'all') 496 for hyp_idx in range(len(hypotheses)): 497 hypotheses[hyp_idx] = self.compute_rnnt_timestamps(hypotheses[hyp_idx], timestamp_type) 498 499 if return_hypotheses: 500 # greedy decoding, can get high-level confidence scores 501 if self.preserve_frame_confidence and ( 502 self.preserve_word_confidence or self.preserve_token_confidence 503 ): 504 hypotheses = self.compute_confidence(hypotheses) 505 return hypotheses, None 506 507 best_hyp_text = [h.text for h in hypotheses] 508 return best_hyp_text, None 509 510 def decode_hypothesis(self, hypotheses_list: List[Hypothesis]) -> List[Union[Hypothesis, NBestHypotheses]]: 511 """ 512 Decode a list of hypotheses into a list of strings. 513 514 Args: 515 hypotheses_list: List of Hypothesis. 516 517 Returns: 518 A list of strings. 519 """ 520 for ind in range(len(hypotheses_list)): 521 # Extract the integer encoded hypothesis 522 prediction = hypotheses_list[ind].y_sequence 523 524 if type(prediction) != list: 525 prediction = prediction.tolist() 526 527 # RNN-T sample level is already preprocessed by implicit RNNT decoding 528 # Simply remove any blank and possibly big blank tokens 529 if self.big_blank_durations is not None: # multi-blank RNNT 530 num_extra_outputs = len(self.big_blank_durations) 531 prediction = [p for p in prediction if p < self.blank_id - num_extra_outputs] 532 elif self.durations is not None: # TDT model. 533 prediction = [p for p in prediction if p < self.blank_id] 534 else: # standard RNN-T 535 prediction = [p for p in prediction if p != self.blank_id] 536 537 # De-tokenize the integer tokens; if not computing timestamps 538 if self.compute_timestamps is True: 539 # keep the original predictions, wrap with the number of repetitions per token and alignments 540 # this is done so that `rnnt_decoder_predictions_tensor()` can process this hypothesis 541 # in order to compute exact time stamps. 542 alignments = copy.deepcopy(hypotheses_list[ind].alignments) 543 token_repetitions = [1] * len(alignments) # preserve number of repetitions per token 544 hypothesis = (prediction, alignments, token_repetitions) 545 else: 546 hypothesis = self.decode_tokens_to_str(prediction) 547 548 # TODO: remove 549 # collapse leading spaces before . , ? for PC models 550 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 551 552 if self.compute_hypothesis_token_set: 553 hypotheses_list[ind].tokens = self.decode_ids_to_tokens(prediction) 554 555 # De-tokenize the integer tokens 556 hypotheses_list[ind].text = hypothesis 557 558 return hypotheses_list 559 560 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 561 """ 562 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 563 Assumes that `frame_confidence` is present in the hypotheses. 564 565 Args: 566 hypotheses_list: List of Hypothesis. 567 568 Returns: 569 A list of hypotheses with high-level confidence scores. 570 """ 571 if self.exclude_blank_from_confidence: 572 for hyp in hypotheses_list: 573 hyp.token_confidence = hyp.non_blank_frame_confidence 574 else: 575 for hyp in hypotheses_list: 576 offset = 0 577 token_confidence = [] 578 if len(hyp.timestep) > 0: 579 for ts, te in zip(hyp.timestep, hyp.timestep[1:] + [len(hyp.frame_confidence)]): 580 if ts != te: 581 # <blank> tokens are considered to belong to the last non-blank token, if any. 582 token_confidence.append( 583 self._aggregate_confidence( 584 [hyp.frame_confidence[ts][offset]] 585 + [fc[0] for fc in hyp.frame_confidence[ts + 1 : te]] 586 ) 587 ) 588 offset = 0 589 else: 590 token_confidence.append(hyp.frame_confidence[ts][offset]) 591 offset += 1 592 hyp.token_confidence = token_confidence 593 if self.preserve_word_confidence: 594 for hyp in hypotheses_list: 595 hyp.word_confidence = self._aggregate_token_confidence(hyp) 596 return hypotheses_list 597 598 @abstractmethod 599 def decode_tokens_to_str(self, tokens: List[int]) -> str: 600 """ 601 Implemented by subclass in order to decoder a token id list into a string. 602 603 Args: 604 tokens: List of int representing the token ids. 605 606 Returns: 607 A decoded string. 608 """ 609 raise NotImplementedError() 610 611 @abstractmethod 612 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 613 """ 614 Implemented by subclass in order to decode a token id list into a token list. 615 A token list is the string representation of each token id. 616 617 Args: 618 tokens: List of int representing the token ids. 619 620 Returns: 621 A list of decoded tokens. 622 """ 623 raise NotImplementedError() 624 625 @abstractmethod 626 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 627 """ 628 Implemented by subclass in order to 629 compute the most likely language ID (LID) string given the tokens. 630 631 Args: 632 tokens: List of int representing the token ids. 633 634 Returns: 635 A decoded LID string. 636 """ 637 raise NotImplementedError() 638 639 @abstractmethod 640 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 641 """ 642 Implemented by subclass in order to 643 decode a token id list into language ID (LID) list. 644 645 Args: 646 tokens: List of int representing the token ids. 647 648 Returns: 649 A list of decoded LIDS. 650 """ 651 raise NotImplementedError() 652 653 def update_joint_fused_batch_size(self): 654 if self.joint_fused_batch_size is None: 655 # do nothing and let the Joint itself handle setting up of the fused batch 656 return 657 658 if not hasattr(self.decoding.joint, 'set_fused_batch_size'): 659 logging.warning( 660 "The joint module does not have `set_fused_batch_size(int)` as a setter function.\n" 661 "Ignoring update of joint fused batch size." 662 ) 663 return 664 665 if not hasattr(self.decoding.joint, 'set_fuse_loss_wer'): 666 logging.warning( 667 "The joint module does not have `set_fuse_loss_wer(bool, RNNTLoss, RNNTWER)` " 668 "as a setter function.\n" 669 "Ignoring update of joint fused batch size." 670 ) 671 return 672 673 if self.joint_fused_batch_size > 0: 674 self.decoding.joint.set_fused_batch_size(self.joint_fused_batch_size) 675 else: 676 logging.info("Joint fused batch size <= 0; Will temporarily disable fused batch step in the Joint.") 677 self.decoding.joint.set_fuse_loss_wer(False) 678 679 def compute_rnnt_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 680 assert timestamp_type in ['char', 'word', 'all'] 681 682 # Unpack the temporary storage 683 decoded_prediction, alignments, token_repetitions = hypothesis.text 684 685 # Retrieve offsets 686 char_offsets = word_offsets = None 687 char_offsets = self._compute_offsets(hypothesis, token_repetitions, self.blank_id) 688 689 # finally, set the flattened decoded predictions to text field for later text decoding 690 hypothesis.text = decoded_prediction 691 692 # Assert number of offsets and hypothesis tokens are 1:1 match. 693 num_flattened_tokens = 0 694 for t in range(len(char_offsets)): 695 # Subtract one here for the extra RNNT BLANK token emitted to designate "End of timestep" 696 num_flattened_tokens += len(char_offsets[t]['char']) - 1 697 698 if num_flattened_tokens != len(hypothesis.text): 699 raise ValueError( 700 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 701 " have to be of the same length, but are: " 702 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 703 f" {len(hypothesis.text)}" 704 ) 705 706 encoded_char_offsets = copy.deepcopy(char_offsets) 707 708 # Correctly process the token ids to chars/subwords. 709 for i, offsets in enumerate(char_offsets): 710 decoded_chars = [] 711 for char in offsets['char'][:-1]: # ignore the RNNT Blank token at end of every timestep with -1 subset 712 decoded_chars.append(self.decode_tokens_to_str([int(char)])) 713 char_offsets[i]["char"] = decoded_chars 714 715 # detect char vs subword models 716 lens = [] 717 for v in char_offsets: 718 tokens = v["char"] 719 # each token may be either 1 unicode token or multiple unicode token 720 # for character based models, only 1 token is used 721 # for subword, more than one token can be used. 722 # Computing max, then summing up total lens is a test to check for char vs subword 723 # For char models, len(lens) == sum(lens) 724 # but this is violated for subword models. 725 max_len = max(len(c) for c in tokens) 726 lens.append(max_len) 727 728 # array of one or more chars implies subword based model with multiple char emitted per TxU step (via subword) 729 if sum(lens) > len(lens): 730 text_type = 'subword' 731 else: 732 # full array of ones implies character based model with 1 char emitted per TxU step 733 text_type = 'char' 734 735 # retrieve word offsets from character offsets 736 word_offsets = None 737 if timestamp_type in ['word', 'all']: 738 if text_type == 'char': 739 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 740 else: 741 # utilize the copy of char offsets with the correct integer ids for tokens 742 # so as to avoid tokenize -> detokenize -> compare -> merge steps. 743 word_offsets = self._get_word_offsets_subwords_sentencepiece( 744 encoded_char_offsets, 745 hypothesis, 746 decode_ids_to_tokens=self.decode_ids_to_tokens, 747 decode_tokens_to_str=self.decode_tokens_to_str, 748 ) 749 750 # attach results 751 if len(hypothesis.timestep) > 0: 752 timestep_info = hypothesis.timestep 753 else: 754 timestep_info = [] 755 756 # Setup defaults 757 hypothesis.timestep = {"timestep": timestep_info} 758 759 # Add char / subword time stamps 760 if char_offsets is not None and timestamp_type in ['char', 'all']: 761 hypothesis.timestep['char'] = char_offsets 762 763 # Add word time stamps 764 if word_offsets is not None and timestamp_type in ['word', 'all']: 765 hypothesis.timestep['word'] = word_offsets 766 767 # Convert the flattened token indices to text 768 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 769 770 return hypothesis 771 772 @staticmethod 773 def _compute_offsets( 774 hypothesis: Hypothesis, token_repetitions: List[int], rnnt_token: int 775 ) -> List[Dict[str, Union[str, int]]]: 776 """ 777 Utility method that calculates the indidual time indices where a token starts and ends. 778 779 Args: 780 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 781 emitted at every time step after rnnt collapse. 782 token_repetitions: A list of ints representing the number of repetitions of each emitted token. 783 rnnt_token: The integer of the rnnt blank token used during rnnt collapse. 784 785 Returns: 786 787 """ 788 start_index = 0 789 790 # If the exact timestep information is available, utilize the 1st non-rnnt blank token timestep 791 # as the start index. 792 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 793 start_index = max(0, hypothesis.timestep[0] - 1) 794 795 # Construct the start and end indices brackets 796 end_indices = np.asarray(token_repetitions).cumsum() 797 start_indices = np.concatenate(([start_index], end_indices[:-1])) 798 799 # Process the TxU dangling alignment tensor, containing pairs of (logits, label) 800 alignment_labels = [al_logits_labels for al_logits_labels in hypothesis.text[1]] 801 for t in range(len(alignment_labels)): 802 for u in range(len(alignment_labels[t])): 803 alignment_labels[t][u] = alignment_labels[t][u][1] # pick label from (logit, label) tuple 804 805 # Merge the results per token into a list of dictionaries 806 offsets = [ 807 {"char": a, "start_offset": s, "end_offset": e} 808 for a, s, e in zip(alignment_labels, start_indices, end_indices) 809 ] 810 811 # Filter out RNNT token (blank at [t][0] position). This is because blank can only occur at end of a 812 # time step for RNNT, so if 0th token is blank, then that timestep is skipped. 813 offsets = list(filter(lambda offsets: offsets["char"][0] != rnnt_token, offsets)) 814 return offsets 815 816 @staticmethod 817 def _get_word_offsets_chars( 818 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 819 ) -> Dict[str, Union[str, float]]: 820 """ 821 Utility method which constructs word time stamps out of character time stamps. 822 823 References: 824 This code is a port of the Hugging Face code for word time stamp construction. 825 826 Args: 827 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 828 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 829 830 Returns: 831 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 832 "end_offset". 833 """ 834 word_offsets = [] 835 836 last_state = "SPACE" 837 word = "" 838 start_offset = 0 839 end_offset = 0 840 for i, offset in enumerate(offsets): 841 chars = offset["char"] 842 for char in chars: 843 state = "SPACE" if char == word_delimiter_char else "WORD" 844 845 if state == last_state: 846 # If we are in the same state as before, we simply repeat what we've done before 847 end_offset = offset["end_offset"] 848 word += char 849 else: 850 # Switching state 851 if state == "SPACE": 852 # Finishing a word 853 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 854 else: 855 # Starting a new word 856 start_offset = offset["start_offset"] 857 end_offset = offset["end_offset"] 858 word = char 859 860 last_state = state 861 862 if last_state == "WORD": 863 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 864 865 return word_offsets 866 867 @staticmethod 868 def _get_word_offsets_subwords_sentencepiece( 869 offsets: Dict[str, Union[str, float]], 870 hypothesis: Hypothesis, 871 decode_ids_to_tokens: Callable[[List[int]], str], 872 decode_tokens_to_str: Callable[[List[int]], str], 873 ) -> Dict[str, Union[str, float]]: 874 """ 875 Utility method which constructs word time stamps out of sub-word time stamps. 876 877 **Note**: Only supports Sentencepiece based tokenizers ! 878 879 Args: 880 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 881 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 882 after rnnt collapse. 883 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 884 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 885 886 Returns: 887 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 888 "end_offset". 889 """ 890 word_offsets = [] 891 built_token = [] 892 previous_token_index = 0 893 # For every offset token 894 for i, offset in enumerate(offsets): 895 # For every subword token in offset token list (ignoring the RNNT Blank token at the end) 896 for char in offset['char'][:-1]: 897 char = int(char) 898 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if built_token: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 # This should only be done when these arrays contain more than one element. 931 if offsets and word_offsets: 932 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 933 934 # If there are any remaining tokens left, inject them all into the final word offset. 935 # The start offset of this token is the start time of the next token to process. 936 # The end offset of this token is the end time of the last token from offsets. 937 # Note that built_token is a flat list; but offsets contains a nested list which 938 # may have different dimensionality. 939 # As such, we can't rely on the length of the list of built_token to index offsets. 940 if built_token: 941 # start from the previous token index as this hasn't been committed to word_offsets yet 942 # if we still have content in built_token 943 start_offset = offsets[previous_token_index]["start_offset"] 944 word_offsets.append( 945 { 946 "word": decode_tokens_to_str(built_token), 947 "start_offset": start_offset, 948 "end_offset": offsets[-1]["end_offset"], 949 } 950 ) 951 built_token.clear() 952 953 return word_offsets 954 955 956 class RNNTDecoding(AbstractRNNTDecoding): 957 """ 958 Used for performing RNN-T auto-regressive decoding of the Decoder+Joint network given the encoder state. 959 960 Args: 961 decoding_cfg: A dict-like object which contains the following key-value pairs. 962 strategy: str value which represents the type of decoding that can occur. 963 Possible values are : 964 - greedy, greedy_batch (for greedy decoding). 965 - beam, tsd, alsd (for beam search decoding). 966 967 compute_hypothesis_token_set: A bool flag, which determines whether to compute a list of decoded 968 tokens as well as the decoded string. Default is False in order to avoid double decoding 969 unless required. 970 971 preserve_alignments: Bool flag which preserves the history of logprobs generated during 972 decoding (sample / batched). When set to true, the Hypothesis will contain 973 the non-null value for `logprobs` in it. Here, `alignments` is a List of List of 974 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 975 976 In order to obtain this hypothesis, please utilize `rnnt_decoder_predictions_tensor` function 977 with the `return_hypotheses` flag set to True. 978 979 The length of the list corresponds to the Acoustic Length (T). 980 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 981 U is the number of target tokens for the current timestep Ti. 982 983 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 984 scores. In order to obtain hypotheses with confidence scores, please utilize 985 `rnnt_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 986 987 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 988 generated during decoding (sample / batched). When set to true, the Hypothesis will contain 989 the non-null value for `frame_confidence` in it. Here, `alignments` is a List of List of floats. 990 991 The length of the list corresponds to the Acoustic Length (T). 992 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 993 U is the number of target tokens for the current timestep Ti. 994 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 995 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 996 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 997 998 The length of the list corresponds to the number of recognized tokens. 999 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1000 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1001 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1002 1003 The length of the list corresponds to the number of recognized words. 1004 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1005 from the `token_confidence`. 1006 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1007 Valid options are `mean`, `min`, `max`, `prod`. 1008 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1009 confidence scores. 1010 1011 name: The method name (str). 1012 Supported values: 1013 - 'max_prob' for using the maximum token probability as a confidence. 1014 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1015 1016 entropy_type: Which type of entropy to use (str). 1017 Used if confidence_method_cfg.name is set to `entropy`. 1018 Supported values: 1019 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1020 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1021 Note that for this entropy, the alpha should comply the following inequality: 1022 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1023 where V is the model vocabulary size. 1024 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1025 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1026 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1027 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1028 - 'renyi' for the Rรฉnyi entropy. 1029 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1030 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1031 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1032 1033 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1034 When the alpha equals one, scaling is not applied to 'max_prob', 1035 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1036 1037 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1038 Supported values: 1039 - 'lin' for using the linear mapping. 1040 - 'exp' for using exponential mapping with linear shift. 1041 1042 The config may further contain the following sub-dictionaries: 1043 "greedy": 1044 max_symbols: int, describing the maximum number of target tokens to decode per 1045 timestep during greedy decoding. Setting to larger values allows longer sentences 1046 to be decoded, at the cost of increased execution time. 1047 1048 preserve_frame_confidence: Same as above, overrides above value. 1049 1050 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 1051 1052 "beam": 1053 beam_size: int, defining the beam size for beam search. Must be >= 1. 1054 If beam_size == 1, will perform cached greedy search. This might be slightly different 1055 results compared to the greedy search above. 1056 1057 score_norm: optional bool, whether to normalize the returned beam score in the hypotheses. 1058 Set to True by default. 1059 1060 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1061 hypotheses after beam search has concluded. This flag is set by default. 1062 1063 tsd_max_sym_exp: optional int, determines number of symmetric expansions of the target symbols 1064 per timestep of the acoustic model. Larger values will allow longer sentences to be decoded, 1065 at increased cost to execution time. 1066 1067 alsd_max_target_len: optional int or float, determines the potential maximum target sequence length. 1068 If an integer is provided, it can decode sequences of that particular maximum length. 1069 If a float is provided, it can decode sequences of int(alsd_max_target_len * seq_len), 1070 where seq_len is the length of the acoustic model output (T). 1071 1072 NOTE: 1073 If a float is provided, it can be greater than 1! 1074 By default, a float of 2.0 is used so that a target sequence can be at most twice 1075 as long as the acoustic model output length T. 1076 1077 maes_num_steps: Number of adaptive steps to take. From the paper, 2 steps is generally sufficient, 1078 and can be reduced to 1 to improve decoding speed while sacrificing some accuracy. int > 0. 1079 1080 maes_prefix_alpha: Maximum prefix length in prefix search. Must be an integer, and is advised to keep this as 1 1081 in order to reduce expensive beam search cost later. int >= 0. 1082 1083 maes_expansion_beta: Maximum number of prefix expansions allowed, in addition to the beam size. 1084 Effectively, the number of hypothesis = beam_size + maes_expansion_beta. Must be an int >= 0, 1085 and affects the speed of inference since large values will perform large beam search in the next step. 1086 1087 maes_expansion_gamma: Float pruning threshold used in the prune-by-value step when computing the expansions. 1088 The default (2.3) is selected from the paper. It performs a comparison (max_log_prob - gamma <= log_prob[v]) 1089 where v is all vocabulary indices in the Vocab set and max_log_prob is the "most" likely token to be 1090 predicted. Gamma therefore provides a margin of additional tokens which can be potential candidates for 1091 expansion apart from the "most likely" candidate. 1092 Lower values will reduce the number of expansions (by increasing pruning-by-value, thereby improving speed 1093 but hurting accuracy). Higher values will increase the number of expansions (by reducing pruning-by-value, 1094 thereby reducing speed but potentially improving accuracy). This is a hyper parameter to be experimentally 1095 tuned on a validation set. 1096 1097 softmax_temperature: Scales the logits of the joint prior to computing log_softmax. 1098 1099 decoder: The Decoder/Prediction network module. 1100 joint: The Joint network module. 1101 vocabulary: The vocabulary (excluding the RNNT blank token) which will be used for decoding. 1102 """ 1103 1104 def __init__( 1105 self, decoding_cfg, decoder, joint, vocabulary, 1106 ): 1107 # we need to ensure blank is the last token in the vocab for the case of RNNT and Multi-blank RNNT. 1108 blank_id = len(vocabulary) + joint.num_extra_outputs 1109 1110 if hasattr(decoding_cfg, 'model_type') and decoding_cfg.model_type == 'tdt': 1111 blank_id = len(vocabulary) 1112 1113 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1114 1115 super(RNNTDecoding, self).__init__( 1116 decoding_cfg=decoding_cfg, decoder=decoder, joint=joint, blank_id=blank_id, 1117 ) 1118 1119 if isinstance(self.decoding, beam_decode.BeamRNNTInfer): 1120 self.decoding.set_decoding_type('char') 1121 1122 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1123 """ 1124 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1125 1126 Args: 1127 hypothesis: Hypothesis 1128 1129 Returns: 1130 A list of word-level confidence scores. 1131 """ 1132 return self._aggregate_token_confidence_chars(hypothesis.words, hypothesis.token_confidence) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c < self.blank_id - self.num_extra_outputs] 1159 return token_list 1160 1161 def decode_tokens_to_lang(self, tokens: List[int]) -> str: 1162 """ 1163 Compute the most likely language ID (LID) string given the tokens. 1164 1165 Args: 1166 tokens: List of int representing the token ids. 1167 1168 Returns: 1169 A decoded LID string. 1170 """ 1171 lang = self.tokenizer.ids_to_lang(tokens) 1172 return lang 1173 1174 def decode_ids_to_langs(self, tokens: List[int]) -> List[str]: 1175 """ 1176 Decode a token id list into language ID (LID) list. 1177 1178 Args: 1179 tokens: List of int representing the token ids. 1180 1181 Returns: 1182 A list of decoded LIDS. 1183 """ 1184 lang_list = self.tokenizer.ids_to_text_and_langs(tokens) 1185 return lang_list 1186 1187 1188 class RNNTWER(Metric): 1189 """ 1190 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference texts. 1191 When doing distributed training/evaluation the result of res=WER(predictions, targets, target_lengths) calls 1192 will be all-reduced between all workers using SUM operations. 1193 Here contains two numbers res=[wer_numerator, wer_denominator]. WER=wer_numerator/wer_denominator. 1194 1195 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step results. 1196 Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1197 1198 Example: 1199 def validation_step(self, batch, batch_idx): 1200 ... 1201 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1202 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1203 return self.val_outputs 1204 1205 def on_validation_epoch_end(self): 1206 ... 1207 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1208 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1209 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1210 self.val_outputs.clear() # free memory 1211 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1212 1213 Args: 1214 decoding: RNNTDecoding object that will perform autoregressive decoding of the RNNT model. 1215 batch_dim_index: Index of the batch dimension. 1216 use_cer: Whether to use Character Error Rate isntead of Word Error Rate. 1217 log_prediction: Whether to log a single decoded sample per call. 1218 1219 Returns: 1220 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenshtein's 1221 distances for all prediction - reference pairs, total number of words in all references. 1222 """ 1223 1224 full_state_update = True 1225 1226 def __init__( 1227 self, decoding: RNNTDecoding, batch_dim_index=0, use_cer=False, log_prediction=True, dist_sync_on_step=False 1228 ): 1229 super(RNNTWER, self).__init__(dist_sync_on_step=dist_sync_on_step) 1230 self.decoding = decoding 1231 self.batch_dim_index = batch_dim_index 1232 self.use_cer = use_cer 1233 self.log_prediction = log_prediction 1234 self.blank_id = self.decoding.blank_id 1235 self.labels_map = self.decoding.labels_map 1236 1237 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1238 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1239 1240 def update( 1241 self, 1242 encoder_output: torch.Tensor, 1243 encoded_lengths: torch.Tensor, 1244 targets: torch.Tensor, 1245 target_lengths: torch.Tensor, 1246 ) -> torch.Tensor: 1247 words = 0 1248 scores = 0 1249 references = [] 1250 with torch.no_grad(): 1251 # prediction_cpu_tensor = tensors[0].long().cpu() 1252 targets_cpu_tensor = targets.long().cpu() 1253 targets_cpu_tensor = move_dimension_to_the_front(targets_cpu_tensor, self.batch_dim_index) 1254 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1255 1256 # iterate over batch 1257 for ind in range(targets_cpu_tensor.shape[0]): 1258 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1259 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1260 1261 reference = self.decoding.decode_tokens_to_str(target) 1262 references.append(reference) 1263 1264 hypotheses, _ = self.decoding.rnnt_decoder_predictions_tensor(encoder_output, encoded_lengths) 1265 1266 if self.log_prediction: 1267 logging.info(f"\n") 1268 logging.info(f"reference :{references[0]}") 1269 logging.info(f"predicted :{hypotheses[0]}") 1270 1271 for h, r in zip(hypotheses, references): 1272 if self.use_cer: 1273 h_list = list(h) 1274 r_list = list(r) 1275 else: 1276 h_list = h.split() 1277 r_list = r.split() 1278 words += len(r_list) 1279 # Compute Levenshtein's distance 1280 scores += editdistance.eval(h_list, r_list) 1281 1282 self.scores += torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1283 self.words += torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1284 # return torch.tensor([scores, words]).to(predictions.device) 1285 1286 def compute(self): 1287 wer = self.scores.float() / self.words 1288 return wer, self.scores.detach(), self.words.detach() 1289 1290 1291 @dataclass 1292 class RNNTDecodingConfig: 1293 model_type: str = "rnnt" # one of "rnnt", "multiblank" or "tdt" 1294 strategy: str = "greedy_batch" 1295 1296 compute_hypothesis_token_set: bool = False 1297 1298 # preserve decoding alignments 1299 preserve_alignments: Optional[bool] = None 1300 1301 # confidence config 1302 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1303 1304 # RNNT Joint fused batch size 1305 fused_batch_size: Optional[int] = None 1306 1307 # compute RNNT time stamps 1308 compute_timestamps: Optional[bool] = None 1309 1310 # compute language IDs 1311 compute_langs: bool = False 1312 1313 # token representing word seperator 1314 word_seperator: str = " " 1315 1316 # type of timestamps to calculate 1317 rnnt_timestamp_type: str = "all" # can be char, word or all for both 1318 1319 # greedy decoding config 1320 greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() 1321 1322 # beam decoding config 1323 beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) 1324 1325 # can be used to change temperature for decoding 1326 temperature: float = 1.0 1327 [end of nemo/collections/asr/metrics/rnnt_wer.py] [start of nemo/collections/asr/metrics/wer.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import re 16 from abc import abstractmethod 17 from dataclasses import dataclass, is_dataclass 18 from typing import Callable, Dict, List, Optional, Tuple, Union 19 20 import editdistance 21 import jiwer 22 import numpy as np 23 import torch 24 from omegaconf import DictConfig, OmegaConf 25 from torchmetrics import Metric 26 27 from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding 28 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig, ConfidenceMixin 29 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis, NBestHypotheses 30 from nemo.utils import logging, logging_mode 31 32 __all__ = ['word_error_rate', 'word_error_rate_detail', 'WER', 'move_dimension_to_the_front'] 33 34 35 def word_error_rate(hypotheses: List[str], references: List[str], use_cer=False) -> float: 36 """ 37 Computes Average Word Error rate between two texts represented as 38 corresponding lists of string. 39 40 Hypotheses and references must have same length. 41 42 Args: 43 hypotheses (list): list of hypotheses 44 references(list) : list of references 45 use_cer (bool): set True to enable cer 46 47 Returns: 48 wer (float): average word error rate 49 """ 50 scores = 0 51 words = 0 52 if len(hypotheses) != len(references): 53 raise ValueError( 54 "In word error rate calculation, hypotheses and reference" 55 " lists must have the same number of elements. But I got:" 56 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 57 ) 58 for h, r in zip(hypotheses, references): 59 if use_cer: 60 h_list = list(h) 61 r_list = list(r) 62 else: 63 h_list = h.split() 64 r_list = r.split() 65 words += len(r_list) 66 # May deprecate using editdistance in future release for here and rest of codebase 67 # once we confirm jiwer is reliable. 68 scores += editdistance.eval(h_list, r_list) 69 if words != 0: 70 wer = 1.0 * scores / words 71 else: 72 wer = float('inf') 73 return wer 74 75 76 def word_error_rate_detail( 77 hypotheses: List[str], references: List[str], use_cer=False 78 ) -> Tuple[float, int, float, float, float]: 79 """ 80 Computes Average Word Error Rate with details (insertion rate, deletion rate, substitution rate) 81 between two texts represented as corresponding lists of string. 82 83 Hypotheses and references must have same length. 84 85 Args: 86 hypotheses (list): list of hypotheses 87 references(list) : list of references 88 use_cer (bool): set True to enable cer 89 90 Returns: 91 wer (float): average word error rate 92 words (int): Total number of words/charactors of given reference texts 93 ins_rate (float): average insertion error rate 94 del_rate (float): average deletion error rate 95 sub_rate (float): average substitution error rate 96 """ 97 scores = 0 98 words = 0 99 ops_count = {'substitutions': 0, 'insertions': 0, 'deletions': 0} 100 101 if len(hypotheses) != len(references): 102 raise ValueError( 103 "In word error rate calculation, hypotheses and reference" 104 " lists must have the same number of elements. But I got:" 105 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 106 ) 107 108 for h, r in zip(hypotheses, references): 109 if use_cer: 110 h_list = list(h) 111 r_list = list(r) 112 else: 113 h_list = h.split() 114 r_list = r.split() 115 116 # To get rid of the issue that jiwer does not allow empty string 117 if len(r_list) == 0: 118 if len(h_list) != 0: 119 errors = len(h_list) 120 ops_count['insertions'] += errors 121 else: 122 errors = 0 123 else: 124 if use_cer: 125 measures = jiwer.cer(r, h, return_dict=True) 126 else: 127 measures = jiwer.compute_measures(r, h) 128 129 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 130 ops_count['insertions'] += measures['insertions'] 131 ops_count['deletions'] += measures['deletions'] 132 ops_count['substitutions'] += measures['substitutions'] 133 134 scores += errors 135 words += len(r_list) 136 137 if words != 0: 138 wer = 1.0 * scores / words 139 ins_rate = 1.0 * ops_count['insertions'] / words 140 del_rate = 1.0 * ops_count['deletions'] / words 141 sub_rate = 1.0 * ops_count['substitutions'] / words 142 else: 143 wer, ins_rate, del_rate, sub_rate = float('inf'), float('inf'), float('inf'), float('inf') 144 145 return wer, words, ins_rate, del_rate, sub_rate 146 147 148 def word_error_rate_per_utt(hypotheses: List[str], references: List[str], use_cer=False) -> Tuple[List[float], float]: 149 """ 150 Computes Word Error Rate per utterance and the average WER 151 between two texts represented as corresponding lists of string. 152 153 Hypotheses and references must have same length. 154 155 Args: 156 hypotheses (list): list of hypotheses 157 references(list) : list of references 158 use_cer (bool): set True to enable cer 159 160 Returns: 161 wer_per_utt (List[float]): word error rate per utterance 162 avg_wer (float): average word error rate 163 """ 164 scores = 0 165 words = 0 166 wer_per_utt = [] 167 168 if len(hypotheses) != len(references): 169 raise ValueError( 170 "In word error rate calculation, hypotheses and reference" 171 " lists must have the same number of elements. But I got:" 172 "{0} and {1} correspondingly".format(len(hypotheses), len(references)) 173 ) 174 175 for h, r in zip(hypotheses, references): 176 if use_cer: 177 h_list = list(h) 178 r_list = list(r) 179 else: 180 h_list = h.split() 181 r_list = r.split() 182 183 # To get rid of the issue that jiwer does not allow empty string 184 if len(r_list) == 0: 185 if len(h_list) != 0: 186 errors = len(h_list) 187 wer_per_utt.append(float('inf')) 188 else: 189 if use_cer: 190 measures = jiwer.cer(r, h, return_dict=True) 191 er = measures['cer'] 192 else: 193 measures = jiwer.compute_measures(r, h) 194 er = measures['wer'] 195 196 errors = measures['insertions'] + measures['deletions'] + measures['substitutions'] 197 wer_per_utt.append(er) 198 199 scores += errors 200 words += len(r_list) 201 202 if words != 0: 203 avg_wer = 1.0 * scores / words 204 else: 205 avg_wer = float('inf') 206 207 return wer_per_utt, avg_wer 208 209 210 def move_dimension_to_the_front(tensor, dim_index): 211 all_dims = list(range(tensor.ndim)) 212 return tensor.permute(*([dim_index] + all_dims[:dim_index] + all_dims[dim_index + 1 :])) 213 214 215 class AbstractCTCDecoding(ConfidenceMixin): 216 """ 217 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs. 218 219 Args: 220 decoding_cfg: A dict-like object which contains the following key-value pairs. 221 strategy: str value which represents the type of decoding that can occur. 222 Possible values are : 223 - greedy (for greedy decoding). 224 - beam (for DeepSpeed KenLM based decoding). 225 226 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 227 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 228 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 229 230 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 231 Can take the following values - "char" for character/subword time stamps, "word" for word level 232 time stamps and "all" (default), for both character level and word level time stamps. 233 234 word_seperator: Str token representing the seperator between words. 235 236 preserve_alignments: Bool flag which preserves the history of logprobs generated during 237 decoding (sample / batched). When set to true, the Hypothesis will contain 238 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 239 240 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 241 scores. In order to obtain hypotheses with confidence scores, please utilize 242 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 243 244 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 245 generated during decoding. When set to true, the Hypothesis will contain 246 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 247 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 248 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 249 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 250 251 The length of the list corresponds to the number of recognized tokens. 252 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 253 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 254 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 255 256 The length of the list corresponds to the number of recognized words. 257 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 258 from the `token_confidence`. 259 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 260 Valid options are `mean`, `min`, `max`, `prod`. 261 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 262 confidence scores. 263 264 name: The method name (str). 265 Supported values: 266 - 'max_prob' for using the maximum token probability as a confidence. 267 - 'entropy' for using a normalized entropy of a log-likelihood vector. 268 269 entropy_type: Which type of entropy to use (str). 270 Used if confidence_method_cfg.name is set to `entropy`. 271 Supported values: 272 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 273 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 274 Note that for this entropy, the alpha should comply the following inequality: 275 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 276 where V is the model vocabulary size. 277 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 278 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 279 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 280 More: https://en.wikipedia.org/wiki/Tsallis_entropy 281 - 'renyi' for the Rรฉnyi entropy. 282 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 285 286 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 287 When the alpha equals one, scaling is not applied to 'max_prob', 288 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 289 290 entropy_norm: A mapping of the entropy value to the interval [0,1]. 291 Supported values: 292 - 'lin' for using the linear mapping. 293 - 'exp' for using exponential mapping with linear shift. 294 295 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 296 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 297 298 The config may further contain the following sub-dictionaries: 299 "greedy": 300 preserve_alignments: Same as above, overrides above value. 301 compute_timestamps: Same as above, overrides above value. 302 preserve_frame_confidence: Same as above, overrides above value. 303 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 304 305 "beam": 306 beam_size: int, defining the beam size for beam search. Must be >= 1. 307 If beam_size == 1, will perform cached greedy search. This might be slightly different 308 results compared to the greedy search above. 309 310 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 311 hypotheses after beam search has concluded. This flag is set by default. 312 313 beam_alpha: float, the strength of the Language model on the final score of a token. 314 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 315 316 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 317 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 318 319 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 320 If the path is invalid (file is not found at path), will raise a deferred error at the moment 321 of calculation of beam search, so that users may update / change the decoding strategy 322 to point to the correct file. 323 324 blank_id: The id of the RNNT blank token. 325 """ 326 327 def __init__(self, decoding_cfg, blank_id: int): 328 super().__init__() 329 330 # Convert dataclas to config 331 if is_dataclass(decoding_cfg): 332 decoding_cfg = OmegaConf.structured(decoding_cfg) 333 334 if not isinstance(decoding_cfg, DictConfig): 335 decoding_cfg = OmegaConf.create(decoding_cfg) 336 337 OmegaConf.set_struct(decoding_cfg, False) 338 339 # update minimal config 340 minimal_cfg = ['greedy'] 341 for item in minimal_cfg: 342 if item not in decoding_cfg: 343 decoding_cfg[item] = OmegaConf.create({}) 344 345 self.cfg = decoding_cfg 346 self.blank_id = blank_id 347 self.preserve_alignments = self.cfg.get('preserve_alignments', None) 348 self.compute_timestamps = self.cfg.get('compute_timestamps', None) 349 self.batch_dim_index = self.cfg.get('batch_dim_index', 0) 350 self.word_seperator = self.cfg.get('word_seperator', ' ') 351 352 possible_strategies = ['greedy', 'beam', 'pyctcdecode', 'flashlight'] 353 if self.cfg.strategy not in possible_strategies: 354 raise ValueError(f"Decoding strategy must be one of {possible_strategies}. Given {self.cfg.strategy}") 355 356 # Update preserve alignments 357 if self.preserve_alignments is None: 358 if self.cfg.strategy in ['greedy']: 359 self.preserve_alignments = self.cfg.greedy.get('preserve_alignments', False) 360 else: 361 self.preserve_alignments = self.cfg.beam.get('preserve_alignments', False) 362 363 # Update compute timestamps 364 if self.compute_timestamps is None: 365 if self.cfg.strategy in ['greedy']: 366 self.compute_timestamps = self.cfg.greedy.get('compute_timestamps', False) 367 elif self.cfg.strategy in ['beam']: 368 self.compute_timestamps = self.cfg.beam.get('compute_timestamps', False) 369 370 # initialize confidence-related fields 371 self._init_confidence(self.cfg.get('confidence_cfg', None)) 372 373 # Confidence estimation is not implemented for strategies other than `greedy` 374 if ( 375 not self.preserve_frame_confidence 376 and self.cfg.strategy != 'greedy' 377 and self.cfg.beam.get('preserve_frame_confidence', False) 378 ): 379 raise NotImplementedError(f"Confidence calculation is not supported for strategy `{self.cfg.strategy}`") 380 381 # we need timestamps to extract non-blank per-frame confidence 382 if self.compute_timestamps is not None: 383 self.compute_timestamps |= self.preserve_frame_confidence 384 385 if self.cfg.strategy == 'greedy': 386 387 self.decoding = ctc_greedy_decoding.GreedyCTCInfer( 388 blank_id=self.blank_id, 389 preserve_alignments=self.preserve_alignments, 390 compute_timestamps=self.compute_timestamps, 391 preserve_frame_confidence=self.preserve_frame_confidence, 392 confidence_method_cfg=self.confidence_method_cfg, 393 ) 394 395 elif self.cfg.strategy == 'beam': 396 397 self.decoding = ctc_beam_decoding.BeamCTCInfer( 398 blank_id=blank_id, 399 beam_size=self.cfg.beam.get('beam_size', 1), 400 search_type='default', 401 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 402 preserve_alignments=self.preserve_alignments, 403 compute_timestamps=self.compute_timestamps, 404 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 405 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 406 kenlm_path=self.cfg.beam.get('kenlm_path', None), 407 ) 408 409 self.decoding.override_fold_consecutive_value = False 410 411 elif self.cfg.strategy == 'pyctcdecode': 412 413 self.decoding = ctc_beam_decoding.BeamCTCInfer( 414 blank_id=blank_id, 415 beam_size=self.cfg.beam.get('beam_size', 1), 416 search_type='pyctcdecode', 417 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 418 preserve_alignments=self.preserve_alignments, 419 compute_timestamps=self.compute_timestamps, 420 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 421 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 422 kenlm_path=self.cfg.beam.get('kenlm_path', None), 423 pyctcdecode_cfg=self.cfg.beam.get('pyctcdecode_cfg', None), 424 ) 425 426 self.decoding.override_fold_consecutive_value = False 427 428 elif self.cfg.strategy == 'flashlight': 429 430 self.decoding = ctc_beam_decoding.BeamCTCInfer( 431 blank_id=blank_id, 432 beam_size=self.cfg.beam.get('beam_size', 1), 433 search_type='flashlight', 434 return_best_hypothesis=self.cfg.beam.get('return_best_hypothesis', True), 435 preserve_alignments=self.preserve_alignments, 436 compute_timestamps=self.compute_timestamps, 437 beam_alpha=self.cfg.beam.get('beam_alpha', 1.0), 438 beam_beta=self.cfg.beam.get('beam_beta', 0.0), 439 kenlm_path=self.cfg.beam.get('kenlm_path', None), 440 flashlight_cfg=self.cfg.beam.get('flashlight_cfg', None), 441 ) 442 443 self.decoding.override_fold_consecutive_value = False 444 445 else: 446 raise ValueError( 447 f"Incorrect decoding strategy supplied. Must be one of {possible_strategies}\n" 448 f"but was provided {self.cfg.strategy}" 449 ) 450 451 def ctc_decoder_predictions_tensor( 452 self, 453 decoder_outputs: torch.Tensor, 454 decoder_lengths: torch.Tensor = None, 455 fold_consecutive: bool = True, 456 return_hypotheses: bool = False, 457 ) -> Tuple[List[str], Optional[List[List[str]]], Optional[Union[Hypothesis, NBestHypotheses]]]: 458 """ 459 Decodes a sequence of labels to words 460 461 Args: 462 decoder_outputs: An integer torch.Tensor of shape [Batch, Time, {Vocabulary}] (if ``batch_index_dim == 0``) or [Time, Batch] 463 (if ``batch_index_dim == 1``) of integer indices that correspond to the index of some character in the 464 label set. 465 decoder_lengths: Optional tensor of length `Batch` which contains the integer lengths 466 of the sequence in the padded `predictions` tensor. 467 fold_consecutive: Bool, determine whether to perform "ctc collapse", folding consecutive tokens 468 into a single token. 469 return_hypotheses: Bool flag whether to return just the decoding predictions of the model 470 or a Hypothesis object that holds information such as the decoded `text`, 471 the `alignment` of emited by the CTC Model, and the `length` of the sequence (if available). 472 May also contain the log-probabilities of the decoder (if this method is called via 473 transcribe()) 474 475 Returns: 476 Either a list of str which represent the CTC decoded strings per sample, 477 or a list of Hypothesis objects containing additional information. 478 """ 479 480 if isinstance(decoder_outputs, torch.Tensor): 481 decoder_outputs = move_dimension_to_the_front(decoder_outputs, self.batch_dim_index) 482 483 if ( 484 hasattr(self.decoding, 'override_fold_consecutive_value') 485 and self.decoding.override_fold_consecutive_value is not None 486 ): 487 logging.info( 488 f"Beam search requires that consecutive ctc tokens are not folded. \n" 489 f"Overriding provided value of `fold_consecutive` = {fold_consecutive} to " 490 f"{self.decoding.override_fold_consecutive_value}", 491 mode=logging_mode.ONCE, 492 ) 493 fold_consecutive = self.decoding.override_fold_consecutive_value 494 495 with torch.inference_mode(): 496 # Resolve the forward step of the decoding strategy 497 hypotheses_list = self.decoding( 498 decoder_output=decoder_outputs, decoder_lengths=decoder_lengths 499 ) # type: List[List[Hypothesis]] 500 501 # extract the hypotheses 502 hypotheses_list = hypotheses_list[0] # type: List[Hypothesis] 503 504 if isinstance(hypotheses_list[0], NBestHypotheses): 505 hypotheses = [] 506 all_hypotheses = [] 507 508 for nbest_hyp in hypotheses_list: # type: NBestHypotheses 509 n_hyps = nbest_hyp.n_best_hypotheses # Extract all hypotheses for this sample 510 decoded_hyps = self.decode_hypothesis( 511 n_hyps, fold_consecutive 512 ) # type: List[Union[Hypothesis, NBestHypotheses]] 513 514 # If computing timestamps 515 if self.compute_timestamps is True: 516 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 517 for hyp_idx in range(len(decoded_hyps)): 518 decoded_hyps[hyp_idx] = self.compute_ctc_timestamps(decoded_hyps[hyp_idx], timestamp_type) 519 520 hypotheses.append(decoded_hyps[0]) # best hypothesis 521 all_hypotheses.append(decoded_hyps) 522 523 if return_hypotheses: 524 return hypotheses, all_hypotheses 525 526 best_hyp_text = [h.text for h in hypotheses] 527 all_hyp_text = [h.text for hh in all_hypotheses for h in hh] 528 return best_hyp_text, all_hyp_text 529 530 else: 531 hypotheses = self.decode_hypothesis( 532 hypotheses_list, fold_consecutive 533 ) # type: List[Union[Hypothesis, NBestHypotheses]] 534 535 # If computing timestamps 536 if self.compute_timestamps is True: 537 # greedy decoding, can get high-level confidence scores 538 if return_hypotheses and (self.preserve_word_confidence or self.preserve_token_confidence): 539 hypotheses = self.compute_confidence(hypotheses) 540 else: 541 # remove unused token_repetitions from Hypothesis.text 542 for hyp in hypotheses: 543 hyp.text = hyp.text[:2] 544 timestamp_type = self.cfg.get('ctc_timestamp_type', 'all') 545 for hyp_idx in range(len(hypotheses)): 546 hypotheses[hyp_idx] = self.compute_ctc_timestamps(hypotheses[hyp_idx], timestamp_type) 547 548 if return_hypotheses: 549 return hypotheses, None 550 551 best_hyp_text = [h.text for h in hypotheses] 552 return best_hyp_text, None 553 554 def decode_hypothesis( 555 self, hypotheses_list: List[Hypothesis], fold_consecutive: bool 556 ) -> List[Union[Hypothesis, NBestHypotheses]]: 557 """ 558 Decode a list of hypotheses into a list of strings. 559 560 Args: 561 hypotheses_list: List of Hypothesis. 562 fold_consecutive: Whether to collapse the ctc blank tokens or not. 563 564 Returns: 565 A list of strings. 566 """ 567 for ind in range(len(hypotheses_list)): 568 # Extract the integer encoded hypothesis 569 hyp = hypotheses_list[ind] 570 prediction = hyp.y_sequence 571 predictions_len = hyp.length if hyp.length > 0 else None 572 573 if fold_consecutive: 574 if type(prediction) != list: 575 prediction = prediction.numpy().tolist() 576 577 if predictions_len is not None: 578 prediction = prediction[:predictions_len] 579 580 # CTC decoding procedure 581 decoded_prediction = [] 582 token_lengths = [] # preserve token lengths 583 token_repetitions = [] # preserve number of repetitions per token 584 585 previous = self.blank_id 586 last_length = 0 587 last_repetition = 1 588 589 for pidx, p in enumerate(prediction): 590 if (p != previous or previous == self.blank_id) and p != self.blank_id: 591 decoded_prediction.append(p) 592 593 token_lengths.append(pidx - last_length) 594 last_length = pidx 595 token_repetitions.append(last_repetition) 596 last_repetition = 1 597 598 if p == previous and previous != self.blank_id: 599 last_repetition += 1 600 601 previous = p 602 603 if len(token_repetitions) > 0: 604 token_repetitions = token_repetitions[1:] + [last_repetition] 605 606 else: 607 if predictions_len is not None: 608 prediction = prediction[:predictions_len] 609 decoded_prediction = prediction[prediction != self.blank_id].tolist() 610 token_lengths = [1] * len(decoded_prediction) # preserve number of repetitions per token 611 token_repetitions = [1] * len(decoded_prediction) # preserve number of repetitions per token 612 613 # De-tokenize the integer tokens; if not computing timestamps 614 if self.compute_timestamps is True: 615 # keep the original predictions, wrap with the number of repetitions per token 616 # this is done so that `ctc_decoder_predictions_tensor()` can process this hypothesis 617 # in order to compute exact time stamps. 618 hypothesis = (decoded_prediction, token_lengths, token_repetitions) 619 else: 620 hypothesis = self.decode_tokens_to_str(decoded_prediction) 621 622 # TODO: remove 623 # collapse leading spaces before . , ? for PC models 624 hypothesis = re.sub(r'(\s+)([\.\,\?])', r'\2', hypothesis) 625 626 # Preserve this wrapped hypothesis or decoded text tokens. 627 hypotheses_list[ind].text = hypothesis 628 629 return hypotheses_list 630 631 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 632 """ 633 Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 634 Assumes that `frame_confidence` is present in the hypotheses. 635 636 Args: 637 hypotheses_list: List of Hypothesis. 638 639 Returns: 640 A list of hypotheses with high-level confidence scores. 641 """ 642 for hyp in hypotheses_list: 643 if not isinstance(hyp.text, tuple) or len(hyp.text) != 3: 644 # the method must have been called in the wrong place 645 raise ValueError( 646 """Wrong format of the `text` attribute of a hypothesis.\n 647 Expected: (decoded_prediction, token_repetitions)\n 648 The method invocation is expected between .decode_hypothesis() and .compute_ctc_timestamps()""" 649 ) 650 token_repetitions = hyp.text[2] 651 hyp.text = hyp.text[:2] 652 token_confidence = [] 653 if self.exclude_blank_from_confidence: 654 non_blank_frame_confidence = hyp.non_blank_frame_confidence 655 i = 0 656 for tr in token_repetitions: 657 # token repetition can be zero 658 j = i + tr 659 token_confidence.append(self._aggregate_confidence(non_blank_frame_confidence[i:j])) 660 i = j 661 else: 662 # <blank> tokens are considered to belong to the last non-blank token, if any. 663 token_lengths = hyp.text[1] 664 if len(token_lengths) > 0: 665 ts = token_lengths[0] 666 for tl in token_lengths[1:] + [len(hyp.frame_confidence)]: 667 token_confidence.append(self._aggregate_confidence(hyp.frame_confidence[ts : ts + tl])) 668 ts += tl 669 hyp.token_confidence = token_confidence 670 if self.preserve_word_confidence: 671 for hyp in hypotheses_list: 672 hyp.word_confidence = self._aggregate_token_confidence(hyp) 673 return hypotheses_list 674 675 @abstractmethod 676 def decode_tokens_to_str(self, tokens: List[int]) -> str: 677 """ 678 Implemented by subclass in order to decoder a token id list into a string. 679 680 Args: 681 tokens: List of int representing the token ids. 682 683 Returns: 684 A decoded string. 685 """ 686 raise NotImplementedError() 687 688 @abstractmethod 689 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 690 """ 691 Implemented by subclass in order to decode a token id list into a token list. 692 A token list is the string representation of each token id. 693 694 Args: 695 tokens: List of int representing the token ids. 696 697 Returns: 698 A list of decoded tokens. 699 """ 700 raise NotImplementedError() 701 702 def compute_ctc_timestamps(self, hypothesis: Hypothesis, timestamp_type: str = "all"): 703 """ 704 Method to compute time stamps at char/subword, and word level given some hypothesis. 705 Requires the input hypothesis to contain a `text` field that is the tuple. The tuple contains - 706 the ctc collapsed integer ids, and the number of repetitions of each token. 707 708 Args: 709 hypothesis: A Hypothesis object, with a wrapped `text` field. 710 The `text` field must contain a tuple with two values - 711 The ctc collapsed integer ids 712 A list of integers that represents the number of repetitions per token. 713 timestamp_type: A str value that represents the type of time stamp calculated. 714 Can be one of "char", "word" or "all" 715 716 Returns: 717 A Hypothesis object with a modified `timestep` value, which is now a dictionary containing 718 the time stamp information. 719 """ 720 assert timestamp_type in ['char', 'word', 'all'] 721 722 # Unpack the temporary storage, and set the decoded predictions 723 decoded_prediction, token_lengths = hypothesis.text 724 hypothesis.text = decoded_prediction 725 726 # Retrieve offsets 727 char_offsets = word_offsets = None 728 char_offsets = self._compute_offsets(hypothesis, token_lengths, self.blank_id) 729 730 # Assert number of offsets and hypothesis tokens are 1:1 match. 731 if len(char_offsets) != len(hypothesis.text): 732 raise ValueError( 733 f"`char_offsets`: {char_offsets} and `processed_tokens`: {hypothesis.text}" 734 " have to be of the same length, but are: " 735 f"`len(offsets)`: {len(char_offsets)} and `len(processed_tokens)`:" 736 f" {len(hypothesis.text)}" 737 ) 738 739 # Correctly process the token ids to chars/subwords. 740 for i, char in enumerate(hypothesis.text): 741 char_offsets[i]["char"] = self.decode_tokens_to_str([char]) 742 743 # detect char vs subword models 744 lens = [len(list(v["char"])) > 1 for v in char_offsets] 745 if any(lens): 746 text_type = 'subword' 747 else: 748 text_type = 'char' 749 750 # retrieve word offsets from character offsets 751 word_offsets = None 752 if timestamp_type in ['word', 'all']: 753 if text_type == 'char': 754 word_offsets = self._get_word_offsets_chars(char_offsets, word_delimiter_char=self.word_seperator) 755 else: 756 word_offsets = self._get_word_offsets_subwords_sentencepiece( 757 char_offsets, 758 hypothesis, 759 decode_ids_to_tokens=self.decode_ids_to_tokens, 760 decode_tokens_to_str=self.decode_tokens_to_str, 761 ) 762 763 # attach results 764 if len(hypothesis.timestep) > 0: 765 timestep_info = hypothesis.timestep 766 else: 767 timestep_info = [] 768 769 # Setup defaults 770 hypothesis.timestep = {"timestep": timestep_info} 771 772 # Add char / subword time stamps 773 if char_offsets is not None and timestamp_type in ['char', 'all']: 774 hypothesis.timestep['char'] = char_offsets 775 776 # Add word time stamps 777 if word_offsets is not None and timestamp_type in ['word', 'all']: 778 hypothesis.timestep['word'] = word_offsets 779 780 # Convert the token indices to text 781 hypothesis.text = self.decode_tokens_to_str(hypothesis.text) 782 783 return hypothesis 784 785 @staticmethod 786 def _compute_offsets( 787 hypothesis: Hypothesis, token_lengths: List[int], ctc_token: int 788 ) -> List[Dict[str, Union[str, int]]]: 789 """ 790 Utility method that calculates the indidual time indices where a token starts and ends. 791 792 Args: 793 hypothesis: A Hypothesis object that contains `text` field that holds the character / subword token 794 emitted at every time step after ctc collapse. 795 token_lengths: A list of ints representing the lengths of each emitted token. 796 ctc_token: The integer of the ctc blank token used during ctc collapse. 797 798 Returns: 799 800 """ 801 start_index = 0 802 803 # If the exact timestep information is available, utilize the 1st non-ctc blank token timestep 804 # as the start index. 805 if hypothesis.timestep is not None and len(hypothesis.timestep) > 0: 806 start_index = max(0, hypothesis.timestep[0] - 1) 807 808 # Construct the start and end indices brackets 809 end_indices = np.asarray(token_lengths).cumsum() 810 start_indices = np.concatenate(([start_index], end_indices[:-1])) 811 812 # Merge the results per token into a list of dictionaries 813 offsets = [ 814 {"char": t, "start_offset": s, "end_offset": e} 815 for t, s, e in zip(hypothesis.text, start_indices, end_indices) 816 ] 817 818 # Filter out CTC token 819 offsets = list(filter(lambda offsets: offsets["char"] != ctc_token, offsets)) 820 return offsets 821 822 @staticmethod 823 def _get_word_offsets_chars( 824 offsets: Dict[str, Union[str, float]], word_delimiter_char: str = " " 825 ) -> Dict[str, Union[str, float]]: 826 """ 827 Utility method which constructs word time stamps out of character time stamps. 828 829 References: 830 This code is a port of the Hugging Face code for word time stamp construction. 831 832 Args: 833 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 834 word_delimiter_char: Character token that represents the word delimiter. By default, " ". 835 836 Returns: 837 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 838 "end_offset". 839 """ 840 word_offsets = [] 841 842 last_state = "SPACE" 843 word = "" 844 start_offset = 0 845 end_offset = 0 846 for i, offset in enumerate(offsets): 847 char = offset["char"] 848 state = "SPACE" if char == word_delimiter_char else "WORD" 849 850 if state == last_state: 851 # If we are in the same state as before, we simply repeat what we've done before 852 end_offset = offset["end_offset"] 853 word += char 854 else: 855 # Switching state 856 if state == "SPACE": 857 # Finishing a word 858 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 859 else: 860 # Starting a new word 861 start_offset = offset["start_offset"] 862 end_offset = offset["end_offset"] 863 word = char 864 865 last_state = state 866 if last_state == "WORD": 867 word_offsets.append({"word": word, "start_offset": start_offset, "end_offset": end_offset}) 868 869 return word_offsets 870 871 @staticmethod 872 def _get_word_offsets_subwords_sentencepiece( 873 offsets: Dict[str, Union[str, float]], 874 hypothesis: Hypothesis, 875 decode_ids_to_tokens: Callable[[List[int]], str], 876 decode_tokens_to_str: Callable[[List[int]], str], 877 ) -> Dict[str, Union[str, float]]: 878 """ 879 Utility method which constructs word time stamps out of sub-word time stamps. 880 881 **Note**: Only supports Sentencepiece based tokenizers ! 882 883 Args: 884 offsets: A list of dictionaries, each containing "char", "start_offset" and "end_offset". 885 hypothesis: Hypothesis object that contains `text` field, where each token is a sub-word id 886 after ctc collapse. 887 decode_ids_to_tokens: A Callable function that accepts a list of integers and maps it to a sub-word. 888 decode_tokens_to_str: A Callable function that accepts a list of integers and maps it to text / str. 889 890 Returns: 891 A list of dictionaries containing the word offsets. Each item contains "word", "start_offset" and 892 "end_offset". 893 """ 894 word_offsets = [] 895 built_token = [] 896 previous_token_index = 0 897 # For every collapsed sub-word token 898 for i, char in enumerate(hypothesis.text): 899 # Compute the sub-word text representation, and the decoded text (stripped of sub-word markers). 900 token = decode_ids_to_tokens([char])[0] 901 token_text = decode_tokens_to_str([char]) 902 903 # It is a sub-word token, or contains an identifier at the beginning such as _ or ## that was stripped 904 # after forcing partial text conversion of the token. 905 if token != token_text: 906 # If there are any partially or fully built sub-word token ids, construct to text. 907 # Note: This is "old" subword, that occurs *after* current sub-word has started. 908 if len(built_token) > 0: 909 word_offsets.append( 910 { 911 "word": decode_tokens_to_str(built_token), 912 "start_offset": offsets[previous_token_index]["start_offset"], 913 "end_offset": offsets[i]["start_offset"], 914 } 915 ) 916 917 # Prepare list of new sub-word ids 918 built_token.clear() 919 built_token.append(char) 920 previous_token_index = i 921 else: 922 # If the token does not contain any sub-word start mark, then the sub-word has not completed yet 923 # Append to current sub-word list. 924 built_token.append(char) 925 926 # Inject the start offset of the first token to word offsets 927 # This is because we always skip the delay the injection of the first sub-word due to the loop 928 # condition and check whether built token is ready or not. 929 # Therefore without this forced injection, the start_offset appears as off by 1. 930 if len(word_offsets) == 0: 931 # alaptev: sometimes word_offsets can be empty 932 if len(built_token) > 0: 933 word_offsets.append( 934 { 935 "word": decode_tokens_to_str(built_token), 936 "start_offset": offsets[0]["start_offset"], 937 "end_offset": offsets[-1]["end_offset"], 938 } 939 ) 940 built_token.clear() 941 else: 942 word_offsets[0]["start_offset"] = offsets[0]["start_offset"] 943 944 # If there are any remaining tokens left, inject them all into the final word offset. 945 # Note: The start offset of this token is the start time of the first token inside build_token. 946 # Note: The end offset of this token is the end time of the last token inside build_token 947 if len(built_token) > 0: 948 word_offsets.append( 949 { 950 "word": decode_tokens_to_str(built_token), 951 "start_offset": offsets[-(len(built_token))]["start_offset"], 952 "end_offset": offsets[-1]["end_offset"], 953 } 954 ) 955 built_token.clear() 956 957 return word_offsets 958 959 @property 960 def preserve_alignments(self): 961 return self._preserve_alignments 962 963 @preserve_alignments.setter 964 def preserve_alignments(self, value): 965 self._preserve_alignments = value 966 967 if hasattr(self, 'decoding'): 968 self.decoding.preserve_alignments = value 969 970 @property 971 def compute_timestamps(self): 972 return self._compute_timestamps 973 974 @compute_timestamps.setter 975 def compute_timestamps(self, value): 976 self._compute_timestamps = value 977 978 if hasattr(self, 'decoding'): 979 self.decoding.compute_timestamps = value 980 981 @property 982 def preserve_frame_confidence(self): 983 return self._preserve_frame_confidence 984 985 @preserve_frame_confidence.setter 986 def preserve_frame_confidence(self, value): 987 self._preserve_frame_confidence = value 988 989 if hasattr(self, 'decoding'): 990 self.decoding.preserve_frame_confidence = value 991 992 993 class CTCDecoding(AbstractCTCDecoding): 994 """ 995 Used for performing CTC auto-regressive / non-auto-regressive decoding of the logprobs for character 996 based models. 997 998 Args: 999 decoding_cfg: A dict-like object which contains the following key-value pairs. 1000 strategy: str value which represents the type of decoding that can occur. 1001 Possible values are : 1002 - greedy (for greedy decoding). 1003 - beam (for DeepSpeed KenLM based decoding). 1004 1005 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 1006 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 1007 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 1008 1009 ctc_timestamp_type: A str value, which represents the types of timestamps that should be calculated. 1010 Can take the following values - "char" for character/subword time stamps, "word" for word level 1011 time stamps and "all" (default), for both character level and word level time stamps. 1012 1013 word_seperator: Str token representing the seperator between words. 1014 1015 preserve_alignments: Bool flag which preserves the history of logprobs generated during 1016 decoding (sample / batched). When set to true, the Hypothesis will contain 1017 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 1018 1019 confidence_cfg: A dict-like object which contains the following key-value pairs related to confidence 1020 scores. In order to obtain hypotheses with confidence scores, please utilize 1021 `ctc_decoder_predictions_tensor` function with the `preserve_frame_confidence` flag set to True. 1022 1023 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 1024 generated during decoding. When set to true, the Hypothesis will contain 1025 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 1026 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 1027 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1028 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 1029 1030 The length of the list corresponds to the number of recognized tokens. 1031 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 1032 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1033 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 1034 1035 The length of the list corresponds to the number of recognized words. 1036 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 1037 from the `token_confidence`. 1038 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 1039 Valid options are `mean`, `min`, `max`, `prod`. 1040 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1041 confidence scores. 1042 1043 name: The method name (str). 1044 Supported values: 1045 - 'max_prob' for using the maximum token probability as a confidence. 1046 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1047 1048 entropy_type: Which type of entropy to use (str). 1049 Used if confidence_method_cfg.name is set to `entropy`. 1050 Supported values: 1051 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1052 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1053 Note that for this entropy, the alpha should comply the following inequality: 1054 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1055 where V is the model vocabulary size. 1056 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1057 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1058 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1059 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1060 - 'renyi' for the Rรฉnyi entropy. 1061 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1062 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1063 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1064 1065 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1066 When the alpha equals one, scaling is not applied to 'max_prob', 1067 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1068 1069 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1070 Supported values: 1071 - 'lin' for using the linear mapping. 1072 - 'exp' for using exponential mapping with linear shift. 1073 1074 batch_dim_index: Index of the batch dimension of ``targets`` and ``predictions`` parameters of 1075 ``ctc_decoder_predictions_tensor`` methods. Can be either 0 or 1. 1076 1077 The config may further contain the following sub-dictionaries: 1078 "greedy": 1079 preserve_alignments: Same as above, overrides above value. 1080 compute_timestamps: Same as above, overrides above value. 1081 preserve_frame_confidence: Same as above, overrides above value. 1082 confidence_method_cfg: Same as above, overrides confidence_cfg.method_cfg. 1083 1084 "beam": 1085 beam_size: int, defining the beam size for beam search. Must be >= 1. 1086 If beam_size == 1, will perform cached greedy search. This might be slightly different 1087 results compared to the greedy search above. 1088 1089 return_best_hypothesis: optional bool, whether to return just the best hypothesis or all of the 1090 hypotheses after beam search has concluded. This flag is set by default. 1091 1092 beam_alpha: float, the strength of the Language model on the final score of a token. 1093 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1094 1095 beam_beta: float, the strength of the sequence length penalty on the final score of a token. 1096 final_score = acoustic_score + beam_alpha * lm_score + beam_beta * seq_length. 1097 1098 kenlm_path: str, path to a KenLM ARPA or .binary file (depending on the strategy chosen). 1099 If the path is invalid (file is not found at path), will raise a deferred error at the moment 1100 of calculation of beam search, so that users may update / change the decoding strategy 1101 to point to the correct file. 1102 1103 blank_id: The id of the RNNT blank token. 1104 """ 1105 1106 def __init__( 1107 self, decoding_cfg, vocabulary, 1108 ): 1109 blank_id = len(vocabulary) 1110 self.vocabulary = vocabulary 1111 self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) 1112 1113 super().__init__(decoding_cfg=decoding_cfg, blank_id=blank_id) 1114 1115 # Finalize Beam Search Decoding framework 1116 if isinstance(self.decoding, ctc_beam_decoding.AbstractBeamCTCInfer): 1117 self.decoding.set_vocabulary(self.vocabulary) 1118 self.decoding.set_decoding_type('char') 1119 1120 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 1121 """ 1122 Implemented by subclass in order to aggregate token confidence to a word-level confidence. 1123 1124 Args: 1125 hypothesis: Hypothesis 1126 1127 Returns: 1128 A list of word-level confidence scores. 1129 """ 1130 return self._aggregate_token_confidence_chars( 1131 self.decode_tokens_to_str(hypothesis.text[0]).split(), hypothesis.token_confidence 1132 ) 1133 1134 def decode_tokens_to_str(self, tokens: List[int]) -> str: 1135 """ 1136 Implemented by subclass in order to decoder a token list into a string. 1137 1138 Args: 1139 tokens: List of int representing the token ids. 1140 1141 Returns: 1142 A decoded string. 1143 """ 1144 hypothesis = ''.join(self.decode_ids_to_tokens(tokens)) 1145 return hypothesis 1146 1147 def decode_ids_to_tokens(self, tokens: List[int]) -> List[str]: 1148 """ 1149 Implemented by subclass in order to decode a token id list into a token list. 1150 A token list is the string representation of each token id. 1151 1152 Args: 1153 tokens: List of int representing the token ids. 1154 1155 Returns: 1156 A list of decoded tokens. 1157 """ 1158 token_list = [self.labels_map[c] for c in tokens if c != self.blank_id] 1159 return token_list 1160 1161 1162 class WER(Metric): 1163 """ 1164 This metric computes numerator and denominator for Overall Word Error Rate (WER) between prediction and reference 1165 texts. When doing distributed training/evaluation the result of ``res=WER(predictions, targets, target_lengths)`` 1166 calls will be all-reduced between all workers using SUM operations. Here ``res`` contains three numbers 1167 ``res=[wer, total_levenstein_distance, total_number_of_words]``. 1168 1169 If used with PytorchLightning LightningModule, include wer_numerator and wer_denominators inside validation_step 1170 results. Then aggregate (sum) then at the end of validation epoch to correctly compute validation WER. 1171 1172 Example: 1173 def validation_step(self, batch, batch_idx): 1174 ... 1175 wer_num, wer_denom = self.__wer(predictions, transcript, transcript_len) 1176 self.val_outputs = {'val_loss': loss_value, 'val_wer_num': wer_num, 'val_wer_denom': wer_denom} 1177 return self.val_outputs 1178 1179 def on_validation_epoch_end(self): 1180 ... 1181 wer_num = torch.stack([x['val_wer_num'] for x in self.val_outputs]).sum() 1182 wer_denom = torch.stack([x['val_wer_denom'] for x in self.val_outputs]).sum() 1183 tensorboard_logs = {'validation_loss': val_loss_mean, 'validation_avg_wer': wer_num / wer_denom} 1184 self.val_outputs.clear() # free memory 1185 return {'val_loss': val_loss_mean, 'log': tensorboard_logs} 1186 1187 Args: 1188 decoding: An instance of CTCDecoding. 1189 use_cer: Whether to use Character Error Rate instead of Word Error Rate. 1190 log_prediction: Whether to log a single decoded sample per call. 1191 fold_consecutive: Whether repeated consecutive characters should be folded into one when decoding. 1192 1193 Returns: 1194 res: a tuple of 3 zero dimensional float32 ``torch.Tensor` objects: a WER score, a sum of Levenstein's 1195 distances for all prediction - reference pairs, total number of words in all references. 1196 """ 1197 1198 full_state_update: bool = True 1199 1200 def __init__( 1201 self, 1202 decoding: CTCDecoding, 1203 use_cer=False, 1204 log_prediction=True, 1205 fold_consecutive=True, 1206 dist_sync_on_step=False, 1207 ): 1208 super().__init__(dist_sync_on_step=dist_sync_on_step) 1209 1210 self.decoding = decoding 1211 self.use_cer = use_cer 1212 self.log_prediction = log_prediction 1213 self.fold_consecutive = fold_consecutive 1214 1215 self.add_state("scores", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1216 self.add_state("words", default=torch.tensor(0), dist_reduce_fx='sum', persistent=False) 1217 1218 def update( 1219 self, 1220 predictions: torch.Tensor, 1221 targets: torch.Tensor, 1222 target_lengths: torch.Tensor, 1223 predictions_lengths: torch.Tensor = None, 1224 ): 1225 """ 1226 Updates metric state. 1227 Args: 1228 predictions: an integer torch.Tensor of shape ``[Batch, Time, {Vocabulary}]`` (if ``batch_dim_index == 0``) or 1229 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1230 targets: an integer torch.Tensor of shape ``[Batch, Time]`` (if ``batch_dim_index == 0``) or 1231 ``[Time, Batch]`` (if ``batch_dim_index == 1``) 1232 target_lengths: an integer torch.Tensor of shape ``[Batch]`` 1233 predictions_lengths: an integer torch.Tensor of shape ``[Batch]`` 1234 """ 1235 words = 0 1236 scores = 0 1237 references = [] 1238 with torch.no_grad(): 1239 # prediction_cpu_tensor = tensors[0].long().cpu() 1240 targets_cpu_tensor = targets.long().cpu() 1241 tgt_lenths_cpu_tensor = target_lengths.long().cpu() 1242 1243 # iterate over batch 1244 for ind in range(targets_cpu_tensor.shape[0]): 1245 tgt_len = tgt_lenths_cpu_tensor[ind].item() 1246 target = targets_cpu_tensor[ind][:tgt_len].numpy().tolist() 1247 reference = self.decoding.decode_tokens_to_str(target) 1248 references.append(reference) 1249 1250 hypotheses, _ = self.decoding.ctc_decoder_predictions_tensor( 1251 predictions, predictions_lengths, fold_consecutive=self.fold_consecutive 1252 ) 1253 1254 if self.log_prediction: 1255 logging.info(f"\n") 1256 logging.info(f"reference:{references[0]}") 1257 logging.info(f"predicted:{hypotheses[0]}") 1258 1259 for h, r in zip(hypotheses, references): 1260 if self.use_cer: 1261 h_list = list(h) 1262 r_list = list(r) 1263 else: 1264 h_list = h.split() 1265 r_list = r.split() 1266 words += len(r_list) 1267 # Compute Levenstein's distance 1268 scores += editdistance.eval(h_list, r_list) 1269 1270 self.scores = torch.tensor(scores, device=self.scores.device, dtype=self.scores.dtype) 1271 self.words = torch.tensor(words, device=self.words.device, dtype=self.words.dtype) 1272 # return torch.tensor([scores, words]).to(predictions.device) 1273 1274 def compute(self): 1275 scores = self.scores.detach().float() 1276 words = self.words.detach().float() 1277 return scores / words, scores, words 1278 1279 1280 @dataclass 1281 class CTCDecodingConfig: 1282 strategy: str = "greedy" 1283 1284 # preserve decoding alignments 1285 preserve_alignments: Optional[bool] = None 1286 1287 # compute ctc time stamps 1288 compute_timestamps: Optional[bool] = None 1289 1290 # token representing word seperator 1291 word_seperator: str = " " 1292 1293 # type of timestamps to calculate 1294 ctc_timestamp_type: str = "all" # can be char, word or all for both 1295 1296 # batch dimension 1297 batch_dim_index: int = 0 1298 1299 # greedy decoding config 1300 greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() 1301 1302 # beam decoding config 1303 beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) 1304 1305 # confidence config 1306 confidence_cfg: ConfidenceConfig = ConfidenceConfig() 1307 1308 # can be used to change temperature for decoding 1309 temperature: float = 1.0 1310 [end of nemo/collections/asr/metrics/wer.py] [start of nemo/collections/asr/models/configs/aligner_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig 18 19 20 @dataclass 21 class AlignerCTCConfig: 22 prob_suppress_index: int = -1 23 prob_suppress_value: float = 1.0 24 25 26 @dataclass 27 class AlignerRNNTConfig: 28 predictor_window_size: int = 0 29 predictor_step_size: int = 1 30 31 32 @dataclass 33 class AlignerWrapperModelConfig: 34 alignment_type: str = "forced" 35 word_output: bool = True 36 cpu_decoding: bool = False 37 decode_batch_size: int = 0 38 ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() 39 rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() 40 41 42 @dataclass 43 class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): 44 decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() 45 [end of nemo/collections/asr/models/configs/aligner_config.py] [start of nemo/collections/asr/models/configs/asr_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 22 from nemo.collections.asr.modules.audio_preprocessing import ( 23 AudioToMelSpectrogramPreprocessorConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class ASRDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[Any] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[Any] = None 40 tarred_shard_strategy: str = "scatter" 41 shard_manifests: bool = False 42 shuffle_n: int = 0 43 44 # Optional 45 int_values: Optional[int] = None 46 augmentor: Optional[Dict[str, Any]] = None 47 max_duration: Optional[float] = None 48 min_duration: Optional[float] = None 49 max_utts: int = 0 50 blank_index: int = -1 51 unk_index: int = -1 52 normalize: bool = False 53 trim: bool = True 54 parser: Optional[str] = 'en' 55 eos_id: Optional[int] = None 56 bos_id: Optional[int] = None 57 pad_id: int = 0 58 use_start_end_token: bool = False 59 return_sample_id: Optional[bool] = False 60 61 # bucketing params 62 bucketing_strategy: str = "synced_randomized" 63 bucketing_batch_size: Optional[Any] = None 64 bucketing_weights: Optional[List[int]] = None 65 66 67 @dataclass 68 class EncDecCTCConfig(model_cfg.ModelConfig): 69 # Model global arguments 70 sample_rate: int = 16000 71 repeat: int = 1 72 dropout: float = 0.0 73 separable: bool = False 74 labels: List[str] = MISSING 75 76 # Dataset configs 77 train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) 78 validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 79 test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) 80 81 # Optimizer / Scheduler config 82 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 83 84 # Model component configs 85 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 86 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 87 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 88 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 89 decoding: CTCDecodingConfig = CTCDecodingConfig() 90 91 92 @dataclass 93 class EncDecCTCModelConfig(model_cfg.NemoConfig): 94 model: EncDecCTCConfig = EncDecCTCConfig() 95 96 97 @dataclass 98 class CacheAwareStreamingConfig: 99 chunk_size: int = 0 # the size of each chunk at each step, it can be a list of two integers to specify different chunk sizes for the first step and others 100 shift_size: int = 0 # the size of the shift in each step, it can be a list of two integers to specify different shift sizes for the first step and others 101 102 cache_drop_size: int = 0 # the number of steps to drop from the cache 103 last_channel_cache_size: int = 0 # the size of the needed cache for last channel layers 104 105 valid_out_len: int = 0 # the number of the steps in the final output which are valid (have the same value as in the offline mode) 106 107 pre_encode_cache_size: int = 0 # the size of the needed cache for the pre-encoding part of the model to avoid caching inside the pre-encoding layers 108 drop_extra_pre_encoded: int = 0 # the number of steps to get dropped after the pre-encoding layer 109 110 last_channel_num: int = 0 # number of the last channel layers (like MHA layers) which need caching in the model 111 last_time_num: int = 0 # number of the last time layers (like convolutions) which need caching in the model 112 [end of nemo/collections/asr/models/configs/asr_models_config.py] [start of nemo/collections/asr/models/configs/classification_models_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, List, Optional 17 18 from omegaconf import MISSING 19 20 import nemo.core.classes.dataset 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderClassificationConfig, ConvASREncoderConfig 27 from nemo.core.config import modelPT as model_cfg 28 29 30 @dataclass 31 class EncDecClassificationDatasetConfig(nemo.core.classes.dataset.DatasetConfig): 32 manifest_filepath: Optional[str] = None 33 sample_rate: int = MISSING 34 labels: List[str] = MISSING 35 trim_silence: bool = False 36 37 # Tarred dataset support 38 is_tarred: bool = False 39 tarred_audio_filepaths: Optional[str] = None 40 tarred_shard_strategy: str = "scatter" 41 shuffle_n: int = 0 42 43 # Optional 44 int_values: Optional[int] = None 45 augmentor: Optional[Dict[str, Any]] = None 46 max_duration: Optional[float] = None 47 min_duration: Optional[float] = None 48 cal_labels_occurrence: Optional[bool] = False 49 50 # VAD Optional 51 vad_stream: Optional[bool] = None 52 window_length_in_sec: float = 0.31 53 shift_length_in_sec: float = 0.01 54 normalize_audio: bool = False 55 is_regression_task: bool = False 56 57 # bucketing params 58 bucketing_strategy: str = "synced_randomized" 59 bucketing_batch_size: Optional[Any] = None 60 bucketing_weights: Optional[List[int]] = None 61 62 63 @dataclass 64 class EncDecClassificationConfig(model_cfg.ModelConfig): 65 # Model global arguments 66 sample_rate: int = 16000 67 repeat: int = 1 68 dropout: float = 0.0 69 separable: bool = True 70 kernel_size_factor: float = 1.0 71 labels: List[str] = MISSING 72 timesteps: int = MISSING 73 74 # Dataset configs 75 train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 76 manifest_filepath=None, shuffle=True, trim_silence=False 77 ) 78 validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 79 manifest_filepath=None, shuffle=False 80 ) 81 test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( 82 manifest_filepath=None, shuffle=False 83 ) 84 85 # Optimizer / Scheduler config 86 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 87 88 # Model component configs 89 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() 90 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 91 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 92 audio_length=timesteps 93 ) 94 95 encoder: ConvASREncoderConfig = ConvASREncoderConfig() 96 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 97 98 99 @dataclass 100 class EncDecClassificationModelConfig(model_cfg.NemoConfig): 101 model: EncDecClassificationConfig = EncDecClassificationConfig() 102 [end of nemo/collections/asr/models/configs/classification_models_config.py] [start of nemo/collections/asr/models/configs/diarizer_config.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import asdict, dataclass 16 from typing import Any, Dict, Optional, Tuple, Union 17 18 19 @dataclass 20 class DiarizerComponentConfig: 21 """Dataclass to imitate HydraConfig dict when accessing parameters.""" 22 23 def get(self, name: str, default: Optional[Any] = None): 24 return getattr(self, name, default) 25 26 def __iter__(self): 27 for key in asdict(self): 28 yield key 29 30 def dict(self) -> Dict: 31 return asdict(self) 32 33 34 @dataclass 35 class ASRDiarizerCTCDecoderParams: 36 pretrained_language_model: Optional[str] = None # KenLM model file: .arpa model file or .bin binary file. 37 beam_width: int = 32 38 alpha: float = 0.5 39 beta: float = 2.5 40 41 42 @dataclass 43 class ASRRealigningLMParams: 44 # Provide a KenLM language model in .arpa format. 45 arpa_language_model: Optional[str] = None 46 # Min number of words for the left context. 47 min_number_of_words: int = 3 48 # Max number of words for the right context. 49 max_number_of_words: int = 10 50 # The threshold for the difference between two log probability values from two hypotheses. 51 logprob_diff_threshold: float = 1.2 52 53 54 @dataclass 55 class ASRDiarizerParams(DiarizerComponentConfig): 56 # if True, speech segmentation for diarization is based on word-timestamps from ASR inference. 57 asr_based_vad: bool = False 58 # Threshold (in sec) that caps the gap between two words when generating VAD timestamps using ASR based VAD. 59 asr_based_vad_threshold: float = 1.0 60 # Batch size can be dependent on each ASR model. Default batch sizes are applied if set to null. 61 asr_batch_size: Optional[int] = None 62 # Native decoder delay. null is recommended to use the default values for each ASR model. 63 decoder_delay_in_sec: Optional[float] = None 64 # Offset to set a reference point from the start of the word. Recommended range of values is [-0.05 0.2]. 65 word_ts_anchor_offset: Optional[float] = None 66 # Select which part of the word timestamp we want to use. The options are: 'start', 'end', 'mid'. 67 word_ts_anchor_pos: str = "start" 68 # Fix the word timestamp using VAD output. You must provide a VAD model to use this feature. 69 fix_word_ts_with_VAD: bool = False 70 # If True, use colored text to distinguish speakers in the output transcript. 71 colored_text: bool = False 72 # If True, the start and end time of each speaker turn is printed in the output transcript. 73 print_time: bool = True 74 # If True, the output transcript breaks the line to fix the line width (default is 90 chars) 75 break_lines: bool = False 76 77 78 @dataclass 79 class ASRDiarizerConfig(DiarizerComponentConfig): 80 model_path: Optional[str] = "stt_en_conformer_ctc_large" 81 parameters: ASRDiarizerParams = ASRDiarizerParams() 82 ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() 83 realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() 84 85 86 @dataclass 87 class VADParams(DiarizerComponentConfig): 88 window_length_in_sec: float = 0.15 # Window length in sec for VAD context input 89 shift_length_in_sec: float = 0.01 # Shift length in sec for generate frame level VAD prediction 90 smoothing: Union[str, bool] = "median" # False or type of smoothing method (eg: median) 91 overlap: float = 0.5 # Overlap ratio for overlapped mean/median smoothing filter 92 onset: float = 0.1 # Onset threshold for detecting the beginning and end of a speech 93 offset: float = 0.1 # Offset threshold for detecting the end of a speech 94 pad_onset: float = 0.1 # Adding durations before each speech segment 95 pad_offset: float = 0 # Adding durations after each speech segment 96 min_duration_on: float = 0 # Threshold for small non_speech deletion 97 min_duration_off: float = 0.2 # Threshold for short speech segment deletion 98 filter_speech_first: bool = True 99 100 101 @dataclass 102 class VADConfig(DiarizerComponentConfig): 103 model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name 104 external_vad_manifest: Optional[str] = None 105 parameters: VADParams = VADParams() 106 107 108 @dataclass 109 class SpeakerEmbeddingsParams(DiarizerComponentConfig): 110 # Window length(s) in sec (floating-point number). either a number or a list. ex) 1.5 or [1.5,1.0,0.5] 111 window_length_in_sec: Tuple[float] = (1.5, 1.25, 1.0, 0.75, 0.5) 112 # Shift length(s) in sec (floating-point number). either a number or a list. ex) 0.75 or [0.75,0.5,0.25] 113 shift_length_in_sec: Tuple[float] = (0.75, 0.625, 0.5, 0.375, 0.25) 114 # Weight for each scale. None (for single scale) or list with window/shift scale count. ex) [0.33,0.33,0.33] 115 multiscale_weights: Tuple[float] = (1, 1, 1, 1, 1) 116 # save speaker embeddings in pickle format. True if clustering result is used for other models, such as MSDD. 117 save_embeddings: bool = True 118 119 120 @dataclass 121 class SpeakerEmbeddingsConfig(DiarizerComponentConfig): 122 # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) 123 model_path: Optional[str] = None 124 parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() 125 126 127 @dataclass 128 class ClusteringParams(DiarizerComponentConfig): 129 # If True, use num of speakers value provided in manifest file. 130 oracle_num_speakers: bool = False 131 # Max number of speakers for each recording. If an oracle number of speakers is passed, this value is ignored. 132 max_num_speakers: int = 8 133 # If the number of segments is lower than this number, enhanced speaker counting is activated. 134 enhanced_count_thres: int = 80 135 # Determines the range of p-value search: 0 < p <= max_rp_threshold. 136 max_rp_threshold: float = 0.25 137 # The higher the number, the more values will be examined with more time. 138 sparse_search_volume: int = 30 139 # If True, take a majority vote on multiple p-values to estimate the number of speakers. 140 maj_vote_spk_count: bool = False 141 142 143 @dataclass 144 class ClusteringConfig(DiarizerComponentConfig): 145 parameters: ClusteringParams = ClusteringParams() 146 147 148 @dataclass 149 class MSDDParams(DiarizerComponentConfig): 150 # If True, use speaker embedding model in checkpoint, else provided speaker embedding model in config will be used. 151 use_speaker_model_from_ckpt: bool = True 152 # Batch size for MSDD inference. 153 infer_batch_size: int = 25 154 # Sigmoid threshold for generating binarized speaker labels. The smaller the more generous on detecting overlaps. 155 sigmoid_threshold: Tuple[float] = (0.7,) 156 # If True, use oracle number of speaker and evaluate F1 score for the given speaker sequences. Default is False. 157 seq_eval_mode: bool = False 158 # If True, break the input audio clip to short sequences and calculate cluster average embeddings for inference. 159 split_infer: bool = True 160 # The length of split short sequence when split_infer is True. 161 diar_window_length: int = 50 162 # If the estimated number of speakers are larger than this number, overlap speech is not estimated. 163 overlap_infer_spk_limit: int = 5 164 165 166 @dataclass 167 class MSDDConfig(DiarizerComponentConfig): 168 model_path: Optional[str] = "diar_msdd_telephonic" 169 parameters: MSDDParams = MSDDParams() 170 171 172 @dataclass 173 class DiarizerConfig(DiarizerComponentConfig): 174 manifest_filepath: Optional[str] = None 175 out_dir: Optional[str] = None 176 oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps 177 collar: float = 0.25 # Collar value for scoring 178 ignore_overlap: bool = True # Consider or ignore overlap segments while scoring 179 vad: VADConfig = VADConfig() 180 speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() 181 clustering: ClusteringConfig = ClusteringConfig() 182 msdd_model: MSDDConfig = MSDDConfig() 183 asr: ASRDiarizerConfig = ASRDiarizerConfig() 184 185 186 @dataclass 187 class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): 188 diarizer: DiarizerConfig = DiarizerConfig() 189 device: str = "cpu" 190 verbose: bool = False 191 batch_size: int = 64 192 num_workers: int = 1 193 sample_rate: int = 16000 194 name: str = "" 195 196 @classmethod 197 def init_config(cls, diar_model_path: str, vad_model_path: str, map_location: str, verbose: bool): 198 return NeuralDiarizerInferenceConfig( 199 DiarizerConfig( 200 vad=VADConfig(model_path=vad_model_path), msdd_model=MSDDConfig(model_path=diar_model_path), 201 ), 202 device=map_location, 203 verbose=verbose, 204 ) 205 [end of nemo/collections/asr/models/configs/diarizer_config.py] [start of nemo/collections/asr/models/configs/k2_sequence_models_config.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 17 from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig 18 from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig 19 from nemo.core.config.modelPT import NemoConfig 20 21 22 @dataclass 23 class GraphModuleConfig: 24 criterion_type: str = "ml" 25 loss_type: str = "ctc" 26 split_batch_size: int = 0 27 dec_type: str = "topo" 28 transcribe_training: bool = True 29 backend_cfg: BackendConfig = BackendConfig() 30 31 32 @dataclass 33 class EncDecK2SeqConfig(EncDecCTCConfig): 34 graph_module_cfg: GraphModuleConfig = GraphModuleConfig() 35 36 37 @dataclass 38 class EncDecK2SeqModelConfig(NemoConfig): 39 model: EncDecK2SeqConfig = EncDecK2SeqConfig() 40 [end of nemo/collections/asr/models/configs/k2_sequence_models_config.py] [start of nemo/collections/asr/models/configs/matchboxnet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import classification_models_config as clf_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMFCCPreprocessorConfig, 23 CropOrPadSpectrogramAugmentationConfig, 24 SpectrogramAugmentationConfig, 25 ) 26 from nemo.collections.asr.modules.conv_asr import ( 27 ConvASRDecoderClassificationConfig, 28 ConvASREncoderConfig, 29 JasperEncoderConfig, 30 ) 31 from nemo.core.config import modelPT as model_cfg 32 33 34 # fmt: off 35 def matchboxnet_3x1x64(): 36 config = [ 37 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 38 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 39 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 40 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 41 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 42 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 43 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 44 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 45 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 46 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 47 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 48 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 49 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 50 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 51 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 52 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 53 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 54 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 55 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 56 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 57 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 58 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 59 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 60 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 61 ] 62 return config 63 64 65 def matchboxnet_3x1x64_vad(): 66 config = [ 67 JasperEncoderConfig(filters=128, repeat=1, kernel=[11], stride=[1], dilation=[1], dropout=0.0, 68 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 69 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 70 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 71 JasperEncoderConfig(filters=64, repeat=1, kernel=[13], stride=[1], dilation=[1], dropout=0.0, 72 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 73 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 74 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 75 JasperEncoderConfig(filters=64, repeat=1, kernel=[15], stride=[1], dilation=[1], dropout=0.0, 76 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 77 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 78 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 79 JasperEncoderConfig(filters=64, repeat=1, kernel=[17], stride=[1], dilation=[1], dropout=0.0, 80 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 81 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 82 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 83 JasperEncoderConfig(filters=128, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.0, 84 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 85 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 86 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 87 JasperEncoderConfig(filters=128, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 88 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 89 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 90 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 91 ] 92 return config 93 94 95 # fmt: on 96 97 98 @dataclass 99 class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): 100 # Model global arguments 101 sample_rate: int = 16000 102 repeat: int = 1 103 dropout: float = 0.0 104 separable: bool = True 105 kernel_size_factor: float = 1.0 106 timesteps: int = 128 107 labels: List[str] = MISSING 108 109 # Dataset configs 110 train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 111 manifest_filepath=None, shuffle=True, trim_silence=False 112 ) 113 validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 114 manifest_filepath=None, shuffle=False 115 ) 116 test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( 117 manifest_filepath=None, shuffle=False 118 ) 119 120 # Optimizer / Scheduler config 121 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 122 123 # Model general component configs 124 preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) 125 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( 126 freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 127 ) 128 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( 129 audio_length=128 130 ) 131 132 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 133 decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() 134 135 136 @dataclass 137 class MatchboxNetVADModelConfig(MatchboxNetModelConfig): 138 timesteps: int = 64 139 labels: List[str] = field(default_factory=lambda: ['background', 'speech']) 140 141 crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = None 142 143 144 class EncDecClassificationModelConfigBuilder(model_cfg.ModelConfigBuilder): 145 VALID_CONFIGS = ['matchboxnet_3x1x64', 'matchboxnet_3x1x64_vad'] 146 147 def __init__(self, name: str = 'matchboxnet_3x1x64', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 148 if name not in EncDecClassificationModelConfigBuilder.VALID_CONFIGS: 149 raise ValueError("`name` must be one of : \n" f"{EncDecClassificationModelConfigBuilder.VALID_CONFIGS}") 150 151 self.name = name 152 153 if 'matchboxnet_3x1x64_vad' in name: 154 if encoder_cfg_func is None: 155 encoder_cfg_func = matchboxnet_3x1x64_vad 156 157 model_cfg = MatchboxNetVADModelConfig( 158 repeat=1, 159 separable=True, 160 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 161 decoder=ConvASRDecoderClassificationConfig(), 162 ) 163 164 elif 'matchboxnet_3x1x64' in name: 165 if encoder_cfg_func is None: 166 encoder_cfg_func = matchboxnet_3x1x64 167 168 model_cfg = MatchboxNetModelConfig( 169 repeat=1, 170 separable=False, 171 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 172 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 173 decoder=ConvASRDecoderClassificationConfig(), 174 ) 175 176 else: 177 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 178 179 super(EncDecClassificationModelConfigBuilder, self).__init__(model_cfg) 180 self.model_cfg: clf_cfg.EncDecClassificationConfig = model_cfg # enable type hinting 181 182 def set_labels(self, labels: List[str]): 183 self.model_cfg.labels = labels 184 185 def set_separable(self, separable: bool): 186 self.model_cfg.separable = separable 187 188 def set_repeat(self, repeat: int): 189 self.model_cfg.repeat = repeat 190 191 def set_sample_rate(self, sample_rate: int): 192 self.model_cfg.sample_rate = sample_rate 193 194 def set_dropout(self, dropout: float = 0.0): 195 self.model_cfg.dropout = dropout 196 197 def set_timesteps(self, timesteps: int): 198 self.model_cfg.timesteps = timesteps 199 200 def set_is_regression_task(self, is_regression_task: bool): 201 self.model_cfg.is_regression_task = is_regression_task 202 203 # Note: Autocomplete for users wont work without these overrides 204 # But practically it is not needed since python will infer at runtime 205 206 # def set_train_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 207 # super().set_train_ds(cfg) 208 # 209 # def set_validation_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 210 # super().set_validation_ds(cfg) 211 # 212 # def set_test_ds(self, cfg: Optional[clf_cfg.EncDecClassificationDatasetConfig] = None): 213 # super().set_test_ds(cfg) 214 215 def _finalize_cfg(self): 216 # propagate labels 217 self.model_cfg.train_ds.labels = self.model_cfg.labels 218 self.model_cfg.validation_ds.labels = self.model_cfg.labels 219 self.model_cfg.test_ds.labels = self.model_cfg.labels 220 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 221 222 # propagate num classes 223 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 224 225 # propagate sample rate 226 self.model_cfg.sample_rate = self.model_cfg.sample_rate 227 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 228 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 229 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 230 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 231 232 # propagate filters 233 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 234 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 235 236 # propagate timeteps 237 if self.model_cfg.crop_or_pad_augment is not None: 238 self.model_cfg.crop_or_pad_augment.audio_length = self.model_cfg.timesteps 239 240 # propagate separable 241 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 242 layer.separable = self.model_cfg.separable 243 244 # propagate repeat 245 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 246 layer.repeat = self.model_cfg.repeat 247 248 # propagate dropout 249 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 250 layer.dropout = self.model_cfg.dropout 251 252 def build(self) -> clf_cfg.EncDecClassificationConfig: 253 return super().build() 254 [end of nemo/collections/asr/models/configs/matchboxnet_config.py] [start of nemo/collections/asr/models/configs/quartznet_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Callable, List, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.collections.asr.models.configs import asr_models_config as ctc_cfg 21 from nemo.collections.asr.modules.audio_preprocessing import ( 22 AudioToMelSpectrogramPreprocessorConfig, 23 SpectrogramAugmentationConfig, 24 ) 25 from nemo.collections.asr.modules.conv_asr import ConvASRDecoderConfig, ConvASREncoderConfig, JasperEncoderConfig 26 from nemo.core.config import modelPT as model_cfg 27 28 29 # fmt: off 30 def qn_15x5(): 31 config = [ 32 JasperEncoderConfig(filters=256, repeat=1, kernel=[33], stride=[2], dilation=[1], dropout=0.0, 33 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 34 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 35 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 36 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 37 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 38 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 39 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 40 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 41 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 42 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 43 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 44 JasperEncoderConfig(filters=256, repeat=5, kernel=[33], stride=[1], dilation=[1], dropout=0.0, 45 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 46 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 47 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 48 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 49 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 50 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 51 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 52 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 53 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 54 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 55 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 56 JasperEncoderConfig(filters=256, repeat=5, kernel=[39], stride=[1], dilation=[1], dropout=0.0, 57 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 58 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 59 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 60 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 61 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 62 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 63 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 64 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 65 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 66 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 67 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 68 JasperEncoderConfig(filters=512, repeat=5, kernel=[51], stride=[1], dilation=[1], dropout=0.0, 69 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 70 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 71 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 72 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 73 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 74 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 75 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 76 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 77 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 78 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 79 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 80 JasperEncoderConfig(filters=512, repeat=5, kernel=[63], stride=[1], dilation=[1], dropout=0.0, 81 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 82 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 83 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 84 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 85 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 86 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 87 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 88 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 89 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 90 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 91 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 92 JasperEncoderConfig(filters=512, repeat=5, kernel=[75], stride=[1], dilation=[1], dropout=0.0, 93 residual=True, groups=1, separable=True, heads=-1, residual_mode='add', 94 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 95 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 96 JasperEncoderConfig(filters=512, repeat=1, kernel=[87], stride=[1], dilation=[2], dropout=0.0, 97 residual=False, groups=1, separable=True, heads=-1, residual_mode='add', 98 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 99 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 100 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.0, 101 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 102 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 103 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 104 ] 105 return config 106 107 108 def jasper_10x5_dr(): 109 config = [ 110 JasperEncoderConfig(filters=256, repeat=1, kernel=[11], stride=[2], dilation=[1], dropout=0.2, 111 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 112 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 113 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 114 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 115 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 116 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 117 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 118 JasperEncoderConfig(filters=256, repeat=5, kernel=[11], stride=[1], dilation=[1], dropout=0.2, 119 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 120 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 121 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 122 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 123 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 124 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 125 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 126 JasperEncoderConfig(filters=384, repeat=5, kernel=[13], stride=[1], dilation=[1], dropout=0.2, 127 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 128 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 129 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 130 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 131 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 132 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 133 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 134 JasperEncoderConfig(filters=512, repeat=5, kernel=[17], stride=[1], dilation=[1], dropout=0.2, 135 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 136 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 137 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 138 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 139 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 140 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 141 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 142 JasperEncoderConfig(filters=640, repeat=5, kernel=[21], stride=[1], dilation=[1], dropout=0.3, 143 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 144 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 145 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 146 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 147 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 148 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 149 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 150 JasperEncoderConfig(filters=768, repeat=5, kernel=[25], stride=[1], dilation=[1], dropout=0.3, 151 residual=True, groups=1, separable=False, heads=-1, residual_mode='add', 152 residual_dense=True, se=False, se_reduction_ratio=8, se_context_size=-1, 153 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 154 JasperEncoderConfig(filters=896, repeat=1, kernel=[29], stride=[1], dilation=[2], dropout=0.4, 155 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 156 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 157 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False), 158 JasperEncoderConfig(filters=1024, repeat=1, kernel=[1], stride=[1], dilation=[1], dropout=0.4, 159 residual=False, groups=1, separable=False, heads=-1, residual_mode='add', 160 residual_dense=False, se=False, se_reduction_ratio=8, se_context_size=-1, 161 se_interpolation_mode='nearest', kernel_size_factor=1.0, stride_last=False) 162 ] 163 return config 164 # fmt: on 165 166 167 @dataclass 168 class JasperModelConfig(ctc_cfg.EncDecCTCConfig): 169 # Model global arguments 170 sample_rate: int = 16000 171 repeat: int = 1 172 dropout: float = 0.0 173 separable: bool = False 174 labels: List[str] = MISSING 175 176 # Dataset configs 177 train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( 178 manifest_filepath=None, shuffle=True, trim_silence=True 179 ) 180 validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 181 test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) 182 183 # Optimizer / Scheduler config 184 optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) 185 186 # Model general component configs 187 preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() 188 spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() 189 encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") 190 decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() 191 192 193 @dataclass 194 class QuartzNetModelConfig(JasperModelConfig): 195 separable: bool = True 196 197 198 class EncDecCTCModelConfigBuilder(model_cfg.ModelConfigBuilder): 199 VALID_CONFIGS = ['quartznet_15x5', 'quartznet_15x5_zh', 'jasper_10x5dr'] 200 201 def __init__(self, name: str = 'quartznet_15x5', encoder_cfg_func: Optional[Callable[[], List[Any]]] = None): 202 if name not in EncDecCTCModelConfigBuilder.VALID_CONFIGS: 203 raise ValueError("`name` must be one of : \n" f"{EncDecCTCModelConfigBuilder.VALID_CONFIGS}") 204 205 self.name = name 206 207 if 'quartznet_15x5' in name: 208 if encoder_cfg_func is None: 209 encoder_cfg_func = qn_15x5 210 211 model_cfg = QuartzNetModelConfig( 212 repeat=5, 213 separable=True, 214 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 215 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 216 decoder=ConvASRDecoderConfig(), 217 ) 218 219 elif 'jasper_10x5' in name: 220 if encoder_cfg_func is None: 221 encoder_cfg_func = jasper_10x5_dr 222 223 model_cfg = JasperModelConfig( 224 repeat=5, 225 separable=False, 226 spec_augment=SpectrogramAugmentationConfig(rect_masks=5, rect_freq=50, rect_time=120), 227 encoder=ConvASREncoderConfig(jasper=encoder_cfg_func(), activation="relu"), 228 decoder=ConvASRDecoderConfig(), 229 ) 230 231 else: 232 raise ValueError(f"Invalid config name submitted to {self.__class__.__name__}") 233 234 super(EncDecCTCModelConfigBuilder, self).__init__(model_cfg) 235 self.model_cfg: ctc_cfg.EncDecCTCConfig = model_cfg # enable type hinting 236 237 if 'zh' in name: 238 self.set_dataset_normalize(normalize=False) 239 240 def set_labels(self, labels: List[str]): 241 self.model_cfg.labels = labels 242 243 def set_separable(self, separable: bool): 244 self.model_cfg.separable = separable 245 246 def set_repeat(self, repeat: int): 247 self.model_cfg.repeat = repeat 248 249 def set_sample_rate(self, sample_rate: int): 250 self.model_cfg.sample_rate = sample_rate 251 252 def set_dropout(self, dropout: float = 0.0): 253 self.model_cfg.dropout = dropout 254 255 def set_dataset_normalize(self, normalize: bool): 256 self.model_cfg.train_ds.normalize = normalize 257 self.model_cfg.validation_ds.normalize = normalize 258 self.model_cfg.test_ds.normalize = normalize 259 260 # Note: Autocomplete for users wont work without these overrides 261 # But practically it is not needed since python will infer at runtime 262 263 # def set_train_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 264 # super().set_train_ds(cfg) 265 # 266 # def set_validation_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 267 # super().set_validation_ds(cfg) 268 # 269 # def set_test_ds(self, cfg: Optional[ctc_cfg.ASRDatasetConfig] = None): 270 # super().set_test_ds(cfg) 271 272 def _finalize_cfg(self): 273 # propagate labels 274 self.model_cfg.train_ds.labels = self.model_cfg.labels 275 self.model_cfg.validation_ds.labels = self.model_cfg.labels 276 self.model_cfg.test_ds.labels = self.model_cfg.labels 277 self.model_cfg.decoder.vocabulary = self.model_cfg.labels 278 279 # propagate num classes 280 self.model_cfg.decoder.num_classes = len(self.model_cfg.labels) 281 282 # propagate sample rate 283 self.model_cfg.sample_rate = self.model_cfg.sample_rate 284 self.model_cfg.preprocessor.sample_rate = self.model_cfg.sample_rate 285 self.model_cfg.train_ds.sample_rate = self.model_cfg.sample_rate 286 self.model_cfg.validation_ds.sample_rate = self.model_cfg.sample_rate 287 self.model_cfg.test_ds.sample_rate = self.model_cfg.sample_rate 288 289 # propagate filters 290 self.model_cfg.encoder.feat_in = self.model_cfg.preprocessor.features 291 self.model_cfg.decoder.feat_in = self.model_cfg.encoder.jasper[-1].filters 292 293 # propagate separable 294 for layer in self.model_cfg.encoder.jasper[:-1]: # type: JasperEncoderConfig 295 layer.separable = self.model_cfg.separable 296 297 # propagate repeat 298 for layer in self.model_cfg.encoder.jasper[1:-2]: # type: JasperEncoderConfig 299 layer.repeat = self.model_cfg.repeat 300 301 # propagate dropout 302 for layer in self.model_cfg.encoder.jasper: # type: JasperEncoderConfig 303 layer.dropout = self.model_cfg.dropout 304 305 def build(self) -> ctc_cfg.EncDecCTCConfig: 306 return super().build() 307 [end of nemo/collections/asr/models/configs/quartznet_config.py] [start of nemo/collections/asr/modules/audio_preprocessing.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import random 17 from abc import ABC, abstractmethod 18 from dataclasses import dataclass 19 from typing import Any, Dict, Optional, Tuple 20 21 import torch 22 from packaging import version 23 24 from nemo.collections.asr.parts.numba.spec_augment import SpecAugmentNumba, spec_augment_launch_heuristics 25 from nemo.collections.asr.parts.preprocessing.features import ( 26 FilterbankFeatures, 27 FilterbankFeaturesTA, 28 make_seq_mask_like, 29 ) 30 from nemo.collections.asr.parts.submodules.spectr_augment import SpecAugment, SpecCutout 31 from nemo.core.classes import Exportable, NeuralModule, typecheck 32 from nemo.core.neural_types import ( 33 AudioSignal, 34 LengthsType, 35 MelSpectrogramType, 36 MFCCSpectrogramType, 37 NeuralType, 38 SpectrogramType, 39 ) 40 from nemo.core.utils import numba_utils 41 from nemo.core.utils.numba_utils import __NUMBA_MINIMUM_VERSION__ 42 from nemo.utils import logging 43 44 try: 45 import torchaudio 46 import torchaudio.functional 47 import torchaudio.transforms 48 49 TORCHAUDIO_VERSION = version.parse(torchaudio.__version__) 50 TORCHAUDIO_VERSION_MIN = version.parse('0.5') 51 52 HAVE_TORCHAUDIO = True 53 except ModuleNotFoundError: 54 HAVE_TORCHAUDIO = False 55 56 __all__ = [ 57 'AudioToMelSpectrogramPreprocessor', 58 'AudioToSpectrogram', 59 'SpectrogramToAudio', 60 'AudioToMFCCPreprocessor', 61 'SpectrogramAugmentation', 62 'MaskedPatchAugmentation', 63 'CropOrPadSpectrogramAugmentation', 64 ] 65 66 67 class AudioPreprocessor(NeuralModule, ABC): 68 """ 69 An interface for Neural Modules that performs audio pre-processing, 70 transforming the wav files to features. 71 """ 72 73 def __init__(self, win_length, hop_length): 74 super().__init__() 75 76 self.win_length = win_length 77 self.hop_length = hop_length 78 79 self.torch_windows = { 80 'hann': torch.hann_window, 81 'hamming': torch.hamming_window, 82 'blackman': torch.blackman_window, 83 'bartlett': torch.bartlett_window, 84 'ones': torch.ones, 85 None: torch.ones, 86 } 87 88 @typecheck() 89 @torch.no_grad() 90 def forward(self, input_signal, length): 91 processed_signal, processed_length = self.get_features(input_signal, length) 92 93 return processed_signal, processed_length 94 95 @abstractmethod 96 def get_features(self, input_signal, length): 97 # Called by forward(). Subclasses should implement this. 98 pass 99 100 101 class AudioToMelSpectrogramPreprocessor(AudioPreprocessor, Exportable): 102 """Featurizer module that converts wavs to mel spectrograms. 103 104 Args: 105 sample_rate (int): Sample rate of the input audio data. 106 Defaults to 16000 107 window_size (float): Size of window for fft in seconds 108 Defaults to 0.02 109 window_stride (float): Stride of window for fft in seconds 110 Defaults to 0.01 111 n_window_size (int): Size of window for fft in samples 112 Defaults to None. Use one of window_size or n_window_size. 113 n_window_stride (int): Stride of window for fft in samples 114 Defaults to None. Use one of window_stride or n_window_stride. 115 window (str): Windowing function for fft. can be one of ['hann', 116 'hamming', 'blackman', 'bartlett'] 117 Defaults to "hann" 118 normalize (str): Can be one of ['per_feature', 'all_features']; all 119 other options disable feature normalization. 'all_features' 120 normalizes the entire spectrogram to be mean 0 with std 1. 121 'pre_features' normalizes per channel / freq instead. 122 Defaults to "per_feature" 123 n_fft (int): Length of FT window. If None, it uses the smallest power 124 of 2 that is larger than n_window_size. 125 Defaults to None 126 preemph (float): Amount of pre emphasis to add to audio. Can be 127 disabled by passing None. 128 Defaults to 0.97 129 features (int): Number of mel spectrogram freq bins to output. 130 Defaults to 64 131 lowfreq (int): Lower bound on mel basis in Hz. 132 Defaults to 0 133 highfreq (int): Lower bound on mel basis in Hz. 134 Defaults to None 135 log (bool): Log features. 136 Defaults to True 137 log_zero_guard_type(str): Need to avoid taking the log of zero. There 138 are two options: "add" or "clamp". 139 Defaults to "add". 140 log_zero_guard_value(float, or str): Add or clamp requires the number 141 to add with or clamp to. log_zero_guard_value can either be a float 142 or "tiny" or "eps". torch.finfo is used if "tiny" or "eps" is 143 passed. 144 Defaults to 2**-24. 145 dither (float): Amount of white-noise dithering. 146 Defaults to 1e-5 147 pad_to (int): Ensures that the output size of the time dimension is 148 a multiple of pad_to. 149 Defaults to 16 150 frame_splicing (int): Defaults to 1 151 exact_pad (bool): If True, sets stft center to False and adds padding, such that num_frames = audio_length 152 // hop_length. Defaults to False. 153 pad_value (float): The value that shorter mels are padded with. 154 Defaults to 0 155 mag_power (float): The power that the linear spectrogram is raised to 156 prior to multiplication with mel basis. 157 Defaults to 2 for a power spec 158 rng : Random number generator 159 nb_augmentation_prob (float) : Probability with which narrowband augmentation would be applied to 160 samples in the batch. 161 Defaults to 0.0 162 nb_max_freq (int) : Frequency above which all frequencies will be masked for narrowband augmentation. 163 Defaults to 4000 164 use_torchaudio: Whether to use the `torchaudio` implementation. 165 mel_norm: Normalization used for mel filterbank weights. 166 Defaults to 'slaney' (area normalization) 167 stft_exact_pad: Deprecated argument, kept for compatibility with older checkpoints. 168 stft_conv: Deprecated argument, kept for compatibility with older checkpoints. 169 """ 170 171 def save_to(self, save_path: str): 172 pass 173 174 @classmethod 175 def restore_from(cls, restore_path: str): 176 pass 177 178 @property 179 def input_types(self): 180 """Returns definitions of module input ports. 181 """ 182 return { 183 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 184 "length": NeuralType( 185 tuple('B'), LengthsType() 186 ), # Please note that length should be in samples not seconds. 187 } 188 189 @property 190 def output_types(self): 191 """Returns definitions of module output ports. 192 193 processed_signal: 194 0: AxisType(BatchTag) 195 1: AxisType(MelSpectrogramSignalTag) 196 2: AxisType(ProcessedTimeTag) 197 processed_length: 198 0: AxisType(BatchTag) 199 """ 200 return { 201 "processed_signal": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 202 "processed_length": NeuralType(tuple('B'), LengthsType()), 203 } 204 205 def __init__( 206 self, 207 sample_rate=16000, 208 window_size=0.02, 209 window_stride=0.01, 210 n_window_size=None, 211 n_window_stride=None, 212 window="hann", 213 normalize="per_feature", 214 n_fft=None, 215 preemph=0.97, 216 features=64, 217 lowfreq=0, 218 highfreq=None, 219 log=True, 220 log_zero_guard_type="add", 221 log_zero_guard_value=2 ** -24, 222 dither=1e-5, 223 pad_to=16, 224 frame_splicing=1, 225 exact_pad=False, 226 pad_value=0, 227 mag_power=2.0, 228 rng=None, 229 nb_augmentation_prob=0.0, 230 nb_max_freq=4000, 231 use_torchaudio: bool = False, 232 mel_norm="slaney", 233 stft_exact_pad=False, # Deprecated arguments; kept for config compatibility 234 stft_conv=False, # Deprecated arguments; kept for config compatibility 235 ): 236 super().__init__(n_window_size, n_window_stride) 237 238 self._sample_rate = sample_rate 239 if window_size and n_window_size: 240 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 241 if window_stride and n_window_stride: 242 raise ValueError( 243 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 244 ) 245 if window_size: 246 n_window_size = int(window_size * self._sample_rate) 247 if window_stride: 248 n_window_stride = int(window_stride * self._sample_rate) 249 250 # Given the long and similar argument list, point to the class and instantiate it by reference 251 if not use_torchaudio: 252 featurizer_class = FilterbankFeatures 253 else: 254 featurizer_class = FilterbankFeaturesTA 255 self.featurizer = featurizer_class( 256 sample_rate=self._sample_rate, 257 n_window_size=n_window_size, 258 n_window_stride=n_window_stride, 259 window=window, 260 normalize=normalize, 261 n_fft=n_fft, 262 preemph=preemph, 263 nfilt=features, 264 lowfreq=lowfreq, 265 highfreq=highfreq, 266 log=log, 267 log_zero_guard_type=log_zero_guard_type, 268 log_zero_guard_value=log_zero_guard_value, 269 dither=dither, 270 pad_to=pad_to, 271 frame_splicing=frame_splicing, 272 exact_pad=exact_pad, 273 pad_value=pad_value, 274 mag_power=mag_power, 275 rng=rng, 276 nb_augmentation_prob=nb_augmentation_prob, 277 nb_max_freq=nb_max_freq, 278 mel_norm=mel_norm, 279 stft_exact_pad=stft_exact_pad, # Deprecated arguments; kept for config compatibility 280 stft_conv=stft_conv, # Deprecated arguments; kept for config compatibility 281 ) 282 283 def input_example(self, max_batch: int = 8, max_dim: int = 32000, min_length: int = 200): 284 batch_size = torch.randint(low=1, high=max_batch, size=[1]).item() 285 max_length = torch.randint(low=min_length, high=max_dim, size=[1]).item() 286 signals = torch.rand(size=[batch_size, max_length]) * 2 - 1 287 lengths = torch.randint(low=min_length, high=max_dim, size=[batch_size]) 288 lengths[0] = max_length 289 return signals, lengths 290 291 def get_features(self, input_signal, length): 292 return self.featurizer(input_signal, length) 293 294 @property 295 def filter_banks(self): 296 return self.featurizer.filter_banks 297 298 299 class AudioToMFCCPreprocessor(AudioPreprocessor): 300 """Preprocessor that converts wavs to MFCCs. 301 Uses torchaudio.transforms.MFCC. 302 303 Args: 304 sample_rate: The sample rate of the audio. 305 Defaults to 16000. 306 window_size: Size of window for fft in seconds. Used to calculate the 307 win_length arg for mel spectrogram. 308 Defaults to 0.02 309 window_stride: Stride of window for fft in seconds. Used to caculate 310 the hop_length arg for mel spect. 311 Defaults to 0.01 312 n_window_size: Size of window for fft in samples 313 Defaults to None. Use one of window_size or n_window_size. 314 n_window_stride: Stride of window for fft in samples 315 Defaults to None. Use one of window_stride or n_window_stride. 316 window: Windowing function for fft. can be one of ['hann', 317 'hamming', 'blackman', 'bartlett', 'none', 'null']. 318 Defaults to 'hann' 319 n_fft: Length of FT window. If None, it uses the smallest power of 2 320 that is larger than n_window_size. 321 Defaults to None 322 lowfreq (int): Lower bound on mel basis in Hz. 323 Defaults to 0 324 highfreq (int): Lower bound on mel basis in Hz. 325 Defaults to None 326 n_mels: Number of mel filterbanks. 327 Defaults to 64 328 n_mfcc: Number of coefficients to retain 329 Defaults to 64 330 dct_type: Type of discrete cosine transform to use 331 norm: Type of norm to use 332 log: Whether to use log-mel spectrograms instead of db-scaled. 333 Defaults to True. 334 """ 335 336 @property 337 def input_types(self): 338 """Returns definitions of module input ports. 339 """ 340 return { 341 "input_signal": NeuralType(('B', 'T'), AudioSignal(freq=self._sample_rate)), 342 "length": NeuralType(tuple('B'), LengthsType()), 343 } 344 345 @property 346 def output_types(self): 347 """Returns definitions of module output ports. 348 """ 349 return { 350 "processed_signal": NeuralType(('B', 'D', 'T'), MFCCSpectrogramType()), 351 "processed_length": NeuralType(tuple('B'), LengthsType()), 352 } 353 354 def save_to(self, save_path: str): 355 pass 356 357 @classmethod 358 def restore_from(cls, restore_path: str): 359 pass 360 361 def __init__( 362 self, 363 sample_rate=16000, 364 window_size=0.02, 365 window_stride=0.01, 366 n_window_size=None, 367 n_window_stride=None, 368 window='hann', 369 n_fft=None, 370 lowfreq=0.0, 371 highfreq=None, 372 n_mels=64, 373 n_mfcc=64, 374 dct_type=2, 375 norm='ortho', 376 log=True, 377 ): 378 self._sample_rate = sample_rate 379 if not HAVE_TORCHAUDIO: 380 logging.error('Could not import torchaudio. Some features might not work.') 381 382 raise ModuleNotFoundError( 383 "torchaudio is not installed but is necessary for " 384 "AudioToMFCCPreprocessor. We recommend you try " 385 "building it from source for the PyTorch version you have." 386 ) 387 if window_size and n_window_size: 388 raise ValueError(f"{self} received both window_size and " f"n_window_size. Only one should be specified.") 389 if window_stride and n_window_stride: 390 raise ValueError( 391 f"{self} received both window_stride and " f"n_window_stride. Only one should be specified." 392 ) 393 # Get win_length (n_window_size) and hop_length (n_window_stride) 394 if window_size: 395 n_window_size = int(window_size * self._sample_rate) 396 if window_stride: 397 n_window_stride = int(window_stride * self._sample_rate) 398 399 super().__init__(n_window_size, n_window_stride) 400 401 mel_kwargs = {} 402 403 mel_kwargs['f_min'] = lowfreq 404 mel_kwargs['f_max'] = highfreq 405 mel_kwargs['n_mels'] = n_mels 406 407 mel_kwargs['n_fft'] = n_fft or 2 ** math.ceil(math.log2(n_window_size)) 408 409 mel_kwargs['win_length'] = n_window_size 410 mel_kwargs['hop_length'] = n_window_stride 411 412 # Set window_fn. None defaults to torch.ones. 413 window_fn = self.torch_windows.get(window, None) 414 if window_fn is None: 415 raise ValueError( 416 f"Window argument for AudioProcessor is invalid: {window}." 417 f"For no window function, use 'ones' or None." 418 ) 419 mel_kwargs['window_fn'] = window_fn 420 421 # Use torchaudio's implementation of MFCCs as featurizer 422 self.featurizer = torchaudio.transforms.MFCC( 423 sample_rate=self._sample_rate, 424 n_mfcc=n_mfcc, 425 dct_type=dct_type, 426 norm=norm, 427 log_mels=log, 428 melkwargs=mel_kwargs, 429 ) 430 431 def get_features(self, input_signal, length): 432 features = self.featurizer(input_signal) 433 seq_len = torch.ceil(length.to(torch.float32) / self.hop_length).to(dtype=torch.long) 434 return features, seq_len 435 436 437 class SpectrogramAugmentation(NeuralModule): 438 """ 439 Performs time and freq cuts in one of two ways. 440 SpecAugment zeroes out vertical and horizontal sections as described in 441 SpecAugment (https://arxiv.org/abs/1904.08779). Arguments for use with 442 SpecAugment are `freq_masks`, `time_masks`, `freq_width`, and `time_width`. 443 SpecCutout zeroes out rectangulars as described in Cutout 444 (https://arxiv.org/abs/1708.04552). Arguments for use with Cutout are 445 `rect_masks`, `rect_freq`, and `rect_time`. 446 447 Args: 448 freq_masks (int): how many frequency segments should be cut. 449 Defaults to 0. 450 time_masks (int): how many time segments should be cut 451 Defaults to 0. 452 freq_width (int): maximum number of frequencies to be cut in one 453 segment. 454 Defaults to 10. 455 time_width (int): maximum number of time steps to be cut in one 456 segment 457 Defaults to 10. 458 rect_masks (int): how many rectangular masks should be cut 459 Defaults to 0. 460 rect_freq (int): maximum size of cut rectangles along the frequency 461 dimension 462 Defaults to 5. 463 rect_time (int): maximum size of cut rectangles along the time 464 dimension 465 Defaults to 25. 466 """ 467 468 @property 469 def input_types(self): 470 """Returns definitions of module input types 471 """ 472 return { 473 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 474 "length": NeuralType(tuple('B'), LengthsType()), 475 } 476 477 @property 478 def output_types(self): 479 """Returns definitions of module output types 480 """ 481 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 482 483 def __init__( 484 self, 485 freq_masks=0, 486 time_masks=0, 487 freq_width=10, 488 time_width=10, 489 rect_masks=0, 490 rect_time=5, 491 rect_freq=20, 492 rng=None, 493 mask_value=0.0, 494 use_numba_spec_augment: bool = True, 495 ): 496 super().__init__() 497 498 if rect_masks > 0: 499 self.spec_cutout = SpecCutout(rect_masks=rect_masks, rect_time=rect_time, rect_freq=rect_freq, rng=rng,) 500 # self.spec_cutout.to(self._device) 501 else: 502 self.spec_cutout = lambda input_spec: input_spec 503 if freq_masks + time_masks > 0: 504 self.spec_augment = SpecAugment( 505 freq_masks=freq_masks, 506 time_masks=time_masks, 507 freq_width=freq_width, 508 time_width=time_width, 509 rng=rng, 510 mask_value=mask_value, 511 ) 512 else: 513 self.spec_augment = lambda input_spec, length: input_spec 514 515 # Check if numba is supported, and use a Numba kernel if it is 516 if use_numba_spec_augment and numba_utils.numba_cuda_is_supported(__NUMBA_MINIMUM_VERSION__): 517 logging.info('Numba CUDA SpecAugment kernel is being used') 518 self.spec_augment_numba = SpecAugmentNumba( 519 freq_masks=freq_masks, 520 time_masks=time_masks, 521 freq_width=freq_width, 522 time_width=time_width, 523 rng=rng, 524 mask_value=mask_value, 525 ) 526 else: 527 self.spec_augment_numba = None 528 529 @typecheck() 530 def forward(self, input_spec, length): 531 augmented_spec = self.spec_cutout(input_spec=input_spec) 532 533 # To run the Numba kernel, correct numba version is required as well as 534 # tensor must be on GPU and length must be provided 535 if self.spec_augment_numba is not None and spec_augment_launch_heuristics(augmented_spec, length): 536 augmented_spec = self.spec_augment_numba(input_spec=augmented_spec, length=length) 537 else: 538 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 539 return augmented_spec 540 541 542 class MaskedPatchAugmentation(NeuralModule): 543 """ 544 Zeroes out fixed size time patches of the spectrogram. 545 All samples in batch are guaranteed to have the same amount of masked time steps. 546 Optionally also performs frequency masking in the same way as SpecAugment. 547 Args: 548 patch_size (int): up to how many time steps does one patch consist of. 549 Defaults to 48. 550 mask_patches (float): how many patches should be masked in each sample. 551 if >= 1., interpreted as number of patches (after converting to int) 552 if <1., interpreted as fraction of total tokens to be masked (number of patches is rounded up) 553 Defaults to 10. 554 freq_masks (int): how many frequency segments should be cut. 555 Defaults to 0. 556 freq_width (int): maximum number of frequencies to be cut in a segment. 557 Defaults to 0. 558 """ 559 560 @property 561 def input_types(self): 562 """Returns definitions of module input types 563 """ 564 return { 565 "input_spec": NeuralType(('B', 'D', 'T'), SpectrogramType()), 566 "length": NeuralType(tuple('B'), LengthsType()), 567 } 568 569 @property 570 def output_types(self): 571 """Returns definitions of module output types 572 """ 573 return {"augmented_spec": NeuralType(('B', 'D', 'T'), SpectrogramType())} 574 575 def __init__( 576 self, patch_size: int = 48, mask_patches: float = 10.0, freq_masks: int = 0, freq_width: int = 0, 577 ): 578 super().__init__() 579 self.patch_size = patch_size 580 if mask_patches >= 1: 581 self.mask_patches = int(mask_patches) 582 elif mask_patches >= 0: 583 self._mask_fraction = mask_patches 584 self.mask_patches = None 585 else: 586 raise ValueError('mask_patches cannot be negative') 587 588 if freq_masks > 0: 589 self.spec_augment = SpecAugment(freq_masks=freq_masks, time_masks=0, freq_width=freq_width, time_width=0,) 590 else: 591 self.spec_augment = None 592 593 @typecheck() 594 def forward(self, input_spec, length): 595 augmented_spec = input_spec 596 597 min_len = torch.min(length) 598 599 if self.mask_patches is None: 600 # masking specified as fraction 601 len_fraction = int(min_len * self._mask_fraction) 602 mask_patches = len_fraction // self.patch_size + int(len_fraction % self.patch_size != 0) 603 else: 604 mask_patches = self.mask_patches 605 606 if min_len < self.patch_size * mask_patches: 607 mask_patches = min_len // self.patch_size 608 609 for idx in range(input_spec.shape[0]): 610 cur_len = length[idx] 611 patches = range(cur_len // self.patch_size) 612 masked_patches = random.sample(patches, mask_patches) 613 614 for mp in masked_patches: 615 augmented_spec[idx, :, mp * self.patch_size : (mp + 1) * self.patch_size] = 0.0 616 617 if self.spec_augment is not None: 618 augmented_spec = self.spec_augment(input_spec=augmented_spec, length=length) 619 620 return augmented_spec 621 622 623 class CropOrPadSpectrogramAugmentation(NeuralModule): 624 """ 625 Pad or Crop the incoming Spectrogram to a certain shape. 626 627 Args: 628 audio_length (int): the final number of timesteps that is required. 629 The signal will be either padded or cropped temporally to this 630 size. 631 """ 632 633 def __init__(self, audio_length): 634 super(CropOrPadSpectrogramAugmentation, self).__init__() 635 self.audio_length = audio_length 636 637 @typecheck() 638 @torch.no_grad() 639 def forward(self, input_signal, length): 640 image = input_signal 641 num_images = image.shape[0] 642 643 audio_length = self.audio_length 644 image_len = image.shape[-1] 645 646 # Crop long signal 647 if image_len > audio_length: # randomly slice 648 cutout_images = [] 649 offset = torch.randint(low=0, high=image_len - audio_length + 1, size=[num_images]) 650 651 for idx, offset in enumerate(offset): 652 cutout_images.append(image[idx : idx + 1, :, offset : offset + audio_length]) 653 654 image = torch.cat(cutout_images, dim=0) 655 del cutout_images 656 657 else: # symmetrically pad short signal with zeros 658 pad_left = (audio_length - image_len) // 2 659 pad_right = (audio_length - image_len) // 2 660 661 if (audio_length - image_len) % 2 == 1: 662 pad_right += 1 663 664 image = torch.nn.functional.pad(image, [pad_left, pad_right], mode="constant", value=0) 665 666 # Replace dynamic length sequences with static number of timesteps 667 length = (length * 0) + audio_length 668 669 return image, length 670 671 @property 672 def input_types(self): 673 """Returns definitions of module output ports. 674 """ 675 return { 676 "input_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 677 "length": NeuralType(tuple('B'), LengthsType()), 678 } 679 680 @property 681 def output_types(self): 682 """Returns definitions of module output ports. 683 """ 684 return { 685 "processed_signal": NeuralType(('B', 'D', 'T'), SpectrogramType()), 686 "processed_length": NeuralType(tuple('B'), LengthsType()), 687 } 688 689 def save_to(self, save_path: str): 690 pass 691 692 @classmethod 693 def restore_from(cls, restore_path: str): 694 pass 695 696 697 class AudioToSpectrogram(NeuralModule): 698 """Transform a batch of input multi-channel signals into a batch of 699 STFT-based spectrograms. 700 701 Args: 702 fft_length: length of FFT 703 hop_length: length of hops/shifts of the sliding window 704 power: exponent for magnitude spectrogram. Default `None` will 705 return a complex-valued spectrogram 706 """ 707 708 def __init__(self, fft_length: int, hop_length: int, power: Optional[float] = None): 709 if not HAVE_TORCHAUDIO: 710 logging.error('Could not import torchaudio. Some features might not work.') 711 712 raise ModuleNotFoundError( 713 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 714 ) 715 716 super().__init__() 717 718 # For now, assume FFT length is divisible by two 719 if fft_length % 2 != 0: 720 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 721 722 self.stft = torchaudio.transforms.Spectrogram( 723 n_fft=fft_length, hop_length=hop_length, power=power, pad_mode='constant' 724 ) 725 726 # number of subbands 727 self.F = fft_length // 2 + 1 728 729 @property 730 def num_subbands(self) -> int: 731 return self.F 732 733 @property 734 def input_types(self) -> Dict[str, NeuralType]: 735 """Returns definitions of module output ports. 736 """ 737 return { 738 "input": NeuralType(('B', 'C', 'T'), AudioSignal()), 739 "input_length": NeuralType(('B',), LengthsType(), optional=True), 740 } 741 742 @property 743 def output_types(self) -> Dict[str, NeuralType]: 744 """Returns definitions of module output ports. 745 """ 746 return { 747 "output": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 748 "output_length": NeuralType(('B',), LengthsType()), 749 } 750 751 @typecheck() 752 def forward( 753 self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None 754 ) -> Tuple[torch.Tensor, torch.Tensor]: 755 """Convert a batch of C-channel input signals 756 into a batch of complex-valued spectrograms. 757 758 Args: 759 input: Time-domain input signal with C channels, shape (B, C, T) 760 input_length: Length of valid entries along the time dimension, shape (B,) 761 762 Returns: 763 Output spectrogram with F subbands and N time frames, shape (B, C, F, N) 764 and output length with shape (B,). 765 """ 766 B, T = input.size(0), input.size(-1) 767 input = input.view(B, -1, T) 768 769 # STFT output (B, C, F, N) 770 with torch.cuda.amp.autocast(enabled=False): 771 output = self.stft(input.float()) 772 773 if input_length is not None: 774 # Mask padded frames 775 output_length = self.get_output_length(input_length=input_length) 776 777 length_mask: torch.Tensor = make_seq_mask_like( 778 lengths=output_length, like=output, time_dim=-1, valid_ones=False 779 ) 780 output = output.masked_fill(length_mask, 0.0) 781 else: 782 # Assume all frames are valid for all examples in the batch 783 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 784 785 return output, output_length 786 787 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 788 """Get length of valid frames for the output. 789 790 Args: 791 input_length: number of valid samples, shape (B,) 792 793 Returns: 794 Number of valid frames, shape (B,) 795 """ 796 output_length = input_length.div(self.stft.hop_length, rounding_mode='floor').add(1).long() 797 return output_length 798 799 800 class SpectrogramToAudio(NeuralModule): 801 """Transform a batch of input multi-channel spectrograms into a batch of 802 time-domain multi-channel signals. 803 804 Args: 805 fft_length: length of FFT 806 hop_length: length of hops/shifts of the sliding window 807 power: exponent for magnitude spectrogram. Default `None` will 808 return a complex-valued spectrogram 809 """ 810 811 def __init__(self, fft_length: int, hop_length: int): 812 if not HAVE_TORCHAUDIO: 813 logging.error('Could not import torchaudio. Some features might not work.') 814 815 raise ModuleNotFoundError( 816 "torchaudio is not installed but is necessary to instantiate a {self.__class__.__name__}" 817 ) 818 819 super().__init__() 820 821 # For now, assume FFT length is divisible by two 822 if fft_length % 2 != 0: 823 raise ValueError(f'fft_length = {fft_length} must be divisible by 2') 824 825 self.istft = torchaudio.transforms.InverseSpectrogram( 826 n_fft=fft_length, hop_length=hop_length, pad_mode='constant' 827 ) 828 829 self.F = fft_length // 2 + 1 830 831 @property 832 def num_subbands(self) -> int: 833 return self.F 834 835 @property 836 def input_types(self) -> Dict[str, NeuralType]: 837 """Returns definitions of module output ports. 838 """ 839 return { 840 "input": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()), 841 "input_length": NeuralType(('B',), LengthsType(), optional=True), 842 } 843 844 @property 845 def output_types(self) -> Dict[str, NeuralType]: 846 """Returns definitions of module output ports. 847 """ 848 return { 849 "output": NeuralType(('B', 'C', 'T'), AudioSignal()), 850 "output_length": NeuralType(('B',), LengthsType()), 851 } 852 853 @typecheck() 854 def forward(self, input: torch.Tensor, input_length: Optional[torch.Tensor] = None) -> torch.Tensor: 855 """Convert input complex-valued spectrogram to a time-domain 856 signal. Multi-channel IO is supported. 857 858 Args: 859 input: Input spectrogram for C channels, shape (B, C, F, N) 860 input_length: Length of valid entries along the time dimension, shape (B,) 861 862 Returns: 863 Time-domain signal with T time-domain samples and C channels, (B, C, T) 864 and output length with shape (B,). 865 """ 866 B, F, N = input.size(0), input.size(-2), input.size(-1) 867 assert F == self.F, f'Number of subbands F={F} not matching self.F={self.F}' 868 input = input.view(B, -1, F, N) 869 870 # iSTFT output (B, C, T) 871 with torch.cuda.amp.autocast(enabled=False): 872 output = self.istft(input.cfloat()) 873 874 if input_length is not None: 875 # Mask padded samples 876 output_length = self.get_output_length(input_length=input_length) 877 878 length_mask: torch.Tensor = make_seq_mask_like( 879 lengths=output_length, like=output, time_dim=-1, valid_ones=False 880 ) 881 output = output.masked_fill(length_mask, 0.0) 882 else: 883 # Assume all frames are valid for all examples in the batch 884 output_length = output.size(-1) * torch.ones(B, device=output.device).long() 885 886 return output, output_length 887 888 def get_output_length(self, input_length: torch.Tensor) -> torch.Tensor: 889 """Get length of valid samples for the output. 890 891 Args: 892 input_length: number of valid frames, shape (B,) 893 894 Returns: 895 Number of valid samples, shape (B,) 896 """ 897 output_length = input_length.sub(1).mul(self.istft.hop_length).long() 898 return output_length 899 900 901 @dataclass 902 class AudioToMelSpectrogramPreprocessorConfig: 903 _target_: str = "nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor" 904 sample_rate: int = 16000 905 window_size: float = 0.02 906 window_stride: float = 0.01 907 n_window_size: Optional[int] = None 908 n_window_stride: Optional[int] = None 909 window: str = "hann" 910 normalize: str = "per_feature" 911 n_fft: Optional[int] = None 912 preemph: float = 0.97 913 features: int = 64 914 lowfreq: int = 0 915 highfreq: Optional[int] = None 916 log: bool = True 917 log_zero_guard_type: str = "add" 918 log_zero_guard_value: float = 2 ** -24 919 dither: float = 1e-5 920 pad_to: int = 16 921 frame_splicing: int = 1 922 exact_pad: bool = False 923 pad_value: int = 0 924 mag_power: float = 2.0 925 rng: Optional[str] = None 926 nb_augmentation_prob: float = 0.0 927 nb_max_freq: int = 4000 928 use_torchaudio: bool = False 929 mel_norm: str = "slaney" 930 stft_exact_pad: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 931 stft_conv: bool = False # Deprecated argument, kept for compatibility with older checkpoints. 932 933 934 @dataclass 935 class AudioToMFCCPreprocessorConfig: 936 _target_: str = 'nemo.collections.asr.modules.AudioToMFCCPreprocessor' 937 sample_rate: int = 16000 938 window_size: float = 0.02 939 window_stride: float = 0.01 940 n_window_size: Optional[int] = None 941 n_window_stride: Optional[int] = None 942 window: str = 'hann' 943 n_fft: Optional[int] = None 944 lowfreq: Optional[float] = 0.0 945 highfreq: Optional[float] = None 946 n_mels: int = 64 947 n_mfcc: int = 64 948 dct_type: int = 2 949 norm: str = 'ortho' 950 log: bool = True 951 952 953 @dataclass 954 class SpectrogramAugmentationConfig: 955 _target_: str = "nemo.collections.asr.modules.SpectrogramAugmentation" 956 freq_masks: int = 0 957 time_masks: int = 0 958 freq_width: int = 0 959 time_width: Optional[Any] = 0 960 rect_masks: int = 0 961 rect_time: int = 0 962 rect_freq: int = 0 963 mask_value: float = 0 964 rng: Optional[Any] = None # random.Random() type 965 use_numba_spec_augment: bool = True 966 967 968 @dataclass 969 class CropOrPadSpectrogramAugmentationConfig: 970 audio_length: int 971 _target_: str = "nemo.collections.asr.modules.CropOrPadSpectrogramAugmentation" 972 973 974 @dataclass 975 class MaskedPatchAugmentationConfig: 976 patch_size: int = 48 977 mask_patches: float = 10.0 978 freq_masks: int = 0 979 freq_width: int = 0 980 _target_: str = "nemo.collections.asr.modules.MaskedPatchAugmentation" 981 [end of nemo/collections/asr/modules/audio_preprocessing.py] [start of nemo/collections/asr/parts/k2/classes.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from abc import ABC 16 from dataclasses import dataclass 17 from typing import Any, Optional, Tuple 18 19 import torch 20 from omegaconf import DictConfig 21 22 from nemo.utils import logging 23 24 25 @dataclass 26 class GraphIntersectDenseConfig: 27 """Graph dense intersection config. 28 """ 29 30 search_beam: float = 20.0 31 output_beam: float = 10.0 32 min_active_states: int = 30 33 max_active_states: int = 10000 34 35 36 @dataclass 37 class GraphModuleConfig: 38 """Config for graph modules. 39 Typically used with graph losses and decoders. 40 """ 41 42 topo_type: str = "default" 43 topo_with_self_loops: bool = True 44 token_lm: Optional[Any] = None 45 intersect_pruned: bool = False 46 intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() 47 boost_coeff: float = 0.0 48 predictor_window_size: int = 0 49 predictor_step_size: int = 1 50 51 52 class ASRK2Mixin(ABC): 53 """k2 Mixin class that simplifies the construction of various models with k2-based losses. 54 55 It does the following: 56 - Sets up the graph loss and decoder (methods _init_k2 and update_k2_modules). 57 - Registers external graphs, if needed. 58 - Augments forward(...) with optional graph decoding to get accurate predictions. 59 """ 60 61 def _init_k2(self): 62 """ 63 k2-related initialization implementation. 64 65 This method is expected to run after the __init__ which sets self._cfg 66 self._cfg is expected to have the attribute graph_module_cfg 67 """ 68 if not hasattr(self, "_cfg"): 69 raise ValueError("self._cfg must be set before calling _init_k2().") 70 if not hasattr(self._cfg, "graph_module_cfg") or self._cfg.graph_module_cfg is None: 71 raise ValueError("self._cfg.graph_module_cfg must be set and cannot be None.") 72 self.graph_module_cfg = self._cfg.graph_module_cfg 73 74 # register token_lm for MAPLoss 75 criterion_type = self.graph_module_cfg.get("criterion_type", "ml") 76 self.use_graph_lm = criterion_type == "map" 77 if self.use_graph_lm: 78 token_lm_path = self.graph_module_cfg.backend_cfg.get("token_lm", None) 79 if token_lm_path is None: 80 raise ValueError( 81 f"graph_module_cfg.backend_cfg.token_lm is empty. It must be set for criterion_type == `{criterion_type}`" 82 ) 83 token_lm_path = self.register_artifact('graph_module_cfg.backend_cfg.token_lm', token_lm_path) 84 self.graph_module_cfg.backend_cfg["token_lm"] = token_lm_path 85 86 self.update_k2_modules(self.graph_module_cfg) 87 88 def update_k2_modules(self, input_cfg: DictConfig): 89 """ 90 Helper function to initialize or update k2 loss and transcribe_decoder. 91 92 Args: 93 input_cfg: DictConfig to take new parameters from. Schema is expected as in 94 nemo.collections.asr.models.configs.k2_sequence_models_config.GraphModuleConfig 95 """ 96 del self.loss 97 if hasattr(self, "transcribe_decoder"): 98 del self.transcribe_decoder 99 100 if hasattr(self, "joint"): 101 # RNNT 102 num_classes = self.joint.num_classes_with_blank - 1 103 else: 104 # CTC, MMI, ... 105 num_classes = self.decoder.num_classes_with_blank - 1 106 remove_consecutive = input_cfg.backend_cfg.get("topo_with_self_loops", True) and input_cfg.backend_cfg.get( 107 "topo_type", "default" 108 ) not in ["forced_blank", "identity",] 109 self._wer.remove_consecutive = remove_consecutive 110 111 from nemo.collections.asr.losses.lattice_losses import LatticeLoss 112 113 self.loss = LatticeLoss( 114 num_classes=num_classes, 115 reduction=self._cfg.get("ctc_reduction", "mean_batch"), 116 backend="k2", 117 criterion_type=input_cfg.get("criterion_type", "ml"), 118 loss_type=input_cfg.get("loss_type", "ctc"), 119 split_batch_size=input_cfg.get("split_batch_size", 0), 120 graph_module_cfg=input_cfg.backend_cfg, 121 ) 122 123 criterion_type = self.loss.criterion_type 124 self.use_graph_lm = criterion_type == "map" 125 transcribe_training = input_cfg.get("transcribe_training", False) 126 if transcribe_training and criterion_type == "ml": 127 logging.warning( 128 f"""You do not need to use transcribe_training=`{transcribe_training}` 129 with criterion_type=`{criterion_type}`. transcribe_training will be set to False.""" 130 ) 131 transcribe_training = False 132 self.transcribe_training = transcribe_training 133 if self.use_graph_lm: 134 from nemo.collections.asr.modules.graph_decoder import ViterbiDecoderWithGraph 135 136 self.transcribe_decoder = ViterbiDecoderWithGraph( 137 num_classes=num_classes, 138 backend="k2", 139 dec_type="token_lm", 140 return_type="1best", 141 return_ilabels=True, 142 output_aligned=True, 143 split_batch_size=input_cfg.get("split_batch_size", 0), 144 graph_module_cfg=input_cfg.backend_cfg, 145 ) 146 147 def _forward_k2_post_processing( 148 self, log_probs: torch.Tensor, encoded_length: torch.Tensor, greedy_predictions: torch.Tensor 149 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: 150 """ 151 k2-related post-processing parf of .forward() 152 153 Args: 154 log_probs: The log probabilities tensor of shape [B, T, D]. 155 encoded_length: The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 156 greedy_predictions: The greedy token predictions of the model of shape [B, T] 157 158 Returns: 159 A tuple of 3 elements - 160 1) The log probabilities tensor of shape [B, T, D]. 161 2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 162 3) The greedy token predictions of the model of shape [B, T] (via argmax) 163 """ 164 # greedy_predictions from .forward() are incorrect for criterion_type=`map` 165 # getting correct greedy_predictions, if needed 166 if self.use_graph_lm and (not self.training or self.transcribe_training): 167 greedy_predictions, encoded_length, _ = self.transcribe_decoder.forward( 168 log_probs=log_probs, log_probs_length=encoded_length 169 ) 170 return log_probs, encoded_length, greedy_predictions 171 [end of nemo/collections/asr/parts/k2/classes.py] [start of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from dataclasses import dataclass 17 from typing import Any, Optional 18 19 import torch 20 from torch import nn as nn 21 22 from nemo.collections.asr.parts.submodules import multi_head_attention as mha 23 from nemo.collections.common.parts import adapter_modules 24 from nemo.core.classes.mixins import adapter_mixin_strategies 25 26 27 class MHAResidualAddAdapterStrategy(adapter_mixin_strategies.ResidualAddAdapterStrategy): 28 """ 29 An implementation of residual addition of an adapter module with its input for the MHA Adapters. 30 """ 31 32 def forward(self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin'): 33 """ 34 A basic strategy, comprising of a residual connection over the input, after forward pass by 35 the underlying adapter. Additional work is done to pack and unpack the dictionary of inputs and outputs. 36 37 Note: The `value` tensor is added to the output of the attention adapter as the residual connection. 38 39 Args: 40 input: A dictionary of multiple input arguments for the adapter module. 41 `query`, `key`, `value`: Original output tensor of the module, or the output of the 42 previous adapter (if more than one adapters are enabled). 43 `mask`: Attention mask. 44 `pos_emb`: Optional positional embedding for relative encoding. 45 adapter: The adapter module that is currently required to perform the forward pass. 46 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 47 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 48 49 Returns: 50 The result tensor, after one of the active adapters has finished its forward passes. 51 """ 52 out = self.compute_output(input, adapter, module=module) 53 54 # If not in training mode, or probability of stochastic depth is 0, skip step. 55 p = self.stochastic_depth 56 if not module.training or p == 0.0: 57 pass 58 else: 59 out = self.apply_stochastic_depth(out, input['value'], adapter, module=module) 60 61 # Return the residual connection output = input + adapter(input) 62 result = input['value'] + out 63 64 # If l2_lambda is activated, register the loss value 65 self.compute_auxiliary_losses(result, input['value'], adapter, module=module) 66 67 return result 68 69 def compute_output( 70 self, input: torch.Tensor, adapter: torch.nn.Module, *, module: 'AdapterModuleMixin' 71 ) -> torch.Tensor: 72 """ 73 Compute the output of a single adapter to some input. 74 75 Args: 76 input: Original output tensor of the module, or the output of the previous adapter (if more than 77 one adapters are enabled). 78 adapter: The adapter module that is currently required to perform the forward pass. 79 module: The calling module, in its entirety. It is a module that implements `AdapterModuleMixin`, 80 therefore the strategy can access all other adapters in this module via `module.adapter_layer`. 81 82 Returns: 83 The result tensor, after one of the active adapters has finished its forward passes. 84 """ 85 if isinstance(input, (list, tuple)): 86 out = adapter(*input) 87 elif isinstance(input, dict): 88 out = adapter(**input) 89 else: 90 out = adapter(input) 91 return out 92 93 94 @dataclass 95 class MHAResidualAddAdapterStrategyConfig(adapter_mixin_strategies.ResidualAddAdapterStrategyConfig): 96 _target_: str = "{0}.{1}".format( 97 MHAResidualAddAdapterStrategy.__module__, MHAResidualAddAdapterStrategy.__name__ 98 ) # mandatory field 99 100 101 class MultiHeadAttentionAdapter(mha.MultiHeadAttention, adapter_modules.AdapterModuleUtil): 102 """Multi-Head Attention layer of Transformer. 103 Args: 104 n_head (int): number of heads 105 n_feat (int): size of the features 106 dropout_rate (float): dropout rate 107 proj_dim (int, optional): Optional integer value for projection before computing attention. 108 If None, then there is no projection (equivalent to proj_dim = n_feat). 109 If > 0, then will project the n_feat to proj_dim before calculating attention. 110 If <0, then will equal n_head, so that each head has a projected dimension of 1. 111 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 112 """ 113 114 def __init__( 115 self, 116 n_head: int, 117 n_feat: int, 118 dropout_rate: float, 119 proj_dim: Optional[int] = None, 120 adapter_strategy: MHAResidualAddAdapterStrategy = None, 121 ): 122 super().__init__(n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, max_cache_len=0) 123 124 self.pre_norm = nn.LayerNorm(n_feat) 125 126 # Set the projection dim to number of heads automatically 127 if proj_dim is not None and proj_dim < 1: 128 proj_dim = n_head 129 130 self.proj_dim = proj_dim 131 132 # Recompute weights for projection dim 133 if self.proj_dim is not None: 134 if self.proj_dim % n_head != 0: 135 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 136 137 self.d_k = self.proj_dim // n_head 138 self.s_d_k = math.sqrt(self.d_k) 139 self.linear_q = nn.Linear(n_feat, self.proj_dim) 140 self.linear_k = nn.Linear(n_feat, self.proj_dim) 141 self.linear_v = nn.Linear(n_feat, self.proj_dim) 142 self.linear_out = nn.Linear(self.proj_dim, n_feat) 143 144 # Setup adapter strategy 145 self.setup_adapter_strategy(adapter_strategy) 146 147 # reset parameters for Q to be identity operation 148 self.reset_parameters() 149 150 def forward(self, query, key, value, mask, pos_emb=None, cache=None): 151 """Compute 'Scaled Dot Product Attention'. 152 Args: 153 query (torch.Tensor): (batch, time1, size) 154 key (torch.Tensor): (batch, time2, size) 155 value(torch.Tensor): (batch, time2, size) 156 mask (torch.Tensor): (batch, time1, time2) 157 cache (torch.Tensor) : (batch, time_cache, size) 158 159 returns: 160 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 161 cache (torch.Tensor) : (batch, time_cache_next, size) 162 """ 163 # Need to perform duplicate computations as at this point the tensors have been 164 # separated by the adapter forward 165 query = self.pre_norm(query) 166 key = self.pre_norm(key) 167 value = self.pre_norm(value) 168 169 return super().forward(query, key, value, mask, pos_emb, cache=cache) 170 171 def reset_parameters(self): 172 with torch.no_grad(): 173 nn.init.zeros_(self.linear_out.weight) 174 nn.init.zeros_(self.linear_out.bias) 175 176 def get_default_strategy_config(self) -> 'dataclass': 177 return MHAResidualAddAdapterStrategyConfig() 178 179 180 @dataclass 181 class MultiHeadAttentionAdapterConfig: 182 n_head: int 183 n_feat: int 184 dropout_rate: float = 0.0 185 proj_dim: Optional[int] = None 186 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 187 _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) 188 189 190 class RelPositionMultiHeadAttentionAdapter(mha.RelPositionMultiHeadAttention, adapter_modules.AdapterModuleUtil): 191 """Multi-Head Attention layer of Transformer-XL with support of relative positional encoding. 192 Paper: https://arxiv.org/abs/1901.02860 193 Args: 194 n_head (int): number of heads 195 n_feat (int): size of the features 196 dropout_rate (float): dropout rate 197 proj_dim (int, optional): Optional integer value for projection before computing attention. 198 If None, then there is no projection (equivalent to proj_dim = n_feat). 199 If > 0, then will project the n_feat to proj_dim before calculating attention. 200 If <0, then will equal n_head, so that each head has a projected dimension of 1. 201 adapter_strategy: By default, MHAResidualAddAdapterStrategyConfig. An adapter composition function object. 202 """ 203 204 def __init__( 205 self, 206 n_head: int, 207 n_feat: int, 208 dropout_rate: float, 209 proj_dim: Optional[int] = None, 210 adapter_strategy: MHAResidualAddAdapterStrategyConfig = None, 211 ): 212 super().__init__( 213 n_head=n_head, n_feat=n_feat, dropout_rate=dropout_rate, pos_bias_u=None, pos_bias_v=None, max_cache_len=0 214 ) 215 216 self.pre_norm = nn.LayerNorm(n_feat) 217 218 # Set the projection dim to number of heads automatically 219 if proj_dim is not None and proj_dim < 1: 220 proj_dim = n_head 221 222 self.proj_dim = proj_dim 223 224 # Recompute weights for projection dim 225 if self.proj_dim is not None: 226 if self.proj_dim % n_head != 0: 227 raise ValueError(f"proj_dim ({proj_dim}) is not divisible by n_head ({n_head})") 228 229 self.d_k = self.proj_dim // n_head 230 self.s_d_k = math.sqrt(self.d_k) 231 self.linear_q = nn.Linear(n_feat, self.proj_dim) 232 self.linear_k = nn.Linear(n_feat, self.proj_dim) 233 self.linear_v = nn.Linear(n_feat, self.proj_dim) 234 self.linear_out = nn.Linear(self.proj_dim, n_feat) 235 self.linear_pos = nn.Linear(n_feat, self.proj_dim, bias=False) 236 self.pos_bias_u = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 237 self.pos_bias_v = nn.Parameter(torch.FloatTensor(self.h, self.d_k)) 238 239 # Setup adapter strategy 240 self.setup_adapter_strategy(adapter_strategy) 241 242 # reset parameters for Q to be identity operation 243 self.reset_parameters() 244 245 def forward(self, query, key, value, mask, pos_emb, cache=None): 246 """Compute 'Scaled Dot Product Attention' with rel. positional encoding. 247 Args: 248 query (torch.Tensor): (batch, time1, size) 249 key (torch.Tensor): (batch, time2, size) 250 value(torch.Tensor): (batch, time2, size) 251 mask (torch.Tensor): (batch, time1, time2) 252 pos_emb (torch.Tensor) : (batch, time1, size) 253 cache (torch.Tensor) : (batch, time_cache, size) 254 Returns: 255 output (torch.Tensor): transformed `value` (batch, time1, d_model) weighted by the query dot key attention 256 cache_next (torch.Tensor) : (batch, time_cache_next, size) 257 """ 258 # Need to perform duplicate computations as at this point the tensors have been 259 # separated by the adapter forward 260 query = self.pre_norm(query) 261 key = self.pre_norm(key) 262 value = self.pre_norm(value) 263 264 return super().forward(query, key, value, mask, pos_emb, cache=cache) 265 266 def reset_parameters(self): 267 with torch.no_grad(): 268 nn.init.zeros_(self.linear_out.weight) 269 nn.init.zeros_(self.linear_out.bias) 270 271 # NOTE: This exact procedure apparently highly important. 272 # Above operation is safe to do as self.linear_out.weight *= 0.0 (similar for bias) 273 # However: 274 # DO NOT REPLACE BELOW WITH self.pos_bias_u *= 0.0 OR self.pos_bias_v *= 0.0 275 # For some reason at init sometimes it will cause the value of the tensor to become NaN 276 # All operations to compute matrix_ac and matrix_bd will then fail. 277 nn.init.zeros_(self.pos_bias_u) 278 nn.init.zeros_(self.pos_bias_v) 279 280 def get_default_strategy_config(self) -> 'dataclass': 281 return MHAResidualAddAdapterStrategyConfig() 282 283 284 @dataclass 285 class RelPositionMultiHeadAttentionAdapterConfig: 286 n_head: int 287 n_feat: int 288 dropout_rate: float = 0.0 289 proj_dim: Optional[int] = None 290 adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() 291 _target_: str = "{0}.{1}".format( 292 RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ 293 ) 294 295 296 class PositionalEncodingAdapter(mha.PositionalEncoding, adapter_modules.AdapterModuleUtil): 297 298 """ 299 Absolute positional embedding adapter. 300 301 .. note:: 302 303 Absolute positional embedding value is added to the input tensor *without residual connection* ! 304 Therefore, the input is changed, if you only require the positional embedding, drop the returned `x` ! 305 306 Args: 307 d_model (int): The input dimension of x. 308 max_len (int): The max sequence length. 309 xscale (float): The input scaling factor. Defaults to 1.0. 310 adapter_strategy (AbstractAdapterStrategy): By default, ReturnResultAdapterStrategyConfig. 311 An adapter composition function object. 312 NOTE: Since this is a positional encoding, it will not add a residual ! 313 """ 314 315 def __init__( 316 self, 317 d_model: int, 318 max_len: int = 5000, 319 xscale=1.0, 320 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 321 ): 322 323 super().__init__( 324 d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0, 325 ) 326 327 # Setup adapter strategy 328 self.setup_adapter_strategy(adapter_strategy) 329 330 def get_default_strategy_config(self) -> 'dataclass': 331 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 332 333 334 @dataclass 335 class PositionalEncodingAdapterConfig: 336 d_model: int 337 max_len: int = 5000 338 xscale: float = 1.0 339 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 340 _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) 341 342 343 class RelPositionalEncodingAdapter(mha.RelPositionalEncoding, adapter_modules.AdapterModuleUtil): 344 """ 345 Relative positional encoding for TransformerXL's layers 346 See : Appendix B in https://arxiv.org/abs/1901.02860 347 348 .. note:: 349 350 Relative positional embedding value is **not** added to the input tensor ! 351 Therefore, the input should be updated changed, if you only require the positional embedding, drop the returned `x` ! 352 353 Args: 354 d_model (int): embedding dim 355 max_len (int): maximum input length 356 xscale (bool): whether to scale the input by sqrt(d_model) 357 adapter_strategy: By default, ReturnResultAdapterStrategyConfig. An adapter composition function object. 358 """ 359 360 def __init__( 361 self, 362 d_model: int, 363 max_len: int = 5000, 364 xscale=1.0, 365 adapter_strategy: adapter_mixin_strategies.ReturnResultAdapterStrategyConfig = None, 366 ): 367 super().__init__(d_model=d_model, dropout_rate=0.0, max_len=max_len, xscale=xscale, dropout_rate_emb=0.0) 368 369 # Setup adapter strategy 370 self.setup_adapter_strategy(adapter_strategy) 371 372 def get_default_strategy_config(self) -> 'dataclass': 373 return adapter_mixin_strategies.ReturnResultAdapterStrategyConfig() 374 375 376 @dataclass 377 class RelPositionalEncodingAdapterConfig: 378 d_model: int 379 max_len: int = 5000 380 xscale: float = 1.0 381 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 382 _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) 383 [end of nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py] [start of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 import os 17 from dataclasses import dataclass 18 from typing import List, Optional, Tuple, Union 19 20 import torch 21 22 from nemo.collections.asr.parts.utils import rnnt_utils 23 from nemo.collections.common.tokenizers.tokenizer_spec import TokenizerSpec 24 from nemo.core.classes import Typing, typecheck 25 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 26 from nemo.utils import logging 27 28 DEFAULT_TOKEN_OFFSET = 100 29 30 31 def pack_hypotheses( 32 hypotheses: List[rnnt_utils.NBestHypotheses], logitlen: torch.Tensor, 33 ) -> List[rnnt_utils.NBestHypotheses]: 34 35 if logitlen is not None: 36 if hasattr(logitlen, 'cpu'): 37 logitlen_cpu = logitlen.to('cpu') 38 else: 39 logitlen_cpu = logitlen 40 41 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.NBestHypotheses 42 for candidate_idx, cand in enumerate(hyp.n_best_hypotheses): 43 cand.y_sequence = torch.tensor(cand.y_sequence, dtype=torch.long) 44 45 if logitlen is not None: 46 cand.length = logitlen_cpu[idx] 47 48 if cand.dec_state is not None: 49 cand.dec_state = _states_to_device(cand.dec_state) 50 51 return hypotheses 52 53 54 def _states_to_device(dec_state, device='cpu'): 55 if torch.is_tensor(dec_state): 56 dec_state = dec_state.to(device) 57 58 elif isinstance(dec_state, (list, tuple)): 59 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 60 61 return dec_state 62 63 64 class AbstractBeamCTCInfer(Typing): 65 """A beam CTC decoder. 66 67 Provides a common abstraction for sample level beam decoding. 68 69 Args: 70 blank_id: int, index of the blank token. Can be 0 or len(vocabulary). 71 beam_size: int, size of the beam used in the underlying beam search engine. 72 73 """ 74 75 @property 76 def input_types(self): 77 """Returns definitions of module input ports. 78 """ 79 return { 80 "decoder_output": NeuralType(('B', 'T', 'D'), LogprobsType()), 81 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 82 } 83 84 @property 85 def output_types(self): 86 """Returns definitions of module output ports. 87 """ 88 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 89 90 def __init__(self, blank_id: int, beam_size: int): 91 self.blank_id = blank_id 92 93 if beam_size < 1: 94 raise ValueError("Beam search size cannot be less than 1!") 95 96 self.beam_size = beam_size 97 98 # Variables set by corresponding setter methods 99 self.vocab = None 100 self.decoding_type = None 101 self.tokenizer = None 102 103 # Utility maps for vocabulary 104 self.vocab_index_map = None 105 self.index_vocab_map = None 106 107 # Internal variable, used to prevent double reduction of consecutive tokens (ctc collapse) 108 self.override_fold_consecutive_value = None 109 110 def set_vocabulary(self, vocab: List[str]): 111 """ 112 Set the vocabulary of the decoding framework. 113 114 Args: 115 vocab: List of str. Each token corresponds to its location in the vocabulary emitted by the model. 116 Note that this vocabulary must NOT contain the "BLANK" token. 117 """ 118 self.vocab = vocab 119 self.vocab_index_map = {v: i for i, v in enumerate(vocab)} 120 self.index_vocab_map = {i: v for i, v in enumerate(vocab)} 121 122 def set_decoding_type(self, decoding_type: str): 123 """ 124 Sets the decoding type of the framework. Can support either char or subword models. 125 126 Args: 127 decoding_type: Str corresponding to decoding type. Only supports "char" and "subword". 128 """ 129 decoding_type = decoding_type.lower() 130 supported_types = ['char', 'subword'] 131 132 if decoding_type not in supported_types: 133 raise ValueError( 134 f"Unsupported decoding type. Supported types = {supported_types}.\n" f"Given = {decoding_type}" 135 ) 136 137 self.decoding_type = decoding_type 138 139 def set_tokenizer(self, tokenizer: TokenizerSpec): 140 """ 141 Set the tokenizer of the decoding framework. 142 143 Args: 144 tokenizer: NeMo tokenizer object, which inherits from TokenizerSpec. 145 """ 146 self.tokenizer = tokenizer 147 148 @typecheck() 149 def forward( 150 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 151 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 152 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 153 Output token is generated auto-repressively. 154 155 Args: 156 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 157 decoder_lengths: list of int representing the length of each sequence 158 output sequence. 159 160 Returns: 161 packed list containing batch number of sentences (Hypotheses). 162 """ 163 raise NotImplementedError() 164 165 def __call__(self, *args, **kwargs): 166 return self.forward(*args, **kwargs) 167 168 169 class BeamCTCInfer(AbstractBeamCTCInfer): 170 """A greedy CTC decoder. 171 172 Provides a common abstraction for sample level and batch level greedy decoding. 173 174 Args: 175 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 176 preserve_alignments: Bool flag which preserves the history of logprobs generated during 177 decoding (sample / batched). When set to true, the Hypothesis will contain 178 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 179 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 180 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 181 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 182 183 """ 184 185 def __init__( 186 self, 187 blank_id: int, 188 beam_size: int, 189 search_type: str = "default", 190 return_best_hypothesis: bool = True, 191 preserve_alignments: bool = False, 192 compute_timestamps: bool = False, 193 beam_alpha: float = 1.0, 194 beam_beta: float = 0.0, 195 kenlm_path: str = None, 196 flashlight_cfg: Optional['FlashlightConfig'] = None, 197 pyctcdecode_cfg: Optional['PyCTCDecodeConfig'] = None, 198 ): 199 super().__init__(blank_id=blank_id, beam_size=beam_size) 200 201 self.search_type = search_type 202 self.return_best_hypothesis = return_best_hypothesis 203 self.preserve_alignments = preserve_alignments 204 self.compute_timestamps = compute_timestamps 205 206 if self.compute_timestamps: 207 raise ValueError(f"Currently this flag is not supported for beam search algorithms.") 208 209 self.vocab = None # This must be set by specific method by user before calling forward() ! 210 211 if search_type == "default" or search_type == "nemo": 212 self.search_algorithm = self.default_beam_search 213 elif search_type == "pyctcdecode": 214 self.search_algorithm = self._pyctcdecode_beam_search 215 elif search_type == "flashlight": 216 self.search_algorithm = self.flashlight_beam_search 217 else: 218 raise NotImplementedError( 219 f"The search type ({search_type}) supplied is not supported!\n" 220 f"Please use one of : (default, nemo, pyctcdecode)" 221 ) 222 223 # Log the beam search algorithm 224 logging.info(f"Beam search algorithm: {search_type}") 225 226 self.beam_alpha = beam_alpha 227 self.beam_beta = beam_beta 228 229 # Default beam search args 230 self.kenlm_path = kenlm_path 231 232 # PyCTCDecode params 233 if pyctcdecode_cfg is None: 234 pyctcdecode_cfg = PyCTCDecodeConfig() 235 self.pyctcdecode_cfg = pyctcdecode_cfg # type: PyCTCDecodeConfig 236 237 if flashlight_cfg is None: 238 flashlight_cfg = FlashlightConfig() 239 self.flashlight_cfg = flashlight_cfg 240 241 # Default beam search scorer functions 242 self.default_beam_scorer = None 243 self.pyctcdecode_beam_scorer = None 244 self.flashlight_beam_scorer = None 245 self.token_offset = 0 246 247 @typecheck() 248 def forward( 249 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 250 ) -> Tuple[List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]]: 251 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 252 Output token is generated auto-repressively. 253 254 Args: 255 decoder_output: A tensor of size (batch, timesteps, features). 256 decoder_lengths: list of int representing the length of each sequence 257 output sequence. 258 259 Returns: 260 packed list containing batch number of sentences (Hypotheses). 261 """ 262 if self.vocab is None: 263 raise RuntimeError("Please set the vocabulary with `set_vocabulary()` before calling this function.") 264 265 if self.decoding_type is None: 266 raise ValueError("Please set the decoding type with `set_decoding_type()` before calling this function.") 267 268 with torch.no_grad(), torch.inference_mode(): 269 # Process each sequence independently 270 prediction_tensor = decoder_output 271 272 if prediction_tensor.ndim != 3: 273 raise ValueError( 274 f"`decoder_output` must be a tensor of shape [B, T, V] (log probs, float). " 275 f"Provided shape = {prediction_tensor.shape}" 276 ) 277 278 # determine type of input - logprobs or labels 279 out_len = decoder_lengths if decoder_lengths is not None else None 280 hypotheses = self.search_algorithm(prediction_tensor, out_len) 281 282 # Pack results into Hypotheses 283 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 284 285 # Pack the result 286 if self.return_best_hypothesis and isinstance(packed_result[0], rnnt_utils.NBestHypotheses): 287 packed_result = [res.n_best_hypotheses[0] for res in packed_result] # type: Hypothesis 288 289 return (packed_result,) 290 291 @torch.no_grad() 292 def default_beam_search( 293 self, x: torch.Tensor, out_len: torch.Tensor 294 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 295 """ 296 Open Seq2Seq Beam Search Algorithm (DeepSpeed) 297 298 Args: 299 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 300 and V is the vocabulary size. The tensor contains log-probabilities. 301 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 302 303 Returns: 304 A list of NBestHypotheses objects, one for each sequence in the batch. 305 """ 306 if self.compute_timestamps: 307 raise ValueError( 308 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 309 ) 310 311 if self.default_beam_scorer is None: 312 # Check for filepath 313 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 314 raise FileNotFoundError( 315 f"KenLM binary file not found at : {self.kenlm_path}. " 316 f"Please set a valid path in the decoding config." 317 ) 318 319 # perform token offset for subword models 320 if self.decoding_type == 'subword': 321 vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 322 else: 323 # char models 324 vocab = self.vocab 325 326 # Must import at runtime to avoid circular dependency due to module level import. 327 from nemo.collections.asr.modules.beam_search_decoder import BeamSearchDecoderWithLM 328 329 self.default_beam_scorer = BeamSearchDecoderWithLM( 330 vocab=vocab, 331 lm_path=self.kenlm_path, 332 beam_width=self.beam_size, 333 alpha=self.beam_alpha, 334 beta=self.beam_beta, 335 num_cpus=max(1, os.cpu_count()), 336 input_tensor=False, 337 ) 338 339 x = x.to('cpu') 340 341 with typecheck.disable_checks(): 342 data = [x[sample_id, : out_len[sample_id], :].softmax(dim=-1) for sample_id in range(len(x))] 343 beams_batch = self.default_beam_scorer.forward(log_probs=data, log_probs_length=None) 344 345 # For each sample in the batch 346 nbest_hypotheses = [] 347 for beams_idx, beams in enumerate(beams_batch): 348 # For each beam candidate / hypothesis in each sample 349 hypotheses = [] 350 for candidate_idx, candidate in enumerate(beams): 351 hypothesis = rnnt_utils.Hypothesis( 352 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 353 ) 354 355 # For subword encoding, NeMo will double encode the subword (multiple tokens) into a 356 # singular unicode id. In doing so, we preserve the semantic of the unicode token, and 357 # compress the size of the final KenLM ARPA / Binary file. 358 # In order to do double encoding, we shift the subword by some token offset. 359 # This step is ignored for character based models. 360 if self.decoding_type == 'subword': 361 pred_token_ids = [ord(c) - self.token_offset for c in candidate[1]] 362 else: 363 # Char models 364 pred_token_ids = [self.vocab_index_map[c] for c in candidate[1]] 365 366 # We preserve the token ids and the score for this hypothesis 367 hypothesis.y_sequence = pred_token_ids 368 hypothesis.score = candidate[0] 369 370 # If alignment must be preserved, we preserve a view of the output logprobs. 371 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 372 # require specific processing for each sample in the beam. 373 # This is done to preserve memory. 374 if self.preserve_alignments: 375 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 376 377 hypotheses.append(hypothesis) 378 379 # Wrap the result in NBestHypothesis. 380 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 381 nbest_hypotheses.append(hypotheses) 382 383 return nbest_hypotheses 384 385 @torch.no_grad() 386 def _pyctcdecode_beam_search( 387 self, x: torch.Tensor, out_len: torch.Tensor 388 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 389 """ 390 PyCTCDecode Beam Search Algorithm. Should support Char and Subword models. 391 392 Args: 393 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 394 and V is the vocabulary size. The tensor contains log-probabilities. 395 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 396 397 Returns: 398 A list of NBestHypotheses objects, one for each sequence in the batch. 399 """ 400 if self.compute_timestamps: 401 raise ValueError( 402 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 403 ) 404 405 try: 406 import pyctcdecode 407 except (ImportError, ModuleNotFoundError): 408 raise ImportError( 409 f"Could not load `pyctcdecode` library. Please install it from pip using :\n" 410 f"pip install --upgrade pyctcdecode" 411 ) 412 413 if self.pyctcdecode_beam_scorer is None: 414 self.pyctcdecode_beam_scorer = pyctcdecode.build_ctcdecoder( 415 labels=self.vocab, kenlm_model_path=self.kenlm_path, alpha=self.beam_alpha, beta=self.beam_beta 416 ) # type: pyctcdecode.BeamSearchDecoderCTC 417 418 x = x.to('cpu').numpy() 419 420 with typecheck.disable_checks(): 421 beams_batch = [] 422 for sample_id in range(len(x)): 423 logprobs = x[sample_id, : out_len[sample_id], :] 424 result = self.pyctcdecode_beam_scorer.decode_beams( 425 logprobs, 426 beam_width=self.beam_size, 427 beam_prune_logp=self.pyctcdecode_cfg.beam_prune_logp, 428 token_min_logp=self.pyctcdecode_cfg.token_min_logp, 429 prune_history=self.pyctcdecode_cfg.prune_history, 430 hotwords=self.pyctcdecode_cfg.hotwords, 431 hotword_weight=self.pyctcdecode_cfg.hotword_weight, 432 lm_start_state=None, 433 ) # Output format: text, last_lm_state, text_frames, logit_score, lm_score 434 beams_batch.append(result) 435 436 nbest_hypotheses = [] 437 for beams_idx, beams in enumerate(beams_batch): 438 hypotheses = [] 439 for candidate_idx, candidate in enumerate(beams): 440 # Candidate = (text, last_lm_state, text_frames, logit_score, lm_score) 441 hypothesis = rnnt_utils.Hypothesis( 442 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 443 ) 444 445 # TODO: Requires token ids to be returned rather than text. 446 if self.decoding_type == 'subword': 447 if self.tokenizer is None: 448 raise ValueError("Tokenizer must be provided for subword decoding. Use set_tokenizer().") 449 450 pred_token_ids = self.tokenizer.text_to_ids(candidate[0]) 451 else: 452 if self.vocab is None: 453 raise ValueError("Vocab must be provided for character decoding. Use set_vocab().") 454 455 chars = list(candidate[0]) 456 pred_token_ids = [self.vocab_index_map[c] for c in chars] 457 458 hypothesis.y_sequence = pred_token_ids 459 hypothesis.text = candidate[0] # text 460 hypothesis.score = candidate[4] # score 461 462 # Inject word level timestamps 463 hypothesis.timestep = candidate[2] # text_frames 464 465 if self.preserve_alignments: 466 hypothesis.alignments = torch.from_numpy(x[beams_idx][: out_len[beams_idx]]) 467 468 hypotheses.append(hypothesis) 469 470 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 471 nbest_hypotheses.append(hypotheses) 472 473 return nbest_hypotheses 474 475 @torch.no_grad() 476 def flashlight_beam_search( 477 self, x: torch.Tensor, out_len: torch.Tensor 478 ) -> List[Union[rnnt_utils.Hypothesis, rnnt_utils.NBestHypotheses]]: 479 """ 480 Flashlight Beam Search Algorithm. Should support Char and Subword models. 481 482 Args: 483 x: Tensor of shape [B, T, V+1], where B is the batch size, T is the maximum sequence length, 484 and V is the vocabulary size. The tensor contains log-probabilities. 485 out_len: Tensor of shape [B], contains lengths of each sequence in the batch. 486 487 Returns: 488 A list of NBestHypotheses objects, one for each sequence in the batch. 489 """ 490 if self.compute_timestamps: 491 raise ValueError( 492 f"Beam Search with strategy `{self.search_type}` does not support time stamp calculation!" 493 ) 494 495 if self.flashlight_beam_scorer is None: 496 # Check for filepath 497 if self.kenlm_path is None or not os.path.exists(self.kenlm_path): 498 raise FileNotFoundError( 499 f"KenLM binary file not found at : {self.kenlm_path}. " 500 f"Please set a valid path in the decoding config." 501 ) 502 503 # perform token offset for subword models 504 # if self.decoding_type == 'subword': 505 # vocab = [chr(idx + self.token_offset) for idx in range(len(self.vocab))] 506 # else: 507 # # char models 508 # vocab = self.vocab 509 510 # Must import at runtime to avoid circular dependency due to module level import. 511 from nemo.collections.asr.modules.flashlight_decoder import FlashLightKenLMBeamSearchDecoder 512 513 self.flashlight_beam_scorer = FlashLightKenLMBeamSearchDecoder( 514 lm_path=self.kenlm_path, 515 vocabulary=self.vocab, 516 tokenizer=self.tokenizer, 517 lexicon_path=self.flashlight_cfg.lexicon_path, 518 boost_path=self.flashlight_cfg.boost_path, 519 beam_size=self.beam_size, 520 beam_size_token=self.flashlight_cfg.beam_size_token, 521 beam_threshold=self.flashlight_cfg.beam_threshold, 522 lm_weight=self.beam_alpha, 523 word_score=self.beam_beta, 524 unk_weight=self.flashlight_cfg.unk_weight, 525 sil_weight=self.flashlight_cfg.sil_weight, 526 ) 527 528 x = x.to('cpu') 529 530 with typecheck.disable_checks(): 531 beams_batch = self.flashlight_beam_scorer.forward(log_probs=x) 532 533 # For each sample in the batch 534 nbest_hypotheses = [] 535 for beams_idx, beams in enumerate(beams_batch): 536 # For each beam candidate / hypothesis in each sample 537 hypotheses = [] 538 for candidate_idx, candidate in enumerate(beams): 539 hypothesis = rnnt_utils.Hypothesis( 540 score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None 541 ) 542 543 # We preserve the token ids and the score for this hypothesis 544 hypothesis.y_sequence = candidate['tokens'].tolist() 545 hypothesis.score = candidate['score'] 546 547 # If alignment must be preserved, we preserve a view of the output logprobs. 548 # Note this view is shared amongst all beams within the sample, be sure to clone it if you 549 # require specific processing for each sample in the beam. 550 # This is done to preserve memory. 551 if self.preserve_alignments: 552 hypothesis.alignments = x[beams_idx][: out_len[beams_idx]] 553 554 hypotheses.append(hypothesis) 555 556 # Wrap the result in NBestHypothesis. 557 hypotheses = rnnt_utils.NBestHypotheses(hypotheses) 558 nbest_hypotheses.append(hypotheses) 559 560 return nbest_hypotheses 561 562 def set_decoding_type(self, decoding_type: str): 563 super().set_decoding_type(decoding_type) 564 565 # Please check train_kenlm.py in scripts/asr_language_modeling/ to find out why we need 566 # TOKEN_OFFSET for BPE-based models 567 if self.decoding_type == 'subword': 568 self.token_offset = DEFAULT_TOKEN_OFFSET 569 570 571 @dataclass 572 class PyCTCDecodeConfig: 573 # These arguments cannot be imported from pyctcdecode (optional dependency) 574 # Therefore we copy the values explicitly 575 # Taken from pyctcdecode.constant 576 beam_prune_logp: float = -10.0 577 token_min_logp: float = -5.0 578 prune_history: bool = False 579 hotwords: Optional[List[str]] = None 580 hotword_weight: float = 10.0 581 582 583 @dataclass 584 class FlashlightConfig: 585 lexicon_path: Optional[str] = None 586 boost_path: Optional[str] = None 587 beam_size_token: int = 16 588 beam_threshold: float = 20.0 589 unk_weight: float = -math.inf 590 sil_weight: float = 0.0 591 592 593 @dataclass 594 class BeamCTCInferConfig: 595 beam_size: int 596 search_type: str = 'default' 597 preserve_alignments: bool = False 598 compute_timestamps: bool = False 599 return_best_hypothesis: bool = True 600 601 beam_alpha: float = 1.0 602 beam_beta: float = 0.0 603 kenlm_path: Optional[str] = None 604 605 flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() 606 pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() 607 [end of nemo/collections/asr/parts/submodules/ctc_beam_decoding.py] [start of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import List, Optional 17 18 import torch 19 from omegaconf import DictConfig, OmegaConf 20 21 from nemo.collections.asr.parts.utils import rnnt_utils 22 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMethodConfig, ConfidenceMethodMixin 23 from nemo.core.classes import Typing, typecheck 24 from nemo.core.neural_types import HypothesisType, LengthsType, LogprobsType, NeuralType 25 from nemo.utils import logging 26 27 28 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 29 30 if logitlen is not None: 31 if hasattr(logitlen, 'cpu'): 32 logitlen_cpu = logitlen.to('cpu') 33 else: 34 logitlen_cpu = logitlen 35 36 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 37 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 38 39 if logitlen is not None: 40 hyp.length = logitlen_cpu[idx] 41 42 if hyp.dec_state is not None: 43 hyp.dec_state = _states_to_device(hyp.dec_state) 44 45 return hypotheses 46 47 48 def _states_to_device(dec_state, device='cpu'): 49 if torch.is_tensor(dec_state): 50 dec_state = dec_state.to(device) 51 52 elif isinstance(dec_state, (list, tuple)): 53 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 54 55 return dec_state 56 57 58 class GreedyCTCInfer(Typing, ConfidenceMethodMixin): 59 """A greedy CTC decoder. 60 61 Provides a common abstraction for sample level and batch level greedy decoding. 62 63 Args: 64 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 65 preserve_alignments: Bool flag which preserves the history of logprobs generated during 66 decoding (sample / batched). When set to true, the Hypothesis will contain 67 the non-null value for `logprobs` in it. Here, `logprobs` is a torch.Tensors. 68 compute_timestamps: A bool flag, which determines whether to compute the character/subword, or 69 word based timestamp mapping the output log-probabilities to discrite intervals of timestamps. 70 The timestamps will be available in the returned Hypothesis.timestep as a dictionary. 71 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 72 generated during decoding. When set to true, the Hypothesis will contain 73 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 74 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 75 confidence scores. 76 77 name: The method name (str). 78 Supported values: 79 - 'max_prob' for using the maximum token probability as a confidence. 80 - 'entropy' for using a normalized entropy of a log-likelihood vector. 81 82 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 83 Supported values: 84 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 85 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 86 Note that for this entropy, the alpha should comply the following inequality: 87 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 88 where V is the model vocabulary size. 89 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 90 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 91 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 92 More: https://en.wikipedia.org/wiki/Tsallis_entropy 93 - 'renyi' for the Rรฉnyi entropy. 94 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 95 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 96 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 97 98 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 99 When the alpha equals one, scaling is not applied to 'max_prob', 100 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 101 102 entropy_norm: A mapping of the entropy value to the interval [0,1]. 103 Supported values: 104 - 'lin' for using the linear mapping. 105 - 'exp' for using exponential mapping with linear shift. 106 107 """ 108 109 @property 110 def input_types(self): 111 """Returns definitions of module input ports. 112 """ 113 # Input can be of dimention - 114 # ('B', 'T', 'D') [Log probs] or ('B', 'T') [Labels] 115 116 return { 117 "decoder_output": NeuralType(None, LogprobsType()), 118 "decoder_lengths": NeuralType(tuple('B'), LengthsType()), 119 } 120 121 @property 122 def output_types(self): 123 """Returns definitions of module output ports. 124 """ 125 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 126 127 def __init__( 128 self, 129 blank_id: int, 130 preserve_alignments: bool = False, 131 compute_timestamps: bool = False, 132 preserve_frame_confidence: bool = False, 133 confidence_method_cfg: Optional[DictConfig] = None, 134 ): 135 super().__init__() 136 137 self.blank_id = blank_id 138 self.preserve_alignments = preserve_alignments 139 # we need timestamps to extract non-blank per-frame confidence 140 self.compute_timestamps = compute_timestamps | preserve_frame_confidence 141 self.preserve_frame_confidence = preserve_frame_confidence 142 143 # set confidence calculation method 144 self._init_confidence_method(confidence_method_cfg) 145 146 @typecheck() 147 def forward( 148 self, decoder_output: torch.Tensor, decoder_lengths: torch.Tensor, 149 ): 150 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 151 Output token is generated auto-repressively. 152 153 Args: 154 decoder_output: A tensor of size (batch, timesteps, features) or (batch, timesteps) (each timestep is a label). 155 decoder_lengths: list of int representing the length of each sequence 156 output sequence. 157 158 Returns: 159 packed list containing batch number of sentences (Hypotheses). 160 """ 161 with torch.inference_mode(): 162 hypotheses = [] 163 # Process each sequence independently 164 prediction_cpu_tensor = decoder_output.cpu() 165 166 if prediction_cpu_tensor.ndim < 2 or prediction_cpu_tensor.ndim > 3: 167 raise ValueError( 168 f"`decoder_output` must be a tensor of shape [B, T] (labels, int) or " 169 f"[B, T, V] (log probs, float). Provided shape = {prediction_cpu_tensor.shape}" 170 ) 171 172 # determine type of input - logprobs or labels 173 if prediction_cpu_tensor.ndim == 2: # labels 174 greedy_decode = self._greedy_decode_labels 175 else: 176 greedy_decode = self._greedy_decode_logprobs 177 178 for ind in range(prediction_cpu_tensor.shape[0]): 179 out_len = decoder_lengths[ind] if decoder_lengths is not None else None 180 hypothesis = greedy_decode(prediction_cpu_tensor[ind], out_len) 181 hypotheses.append(hypothesis) 182 183 # Pack results into Hypotheses 184 packed_result = pack_hypotheses(hypotheses, decoder_lengths) 185 186 return (packed_result,) 187 188 @torch.no_grad() 189 def _greedy_decode_logprobs(self, x: torch.Tensor, out_len: torch.Tensor): 190 # x: [T, D] 191 # out_len: [seq_len] 192 193 # Initialize blank state and empty label set in Hypothesis 194 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 195 prediction = x.detach().cpu() 196 197 if out_len is not None: 198 prediction = prediction[:out_len] 199 200 prediction_logprobs, prediction_labels = prediction.max(dim=-1) 201 202 non_blank_ids = prediction_labels != self.blank_id 203 hypothesis.y_sequence = prediction_labels.numpy().tolist() 204 hypothesis.score = (prediction_logprobs[non_blank_ids]).sum() 205 206 if self.preserve_alignments: 207 # Preserve the logprobs, as well as labels after argmax 208 hypothesis.alignments = (prediction.clone(), prediction_labels.clone()) 209 210 if self.compute_timestamps: 211 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 212 213 if self.preserve_frame_confidence: 214 hypothesis.frame_confidence = self._get_confidence(prediction) 215 216 return hypothesis 217 218 @torch.no_grad() 219 def _greedy_decode_labels(self, x: torch.Tensor, out_len: torch.Tensor): 220 # x: [T] 221 # out_len: [seq_len] 222 223 # Initialize blank state and empty label set in Hypothesis 224 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 225 prediction_labels = x.detach().cpu() 226 227 if out_len is not None: 228 prediction_labels = prediction_labels[:out_len] 229 230 non_blank_ids = prediction_labels != self.blank_id 231 hypothesis.y_sequence = prediction_labels.numpy().tolist() 232 hypothesis.score = -1.0 233 234 if self.preserve_alignments: 235 raise ValueError("Requested for alignments, but predictions provided were labels, not log probabilities.") 236 237 if self.compute_timestamps: 238 hypothesis.timestep = torch.nonzero(non_blank_ids, as_tuple=False)[:, 0].numpy().tolist() 239 240 if self.preserve_frame_confidence: 241 raise ValueError( 242 "Requested for per-frame confidence, but predictions provided were labels, not log probabilities." 243 ) 244 245 return hypothesis 246 247 def __call__(self, *args, **kwargs): 248 return self.forward(*args, **kwargs) 249 250 251 @dataclass 252 class GreedyCTCInferConfig: 253 preserve_alignments: bool = False 254 compute_timestamps: bool = False 255 preserve_frame_confidence: bool = False 256 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 257 258 def __post_init__(self): 259 # OmegaConf.structured ensures that post_init check is always executed 260 self.confidence_method_cfg = OmegaConf.structured( 261 self.confidence_method_cfg 262 if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) 263 else ConfidenceMethodConfig(**self.confidence_method_cfg) 264 ) 265 [end of nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py] [start of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 # Copyright 2017 Johns Hopkins University (Shinji Watanabe) 16 # 17 # Licensed under the Apache License, Version 2.0 (the "License"); 18 # you may not use this file except in compliance with the License. 19 # You may obtain a copy of the License at 20 # 21 # http://www.apache.org/licenses/LICENSE-2.0 22 # 23 # Unless required by applicable law or agreed to in writing, software 24 # distributed under the License is distributed on an "AS IS" BASIS, 25 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 26 # See the License for the specific language governing permissions and 27 # limitations under the License. 28 29 from dataclasses import dataclass 30 from typing import List, Optional, Tuple, Union 31 32 import numpy as np 33 import torch 34 from omegaconf import DictConfig, OmegaConf 35 36 from nemo.collections.asr.modules import rnnt_abstract 37 from nemo.collections.asr.parts.utils import rnnt_utils 38 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceMethodConfig, ConfidenceMethodMixin 39 from nemo.collections.common.parts.rnn import label_collate 40 from nemo.core.classes import Typing, typecheck 41 from nemo.core.neural_types import AcousticEncodedRepresentation, ElementType, HypothesisType, LengthsType, NeuralType 42 from nemo.utils import logging 43 44 45 def pack_hypotheses(hypotheses: List[rnnt_utils.Hypothesis], logitlen: torch.Tensor,) -> List[rnnt_utils.Hypothesis]: 46 47 if hasattr(logitlen, 'cpu'): 48 logitlen_cpu = logitlen.to('cpu') 49 else: 50 logitlen_cpu = logitlen 51 52 for idx, hyp in enumerate(hypotheses): # type: rnnt_utils.Hypothesis 53 hyp.y_sequence = torch.tensor(hyp.y_sequence, dtype=torch.long) 54 hyp.length = logitlen_cpu[idx] 55 56 if hyp.dec_state is not None: 57 hyp.dec_state = _states_to_device(hyp.dec_state) 58 59 return hypotheses 60 61 62 def _states_to_device(dec_state, device='cpu'): 63 if torch.is_tensor(dec_state): 64 dec_state = dec_state.to(device) 65 66 elif isinstance(dec_state, (list, tuple)): 67 dec_state = tuple(_states_to_device(dec_i, device) for dec_i in dec_state) 68 69 return dec_state 70 71 72 class _GreedyRNNTInfer(Typing, ConfidenceMethodMixin): 73 """A greedy transducer decoder. 74 75 Provides a common abstraction for sample level and batch level greedy decoding. 76 77 Args: 78 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 79 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 80 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 81 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 82 to a sequence in a single time step; if set to None then there is 83 no limit. 84 preserve_alignments: Bool flag which preserves the history of alignments generated during 85 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 86 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 87 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 88 89 The length of the list corresponds to the Acoustic Length (T). 90 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 91 U is the number of target tokens for the current timestep Ti. 92 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 93 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 94 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 95 96 The length of the list corresponds to the Acoustic Length (T). 97 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 98 U is the number of target tokens for the current timestep Ti. 99 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 100 confidence scores. 101 102 name: The method name (str). 103 Supported values: 104 - 'max_prob' for using the maximum token probability as a confidence. 105 - 'entropy' for using a normalized entropy of a log-likelihood vector. 106 107 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 108 Supported values: 109 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 110 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 111 Note that for this entropy, the alpha should comply the following inequality: 112 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 113 where V is the model vocabulary size. 114 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 115 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 116 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 117 More: https://en.wikipedia.org/wiki/Tsallis_entropy 118 - 'renyi' for the Rรฉnyi entropy. 119 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 120 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 121 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 122 123 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 124 When the alpha equals one, scaling is not applied to 'max_prob', 125 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 126 127 entropy_norm: A mapping of the entropy value to the interval [0,1]. 128 Supported values: 129 - 'lin' for using the linear mapping. 130 - 'exp' for using exponential mapping with linear shift. 131 """ 132 133 @property 134 def input_types(self): 135 """Returns definitions of module input ports. 136 """ 137 return { 138 "encoder_output": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()), 139 "encoded_lengths": NeuralType(tuple('B'), LengthsType()), 140 "partial_hypotheses": [NeuralType(elements_type=HypothesisType(), optional=True)], # must always be last 141 } 142 143 @property 144 def output_types(self): 145 """Returns definitions of module output ports. 146 """ 147 return {"predictions": [NeuralType(elements_type=HypothesisType())]} 148 149 def __init__( 150 self, 151 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 152 joint_model: rnnt_abstract.AbstractRNNTJoint, 153 blank_index: int, 154 max_symbols_per_step: Optional[int] = None, 155 preserve_alignments: bool = False, 156 preserve_frame_confidence: bool = False, 157 confidence_method_cfg: Optional[DictConfig] = None, 158 ): 159 super().__init__() 160 self.decoder = decoder_model 161 self.joint = joint_model 162 163 self._blank_index = blank_index 164 self._SOS = blank_index # Start of single index 165 self.max_symbols = max_symbols_per_step 166 self.preserve_alignments = preserve_alignments 167 self.preserve_frame_confidence = preserve_frame_confidence 168 169 # set confidence calculation method 170 self._init_confidence_method(confidence_method_cfg) 171 172 def __call__(self, *args, **kwargs): 173 return self.forward(*args, **kwargs) 174 175 @torch.no_grad() 176 def _pred_step( 177 self, 178 label: Union[torch.Tensor, int], 179 hidden: Optional[torch.Tensor], 180 add_sos: bool = False, 181 batch_size: Optional[int] = None, 182 ) -> Tuple[torch.Tensor, torch.Tensor]: 183 """ 184 Common prediction step based on the AbstractRNNTDecoder implementation. 185 186 Args: 187 label: (int/torch.Tensor): Label or "Start-of-Signal" token. 188 hidden: (Optional torch.Tensor): RNN State vector 189 add_sos (bool): Whether to add a zero vector at the begging as "start of sentence" token. 190 batch_size: Batch size of the output tensor. 191 192 Returns: 193 g: (B, U, H) if add_sos is false, else (B, U + 1, H) 194 hid: (h, c) where h is the final sequence hidden state and c is 195 the final cell state: 196 h (tensor), shape (L, B, H) 197 c (tensor), shape (L, B, H) 198 """ 199 if isinstance(label, torch.Tensor): 200 # label: [batch, 1] 201 if label.dtype != torch.long: 202 label = label.long() 203 204 else: 205 # Label is an integer 206 if label == self._SOS: 207 return self.decoder.predict(None, hidden, add_sos=add_sos, batch_size=batch_size) 208 209 label = label_collate([[label]]) 210 211 # output: [B, 1, K] 212 return self.decoder.predict(label, hidden, add_sos=add_sos, batch_size=batch_size) 213 214 def _joint_step(self, enc, pred, log_normalize: Optional[bool] = None): 215 """ 216 Common joint step based on AbstractRNNTJoint implementation. 217 218 Args: 219 enc: Output of the Encoder model. A torch.Tensor of shape [B, 1, H1] 220 pred: Output of the Decoder model. A torch.Tensor of shape [B, 1, H2] 221 log_normalize: Whether to log normalize or not. None will log normalize only for CPU. 222 223 Returns: 224 logits of shape (B, T=1, U=1, V + 1) 225 """ 226 with torch.no_grad(): 227 logits = self.joint.joint(enc, pred) 228 229 if log_normalize is None: 230 if not logits.is_cuda: # Use log softmax only if on CPU 231 logits = logits.log_softmax(dim=len(logits.shape) - 1) 232 else: 233 if log_normalize: 234 logits = logits.log_softmax(dim=len(logits.shape) - 1) 235 236 return logits 237 238 239 class GreedyRNNTInfer(_GreedyRNNTInfer): 240 """A greedy transducer decoder. 241 242 Sequence level greedy decoding, performed auto-regressively. 243 244 Args: 245 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 246 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 247 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 248 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 249 to a sequence in a single time step; if set to None then there is 250 no limit. 251 preserve_alignments: Bool flag which preserves the history of alignments generated during 252 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 253 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 254 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 255 256 The length of the list corresponds to the Acoustic Length (T). 257 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 258 U is the number of target tokens for the current timestep Ti. 259 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 260 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 261 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 262 263 The length of the list corresponds to the Acoustic Length (T). 264 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 265 U is the number of target tokens for the current timestep Ti. 266 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 267 confidence scores. 268 269 name: The method name (str). 270 Supported values: 271 - 'max_prob' for using the maximum token probability as a confidence. 272 - 'entropy' for using a normalized entropy of a log-likelihood vector. 273 274 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 275 Supported values: 276 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 277 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 278 Note that for this entropy, the alpha should comply the following inequality: 279 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 280 where V is the model vocabulary size. 281 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 282 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 283 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 284 More: https://en.wikipedia.org/wiki/Tsallis_entropy 285 - 'renyi' for the Rรฉnyi entropy. 286 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 287 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 288 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 289 290 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 291 When the alpha equals one, scaling is not applied to 'max_prob', 292 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 293 294 entropy_norm: A mapping of the entropy value to the interval [0,1]. 295 Supported values: 296 - 'lin' for using the linear mapping. 297 - 'exp' for using exponential mapping with linear shift. 298 """ 299 300 def __init__( 301 self, 302 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 303 joint_model: rnnt_abstract.AbstractRNNTJoint, 304 blank_index: int, 305 max_symbols_per_step: Optional[int] = None, 306 preserve_alignments: bool = False, 307 preserve_frame_confidence: bool = False, 308 confidence_method_cfg: Optional[DictConfig] = None, 309 ): 310 super().__init__( 311 decoder_model=decoder_model, 312 joint_model=joint_model, 313 blank_index=blank_index, 314 max_symbols_per_step=max_symbols_per_step, 315 preserve_alignments=preserve_alignments, 316 preserve_frame_confidence=preserve_frame_confidence, 317 confidence_method_cfg=confidence_method_cfg, 318 ) 319 320 @typecheck() 321 def forward( 322 self, 323 encoder_output: torch.Tensor, 324 encoded_lengths: torch.Tensor, 325 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 326 ): 327 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 328 Output token is generated auto-regressively. 329 330 Args: 331 encoder_output: A tensor of size (batch, features, timesteps). 332 encoded_lengths: list of int representing the length of each sequence 333 output sequence. 334 335 Returns: 336 packed list containing batch number of sentences (Hypotheses). 337 """ 338 # Preserve decoder and joint training state 339 decoder_training_state = self.decoder.training 340 joint_training_state = self.joint.training 341 342 with torch.inference_mode(): 343 # Apply optional preprocessing 344 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 345 346 self.decoder.eval() 347 self.joint.eval() 348 349 hypotheses = [] 350 # Process each sequence independently 351 with self.decoder.as_frozen(), self.joint.as_frozen(): 352 for batch_idx in range(encoder_output.size(0)): 353 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 354 logitlen = encoded_lengths[batch_idx] 355 356 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 357 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 358 hypotheses.append(hypothesis) 359 360 # Pack results into Hypotheses 361 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 362 363 self.decoder.train(decoder_training_state) 364 self.joint.train(joint_training_state) 365 366 return (packed_result,) 367 368 @torch.no_grad() 369 def _greedy_decode( 370 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 371 ): 372 # x: [T, 1, D] 373 # out_len: [seq_len] 374 375 # Initialize blank state and empty label set in Hypothesis 376 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 377 378 if partial_hypotheses is not None: 379 hypothesis.last_token = partial_hypotheses.last_token 380 hypothesis.y_sequence = ( 381 partial_hypotheses.y_sequence.cpu().tolist() 382 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 383 else partial_hypotheses.y_sequence 384 ) 385 if partial_hypotheses.dec_state is not None: 386 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 387 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 388 389 if self.preserve_alignments: 390 # Alignments is a 2-dimensional dangling list representing T x U 391 hypothesis.alignments = [[]] 392 393 if self.preserve_frame_confidence: 394 hypothesis.frame_confidence = [[]] 395 396 # For timestep t in X_t 397 for time_idx in range(out_len): 398 # Extract encoder embedding at timestep t 399 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 400 f = x.narrow(dim=0, start=time_idx, length=1) 401 402 # Setup exit flags and counter 403 not_blank = True 404 symbols_added = 0 405 # While blank is not predicted, or we dont run out of max symbols per timestep 406 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 407 # In the first timestep, we initialize the network with RNNT Blank 408 # In later timesteps, we provide previous predicted label as input. 409 if hypothesis.last_token is None and hypothesis.dec_state is None: 410 last_label = self._SOS 411 else: 412 last_label = label_collate([[hypothesis.last_token]]) 413 414 # Perform prediction network and joint network steps. 415 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 416 # If preserving per-frame confidence, log_normalize must be true 417 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 418 0, 0, 0, : 419 ] 420 421 del g 422 423 # torch.max(0) op doesnt exist for FP 16. 424 if logp.dtype != torch.float32: 425 logp = logp.float() 426 427 # get index k, of max prob 428 v, k = logp.max(0) 429 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 430 431 if self.preserve_alignments: 432 # insert logprobs into last timestep 433 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 434 435 if self.preserve_frame_confidence: 436 # insert confidence into last timestep 437 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 438 439 del logp 440 441 # If blank token is predicted, exit inner loop, move onto next timestep t 442 if k == self._blank_index: 443 not_blank = False 444 else: 445 # Append token to label set, update RNN state. 446 hypothesis.y_sequence.append(k) 447 hypothesis.score += float(v) 448 hypothesis.timestep.append(time_idx) 449 hypothesis.dec_state = hidden_prime 450 hypothesis.last_token = k 451 452 # Increment token counter. 453 symbols_added += 1 454 455 if self.preserve_alignments: 456 # convert Ti-th logits into a torch array 457 hypothesis.alignments.append([]) # blank buffer for next timestep 458 459 if self.preserve_frame_confidence: 460 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 461 462 # Remove trailing empty list of Alignments 463 if self.preserve_alignments: 464 if len(hypothesis.alignments[-1]) == 0: 465 del hypothesis.alignments[-1] 466 467 # Remove trailing empty list of per-frame confidence 468 if self.preserve_frame_confidence: 469 if len(hypothesis.frame_confidence[-1]) == 0: 470 del hypothesis.frame_confidence[-1] 471 472 # Unpack the hidden states 473 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 474 475 return hypothesis 476 477 478 class GreedyBatchedRNNTInfer(_GreedyRNNTInfer): 479 """A batch level greedy transducer decoder. 480 481 Batch level greedy decoding, performed auto-regressively. 482 483 Args: 484 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 485 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 486 blank_index: int index of the blank token. Can be 0 or len(vocabulary). 487 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 488 to a sequence in a single time step; if set to None then there is 489 no limit. 490 preserve_alignments: Bool flag which preserves the history of alignments generated during 491 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 492 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 493 Tuple(Tensor (of length V + 1), Tensor(scalar, label after argmax)). 494 495 The length of the list corresponds to the Acoustic Length (T). 496 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 497 U is the number of target tokens for the current timestep Ti. 498 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 499 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 500 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 501 502 The length of the list corresponds to the Acoustic Length (T). 503 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 504 U is the number of target tokens for the current timestep Ti. 505 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 506 confidence scores. 507 508 name: The method name (str). 509 Supported values: 510 - 'max_prob' for using the maximum token probability as a confidence. 511 - 'entropy' for using a normalized entropy of a log-likelihood vector. 512 513 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 514 Supported values: 515 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 516 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 517 Note that for this entropy, the alpha should comply the following inequality: 518 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 519 where V is the model vocabulary size. 520 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 521 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 522 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 523 More: https://en.wikipedia.org/wiki/Tsallis_entropy 524 - 'renyi' for the Rรฉnyi entropy. 525 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 526 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 527 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 528 529 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 530 When the alpha equals one, scaling is not applied to 'max_prob', 531 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 532 533 entropy_norm: A mapping of the entropy value to the interval [0,1]. 534 Supported values: 535 - 'lin' for using the linear mapping. 536 - 'exp' for using exponential mapping with linear shift. 537 """ 538 539 def __init__( 540 self, 541 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 542 joint_model: rnnt_abstract.AbstractRNNTJoint, 543 blank_index: int, 544 max_symbols_per_step: Optional[int] = None, 545 preserve_alignments: bool = False, 546 preserve_frame_confidence: bool = False, 547 confidence_method_cfg: Optional[DictConfig] = None, 548 ): 549 super().__init__( 550 decoder_model=decoder_model, 551 joint_model=joint_model, 552 blank_index=blank_index, 553 max_symbols_per_step=max_symbols_per_step, 554 preserve_alignments=preserve_alignments, 555 preserve_frame_confidence=preserve_frame_confidence, 556 confidence_method_cfg=confidence_method_cfg, 557 ) 558 559 # Depending on availability of `blank_as_pad` support 560 # switch between more efficient batch decoding technique 561 if self.decoder.blank_as_pad: 562 self._greedy_decode = self._greedy_decode_blank_as_pad 563 else: 564 self._greedy_decode = self._greedy_decode_masked 565 566 @typecheck() 567 def forward( 568 self, 569 encoder_output: torch.Tensor, 570 encoded_lengths: torch.Tensor, 571 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 572 ): 573 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 574 Output token is generated auto-regressively. 575 576 Args: 577 encoder_output: A tensor of size (batch, features, timesteps). 578 encoded_lengths: list of int representing the length of each sequence 579 output sequence. 580 581 Returns: 582 packed list containing batch number of sentences (Hypotheses). 583 """ 584 # Preserve decoder and joint training state 585 decoder_training_state = self.decoder.training 586 joint_training_state = self.joint.training 587 588 with torch.inference_mode(): 589 # Apply optional preprocessing 590 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 591 logitlen = encoded_lengths 592 593 self.decoder.eval() 594 self.joint.eval() 595 596 with self.decoder.as_frozen(), self.joint.as_frozen(): 597 inseq = encoder_output # [B, T, D] 598 hypotheses = self._greedy_decode( 599 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 600 ) 601 602 # Pack the hypotheses results 603 packed_result = pack_hypotheses(hypotheses, logitlen) 604 605 self.decoder.train(decoder_training_state) 606 self.joint.train(joint_training_state) 607 608 return (packed_result,) 609 610 def _greedy_decode_blank_as_pad( 611 self, 612 x: torch.Tensor, 613 out_len: torch.Tensor, 614 device: torch.device, 615 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 616 ): 617 if partial_hypotheses is not None: 618 raise NotImplementedError("`partial_hypotheses` support is not supported") 619 620 with torch.inference_mode(): 621 # x: [B, T, D] 622 # out_len: [B] 623 # device: torch.device 624 625 # Initialize list of Hypothesis 626 batchsize = x.shape[0] 627 hypotheses = [ 628 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 629 ] 630 631 # Initialize Hidden state matrix (shared by entire batch) 632 hidden = None 633 634 # If alignments need to be preserved, register a dangling list to hold the values 635 if self.preserve_alignments: 636 # alignments is a 3-dimensional dangling list representing B x T x U 637 for hyp in hypotheses: 638 hyp.alignments = [[]] 639 640 # If confidence scores need to be preserved, register a dangling list to hold the values 641 if self.preserve_frame_confidence: 642 # frame_confidence is a 3-dimensional dangling list representing B x T x U 643 for hyp in hypotheses: 644 hyp.frame_confidence = [[]] 645 646 # Last Label buffer + Last Label without blank buffer 647 # batch level equivalent of the last_label 648 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 649 650 # Mask buffers 651 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 652 653 # Get max sequence length 654 max_out_len = out_len.max() 655 for time_idx in range(max_out_len): 656 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 657 658 # Prepare t timestamp batch variables 659 not_blank = True 660 symbols_added = 0 661 662 # Reset blank mask 663 blank_mask.mul_(False) 664 665 # Update blank mask with time mask 666 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 667 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 668 blank_mask = time_idx >= out_len 669 # Start inner loop 670 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 671 # Batch prediction and joint network steps 672 # If very first prediction step, submit SOS tag (blank) to pred_step. 673 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 674 if time_idx == 0 and symbols_added == 0 and hidden is None: 675 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 676 else: 677 # Perform batch step prediction of decoder, getting new states and scores ("g") 678 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 679 680 # Batched joint step - Output = [B, V + 1] 681 # If preserving per-frame confidence, log_normalize must be true 682 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 683 :, 0, 0, : 684 ] 685 686 if logp.dtype != torch.float32: 687 logp = logp.float() 688 689 # Get index k, of max prob for batch 690 v, k = logp.max(1) 691 del g 692 693 # Update blank mask with current predicted blanks 694 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 695 k_is_blank = k == self._blank_index 696 blank_mask.bitwise_or_(k_is_blank) 697 all_blanks = torch.all(blank_mask) 698 699 del k_is_blank 700 701 # If preserving alignments, check if sequence length of sample has been reached 702 # before adding alignment 703 if self.preserve_alignments: 704 # Insert logprobs into last timestep per sample 705 logp_vals = logp.to('cpu') 706 logp_ids = logp_vals.max(1)[1] 707 for batch_idx, is_blank in enumerate(blank_mask): 708 # we only want to update non-blanks, unless we are at the last step in the loop where 709 # all elements produced blanks, otherwise there will be duplicate predictions 710 # saved in alignments 711 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 712 hypotheses[batch_idx].alignments[-1].append( 713 (logp_vals[batch_idx], logp_ids[batch_idx]) 714 ) 715 del logp_vals 716 717 # If preserving per-frame confidence, check if sequence length of sample has been reached 718 # before adding confidence scores 719 if self.preserve_frame_confidence: 720 # Insert probabilities into last timestep per sample 721 confidence = self._get_confidence(logp) 722 for batch_idx, is_blank in enumerate(blank_mask): 723 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 724 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 725 del logp 726 727 # If all samples predict / have predicted prior blanks, exit loop early 728 # This is equivalent to if single sample predicted k 729 if all_blanks: 730 not_blank = False 731 else: 732 # Collect batch indices where blanks occurred now/past 733 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 734 735 # Recover prior state for all samples which predicted blank now/past 736 if hidden is not None: 737 # LSTM has 2 states 738 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 739 740 elif len(blank_indices) > 0 and hidden is None: 741 # Reset state if there were some blank and other non-blank predictions in batch 742 # Original state is filled with zeros so we just multiply 743 # LSTM has 2 states 744 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 745 746 # Recover prior predicted label for all samples which predicted blank now/past 747 k[blank_indices] = last_label[blank_indices, 0] 748 749 # Update new label and hidden state for next iteration 750 last_label = k.clone().view(-1, 1) 751 hidden = hidden_prime 752 753 # Update predicted labels, accounting for time mask 754 # If blank was predicted even once, now or in the past, 755 # Force the current predicted label to also be blank 756 # This ensures that blanks propogate across all timesteps 757 # once they have occured (normally stopping condition of sample level loop). 758 for kidx, ki in enumerate(k): 759 if blank_mask[kidx] == 0: 760 hypotheses[kidx].y_sequence.append(ki) 761 hypotheses[kidx].timestep.append(time_idx) 762 hypotheses[kidx].score += float(v[kidx]) 763 symbols_added += 1 764 765 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 766 # Then preserve U at current timestep Ti 767 # Finally, forward the timestep history to Ti+1 for that sample 768 # All of this should only be done iff the current time index <= sample-level AM length. 769 # Otherwise ignore and move to next sample / next timestep. 770 if self.preserve_alignments: 771 772 # convert Ti-th logits into a torch array 773 for batch_idx in range(batchsize): 774 775 # this checks if current timestep <= sample-level AM length 776 # If current timestep > sample-level AM length, no alignments will be added 777 # Therefore the list of Uj alignments is empty here. 778 if len(hypotheses[batch_idx].alignments[-1]) > 0: 779 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 780 781 # Do the same if preserving per-frame confidence 782 if self.preserve_frame_confidence: 783 784 for batch_idx in range(batchsize): 785 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 786 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 787 788 # Remove trailing empty list of alignments at T_{am-len} x Uj 789 if self.preserve_alignments: 790 for batch_idx in range(batchsize): 791 if len(hypotheses[batch_idx].alignments[-1]) == 0: 792 del hypotheses[batch_idx].alignments[-1] 793 794 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 795 if self.preserve_frame_confidence: 796 for batch_idx in range(batchsize): 797 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 798 del hypotheses[batch_idx].frame_confidence[-1] 799 800 # Preserve states 801 for batch_idx in range(batchsize): 802 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 803 804 return hypotheses 805 806 def _greedy_decode_masked( 807 self, 808 x: torch.Tensor, 809 out_len: torch.Tensor, 810 device: torch.device, 811 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 812 ): 813 if partial_hypotheses is not None: 814 raise NotImplementedError("`partial_hypotheses` support is not supported") 815 816 # x: [B, T, D] 817 # out_len: [B] 818 # device: torch.device 819 820 # Initialize state 821 batchsize = x.shape[0] 822 hypotheses = [ 823 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 824 ] 825 826 # Initialize Hidden state matrix (shared by entire batch) 827 hidden = None 828 829 # If alignments need to be preserved, register a danling list to hold the values 830 if self.preserve_alignments: 831 # alignments is a 3-dimensional dangling list representing B x T x U 832 for hyp in hypotheses: 833 hyp.alignments = [[]] 834 else: 835 alignments = None 836 837 # If confidence scores need to be preserved, register a danling list to hold the values 838 if self.preserve_frame_confidence: 839 # frame_confidence is a 3-dimensional dangling list representing B x T x U 840 for hyp in hypotheses: 841 hyp.frame_confidence = [[]] 842 843 # Last Label buffer + Last Label without blank buffer 844 # batch level equivalent of the last_label 845 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 846 last_label_without_blank = last_label.clone() 847 848 # Mask buffers 849 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 850 851 # Get max sequence length 852 max_out_len = out_len.max() 853 854 with torch.inference_mode(): 855 for time_idx in range(max_out_len): 856 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 857 858 # Prepare t timestamp batch variables 859 not_blank = True 860 symbols_added = 0 861 862 # Reset blank mask 863 blank_mask.mul_(False) 864 865 # Update blank mask with time mask 866 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 867 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 868 blank_mask = time_idx >= out_len 869 870 # Start inner loop 871 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 872 # Batch prediction and joint network steps 873 # If very first prediction step, submit SOS tag (blank) to pred_step. 874 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 875 if time_idx == 0 and symbols_added == 0 and hidden is None: 876 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 877 else: 878 # Set a dummy label for the blank value 879 # This value will be overwritten by "blank" again the last label update below 880 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 881 last_label_without_blank_mask = last_label == self._blank_index 882 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 883 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 884 ~last_label_without_blank_mask 885 ] 886 887 # Perform batch step prediction of decoder, getting new states and scores ("g") 888 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 889 890 # Batched joint step - Output = [B, V + 1] 891 # If preserving per-frame confidence, log_normalize must be true 892 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 893 :, 0, 0, : 894 ] 895 896 if logp.dtype != torch.float32: 897 logp = logp.float() 898 899 # Get index k, of max prob for batch 900 v, k = logp.max(1) 901 del g 902 903 # Update blank mask with current predicted blanks 904 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 905 k_is_blank = k == self._blank_index 906 blank_mask.bitwise_or_(k_is_blank) 907 all_blanks = torch.all(blank_mask) 908 909 # If preserving alignments, check if sequence length of sample has been reached 910 # before adding alignment 911 if self.preserve_alignments: 912 # Insert logprobs into last timestep per sample 913 logp_vals = logp.to('cpu') 914 logp_ids = logp_vals.max(1)[1] 915 for batch_idx, is_blank in enumerate(blank_mask): 916 # we only want to update non-blanks, unless we are at the last step in the loop where 917 # all elements produced blanks, otherwise there will be duplicate predictions 918 # saved in alignments 919 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 920 hypotheses[batch_idx].alignments[-1].append( 921 (logp_vals[batch_idx], logp_ids[batch_idx]) 922 ) 923 924 del logp_vals 925 926 # If preserving per-frame confidence, check if sequence length of sample has been reached 927 # before adding confidence scores 928 if self.preserve_frame_confidence: 929 # Insert probabilities into last timestep per sample 930 confidence = self._get_confidence(logp) 931 for batch_idx, is_blank in enumerate(blank_mask): 932 if time_idx < out_len[batch_idx] and (all_blanks or not is_blank): 933 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 934 del logp 935 936 # If all samples predict / have predicted prior blanks, exit loop early 937 # This is equivalent to if single sample predicted k 938 if blank_mask.all(): 939 not_blank = False 940 else: 941 # Collect batch indices where blanks occurred now/past 942 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 943 944 # Recover prior state for all samples which predicted blank now/past 945 if hidden is not None: 946 # LSTM has 2 states 947 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 948 949 elif len(blank_indices) > 0 and hidden is None: 950 # Reset state if there were some blank and other non-blank predictions in batch 951 # Original state is filled with zeros so we just multiply 952 # LSTM has 2 states 953 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 954 955 # Recover prior predicted label for all samples which predicted blank now/past 956 k[blank_indices] = last_label[blank_indices, 0] 957 958 # Update new label and hidden state for next iteration 959 last_label = k.view(-1, 1) 960 hidden = hidden_prime 961 962 # Update predicted labels, accounting for time mask 963 # If blank was predicted even once, now or in the past, 964 # Force the current predicted label to also be blank 965 # This ensures that blanks propogate across all timesteps 966 # once they have occured (normally stopping condition of sample level loop). 967 for kidx, ki in enumerate(k): 968 if blank_mask[kidx] == 0: 969 hypotheses[kidx].y_sequence.append(ki) 970 hypotheses[kidx].timestep.append(time_idx) 971 hypotheses[kidx].score += float(v[kidx]) 972 973 symbols_added += 1 974 975 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 976 # Then preserve U at current timestep Ti 977 # Finally, forward the timestep history to Ti+1 for that sample 978 # All of this should only be done iff the current time index <= sample-level AM length. 979 # Otherwise ignore and move to next sample / next timestep. 980 if self.preserve_alignments: 981 982 # convert Ti-th logits into a torch array 983 for batch_idx in range(batchsize): 984 985 # this checks if current timestep <= sample-level AM length 986 # If current timestep > sample-level AM length, no alignments will be added 987 # Therefore the list of Uj alignments is empty here. 988 if len(hypotheses[batch_idx].alignments[-1]) > 0: 989 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 990 991 # Do the same if preserving per-frame confidence 992 if self.preserve_frame_confidence: 993 994 for batch_idx in range(batchsize): 995 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 996 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 997 998 # Remove trailing empty list of alignments at T_{am-len} x Uj 999 if self.preserve_alignments: 1000 for batch_idx in range(batchsize): 1001 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1002 del hypotheses[batch_idx].alignments[-1] 1003 1004 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1005 if self.preserve_frame_confidence: 1006 for batch_idx in range(batchsize): 1007 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1008 del hypotheses[batch_idx].frame_confidence[-1] 1009 1010 # Preserve states 1011 for batch_idx in range(batchsize): 1012 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1013 1014 return hypotheses 1015 1016 1017 class ExportedModelGreedyBatchedRNNTInfer: 1018 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = None): 1019 self.encoder_model_path = encoder_model 1020 self.decoder_joint_model_path = decoder_joint_model 1021 self.max_symbols_per_step = max_symbols_per_step 1022 1023 # Will be populated at runtime 1024 self._blank_index = None 1025 1026 def __call__(self, audio_signal: torch.Tensor, length: torch.Tensor): 1027 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 1028 Output token is generated auto-regressively. 1029 1030 Args: 1031 encoder_output: A tensor of size (batch, features, timesteps). 1032 encoded_lengths: list of int representing the length of each sequence 1033 output sequence. 1034 1035 Returns: 1036 packed list containing batch number of sentences (Hypotheses). 1037 """ 1038 with torch.no_grad(): 1039 # Apply optional preprocessing 1040 encoder_output, encoded_lengths = self.run_encoder(audio_signal=audio_signal, length=length) 1041 1042 if torch.is_tensor(encoder_output): 1043 encoder_output = encoder_output.transpose(1, 2) 1044 else: 1045 encoder_output = encoder_output.transpose([0, 2, 1]) # (B, T, D) 1046 logitlen = encoded_lengths 1047 1048 inseq = encoder_output # [B, T, D] 1049 hypotheses, timestamps = self._greedy_decode(inseq, logitlen) 1050 1051 # Pack the hypotheses results 1052 packed_result = [rnnt_utils.Hypothesis(score=-1.0, y_sequence=[]) for _ in range(len(hypotheses))] 1053 for i in range(len(packed_result)): 1054 packed_result[i].y_sequence = torch.tensor(hypotheses[i], dtype=torch.long) 1055 packed_result[i].length = timestamps[i] 1056 1057 del hypotheses 1058 1059 return packed_result 1060 1061 def _greedy_decode(self, x, out_len): 1062 # x: [B, T, D] 1063 # out_len: [B] 1064 1065 # Initialize state 1066 batchsize = x.shape[0] 1067 hidden = self._get_initial_states(batchsize) 1068 target_lengths = torch.ones(batchsize, dtype=torch.int32) 1069 1070 # Output string buffer 1071 label = [[] for _ in range(batchsize)] 1072 timesteps = [[] for _ in range(batchsize)] 1073 1074 # Last Label buffer + Last Label without blank buffer 1075 # batch level equivalent of the last_label 1076 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long).numpy() 1077 if torch.is_tensor(x): 1078 last_label = torch.from_numpy(last_label).to(self.device) 1079 1080 # Mask buffers 1081 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool).numpy() 1082 1083 # Get max sequence length 1084 max_out_len = out_len.max() 1085 for time_idx in range(max_out_len): 1086 f = x[:, time_idx : time_idx + 1, :] # [B, 1, D] 1087 1088 if torch.is_tensor(f): 1089 f = f.transpose(1, 2) 1090 else: 1091 f = f.transpose([0, 2, 1]) 1092 1093 # Prepare t timestamp batch variables 1094 not_blank = True 1095 symbols_added = 0 1096 1097 # Reset blank mask 1098 blank_mask *= False 1099 1100 # Update blank mask with time mask 1101 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1102 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1103 blank_mask = time_idx >= out_len 1104 # Start inner loop 1105 while not_blank and (self.max_symbols_per_step is None or symbols_added < self.max_symbols_per_step): 1106 1107 # Batch prediction and joint network steps 1108 # If very first prediction step, submit SOS tag (blank) to pred_step. 1109 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1110 if time_idx == 0 and symbols_added == 0: 1111 g = torch.tensor([self._blank_index] * batchsize, dtype=torch.int32).view(-1, 1) 1112 else: 1113 if torch.is_tensor(last_label): 1114 g = last_label.type(torch.int32) 1115 else: 1116 g = last_label.astype(np.int32) 1117 1118 # Batched joint step - Output = [B, V + 1] 1119 joint_out, hidden_prime = self.run_decoder_joint(f, g, target_lengths, *hidden) 1120 logp, pred_lengths = joint_out 1121 logp = logp[:, 0, 0, :] 1122 1123 # Get index k, of max prob for batch 1124 if torch.is_tensor(logp): 1125 v, k = logp.max(1) 1126 else: 1127 k = np.argmax(logp, axis=1).astype(np.int32) 1128 1129 # Update blank mask with current predicted blanks 1130 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1131 k_is_blank = k == self._blank_index 1132 blank_mask |= k_is_blank 1133 1134 del k_is_blank 1135 del logp 1136 1137 # If all samples predict / have predicted prior blanks, exit loop early 1138 # This is equivalent to if single sample predicted k 1139 if blank_mask.all(): 1140 not_blank = False 1141 1142 else: 1143 # Collect batch indices where blanks occurred now/past 1144 if torch.is_tensor(blank_mask): 1145 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1146 else: 1147 blank_indices = blank_mask.astype(np.int32).nonzero() 1148 1149 if type(blank_indices) in (list, tuple): 1150 blank_indices = blank_indices[0] 1151 1152 # Recover prior state for all samples which predicted blank now/past 1153 if hidden is not None: 1154 # LSTM has 2 states 1155 for state_id in range(len(hidden)): 1156 hidden_prime[state_id][:, blank_indices, :] = hidden[state_id][:, blank_indices, :] 1157 1158 elif len(blank_indices) > 0 and hidden is None: 1159 # Reset state if there were some blank and other non-blank predictions in batch 1160 # Original state is filled with zeros so we just multiply 1161 # LSTM has 2 states 1162 for state_id in range(len(hidden_prime)): 1163 hidden_prime[state_id][:, blank_indices, :] *= 0.0 1164 1165 # Recover prior predicted label for all samples which predicted blank now/past 1166 k[blank_indices] = last_label[blank_indices, 0] 1167 1168 # Update new label and hidden state for next iteration 1169 if torch.is_tensor(k): 1170 last_label = k.clone().reshape(-1, 1) 1171 else: 1172 last_label = k.copy().reshape(-1, 1) 1173 hidden = hidden_prime 1174 1175 # Update predicted labels, accounting for time mask 1176 # If blank was predicted even once, now or in the past, 1177 # Force the current predicted label to also be blank 1178 # This ensures that blanks propogate across all timesteps 1179 # once they have occured (normally stopping condition of sample level loop). 1180 for kidx, ki in enumerate(k): 1181 if blank_mask[kidx] == 0: 1182 label[kidx].append(ki) 1183 timesteps[kidx].append(time_idx) 1184 1185 symbols_added += 1 1186 1187 return label, timesteps 1188 1189 def _setup_blank_index(self): 1190 raise NotImplementedError() 1191 1192 def run_encoder(self, audio_signal, length): 1193 raise NotImplementedError() 1194 1195 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1196 raise NotImplementedError() 1197 1198 def _get_initial_states(self, batchsize): 1199 raise NotImplementedError() 1200 1201 1202 class ONNXGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1203 def __init__(self, encoder_model: str, decoder_joint_model: str, max_symbols_per_step: Optional[int] = 10): 1204 super().__init__( 1205 encoder_model=encoder_model, 1206 decoder_joint_model=decoder_joint_model, 1207 max_symbols_per_step=max_symbols_per_step, 1208 ) 1209 1210 try: 1211 import onnx 1212 import onnxruntime 1213 except (ModuleNotFoundError, ImportError): 1214 raise ImportError(f"`onnx` or `onnxruntime` could not be imported, please install the libraries.\n") 1215 1216 if torch.cuda.is_available(): 1217 # Try to use onnxruntime-gpu 1218 providers = ['TensorrtExecutionProvider', 'CUDAExecutionProvider'] 1219 else: 1220 # Fall back to CPU and onnxruntime-cpu 1221 providers = ['CPUExecutionProvider'] 1222 1223 onnx_session_opt = onnxruntime.SessionOptions() 1224 onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL 1225 1226 onnx_model = onnx.load(self.encoder_model_path) 1227 onnx.checker.check_model(onnx_model, full_check=True) 1228 self.encoder_model = onnx_model 1229 self.encoder = onnxruntime.InferenceSession( 1230 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1231 ) 1232 1233 onnx_model = onnx.load(self.decoder_joint_model_path) 1234 onnx.checker.check_model(onnx_model, full_check=True) 1235 self.decoder_joint_model = onnx_model 1236 self.decoder_joint = onnxruntime.InferenceSession( 1237 onnx_model.SerializeToString(), providers=providers, provider_options=onnx_session_opt 1238 ) 1239 1240 logging.info("Successfully loaded encoder, decoder and joint onnx models !") 1241 1242 # Will be populated at runtime 1243 self._blank_index = None 1244 self.max_symbols_per_step = max_symbols_per_step 1245 1246 self._setup_encoder_input_output_keys() 1247 self._setup_decoder_joint_input_output_keys() 1248 self._setup_blank_index() 1249 1250 def _setup_encoder_input_output_keys(self): 1251 self.encoder_inputs = list(self.encoder_model.graph.input) 1252 self.encoder_outputs = list(self.encoder_model.graph.output) 1253 1254 def _setup_decoder_joint_input_output_keys(self): 1255 self.decoder_joint_inputs = list(self.decoder_joint_model.graph.input) 1256 self.decoder_joint_outputs = list(self.decoder_joint_model.graph.output) 1257 1258 def _setup_blank_index(self): 1259 # ASSUME: Single input with no time length information 1260 dynamic_dim = 257 1261 shapes = self.encoder_inputs[0].type.tensor_type.shape.dim 1262 ip_shape = [] 1263 for shape in shapes: 1264 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1265 ip_shape.append(dynamic_dim) # replace dynamic axes with constant 1266 else: 1267 ip_shape.append(int(shape.dim_value)) 1268 1269 enc_logits, encoded_length = self.run_encoder( 1270 audio_signal=torch.randn(*ip_shape), length=torch.randint(0, 1, size=(dynamic_dim,)) 1271 ) 1272 1273 # prepare states 1274 states = self._get_initial_states(batchsize=dynamic_dim) 1275 1276 # run decoder 1 step 1277 joint_out, states = self.run_decoder_joint(enc_logits, None, None, *states) 1278 log_probs, lengths = joint_out 1279 1280 self._blank_index = log_probs.shape[-1] - 1 # last token of vocab size is blank token 1281 logging.info( 1282 f"Enc-Dec-Joint step was evaluated, blank token id = {self._blank_index}; vocab size = {log_probs.shape[-1]}" 1283 ) 1284 1285 def run_encoder(self, audio_signal, length): 1286 if hasattr(audio_signal, 'cpu'): 1287 audio_signal = audio_signal.cpu().numpy() 1288 1289 if hasattr(length, 'cpu'): 1290 length = length.cpu().numpy() 1291 1292 ip = { 1293 self.encoder_inputs[0].name: audio_signal, 1294 self.encoder_inputs[1].name: length, 1295 } 1296 enc_out = self.encoder.run(None, ip) 1297 enc_out, encoded_length = enc_out # ASSUME: single output 1298 return enc_out, encoded_length 1299 1300 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1301 # ASSUME: Decoder is RNN Transducer 1302 if targets is None: 1303 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32) 1304 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32) 1305 1306 if hasattr(targets, 'cpu'): 1307 targets = targets.cpu().numpy() 1308 1309 if hasattr(target_length, 'cpu'): 1310 target_length = target_length.cpu().numpy() 1311 1312 ip = { 1313 self.decoder_joint_inputs[0].name: enc_logits, 1314 self.decoder_joint_inputs[1].name: targets, 1315 self.decoder_joint_inputs[2].name: target_length, 1316 } 1317 1318 num_states = 0 1319 if states is not None and len(states) > 0: 1320 num_states = len(states) 1321 for idx, state in enumerate(states): 1322 if hasattr(state, 'cpu'): 1323 state = state.cpu().numpy() 1324 1325 ip[self.decoder_joint_inputs[len(ip)].name] = state 1326 1327 dec_out = self.decoder_joint.run(None, ip) 1328 1329 # unpack dec output 1330 if num_states > 0: 1331 new_states = dec_out[-num_states:] 1332 dec_out = dec_out[:-num_states] 1333 else: 1334 new_states = None 1335 1336 return dec_out, new_states 1337 1338 def _get_initial_states(self, batchsize): 1339 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1340 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1341 num_states = len(input_state_nodes) 1342 if num_states == 0: 1343 return 1344 1345 input_states = [] 1346 for state_id in range(num_states): 1347 node = input_state_nodes[state_id] 1348 ip_shape = [] 1349 for shape_idx, shape in enumerate(node.type.tensor_type.shape.dim): 1350 if hasattr(shape, 'dim_param') and 'dynamic' in shape.dim_param: 1351 ip_shape.append(batchsize) # replace dynamic axes with constant 1352 else: 1353 ip_shape.append(int(shape.dim_value)) 1354 1355 input_states.append(torch.zeros(*ip_shape)) 1356 1357 return input_states 1358 1359 1360 class TorchscriptGreedyBatchedRNNTInfer(ExportedModelGreedyBatchedRNNTInfer): 1361 def __init__( 1362 self, 1363 encoder_model: str, 1364 decoder_joint_model: str, 1365 cfg: DictConfig, 1366 device: str, 1367 max_symbols_per_step: Optional[int] = 10, 1368 ): 1369 super().__init__( 1370 encoder_model=encoder_model, 1371 decoder_joint_model=decoder_joint_model, 1372 max_symbols_per_step=max_symbols_per_step, 1373 ) 1374 1375 self.cfg = cfg 1376 self.device = device 1377 1378 self.encoder = torch.jit.load(self.encoder_model_path, map_location=self.device) 1379 self.decoder_joint = torch.jit.load(self.decoder_joint_model_path, map_location=self.device) 1380 1381 logging.info("Successfully loaded encoder, decoder and joint torchscript models !") 1382 1383 # Will be populated at runtime 1384 self._blank_index = None 1385 self.max_symbols_per_step = max_symbols_per_step 1386 1387 self._setup_encoder_input_keys() 1388 self._setup_decoder_joint_input_keys() 1389 self._setup_blank_index() 1390 1391 def _setup_encoder_input_keys(self): 1392 arguments = self.encoder.forward.schema.arguments[1:] 1393 self.encoder_inputs = [arg for arg in arguments] 1394 1395 def _setup_decoder_joint_input_keys(self): 1396 arguments = self.decoder_joint.forward.schema.arguments[1:] 1397 self.decoder_joint_inputs = [arg for arg in arguments] 1398 1399 def _setup_blank_index(self): 1400 self._blank_index = len(self.cfg.joint.vocabulary) 1401 1402 logging.info(f"Blank token id = {self._blank_index}; vocab size = {len(self.cfg.joint.vocabulary) + 1}") 1403 1404 def run_encoder(self, audio_signal, length): 1405 enc_out = self.encoder(audio_signal, length) 1406 enc_out, encoded_length = enc_out # ASSUME: single output 1407 return enc_out, encoded_length 1408 1409 def run_decoder_joint(self, enc_logits, targets, target_length, *states): 1410 # ASSUME: Decoder is RNN Transducer 1411 if targets is None: 1412 targets = torch.zeros(enc_logits.shape[0], 1, dtype=torch.int32, device=enc_logits.device) 1413 target_length = torch.ones(enc_logits.shape[0], dtype=torch.int32, device=enc_logits.device) 1414 1415 num_states = 0 1416 if states is not None and len(states) > 0: 1417 num_states = len(states) 1418 1419 dec_out = self.decoder_joint(enc_logits, targets, target_length, *states) 1420 1421 # unpack dec output 1422 if num_states > 0: 1423 new_states = dec_out[-num_states:] 1424 dec_out = dec_out[:-num_states] 1425 else: 1426 new_states = None 1427 1428 return dec_out, new_states 1429 1430 def _get_initial_states(self, batchsize): 1431 # ASSUME: LSTM STATES of shape (layers, batchsize, dim) 1432 input_state_nodes = [ip for ip in self.decoder_joint_inputs if 'state' in ip.name] 1433 num_states = len(input_state_nodes) 1434 if num_states == 0: 1435 return 1436 1437 input_states = [] 1438 for state_id in range(num_states): 1439 # Hardcode shape size for LSTM (1 is for num layers in LSTM, which is flattened for export) 1440 ip_shape = [1, batchsize, self.cfg.model_defaults.pred_hidden] 1441 input_states.append(torch.zeros(*ip_shape, device=self.device)) 1442 1443 return input_states 1444 1445 1446 class GreedyMultiblankRNNTInfer(GreedyRNNTInfer): 1447 """A greedy transducer decoder for multi-blank RNN-T. 1448 1449 Sequence level greedy decoding, performed auto-regressively. 1450 1451 Args: 1452 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1453 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1454 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1455 big_blank_durations: a list containing durations for big blanks the model supports. 1456 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1457 to a sequence in a single time step; if set to None then there is 1458 no limit. 1459 preserve_alignments: Bool flag which preserves the history of alignments generated during 1460 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1461 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1462 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1463 The length of the list corresponds to the Acoustic Length (T). 1464 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1465 U is the number of target tokens for the current timestep Ti. 1466 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1467 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1468 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1469 The length of the list corresponds to the Acoustic Length (T). 1470 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1471 U is the number of target tokens for the current timestep Ti. 1472 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1473 confidence scores. 1474 1475 name: The method name (str). 1476 Supported values: 1477 - 'max_prob' for using the maximum token probability as a confidence. 1478 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1479 1480 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 1481 Supported values: 1482 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1483 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1484 Note that for this entropy, the alpha should comply the following inequality: 1485 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1486 where V is the model vocabulary size. 1487 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1488 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1489 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1490 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1491 - 'renyi' for the Rรฉnyi entropy. 1492 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1493 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1494 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1495 1496 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1497 When the alpha equals one, scaling is not applied to 'max_prob', 1498 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1499 1500 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1501 Supported values: 1502 - 'lin' for using the linear mapping. 1503 - 'exp' for using exponential mapping with linear shift. 1504 """ 1505 1506 def __init__( 1507 self, 1508 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1509 joint_model: rnnt_abstract.AbstractRNNTJoint, 1510 blank_index: int, 1511 big_blank_durations: list, 1512 max_symbols_per_step: Optional[int] = None, 1513 preserve_alignments: bool = False, 1514 preserve_frame_confidence: bool = False, 1515 confidence_method_cfg: Optional[DictConfig] = None, 1516 ): 1517 super().__init__( 1518 decoder_model=decoder_model, 1519 joint_model=joint_model, 1520 blank_index=blank_index, 1521 max_symbols_per_step=max_symbols_per_step, 1522 preserve_alignments=preserve_alignments, 1523 preserve_frame_confidence=preserve_frame_confidence, 1524 confidence_method_cfg=confidence_method_cfg, 1525 ) 1526 self.big_blank_durations = big_blank_durations 1527 self._SOS = blank_index - len(big_blank_durations) 1528 1529 @torch.no_grad() 1530 def _greedy_decode( 1531 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 1532 ): 1533 # x: [T, 1, D] 1534 # out_len: [seq_len] 1535 1536 # Initialize blank state and empty label set in Hypothesis 1537 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 1538 1539 if partial_hypotheses is not None: 1540 hypothesis.last_token = partial_hypotheses.last_token 1541 hypothesis.y_sequence = ( 1542 partial_hypotheses.y_sequence.cpu().tolist() 1543 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 1544 else partial_hypotheses.y_sequence 1545 ) 1546 if partial_hypotheses.dec_state is not None: 1547 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 1548 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 1549 1550 if self.preserve_alignments: 1551 # Alignments is a 2-dimensional dangling list representing T x U 1552 hypothesis.alignments = [[]] 1553 1554 if self.preserve_frame_confidence: 1555 hypothesis.frame_confidence = [[]] 1556 1557 # if this variable is > 1, it means the last emission was a big-blank and we need to skip frames. 1558 big_blank_duration = 1 1559 1560 # For timestep t in X_t 1561 for time_idx in range(out_len): 1562 if big_blank_duration > 1: 1563 # skip frames until big_blank_duration == 1. 1564 big_blank_duration -= 1 1565 continue 1566 # Extract encoder embedding at timestep t 1567 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 1568 f = x.narrow(dim=0, start=time_idx, length=1) 1569 1570 # Setup exit flags and counter 1571 not_blank = True 1572 symbols_added = 0 1573 1574 # While blank is not predicted, or we dont run out of max symbols per timestep 1575 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1576 # In the first timestep, we initialize the network with RNNT Blank 1577 # In later timesteps, we provide previous predicted label as input. 1578 if hypothesis.last_token is None and hypothesis.dec_state is None: 1579 last_label = self._SOS 1580 else: 1581 last_label = label_collate([[hypothesis.last_token]]) 1582 1583 # Perform prediction network and joint network steps. 1584 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 1585 # If preserving per-frame confidence, log_normalize must be true 1586 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1587 0, 0, 0, : 1588 ] 1589 1590 del g 1591 1592 # torch.max(0) op doesnt exist for FP 16. 1593 if logp.dtype != torch.float32: 1594 logp = logp.float() 1595 1596 # get index k, of max prob 1597 v, k = logp.max(0) 1598 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 1599 1600 # Note, we have non-blanks in the vocab first, followed by big blanks, and standard blank at last. 1601 # here we check if it's a big blank and if yes, set the duration variable. 1602 if k >= self._blank_index - len(self.big_blank_durations) and k < self._blank_index: 1603 big_blank_duration = self.big_blank_durations[self._blank_index - k - 1] 1604 1605 if self.preserve_alignments: 1606 # insert logprobs into last timestep 1607 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 1608 1609 if self.preserve_frame_confidence: 1610 # insert confidence into last timestep 1611 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 1612 1613 del logp 1614 1615 # If any type of blank token is predicted, exit inner loop, move onto next timestep t 1616 if k >= self._blank_index - len(self.big_blank_durations): 1617 not_blank = False 1618 else: 1619 # Append token to label set, update RNN state. 1620 hypothesis.y_sequence.append(k) 1621 hypothesis.score += float(v) 1622 hypothesis.timestep.append(time_idx) 1623 hypothesis.dec_state = hidden_prime 1624 hypothesis.last_token = k 1625 1626 # Increment token counter. 1627 symbols_added += 1 1628 1629 if self.preserve_alignments: 1630 # convert Ti-th logits into a torch array 1631 hypothesis.alignments.append([]) # blank buffer for next timestep 1632 1633 if self.preserve_frame_confidence: 1634 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 1635 1636 # Remove trailing empty list of Alignments 1637 if self.preserve_alignments: 1638 if len(hypothesis.alignments[-1]) == 0: 1639 del hypothesis.alignments[-1] 1640 1641 # Remove trailing empty list of per-frame confidence 1642 if self.preserve_frame_confidence: 1643 if len(hypothesis.frame_confidence[-1]) == 0: 1644 del hypothesis.frame_confidence[-1] 1645 1646 # Unpack the hidden states 1647 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 1648 1649 return hypothesis 1650 1651 1652 class GreedyBatchedMultiblankRNNTInfer(GreedyBatchedRNNTInfer): 1653 """A batch level greedy transducer decoder. 1654 Batch level greedy decoding, performed auto-regressively. 1655 Args: 1656 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 1657 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 1658 blank_index: int index of the blank token. Must be len(vocabulary) for multi-blank RNNTs. 1659 big_blank_durations: a list containing durations for big blanks the model supports. 1660 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 1661 to a sequence in a single time step; if set to None then there is 1662 no limit. 1663 preserve_alignments: Bool flag which preserves the history of alignments generated during 1664 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1665 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 1666 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 1667 The length of the list corresponds to the Acoustic Length (T). 1668 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 1669 U is the number of target tokens for the current timestep Ti. 1670 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 1671 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 1672 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 1673 The length of the list corresponds to the Acoustic Length (T). 1674 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 1675 U is the number of target tokens for the current timestep Ti. 1676 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 1677 confidence scores. 1678 1679 name: The method name (str). 1680 Supported values: 1681 - 'max_prob' for using the maximum token probability as a confidence. 1682 - 'entropy' for using a normalized entropy of a log-likelihood vector. 1683 1684 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 1685 Supported values: 1686 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 1687 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 1688 Note that for this entropy, the alpha should comply the following inequality: 1689 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 1690 where V is the model vocabulary size. 1691 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 1692 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 1693 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1694 More: https://en.wikipedia.org/wiki/Tsallis_entropy 1695 - 'renyi' for the Rรฉnyi entropy. 1696 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 1697 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 1698 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 1699 1700 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 1701 When the alpha equals one, scaling is not applied to 'max_prob', 1702 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 1703 1704 entropy_norm: A mapping of the entropy value to the interval [0,1]. 1705 Supported values: 1706 - 'lin' for using the linear mapping. 1707 - 'exp' for using exponential mapping with linear shift. 1708 """ 1709 1710 def __init__( 1711 self, 1712 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 1713 joint_model: rnnt_abstract.AbstractRNNTJoint, 1714 blank_index: int, 1715 big_blank_durations: List[int], 1716 max_symbols_per_step: Optional[int] = None, 1717 preserve_alignments: bool = False, 1718 preserve_frame_confidence: bool = False, 1719 confidence_method_cfg: Optional[DictConfig] = None, 1720 ): 1721 super().__init__( 1722 decoder_model=decoder_model, 1723 joint_model=joint_model, 1724 blank_index=blank_index, 1725 max_symbols_per_step=max_symbols_per_step, 1726 preserve_alignments=preserve_alignments, 1727 preserve_frame_confidence=preserve_frame_confidence, 1728 confidence_method_cfg=confidence_method_cfg, 1729 ) 1730 self.big_blank_durations = big_blank_durations 1731 1732 # Depending on availability of `blank_as_pad` support 1733 # switch between more efficient batch decoding technique 1734 if self.decoder.blank_as_pad: 1735 self._greedy_decode = self._greedy_decode_blank_as_pad 1736 else: 1737 self._greedy_decode = self._greedy_decode_masked 1738 self._SOS = blank_index - len(big_blank_durations) 1739 1740 def _greedy_decode_blank_as_pad( 1741 self, 1742 x: torch.Tensor, 1743 out_len: torch.Tensor, 1744 device: torch.device, 1745 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1746 ): 1747 if partial_hypotheses is not None: 1748 raise NotImplementedError("`partial_hypotheses` support is not supported") 1749 1750 with torch.inference_mode(): 1751 # x: [B, T, D] 1752 # out_len: [B] 1753 # device: torch.device 1754 1755 # Initialize list of Hypothesis 1756 batchsize = x.shape[0] 1757 hypotheses = [ 1758 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1759 ] 1760 1761 # Initialize Hidden state matrix (shared by entire batch) 1762 hidden = None 1763 1764 # If alignments need to be preserved, register a danling list to hold the values 1765 if self.preserve_alignments: 1766 # alignments is a 3-dimensional dangling list representing B x T x U 1767 for hyp in hypotheses: 1768 hyp.alignments = [[]] 1769 1770 # If confidence scores need to be preserved, register a danling list to hold the values 1771 if self.preserve_frame_confidence: 1772 # frame_confidence is a 3-dimensional dangling list representing B x T x U 1773 for hyp in hypotheses: 1774 hyp.frame_confidence = [[]] 1775 1776 # Last Label buffer + Last Label without blank buffer 1777 # batch level equivalent of the last_label 1778 last_label = torch.full([batchsize, 1], fill_value=self._SOS, dtype=torch.long, device=device) 1779 1780 # this mask is true for if the emission is *any type* of blank. 1781 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 1782 1783 # Get max sequence length 1784 max_out_len = out_len.max() 1785 1786 # We have a mask for each big blank. A mask is "true" means: the previous emission is exactly the big-blank 1787 # with the corresponding duration, or has larger duration. E.g., for big_blank_mask for duration 2, it will 1788 # be set true if the previous emission was a big blank with duration 4, or 3 or 2; but false if prevoius 1789 # emission was a standard blank (with duration = 1). 1790 big_blank_masks = [torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device)] * len( 1791 self.big_blank_durations 1792 ) 1793 1794 # if this variable > 1, it means the previous emission was big-blank and we need to skip frames. 1795 big_blank_duration = 1 1796 1797 for time_idx in range(max_out_len): 1798 if big_blank_duration > 1: 1799 # skip frames until big_blank_duration == 1 1800 big_blank_duration -= 1 1801 continue 1802 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 1803 1804 # Prepare t timestamp batch variables 1805 not_blank = True 1806 symbols_added = 0 1807 1808 # Reset all blank masks 1809 blank_mask.mul_(False) 1810 for i in range(len(big_blank_masks)): 1811 big_blank_masks[i].mul_(False) 1812 1813 # Update blank mask with time mask 1814 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 1815 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 1816 blank_mask = time_idx >= out_len 1817 for i in range(len(big_blank_masks)): 1818 big_blank_masks[i] = time_idx >= out_len 1819 1820 # Start inner loop 1821 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 1822 # Batch prediction and joint network steps 1823 # If very first prediction step, submit SOS tag (blank) to pred_step. 1824 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 1825 if time_idx == 0 and symbols_added == 0 and hidden is None: 1826 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 1827 else: 1828 # Perform batch step prediction of decoder, getting new states and scores ("g") 1829 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 1830 1831 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 1832 # If preserving per-frame confidence, log_normalize must be true 1833 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 1834 :, 0, 0, : 1835 ] 1836 1837 if logp.dtype != torch.float32: 1838 logp = logp.float() 1839 1840 # Get index k, of max prob for batch 1841 v, k = logp.max(1) 1842 del g 1843 1844 # Update blank mask with current predicted blanks 1845 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 1846 k_is_blank = k >= self._blank_index - len(self.big_blank_durations) 1847 blank_mask.bitwise_or_(k_is_blank) 1848 1849 for i in range(len(big_blank_masks)): 1850 # using <= since as we mentioned before, the mask doesn't store exact matches. 1851 # instead, it is True when the predicted blank's duration is >= the duration that the 1852 # mask corresponds to. 1853 k_is_big_blank = k <= self._blank_index - 1 - i 1854 1855 # need to do a bitwise_and since it could also be a non-blank. 1856 k_is_big_blank.bitwise_and_(k_is_blank) 1857 big_blank_masks[i].bitwise_or_(k_is_big_blank) 1858 1859 del k_is_blank 1860 1861 # If preserving alignments, check if sequence length of sample has been reached 1862 # before adding alignment 1863 if self.preserve_alignments: 1864 # Insert logprobs into last timestep per sample 1865 logp_vals = logp.to('cpu') 1866 logp_ids = logp_vals.max(1)[1] 1867 for batch_idx in range(batchsize): 1868 if time_idx < out_len[batch_idx]: 1869 hypotheses[batch_idx].alignments[-1].append( 1870 (logp_vals[batch_idx], logp_ids[batch_idx]) 1871 ) 1872 del logp_vals 1873 1874 # If preserving per-frame confidence, check if sequence length of sample has been reached 1875 # before adding confidence scores 1876 if self.preserve_frame_confidence: 1877 # Insert probabilities into last timestep per sample 1878 confidence = self._get_confidence(logp) 1879 for batch_idx in range(batchsize): 1880 if time_idx < out_len[batch_idx]: 1881 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 1882 del logp 1883 1884 # If all samples predict / have predicted prior blanks, exit loop early 1885 # This is equivalent to if single sample predicted k 1886 if blank_mask.all(): 1887 not_blank = False 1888 else: 1889 # Collect batch indices where blanks occurred now/past 1890 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 1891 1892 # Recover prior state for all samples which predicted blank now/past 1893 if hidden is not None: 1894 # LSTM has 2 states 1895 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 1896 1897 elif len(blank_indices) > 0 and hidden is None: 1898 # Reset state if there were some blank and other non-blank predictions in batch 1899 # Original state is filled with zeros so we just multiply 1900 # LSTM has 2 states 1901 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 1902 1903 # Recover prior predicted label for all samples which predicted blank now/past 1904 k[blank_indices] = last_label[blank_indices, 0] 1905 1906 # Update new label and hidden state for next iteration 1907 last_label = k.clone().view(-1, 1) 1908 hidden = hidden_prime 1909 1910 # Update predicted labels, accounting for time mask 1911 # If blank was predicted even once, now or in the past, 1912 # Force the current predicted label to also be blank 1913 # This ensures that blanks propogate across all timesteps 1914 # once they have occured (normally stopping condition of sample level loop). 1915 for kidx, ki in enumerate(k): 1916 if blank_mask[kidx] == 0: 1917 hypotheses[kidx].y_sequence.append(ki) 1918 hypotheses[kidx].timestep.append(time_idx) 1919 hypotheses[kidx].score += float(v[kidx]) 1920 1921 symbols_added += 1 1922 1923 for i in range(len(big_blank_masks) + 1): 1924 # The task here is find the shortest blank duration of all batches. 1925 # so we start from the shortest blank duration and go up, 1926 # and stop once we found the duration whose corresponding mask isn't all True. 1927 if i == len(big_blank_masks) or not big_blank_masks[i].all(): 1928 big_blank_duration = self.big_blank_durations[i - 1] if i > 0 else 1 1929 break 1930 1931 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 1932 # Then preserve U at current timestep Ti 1933 # Finally, forward the timestep history to Ti+1 for that sample 1934 # All of this should only be done iff the current time index <= sample-level AM length. 1935 # Otherwise ignore and move to next sample / next timestep. 1936 if self.preserve_alignments: 1937 1938 # convert Ti-th logits into a torch array 1939 for batch_idx in range(batchsize): 1940 1941 # this checks if current timestep <= sample-level AM length 1942 # If current timestep > sample-level AM length, no alignments will be added 1943 # Therefore the list of Uj alignments is empty here. 1944 if len(hypotheses[batch_idx].alignments[-1]) > 0: 1945 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 1946 1947 # Do the same if preserving per-frame confidence 1948 if self.preserve_frame_confidence: 1949 1950 for batch_idx in range(batchsize): 1951 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 1952 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 1953 1954 # Remove trailing empty list of alignments at T_{am-len} x Uj 1955 if self.preserve_alignments: 1956 for batch_idx in range(batchsize): 1957 if len(hypotheses[batch_idx].alignments[-1]) == 0: 1958 del hypotheses[batch_idx].alignments[-1] 1959 1960 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 1961 if self.preserve_frame_confidence: 1962 for batch_idx in range(batchsize): 1963 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 1964 del hypotheses[batch_idx].frame_confidence[-1] 1965 1966 # Preserve states 1967 for batch_idx in range(batchsize): 1968 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 1969 1970 return hypotheses 1971 1972 def _greedy_decode_masked( 1973 self, 1974 x: torch.Tensor, 1975 out_len: torch.Tensor, 1976 device: torch.device, 1977 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 1978 ): 1979 if partial_hypotheses is not None: 1980 raise NotImplementedError("`partial_hypotheses` support is not supported") 1981 1982 if self.big_blank_durations != [1] * len(self.big_blank_durations): 1983 raise NotImplementedError( 1984 "Efficient frame-skipping version for multi-blank masked decoding is not supported." 1985 ) 1986 1987 # x: [B, T, D] 1988 # out_len: [B] 1989 # device: torch.device 1990 1991 # Initialize state 1992 batchsize = x.shape[0] 1993 hypotheses = [ 1994 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 1995 ] 1996 1997 # Initialize Hidden state matrix (shared by entire batch) 1998 hidden = None 1999 2000 # If alignments need to be preserved, register a danling list to hold the values 2001 if self.preserve_alignments: 2002 # alignments is a 3-dimensional dangling list representing B x T x U 2003 for hyp in hypotheses: 2004 hyp.alignments = [[]] 2005 else: 2006 hyp.alignments = None 2007 2008 # If confidence scores need to be preserved, register a danling list to hold the values 2009 if self.preserve_frame_confidence: 2010 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2011 for hyp in hypotheses: 2012 hyp.frame_confidence = [[]] 2013 2014 # Last Label buffer + Last Label without blank buffer 2015 # batch level equivalent of the last_label 2016 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2017 last_label_without_blank = last_label.clone() 2018 2019 # Mask buffers 2020 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2021 2022 # Get max sequence length 2023 max_out_len = out_len.max() 2024 2025 with torch.inference_mode(): 2026 for time_idx in range(max_out_len): 2027 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2028 2029 # Prepare t timestamp batch variables 2030 not_blank = True 2031 symbols_added = 0 2032 2033 # Reset blank mask 2034 blank_mask.mul_(False) 2035 2036 # Update blank mask with time mask 2037 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2038 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2039 blank_mask = time_idx >= out_len 2040 2041 # Start inner loop 2042 while not_blank and (self.max_symbols is None or symbols_added < self.max_symbols): 2043 # Batch prediction and joint network steps 2044 # If very first prediction step, submit SOS tag (blank) to pred_step. 2045 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2046 if time_idx == 0 and symbols_added == 0 and hidden is None: 2047 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2048 else: 2049 # Set a dummy label for the blank value 2050 # This value will be overwritten by "blank" again the last label update below 2051 # This is done as vocabulary of prediction network does not contain "blank" token of RNNT 2052 last_label_without_blank_mask = last_label >= self._blank_index 2053 last_label_without_blank[last_label_without_blank_mask] = 0 # temp change of label 2054 last_label_without_blank[~last_label_without_blank_mask] = last_label[ 2055 ~last_label_without_blank_mask 2056 ] 2057 2058 # Perform batch step prediction of decoder, getting new states and scores ("g") 2059 g, hidden_prime = self._pred_step(last_label_without_blank, hidden, batch_size=batchsize) 2060 2061 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2062 # If preserving per-frame confidence, log_normalize must be true 2063 logp = self._joint_step(f, g, log_normalize=True if self.preserve_frame_confidence else None)[ 2064 :, 0, 0, : 2065 ] 2066 2067 if logp.dtype != torch.float32: 2068 logp = logp.float() 2069 2070 # Get index k, of max prob for batch 2071 v, k = logp.max(1) 2072 del g 2073 2074 # Update blank mask with current predicted blanks 2075 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2076 k_is_blank = k == self._blank_index 2077 blank_mask.bitwise_or_(k_is_blank) 2078 2079 # If preserving alignments, check if sequence length of sample has been reached 2080 # before adding alignment 2081 if self.preserve_alignments: 2082 # Insert logprobs into last timestep per sample 2083 logp_vals = logp.to('cpu') 2084 logp_ids = logp_vals.max(1)[1] 2085 for batch_idx in range(batchsize): 2086 if time_idx < out_len[batch_idx]: 2087 hypotheses[batch_idx].alignments[-1].append( 2088 (logp_vals[batch_idx], logp_ids[batch_idx]) 2089 ) 2090 del logp_vals 2091 2092 # If preserving per-frame confidence, check if sequence length of sample has been reached 2093 # before adding confidence scores 2094 if self.preserve_frame_confidence: 2095 # Insert probabilities into last timestep per sample 2096 confidence = self._get_confidence(logp) 2097 for batch_idx in range(batchsize): 2098 if time_idx < out_len[batch_idx]: 2099 hypotheses[batch_idx].frame_confidence[-1].append(confidence[batch_idx]) 2100 del logp 2101 2102 # If all samples predict / have predicted prior blanks, exit loop early 2103 # This is equivalent to if single sample predicted k 2104 if blank_mask.all(): 2105 not_blank = False 2106 else: 2107 # Collect batch indices where blanks occurred now/past 2108 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2109 2110 # Recover prior state for all samples which predicted blank now/past 2111 if hidden is not None: 2112 # LSTM has 2 states 2113 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2114 2115 elif len(blank_indices) > 0 and hidden is None: 2116 # Reset state if there were some blank and other non-blank predictions in batch 2117 # Original state is filled with zeros so we just multiply 2118 # LSTM has 2 states 2119 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2120 2121 # Recover prior predicted label for all samples which predicted blank now/past 2122 k[blank_indices] = last_label[blank_indices, 0] 2123 2124 # Update new label and hidden state for next iteration 2125 last_label = k.view(-1, 1) 2126 hidden = hidden_prime 2127 2128 # Update predicted labels, accounting for time mask 2129 # If blank was predicted even once, now or in the past, 2130 # Force the current predicted label to also be blank 2131 # This ensures that blanks propogate across all timesteps 2132 # once they have occured (normally stopping condition of sample level loop). 2133 for kidx, ki in enumerate(k): 2134 if blank_mask[kidx] == 0: 2135 hypotheses[kidx].y_sequence.append(ki) 2136 hypotheses[kidx].timestep.append(time_idx) 2137 hypotheses[kidx].score += float(v[kidx]) 2138 2139 symbols_added += 1 2140 2141 # If preserving alignments, convert the current Uj alignments into a torch.Tensor 2142 # Then preserve U at current timestep Ti 2143 # Finally, forward the timestep history to Ti+1 for that sample 2144 # All of this should only be done iff the current time index <= sample-level AM length. 2145 # Otherwise ignore and move to next sample / next timestep. 2146 if self.preserve_alignments: 2147 2148 # convert Ti-th logits into a torch array 2149 for batch_idx in range(batchsize): 2150 2151 # this checks if current timestep <= sample-level AM length 2152 # If current timestep > sample-level AM length, no alignments will be added 2153 # Therefore the list of Uj alignments is empty here. 2154 if len(hypotheses[batch_idx].alignments[-1]) > 0: 2155 hypotheses[batch_idx].alignments.append([]) # blank buffer for next timestep 2156 2157 # Do the same if preserving per-frame confidence 2158 if self.preserve_frame_confidence: 2159 2160 for batch_idx in range(batchsize): 2161 if len(hypotheses[batch_idx].frame_confidence[-1]) > 0: 2162 hypotheses[batch_idx].frame_confidence.append([]) # blank buffer for next timestep 2163 2164 # Remove trailing empty list of alignments at T_{am-len} x Uj 2165 if self.preserve_alignments: 2166 for batch_idx in range(batchsize): 2167 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2168 del hypotheses[batch_idx].alignments[-1] 2169 2170 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2171 if self.preserve_frame_confidence: 2172 for batch_idx in range(batchsize): 2173 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2174 del hypotheses[batch_idx].frame_confidence[-1] 2175 2176 # Preserve states 2177 for batch_idx in range(batchsize): 2178 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2179 2180 return hypotheses 2181 2182 2183 @dataclass 2184 class GreedyRNNTInferConfig: 2185 max_symbols_per_step: Optional[int] = 10 2186 preserve_alignments: bool = False 2187 preserve_frame_confidence: bool = False 2188 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 2189 2190 def __post_init__(self): 2191 # OmegaConf.structured ensures that post_init check is always executed 2192 self.confidence_method_cfg = OmegaConf.structured( 2193 self.confidence_method_cfg 2194 if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) 2195 else ConfidenceMethodConfig(**self.confidence_method_cfg) 2196 ) 2197 2198 2199 @dataclass 2200 class GreedyBatchedRNNTInferConfig: 2201 max_symbols_per_step: Optional[int] = 10 2202 preserve_alignments: bool = False 2203 preserve_frame_confidence: bool = False 2204 confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() 2205 2206 def __post_init__(self): 2207 # OmegaConf.structured ensures that post_init check is always executed 2208 self.confidence_method_cfg = OmegaConf.structured( 2209 self.confidence_method_cfg 2210 if isinstance(self.confidence_method_cfg, ConfidenceMethodConfig) 2211 else ConfidenceMethodConfig(**self.confidence_method_cfg) 2212 ) 2213 2214 2215 class GreedyTDTInfer(_GreedyRNNTInfer): 2216 """A greedy TDT decoder. 2217 2218 Sequence level greedy decoding, performed auto-regressively. 2219 2220 Args: 2221 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2222 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2223 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2224 durations: a list containing durations for TDT. 2225 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2226 to a sequence in a single time step; if set to None then there is 2227 no limit. 2228 preserve_alignments: Bool flag which preserves the history of alignments generated during 2229 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2230 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2231 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2232 The length of the list corresponds to the Acoustic Length (T). 2233 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2234 U is the number of target tokens for the current timestep Ti. 2235 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2236 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2237 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2238 The length of the list corresponds to the Acoustic Length (T). 2239 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2240 U is the number of target tokens for the current timestep Ti. 2241 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 2242 confidence scores. 2243 2244 name: The method name (str). 2245 Supported values: 2246 - 'max_prob' for using the maximum token probability as a confidence. 2247 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2248 2249 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 2250 Supported values: 2251 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2252 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2253 Note that for this entropy, the alpha should comply the following inequality: 2254 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2255 where V is the model vocabulary size. 2256 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2257 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2258 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2259 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2260 - 'renyi' for the Rรฉnyi entropy. 2261 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2262 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2263 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2264 2265 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2266 When the alpha equals one, scaling is not applied to 'max_prob', 2267 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2268 2269 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2270 Supported values: 2271 - 'lin' for using the linear mapping. 2272 - 'exp' for using exponential mapping with linear shift. 2273 """ 2274 2275 def __init__( 2276 self, 2277 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2278 joint_model: rnnt_abstract.AbstractRNNTJoint, 2279 blank_index: int, 2280 durations: list, 2281 max_symbols_per_step: Optional[int] = None, 2282 preserve_alignments: bool = False, 2283 preserve_frame_confidence: bool = False, 2284 confidence_method_cfg: Optional[DictConfig] = None, 2285 ): 2286 super().__init__( 2287 decoder_model=decoder_model, 2288 joint_model=joint_model, 2289 blank_index=blank_index, 2290 max_symbols_per_step=max_symbols_per_step, 2291 preserve_alignments=preserve_alignments, 2292 preserve_frame_confidence=preserve_frame_confidence, 2293 confidence_method_cfg=confidence_method_cfg, 2294 ) 2295 self.durations = durations 2296 2297 @typecheck() 2298 def forward( 2299 self, 2300 encoder_output: torch.Tensor, 2301 encoded_lengths: torch.Tensor, 2302 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2303 ): 2304 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2305 Output token is generated auto-regressively. 2306 Args: 2307 encoder_output: A tensor of size (batch, features, timesteps). 2308 encoded_lengths: list of int representing the length of each sequence 2309 output sequence. 2310 Returns: 2311 packed list containing batch number of sentences (Hypotheses). 2312 """ 2313 # Preserve decoder and joint training state 2314 decoder_training_state = self.decoder.training 2315 joint_training_state = self.joint.training 2316 2317 with torch.inference_mode(): 2318 # Apply optional preprocessing 2319 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2320 2321 self.decoder.eval() 2322 self.joint.eval() 2323 2324 hypotheses = [] 2325 # Process each sequence independently 2326 with self.decoder.as_frozen(), self.joint.as_frozen(): 2327 for batch_idx in range(encoder_output.size(0)): 2328 inseq = encoder_output[batch_idx, :, :].unsqueeze(1) # [T, 1, D] 2329 logitlen = encoded_lengths[batch_idx] 2330 2331 partial_hypothesis = partial_hypotheses[batch_idx] if partial_hypotheses is not None else None 2332 hypothesis = self._greedy_decode(inseq, logitlen, partial_hypotheses=partial_hypothesis) 2333 hypotheses.append(hypothesis) 2334 2335 # Pack results into Hypotheses 2336 packed_result = pack_hypotheses(hypotheses, encoded_lengths) 2337 2338 self.decoder.train(decoder_training_state) 2339 self.joint.train(joint_training_state) 2340 2341 return (packed_result,) 2342 2343 @torch.no_grad() 2344 def _greedy_decode( 2345 self, x: torch.Tensor, out_len: torch.Tensor, partial_hypotheses: Optional[rnnt_utils.Hypothesis] = None 2346 ): 2347 # x: [T, 1, D] 2348 # out_len: [seq_len] 2349 2350 # Initialize blank state and empty label set in Hypothesis 2351 hypothesis = rnnt_utils.Hypothesis(score=0.0, y_sequence=[], dec_state=None, timestep=[], last_token=None) 2352 2353 if partial_hypotheses is not None: 2354 hypothesis.last_token = partial_hypotheses.last_token 2355 hypothesis.y_sequence = ( 2356 partial_hypotheses.y_sequence.cpu().tolist() 2357 if isinstance(partial_hypotheses.y_sequence, torch.Tensor) 2358 else partial_hypotheses.y_sequence 2359 ) 2360 if partial_hypotheses.dec_state is not None: 2361 hypothesis.dec_state = self.decoder.batch_concat_states([partial_hypotheses.dec_state]) 2362 hypothesis.dec_state = _states_to_device(hypothesis.dec_state, x.device) 2363 2364 if self.preserve_alignments: 2365 # Alignments is a 2-dimensional dangling list representing T x U 2366 hypothesis.alignments = [[]] 2367 2368 if self.preserve_frame_confidence: 2369 hypothesis.frame_confidence = [[]] 2370 2371 time_idx = 0 2372 while time_idx < out_len: 2373 # Extract encoder embedding at timestep t 2374 # f = x[time_idx, :, :].unsqueeze(0) # [1, 1, D] 2375 f = x.narrow(dim=0, start=time_idx, length=1) 2376 2377 # Setup exit flags and counter 2378 not_blank = True 2379 symbols_added = 0 2380 2381 need_loop = True 2382 # While blank is not predicted, or we dont run out of max symbols per timestep 2383 while need_loop and (self.max_symbols is None or symbols_added < self.max_symbols): 2384 # In the first timestep, we initialize the network with RNNT Blank 2385 # In later timesteps, we provide previous predicted label as input. 2386 if hypothesis.last_token is None and hypothesis.dec_state is None: 2387 last_label = self._SOS 2388 else: 2389 last_label = label_collate([[hypothesis.last_token]]) 2390 2391 # Perform prediction network and joint network steps. 2392 g, hidden_prime = self._pred_step(last_label, hypothesis.dec_state) 2393 # If preserving per-frame confidence, log_normalize must be true 2394 logits = self._joint_step(f, g, log_normalize=False) 2395 logp = logits[0, 0, 0, : -len(self.durations)] 2396 if self.preserve_frame_confidence: 2397 logp = torch.log_softmax(logp, -1) 2398 2399 duration_logp = torch.log_softmax(logits[0, 0, 0, -len(self.durations) :], dim=-1) 2400 del g 2401 2402 # torch.max(0) op doesnt exist for FP 16. 2403 if logp.dtype != torch.float32: 2404 logp = logp.float() 2405 2406 # get index k, of max prob 2407 v, k = logp.max(0) 2408 k = k.item() # K is the label at timestep t_s in inner loop, s >= 0. 2409 2410 d_v, d_k = duration_logp.max(0) 2411 d_k = d_k.item() 2412 2413 skip = self.durations[d_k] 2414 2415 if self.preserve_alignments: 2416 # insert logprobs into last timestep 2417 hypothesis.alignments[-1].append((logp.to('cpu'), torch.tensor(k, dtype=torch.int32))) 2418 2419 if self.preserve_frame_confidence: 2420 # insert confidence into last timestep 2421 hypothesis.frame_confidence[-1].append(self._get_confidence(logp)) 2422 2423 del logp 2424 2425 # If blank token is predicted, exit inner loop, move onto next timestep t 2426 if k == self._blank_index: 2427 not_blank = False 2428 else: 2429 # Append token to label set, update RNN state. 2430 hypothesis.y_sequence.append(k) 2431 hypothesis.score += float(v) 2432 hypothesis.timestep.append(time_idx) 2433 hypothesis.dec_state = hidden_prime 2434 hypothesis.last_token = k 2435 2436 # Increment token counter. 2437 symbols_added += 1 2438 time_idx += skip 2439 need_loop = skip == 0 2440 2441 # this rarely happens, but we manually increment the `skip` number 2442 # if blank is emitted and duration=0 is predicted. This prevents possible 2443 # infinite loops. 2444 if skip == 0: 2445 skip = 1 2446 2447 if self.preserve_alignments: 2448 # convert Ti-th logits into a torch array 2449 hypothesis.alignments.append([]) # blank buffer for next timestep 2450 2451 if self.preserve_frame_confidence: 2452 hypothesis.frame_confidence.append([]) # blank buffer for next timestep 2453 2454 if symbols_added == self.max_symbols: 2455 time_idx += 1 2456 2457 # Remove trailing empty list of Alignments 2458 if self.preserve_alignments: 2459 if len(hypothesis.alignments[-1]) == 0: 2460 del hypothesis.alignments[-1] 2461 2462 # Remove trailing empty list of per-frame confidence 2463 if self.preserve_frame_confidence: 2464 if len(hypothesis.frame_confidence[-1]) == 0: 2465 del hypothesis.frame_confidence[-1] 2466 2467 # Unpack the hidden states 2468 hypothesis.dec_state = self.decoder.batch_select_state(hypothesis.dec_state, 0) 2469 2470 return hypothesis 2471 2472 2473 class GreedyBatchedTDTInfer(_GreedyRNNTInfer): 2474 """A batch level greedy TDT decoder. 2475 Batch level greedy decoding, performed auto-regressively. 2476 Args: 2477 decoder_model: rnnt_utils.AbstractRNNTDecoder implementation. 2478 joint_model: rnnt_utils.AbstractRNNTJoint implementation. 2479 blank_index: int index of the blank token. Must be len(vocabulary) for TDT models. 2480 durations: a list containing durations. 2481 max_symbols_per_step: Optional int. The maximum number of symbols that can be added 2482 to a sequence in a single time step; if set to None then there is 2483 no limit. 2484 preserve_alignments: Bool flag which preserves the history of alignments generated during 2485 greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2486 the non-null value for `alignments` in it. Here, `alignments` is a List of List of 2487 Tuple(Tensor (of length V + 1 + num-big-blanks), Tensor(scalar, label after argmax)). 2488 The length of the list corresponds to the Acoustic Length (T). 2489 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more targets from a vocabulary. 2490 U is the number of target tokens for the current timestep Ti. 2491 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores generated 2492 during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 2493 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of List of floats. 2494 The length of the list corresponds to the Acoustic Length (T). 2495 Each value in the list (Ti) is a torch.Tensor (U), representing 1 or more confidence scores. 2496 U is the number of target tokens for the current timestep Ti. 2497 confidence_method_cfg: A dict-like object which contains the method name and settings to compute per-frame 2498 confidence scores. 2499 2500 name: The method name (str). 2501 Supported values: 2502 - 'max_prob' for using the maximum token probability as a confidence. 2503 - 'entropy' for using a normalized entropy of a log-likelihood vector. 2504 2505 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 2506 Supported values: 2507 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 2508 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 2509 Note that for this entropy, the alpha should comply the following inequality: 2510 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 2511 where V is the model vocabulary size. 2512 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 2513 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 2514 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2515 More: https://en.wikipedia.org/wiki/Tsallis_entropy 2516 - 'renyi' for the Rรฉnyi entropy. 2517 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 2518 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 2519 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 2520 2521 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 2522 When the alpha equals one, scaling is not applied to 'max_prob', 2523 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 2524 2525 entropy_norm: A mapping of the entropy value to the interval [0,1]. 2526 Supported values: 2527 - 'lin' for using the linear mapping. 2528 - 'exp' for using exponential mapping with linear shift. 2529 """ 2530 2531 def __init__( 2532 self, 2533 decoder_model: rnnt_abstract.AbstractRNNTDecoder, 2534 joint_model: rnnt_abstract.AbstractRNNTJoint, 2535 blank_index: int, 2536 durations: List[int], 2537 max_symbols_per_step: Optional[int] = None, 2538 preserve_alignments: bool = False, 2539 preserve_frame_confidence: bool = False, 2540 confidence_method_cfg: Optional[DictConfig] = None, 2541 ): 2542 super().__init__( 2543 decoder_model=decoder_model, 2544 joint_model=joint_model, 2545 blank_index=blank_index, 2546 max_symbols_per_step=max_symbols_per_step, 2547 preserve_alignments=preserve_alignments, 2548 preserve_frame_confidence=preserve_frame_confidence, 2549 confidence_method_cfg=confidence_method_cfg, 2550 ) 2551 self.durations = durations 2552 2553 # Depending on availability of `blank_as_pad` support 2554 # switch between more efficient batch decoding technique 2555 if self.decoder.blank_as_pad: 2556 self._greedy_decode = self._greedy_decode_blank_as_pad 2557 else: 2558 self._greedy_decode = self._greedy_decode_masked 2559 2560 @typecheck() 2561 def forward( 2562 self, 2563 encoder_output: torch.Tensor, 2564 encoded_lengths: torch.Tensor, 2565 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2566 ): 2567 """Returns a list of hypotheses given an input batch of the encoder hidden embedding. 2568 Output token is generated auto-regressively. 2569 Args: 2570 encoder_output: A tensor of size (batch, features, timesteps). 2571 encoded_lengths: list of int representing the length of each sequence 2572 output sequence. 2573 Returns: 2574 packed list containing batch number of sentences (Hypotheses). 2575 """ 2576 # Preserve decoder and joint training state 2577 decoder_training_state = self.decoder.training 2578 joint_training_state = self.joint.training 2579 2580 with torch.inference_mode(): 2581 # Apply optional preprocessing 2582 encoder_output = encoder_output.transpose(1, 2) # (B, T, D) 2583 logitlen = encoded_lengths 2584 2585 self.decoder.eval() 2586 self.joint.eval() 2587 2588 with self.decoder.as_frozen(), self.joint.as_frozen(): 2589 inseq = encoder_output # [B, T, D] 2590 hypotheses = self._greedy_decode( 2591 inseq, logitlen, device=inseq.device, partial_hypotheses=partial_hypotheses 2592 ) 2593 2594 # Pack the hypotheses results 2595 packed_result = pack_hypotheses(hypotheses, logitlen) 2596 2597 self.decoder.train(decoder_training_state) 2598 self.joint.train(joint_training_state) 2599 2600 return (packed_result,) 2601 2602 def _greedy_decode_blank_as_pad( 2603 self, 2604 x: torch.Tensor, 2605 out_len: torch.Tensor, 2606 device: torch.device, 2607 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2608 ): 2609 if partial_hypotheses is not None: 2610 raise NotImplementedError("`partial_hypotheses` support is not supported") 2611 2612 with torch.inference_mode(): 2613 # x: [B, T, D] 2614 # out_len: [B] 2615 # device: torch.device 2616 2617 # Initialize list of Hypothesis 2618 batchsize = x.shape[0] 2619 hypotheses = [ 2620 rnnt_utils.Hypothesis(score=0.0, y_sequence=[], timestep=[], dec_state=None) for _ in range(batchsize) 2621 ] 2622 2623 # Initialize Hidden state matrix (shared by entire batch) 2624 hidden = None 2625 2626 # If alignments need to be preserved, register a danling list to hold the values 2627 if self.preserve_alignments: 2628 # alignments is a 3-dimensional dangling list representing B x T x U 2629 for hyp in hypotheses: 2630 hyp.alignments = [[]] 2631 2632 # If confidence scores need to be preserved, register a danling list to hold the values 2633 if self.preserve_frame_confidence: 2634 # frame_confidence is a 3-dimensional dangling list representing B x T x U 2635 for hyp in hypotheses: 2636 hyp.frame_confidence = [[]] 2637 2638 # Last Label buffer + Last Label without blank buffer 2639 # batch level equivalent of the last_label 2640 last_label = torch.full([batchsize, 1], fill_value=self._blank_index, dtype=torch.long, device=device) 2641 2642 # Mask buffers 2643 blank_mask = torch.full([batchsize], fill_value=0, dtype=torch.bool, device=device) 2644 2645 # Get max sequence length 2646 max_out_len = out_len.max() 2647 2648 # skip means the number of frames the next decoding step should "jump" to. When skip == 1 2649 # it means the next decoding step will just use the next input frame. 2650 skip = 1 2651 for time_idx in range(max_out_len): 2652 if skip > 1: # if skip > 1 at the current step, we decrement it and skip the current frame. 2653 skip -= 1 2654 continue 2655 f = x.narrow(dim=1, start=time_idx, length=1) # [B, 1, D] 2656 2657 # need_to_stay is a boolean indicates whether the next decoding step should remain in the same frame. 2658 need_to_stay = True 2659 symbols_added = 0 2660 2661 # Reset blank mask 2662 blank_mask.mul_(False) 2663 2664 # Update blank mask with time mask 2665 # Batch: [B, T, D], but Bi may have seq len < max(seq_lens_in_batch) 2666 # Forcibly mask with "blank" tokens, for all sample where current time step T > seq_len 2667 blank_mask = time_idx >= out_len 2668 2669 # Start inner loop 2670 while need_to_stay and (self.max_symbols is None or symbols_added < self.max_symbols): 2671 # Batch prediction and joint network steps 2672 # If very first prediction step, submit SOS tag (blank) to pred_step. 2673 # This feeds a zero tensor as input to AbstractRNNTDecoder to prime the state 2674 if time_idx == 0 and symbols_added == 0 and hidden is None: 2675 g, hidden_prime = self._pred_step(self._SOS, hidden, batch_size=batchsize) 2676 else: 2677 # Perform batch step prediction of decoder, getting new states and scores ("g") 2678 g, hidden_prime = self._pred_step(last_label, hidden, batch_size=batchsize) 2679 2680 # Batched joint step - Output = [B, V + 1 + num-big-blanks] 2681 # Note: log_normalize must not be True here since the joiner output is contanetation of both token logits and duration logits, 2682 # and they need to be normalized independently. 2683 joined = self._joint_step(f, g, log_normalize=None) 2684 logp = joined[:, 0, 0, : -len(self.durations)] 2685 duration_logp = joined[:, 0, 0, -len(self.durations) :] 2686 2687 if logp.dtype != torch.float32: 2688 logp = logp.float() 2689 duration_logp = duration_logp.float() 2690 2691 # get the max for both token and duration predictions. 2692 v, k = logp.max(1) 2693 dv, dk = duration_logp.max(1) 2694 2695 # here we set the skip value to be the minimum of all predicted durations, hense the "torch.min(dk)" call there. 2696 # Please refer to Section 5.2 of our paper https://arxiv.org/pdf/2304.06795.pdf for explanation of this. 2697 skip = self.durations[int(torch.min(dk))] 2698 2699 # this is a special case: if all batches emit blanks, we require that skip be at least 1 2700 # so we don't loop forever at the current frame. 2701 if blank_mask.all(): 2702 if skip == 0: 2703 skip = 1 2704 2705 need_to_stay = skip == 0 2706 del g 2707 2708 # Update blank mask with current predicted blanks 2709 # This is accumulating blanks over all time steps T and all target steps min(max_symbols, U) 2710 k_is_blank = k == self._blank_index 2711 blank_mask.bitwise_or_(k_is_blank) 2712 2713 del k_is_blank 2714 del logp, duration_logp 2715 2716 # If all samples predict / have predicted prior blanks, exit loop early 2717 # This is equivalent to if single sample predicted k 2718 if not blank_mask.all(): 2719 # Collect batch indices where blanks occurred now/past 2720 blank_indices = (blank_mask == 1).nonzero(as_tuple=False) 2721 2722 # Recover prior state for all samples which predicted blank now/past 2723 if hidden is not None: 2724 hidden_prime = self.decoder.batch_copy_states(hidden_prime, hidden, blank_indices) 2725 2726 elif len(blank_indices) > 0 and hidden is None: 2727 # Reset state if there were some blank and other non-blank predictions in batch 2728 # Original state is filled with zeros so we just multiply 2729 # LSTM has 2 states 2730 hidden_prime = self.decoder.batch_copy_states(hidden_prime, None, blank_indices, value=0.0) 2731 2732 # Recover prior predicted label for all samples which predicted blank now/past 2733 k[blank_indices] = last_label[blank_indices, 0] 2734 2735 # Update new label and hidden state for next iteration 2736 last_label = k.clone().view(-1, 1) 2737 hidden = hidden_prime 2738 2739 # Update predicted labels, accounting for time mask 2740 # If blank was predicted even once, now or in the past, 2741 # Force the current predicted label to also be blank 2742 # This ensures that blanks propogate across all timesteps 2743 # once they have occured (normally stopping condition of sample level loop). 2744 for kidx, ki in enumerate(k): 2745 if blank_mask[kidx] == 0: 2746 hypotheses[kidx].y_sequence.append(ki) 2747 hypotheses[kidx].timestep.append(time_idx) 2748 hypotheses[kidx].score += float(v[kidx]) 2749 2750 symbols_added += 1 2751 2752 # Remove trailing empty list of alignments at T_{am-len} x Uj 2753 if self.preserve_alignments: 2754 for batch_idx in range(batchsize): 2755 if len(hypotheses[batch_idx].alignments[-1]) == 0: 2756 del hypotheses[batch_idx].alignments[-1] 2757 2758 # Remove trailing empty list of confidence scores at T_{am-len} x Uj 2759 if self.preserve_frame_confidence: 2760 for batch_idx in range(batchsize): 2761 if len(hypotheses[batch_idx].frame_confidence[-1]) == 0: 2762 del hypotheses[batch_idx].frame_confidence[-1] 2763 2764 # Preserve states 2765 for batch_idx in range(batchsize): 2766 hypotheses[batch_idx].dec_state = self.decoder.batch_select_state(hidden, batch_idx) 2767 2768 return hypotheses 2769 2770 def _greedy_decode_masked( 2771 self, 2772 x: torch.Tensor, 2773 out_len: torch.Tensor, 2774 device: torch.device, 2775 partial_hypotheses: Optional[List[rnnt_utils.Hypothesis]] = None, 2776 ): 2777 raise NotImplementedError("masked greedy-batched decode is not supported for TDT models.") 2778 [end of nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py] [start of nemo/collections/asr/parts/utils/asr_confidence_utils.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import math 16 from abc import ABC, abstractmethod 17 from dataclasses import dataclass 18 from functools import partial 19 from typing import List, Optional 20 21 import torch 22 from omegaconf import DictConfig, OmegaConf 23 24 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 25 from nemo.utils import logging 26 27 28 class ConfidenceMethodConstants: 29 NAMES = ("max_prob", "entropy") 30 ENTROPY_TYPES = ("gibbs", "tsallis", "renyi") 31 ENTROPY_NORMS = ("lin", "exp") 32 33 @classmethod 34 def print(cls): 35 return ( 36 cls.__name__ 37 + ": " 38 + str({"NAMES": cls.NAMES, "ENTROPY_TYPES": cls.ENTROPY_TYPES, "ENTROPY_NORMS": cls.ENTROPY_NORMS}) 39 ) 40 41 42 class ConfidenceConstants: 43 AGGREGATIONS = ("mean", "min", "max", "prod") 44 45 @classmethod 46 def print(cls): 47 return cls.__name__ + ": " + str({"AGGREGATIONS": cls.AGGREGATIONS}) 48 49 50 @dataclass 51 class ConfidenceMethodConfig: 52 """A Config which contains the method name and settings to compute per-frame confidence scores. 53 54 Args: 55 name: The method name (str). 56 Supported values: 57 - 'max_prob' for using the maximum token probability as a confidence. 58 - 'entropy' for using a normalized entropy of a log-likelihood vector. 59 60 entropy_type: Which type of entropy to use (str). 61 Used if confidence_method_cfg.name is set to `entropy`. 62 Supported values: 63 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 64 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 65 Note that for this entropy, the alpha should comply the following inequality: 66 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 67 where V is the model vocabulary size. 68 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 69 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 70 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 71 More: https://en.wikipedia.org/wiki/Tsallis_entropy 72 - 'renyi' for the Rรฉnyi entropy. 73 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 74 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 75 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 76 77 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 78 When the alpha equals one, scaling is not applied to 'max_prob', 79 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 80 81 entropy_norm: A mapping of the entropy value to the interval [0,1]. 82 Supported values: 83 - 'lin' for using the linear mapping. 84 - 'exp' for using exponential mapping with linear shift. 85 """ 86 87 name: str = "entropy" 88 entropy_type: str = "tsallis" 89 alpha: float = 0.33 90 entropy_norm: str = "exp" 91 temperature: str = "DEPRECATED" 92 93 def __post_init__(self): 94 if self.temperature != "DEPRECATED": 95 # self.temperature has type str 96 self.alpha = float(self.temperature) 97 self.temperature = "DEPRECATED" 98 if self.name not in ConfidenceMethodConstants.NAMES: 99 raise ValueError( 100 f"`name` must be one of the following: " 101 f"{'`' + '`, `'.join(ConfidenceMethodConstants.NAMES) + '`'}. Provided: `{self.name}`" 102 ) 103 if self.entropy_type not in ConfidenceMethodConstants.ENTROPY_TYPES: 104 raise ValueError( 105 f"`entropy_type` must be one of the following: " 106 f"{'`' + '`, `'.join(ConfidenceMethodConstants.ENTROPY_TYPES) + '`'}. Provided: `{self.entropy_type}`" 107 ) 108 if self.alpha <= 0.0: 109 raise ValueError(f"`alpha` must be > 0. Provided: {self.alpha}") 110 if self.entropy_norm not in ConfidenceMethodConstants.ENTROPY_NORMS: 111 raise ValueError( 112 f"`entropy_norm` must be one of the following: " 113 f"{'`' + '`, `'.join(ConfidenceMethodConstants.ENTROPY_NORMS) + '`'}. Provided: `{self.entropy_norm}`" 114 ) 115 116 117 @dataclass 118 class ConfidenceConfig: 119 """A config which contains the following key-value pairs related to confidence scores. 120 121 Args: 122 preserve_frame_confidence: Bool flag which preserves the history of per-frame confidence scores 123 generated during decoding. When set to true, the Hypothesis will contain 124 the non-null value for `frame_confidence` in it. Here, `frame_confidence` is a List of floats. 125 preserve_token_confidence: Bool flag which preserves the history of per-token confidence scores 126 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 127 the non-null value for `token_confidence` in it. Here, `token_confidence` is a List of floats. 128 129 The length of the list corresponds to the number of recognized tokens. 130 preserve_word_confidence: Bool flag which preserves the history of per-word confidence scores 131 generated during greedy decoding (sample / batched). When set to true, the Hypothesis will contain 132 the non-null value for `word_confidence` in it. Here, `word_confidence` is a List of floats. 133 134 The length of the list corresponds to the number of recognized words. 135 exclude_blank: Bool flag indicating that blank token confidence scores are to be excluded 136 from the `token_confidence`. 137 aggregation: Which aggregation type to use for collapsing per-token confidence into per-word confidence. 138 Valid options are `mean`, `min`, `max`, `prod`. 139 method_cfg: A dict-like object which contains the method name and settings to compute per-frame 140 confidence scores. 141 142 name: The method name (str). 143 Supported values: 144 - 'max_prob' for using the maximum token probability as a confidence. 145 - 'entropy' for using a normalized entropy of a log-likelihood vector. 146 147 entropy_type: Which type of entropy to use (str). Used if confidence_method_cfg.name is set to `entropy`. 148 Supported values: 149 - 'gibbs' for the (standard) Gibbs entropy. If the alpha (ฮฑ) is provided, 150 the formula is the following: H_ฮฑ = -sum_i((p^ฮฑ_i)*log(p^ฮฑ_i)). 151 Note that for this entropy, the alpha should comply the following inequality: 152 (log(V)+2-sqrt(log^2(V)+4))/(2*log(V)) <= ฮฑ <= (1+log(V-1))/log(V-1) 153 where V is the model vocabulary size. 154 - 'tsallis' for the Tsallis entropy with the Boltzmann constant one. 155 Tsallis entropy formula is the following: H_ฮฑ = 1/(ฮฑ-1)*(1-sum_i(p^ฮฑ_i)), 156 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 157 More: https://en.wikipedia.org/wiki/Tsallis_entropy 158 - 'renyi' for the Rรฉnyi entropy. 159 Rรฉnyi entropy formula is the following: H_ฮฑ = 1/(1-ฮฑ)*log_2(sum_i(p^ฮฑ_i)), 160 where ฮฑ is a parameter. When ฮฑ == 1, it works like the Gibbs entropy. 161 More: https://en.wikipedia.org/wiki/R%C3%A9nyi_entropy 162 163 alpha: Power scale for logsoftmax (ฮฑ for entropies). Here we restrict it to be > 0. 164 When the alpha equals one, scaling is not applied to 'max_prob', 165 and any entropy type behaves like the Shannon entropy: H = -sum_i(p_i*log(p_i)) 166 167 entropy_norm: A mapping of the entropy value to the interval [0,1]. 168 Supported values: 169 - 'lin' for using the linear mapping. 170 - 'exp' for using exponential mapping with linear shift. 171 """ 172 173 preserve_frame_confidence: bool = False 174 preserve_token_confidence: bool = False 175 preserve_word_confidence: bool = False 176 exclude_blank: bool = True 177 aggregation: str = "min" 178 method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() 179 180 def __post_init__(self): 181 # OmegaConf.structured ensures that post_init check is always executed 182 self.method_cfg = OmegaConf.structured( 183 self.method_cfg 184 if isinstance(self.method_cfg, ConfidenceMethodConfig) 185 else ConfidenceMethodConfig(**self.method_cfg) 186 ) 187 if self.aggregation not in ConfidenceConstants.AGGREGATIONS: 188 raise ValueError( 189 f"`aggregation` has to be one of the following: " 190 f"{'`' + '`, `'.join(ConfidenceConstants.AGGREGATIONS) + '`'}. Provided: `{self.aggregation}`" 191 ) 192 193 194 def get_confidence_measure_bank(): 195 """Generate a dictionary with confidence measure functionals. 196 197 Supported confidence measures: 198 max_prob: normalized maximum probability 199 entropy_gibbs_lin: Gibbs entropy with linear normalization 200 entropy_gibbs_exp: Gibbs entropy with exponential normalization 201 entropy_tsallis_lin: Tsallis entropy with linear normalization 202 entropy_tsallis_exp: Tsallis entropy with exponential normalization 203 entropy_renyi_lin: Rรฉnyi entropy with linear normalization 204 entropy_renyi_exp: Rรฉnyi entropy with exponential normalization 205 206 Returns: 207 dictionary with lambda functions. 208 """ 209 # helper functions 210 # Gibbs entropy is implemented without alpha 211 neg_entropy_gibbs = lambda x: (x.exp() * x).sum(-1) 212 neg_entropy_alpha = lambda x, t: (x * t).exp().sum(-1) 213 neg_entropy_alpha_gibbs = lambda x, t: ((x * t).exp() * x).sum(-1) 214 # too big for a lambda 215 def entropy_tsallis_exp(x, v, t): 216 exp_neg_max_ent = math.exp((1 - math.pow(v, 1 - t)) / (1 - t)) 217 return (((1 - neg_entropy_alpha(x, t)) / (1 - t)).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 218 219 def entropy_gibbs_exp(x, v, t): 220 exp_neg_max_ent = math.pow(v, -t * math.pow(v, 1 - t)) 221 return ((neg_entropy_alpha_gibbs(x, t) * t).exp() - exp_neg_max_ent) / (1 - exp_neg_max_ent) 222 223 # use Gibbs entropies for Tsallis and Rรฉnyi with t == 1.0 224 entropy_gibbs_lin_baseline = lambda x, v: 1 + neg_entropy_gibbs(x) / math.log(v) 225 entropy_gibbs_exp_baseline = lambda x, v: (neg_entropy_gibbs(x).exp() * v - 1) / (v - 1) 226 # fill the measure bank 227 confidence_measure_bank = {} 228 # Maximum probability measure is implemented without alpha 229 confidence_measure_bank["max_prob"] = ( 230 lambda x, v, t: (x.max(dim=-1)[0].exp() * v - 1) / (v - 1) 231 if t == 1.0 232 else ((x.max(dim=-1)[0] * t).exp() * math.pow(v, t) - 1) / (math.pow(v, t) - 1) 233 ) 234 confidence_measure_bank["entropy_gibbs_lin"] = ( 235 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 236 if t == 1.0 237 else 1 + neg_entropy_alpha_gibbs(x, t) / math.log(v) / math.pow(v, 1 - t) 238 ) 239 confidence_measure_bank["entropy_gibbs_exp"] = ( 240 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_gibbs_exp(x, v, t) 241 ) 242 confidence_measure_bank["entropy_tsallis_lin"] = ( 243 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 244 if t == 1.0 245 else 1 + (1 - neg_entropy_alpha(x, t)) / (math.pow(v, 1 - t) - 1) 246 ) 247 confidence_measure_bank["entropy_tsallis_exp"] = ( 248 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) if t == 1.0 else entropy_tsallis_exp(x, v, t) 249 ) 250 confidence_measure_bank["entropy_renyi_lin"] = ( 251 lambda x, v, t: entropy_gibbs_lin_baseline(x, v) 252 if t == 1.0 253 else 1 + neg_entropy_alpha(x, t).log2() / (t - 1) / math.log(v, 2) 254 ) 255 confidence_measure_bank["entropy_renyi_exp"] = ( 256 lambda x, v, t: entropy_gibbs_exp_baseline(x, v) 257 if t == 1.0 258 else (neg_entropy_alpha(x, t).pow(1 / (t - 1)) * v - 1) / (v - 1) 259 ) 260 return confidence_measure_bank 261 262 263 def get_confidence_aggregation_bank(): 264 """Generate a dictionary with confidence aggregation functions. 265 266 Supported confidence aggregation functions: 267 min: minimum 268 max: maximum 269 mean: arithmetic mean 270 prod: product 271 272 Returns: 273 dictionary with functions. 274 """ 275 confidence_aggregation_bank = {"mean": lambda x: sum(x) / len(x), "min": min, "max": max} 276 # python 3.7 and earlier do not have math.prod 277 if hasattr(math, "prod"): 278 confidence_aggregation_bank["prod"] = math.prod 279 else: 280 import operator 281 from functools import reduce 282 283 confidence_aggregation_bank["prod"] = lambda x: reduce(operator.mul, x, 1) 284 return confidence_aggregation_bank 285 286 287 class ConfidenceMethodMixin(ABC): 288 """Confidence Method Mixin class. 289 290 It initializes per-frame confidence method. 291 """ 292 293 def _init_confidence_method(self, confidence_method_cfg: Optional[DictConfig] = None): 294 """Initialize per-frame confidence method from config. 295 """ 296 # OmegaConf.structured ensures that post_init check is always executed 297 confidence_method_cfg = OmegaConf.structured( 298 ConfidenceMethodConfig() 299 if confidence_method_cfg is None 300 else ConfidenceMethodConfig(**confidence_method_cfg) 301 ) 302 303 # set confidence calculation method 304 # we suppose that self.blank_id == len(vocabulary) 305 self.num_tokens = (self.blank_id if hasattr(self, "blank_id") else self._blank_index) + 1 306 self.alpha = confidence_method_cfg.alpha 307 308 # init confidence measure bank 309 self.confidence_measure_bank = get_confidence_measure_bank() 310 311 measure = None 312 # construct measure_name 313 measure_name = "" 314 if confidence_method_cfg.name == "max_prob": 315 measure_name = "max_prob" 316 elif confidence_method_cfg.name == "entropy": 317 measure_name = '_'.join( 318 [confidence_method_cfg.name, confidence_method_cfg.entropy_type, confidence_method_cfg.entropy_norm] 319 ) 320 else: 321 raise ValueError(f"Unsupported `confidence_method_cfg.name`: `{confidence_method_cfg.name}`") 322 if measure_name not in self.confidence_measure_bank: 323 raise ValueError(f"Unsupported measure setup: `{measure_name}`") 324 measure = partial(self.confidence_measure_bank[measure_name], v=self.num_tokens, t=self.alpha) 325 self._get_confidence = lambda x: measure(torch.nan_to_num(x)).tolist() 326 327 328 class ConfidenceMixin(ABC): 329 """Confidence Mixin class. 330 331 It is responsible for confidence estimation method initialization and high-level confidence score calculation. 332 """ 333 334 def _init_confidence(self, confidence_cfg: Optional[DictConfig] = None): 335 """Initialize confidence-related fields and confidence aggregation function from config. 336 """ 337 # OmegaConf.structured ensures that post_init check is always executed 338 confidence_cfg = OmegaConf.structured( 339 ConfidenceConfig() if confidence_cfg is None else ConfidenceConfig(**confidence_cfg) 340 ) 341 self.confidence_method_cfg = confidence_cfg.method_cfg 342 343 # extract the config 344 self.preserve_word_confidence = confidence_cfg.get('preserve_word_confidence', False) 345 # set preserve_frame_confidence and preserve_token_confidence to True 346 # if preserve_word_confidence is True 347 self.preserve_token_confidence = ( 348 confidence_cfg.get('preserve_token_confidence', False) | self.preserve_word_confidence 349 ) 350 # set preserve_frame_confidence to True if preserve_token_confidence is True 351 self.preserve_frame_confidence = ( 352 confidence_cfg.get('preserve_frame_confidence', False) | self.preserve_token_confidence 353 ) 354 self.exclude_blank_from_confidence = confidence_cfg.get('exclude_blank', True) 355 self.word_confidence_aggregation = confidence_cfg.get('aggregation', "min") 356 357 # define aggregation functions 358 self.confidence_aggregation_bank = get_confidence_aggregation_bank() 359 self._aggregate_confidence = self.confidence_aggregation_bank[self.word_confidence_aggregation] 360 361 # Update preserve frame confidence 362 if self.preserve_frame_confidence is False: 363 if self.cfg.strategy in ['greedy', 'greedy_batch']: 364 self.preserve_frame_confidence = self.cfg.greedy.get('preserve_frame_confidence', False) 365 # OmegaConf.structured ensures that post_init check is always executed 366 confidence_method_cfg = OmegaConf.structured(self.cfg.greedy).get('confidence_method_cfg', None) 367 self.confidence_method_cfg = ( 368 OmegaConf.structured(ConfidenceMethodConfig()) 369 if confidence_method_cfg is None 370 else OmegaConf.structured(ConfidenceMethodConfig(**confidence_method_cfg)) 371 ) 372 373 @abstractmethod 374 def compute_confidence(self, hypotheses_list: List[Hypothesis]) -> List[Hypothesis]: 375 """Computes high-level (per-token and/or per-word) confidence scores for a list of hypotheses. 376 Assumes that `frame_confidence` is present in the hypotheses. 377 378 Args: 379 hypotheses_list: List of Hypothesis. 380 381 Returns: 382 A list of hypotheses with high-level confidence scores. 383 """ 384 raise NotImplementedError() 385 386 @abstractmethod 387 def _aggregate_token_confidence(self, hypothesis: Hypothesis) -> List[float]: 388 """Implemented by subclass in order to aggregate token confidence to a word-level confidence. 389 390 Args: 391 hypothesis: Hypothesis 392 393 Returns: 394 A list of word-level confidence scores. 395 """ 396 raise NotImplementedError() 397 398 def _aggregate_token_confidence_chars(self, words: List[str], token_confidence: List[float]) -> List[float]: 399 """Implementation of token confidence aggregation for character-based models. 400 401 Args: 402 words: List of words of a hypothesis. 403 token_confidence: List of token-level confidence scores of a hypothesis. 404 405 Returns: 406 A list of word-level confidence scores. 407 """ 408 word_confidence = [] 409 i = 0 410 for word in words: 411 word_len = len(word) 412 word_confidence.append(self._aggregate_confidence(token_confidence[i : i + word_len])) 413 # we assume that there is exactly one space token between words and exclude it from word confidence 414 i += word_len + 1 415 return word_confidence 416 417 def _aggregate_token_confidence_subwords_sentencepiece( 418 self, words: List[str], token_confidence: List[float], token_ids: List[int] 419 ) -> List[float]: 420 """Implementation of token confidence aggregation for subword-based models. 421 422 **Note**: Only supports Sentencepiece based tokenizers ! 423 424 Args: 425 words: List of words of a hypothesis. 426 token_confidence: List of token-level confidence scores of a hypothesis. 427 token_ids: List of token ids of a hypothesis. 428 429 Returns: 430 A list of word-level confidence scores. 431 """ 432 word_confidence = [] 433 # run only if there are final words 434 if len(words) > 0: 435 j = 0 436 prev_unk = False 437 prev_underline = False 438 for i, token_id in enumerate(token_ids): 439 token = self.decode_ids_to_tokens([int(token_id)])[0] 440 token_text = self.decode_tokens_to_str([int(token_id)]) 441 # treat `<unk>` as a separate word regardless of the next token 442 # to match the result of `tokenizer.ids_to_text` 443 if (token != token_text or prev_unk) and i > j: 444 # do not add confidence for `โ–` if the current token starts with `โ–` 445 # to match the result of `tokenizer.ids_to_text` 446 if not prev_underline: 447 word_confidence.append(self._aggregate_confidence(token_confidence[j:i])) 448 j = i 449 prev_unk = token == '<unk>' 450 prev_underline = token == 'โ–' 451 if not prev_underline: 452 word_confidence.append(self._aggregate_confidence(token_confidence[j : len(token_ids)])) 453 if len(words) != len(word_confidence): 454 raise RuntimeError( 455 f"""Something went wrong with word-level confidence aggregation.\n 456 Please check these values for debugging:\n 457 len(words): {len(words)},\n 458 len(word_confidence): {len(word_confidence)},\n 459 recognized text: `{' '.join(words)}`""" 460 ) 461 return word_confidence 462 [end of nemo/collections/asr/parts/utils/asr_confidence_utils.py] [start of nemo/collections/common/parts/adapter_modules.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, is_dataclass 16 from typing import Any, Optional 17 18 from hydra.utils import instantiate 19 from omegaconf import OmegaConf 20 from torch import nn as nn 21 22 from nemo.collections.common.parts.utils import activation_registry 23 from nemo.core.classes.mixins import access_mixins, adapter_mixin_strategies 24 25 26 class AdapterModuleUtil(access_mixins.AccessMixin): 27 """ 28 Base class of Adapter Modules, providing common functionality to all Adapter Modules. 29 """ 30 31 def setup_adapter_strategy(self, adapter_strategy: Optional[adapter_mixin_strategies.AbstractAdapterStrategy]): 32 """ 33 Setup adapter strategy of this class, enabling dynamic change in the way the adapter output is 34 merged with the input. 35 36 When called successfully, will assign the variable `adapter_strategy` to the module. 37 38 Args: 39 adapter_strategy: Can be a None or an implementation of AbstractAdapterStrategy. 40 """ 41 # set default adapter strategy 42 if adapter_strategy is None: 43 adapter_strategy = self.get_default_strategy_config() 44 45 if is_dataclass(adapter_strategy): 46 adapter_strategy = OmegaConf.structured(adapter_strategy) 47 OmegaConf.set_struct(adapter_strategy, False) 48 49 # The config must have the `_target_` field pointing to the actual adapter strategy class 50 # which will load that strategy dynamically to this module. 51 if isinstance(adapter_strategy, dict) or OmegaConf.is_config(adapter_strategy): 52 self.adapter_strategy = instantiate(adapter_strategy) 53 elif isinstance(adapter_strategy, adapter_mixin_strategies.AbstractAdapterStrategy): 54 self.adapter_strategy = adapter_strategy 55 else: 56 raise AttributeError(f'`adapter_strategy` provided is invalid : {adapter_strategy}') 57 58 def get_default_strategy_config(self) -> 'dataclass': 59 """ 60 Returns a default adapter module strategy. 61 """ 62 return adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 63 64 def adapter_unfreeze(self,): 65 """ 66 Sets the requires grad for all parameters in the adapter to True. 67 This method should be overridden for any custom unfreeze behavior that is required. 68 For example, if not all params of the adapter should be unfrozen. 69 """ 70 for param in self.parameters(): 71 param.requires_grad_(True) 72 73 74 class LinearAdapter(nn.Module, AdapterModuleUtil): 75 76 """ 77 Simple Linear Feedforward Adapter module with LayerNorm and singe hidden layer with activation function. 78 Note: The adapter explicitly initializes its final layer with all zeros in order to avoid affecting the 79 original model when all adapters are disabled. 80 81 Args: 82 in_features: Input dimension of the module. Note that for adapters, input_dim == output_dim. 83 dim: Hidden dimension of the feed forward network. 84 activation: Str name for an activation function. 85 norm_position: Str, can be `pre` or `post`. Defaults to `pre`. Determines whether the normalization 86 will occur in the first layer or the last layer. Certain architectures may prefer one over the other. 87 dropout: float value, whether to perform dropout on the output of the last layer of the adapter. 88 adapter_strategy: By default, ResidualAddAdapterStrategyConfig. An adapter composition function object. 89 """ 90 91 def __init__( 92 self, 93 in_features: int, 94 dim: int, 95 activation: str = 'swish', 96 norm_position: str = 'pre', 97 dropout: float = 0.0, 98 adapter_strategy: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig = None, 99 ): 100 super().__init__() 101 102 activation = activation_registry[activation]() 103 # If the activation can be executed in place, do so. 104 if hasattr(activation, 'inplace'): 105 activation.inplace = True 106 107 assert norm_position in ['pre', 'post'] 108 self.norm_position = norm_position 109 110 if norm_position == 'pre': 111 self.module = nn.Sequential( 112 nn.LayerNorm(in_features), 113 nn.Linear(in_features, dim, bias=False), 114 activation, 115 nn.Linear(dim, in_features, bias=False), 116 ) 117 118 elif norm_position == 'post': 119 self.module = nn.Sequential( 120 nn.Linear(in_features, dim, bias=False), 121 activation, 122 nn.Linear(dim, in_features, bias=False), 123 nn.LayerNorm(in_features), 124 ) 125 126 if dropout > 0.0: 127 self.dropout = nn.Dropout(dropout) 128 else: 129 self.dropout = None 130 131 # Setup adapter strategy 132 self.setup_adapter_strategy(adapter_strategy) 133 134 # reset parameters 135 self.reset_parameters() 136 137 def reset_parameters(self): 138 # Final layer initializations must be 0 139 if self.norm_position == 'pre': 140 self.module[-1].weight.data *= 0 141 142 elif self.norm_position == 'post': 143 self.module[-1].weight.data *= 0 144 self.module[-1].bias.data *= 0 145 146 def forward(self, x): 147 x = self.module(x) 148 149 # Add dropout if available 150 if self.dropout is not None: 151 x = self.dropout(x) 152 153 return x 154 155 156 @dataclass 157 class LinearAdapterConfig: 158 in_features: int 159 dim: int 160 activation: str = 'swish' 161 norm_position: str = 'pre' 162 dropout: float = 0.0 163 adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() 164 _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) 165 [end of nemo/collections/common/parts/adapter_modules.py] [start of nemo/collections/common/tokenizers/en_ja_tokenizers.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import re 15 from typing import List 16 17 import ipadic 18 import MeCab 19 from pangu import spacing 20 from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer 21 22 23 class EnJaProcessor: 24 """ 25 Tokenizer, Detokenizer and Normalizer utilities for Japanese & English 26 Args: 27 lang_id: One of ['en', 'ja']. 28 """ 29 30 def __init__(self, lang_id: str): 31 self.lang_id = lang_id 32 self.moses_tokenizer = MosesTokenizer(lang=lang_id) 33 self.moses_detokenizer = MosesDetokenizer(lang=lang_id) 34 self.normalizer = MosesPunctNormalizer( 35 lang=lang_id, pre_replace_unicode_punct=True, post_remove_control_chars=True 36 ) 37 38 def detokenize(self, tokens: List[str]) -> str: 39 """ 40 Detokenizes a list of tokens 41 Args: 42 tokens: list of strings as tokens 43 Returns: 44 detokenized Japanese or English string 45 """ 46 return self.moses_detokenizer.detokenize(tokens) 47 48 def tokenize(self, text) -> str: 49 """ 50 Tokenizes text using Moses. Returns a string of tokens. 51 """ 52 tokens = self.moses_tokenizer.tokenize(text) 53 return ' '.join(tokens) 54 55 def normalize(self, text) -> str: 56 # Normalization doesn't handle Japanese periods correctly; 57 # 'ใ€‚'becomes '.'. 58 if self.lang_id == 'en': 59 return self.normalizer.normalize(text) 60 else: 61 return text 62 63 64 class JaMecabProcessor: 65 """ 66 Tokenizer, Detokenizer and Normalizer utilities for Japanese MeCab & English 67 """ 68 69 def __init__(self): 70 self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") 71 72 def detokenize(self, text: List[str]) -> str: 73 RE_WS_IN_FW = re.compile( 74 r'([\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])\s+(?=[\u2018\u2019\u201c\u201d\u2e80-\u312f\u3200-\u32ff\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff\uff00-\uffef])' 75 ) 76 77 detokenize = lambda s: spacing(RE_WS_IN_FW.sub(r'\1', s)).strip() 78 return detokenize(' '.join(text)) 79 80 def tokenize(self, text) -> str: 81 """ 82 Tokenizes text using Moses. Returns a string of tokens. 83 """ 84 return self.mecab_tokenizer.parse(text).strip() 85 86 def normalize(self, text) -> str: 87 return text 88 [end of nemo/collections/common/tokenizers/en_ja_tokenizers.py] [start of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Optional, Tuple 17 18 from omegaconf.omegaconf import MISSING 19 20 from nemo.collections.nlp.data.machine_translation.machine_translation_dataset import TranslationDataConfig 21 from nemo.collections.nlp.models.enc_dec_nlp_model import EncDecNLPModelConfig 22 from nemo.collections.nlp.modules.common.token_classifier import TokenClassifierConfig 23 from nemo.collections.nlp.modules.common.tokenizer_utils import TokenizerConfig 24 from nemo.collections.nlp.modules.common.transformer.transformer import ( 25 NeMoTransformerConfig, 26 NeMoTransformerEncoderConfig, 27 ) 28 from nemo.collections.nlp.modules.common.transformer.transformer_bottleneck import ( 29 NeMoTransformerBottleneckDecoderConfig, 30 NeMoTransformerBottleneckEncoderConfig, 31 ) 32 from nemo.core.config.modelPT import OptimConfig, SchedConfig 33 34 35 @dataclass 36 class MTSchedConfig(SchedConfig): 37 name: str = 'InverseSquareRootAnnealing' 38 warmup_ratio: Optional[float] = None 39 last_epoch: int = -1 40 41 42 # TODO: Refactor this dataclass to to support more optimizers (it pins the optimizer to Adam-like optimizers). 43 @dataclass 44 class MTOptimConfig(OptimConfig): 45 name: str = 'adam' 46 lr: float = 1e-3 47 betas: Tuple[float, float] = (0.9, 0.98) 48 weight_decay: float = 0.0 49 sched: Optional[MTSchedConfig] = MTSchedConfig() 50 51 52 @dataclass 53 class MTEncDecModelConfig(EncDecNLPModelConfig): 54 # machine translation configurations 55 num_val_examples: int = 3 56 num_test_examples: int = 3 57 max_generation_delta: int = 10 58 label_smoothing: Optional[float] = 0.0 59 beam_size: int = 4 60 len_pen: float = 0.0 61 src_language: Any = 'en' # Any = str or List[str] 62 tgt_language: Any = 'en' # Any = str or List[str] 63 find_unused_parameters: Optional[bool] = True 64 shared_tokenizer: Optional[bool] = True 65 multilingual: Optional[bool] = False 66 preproc_out_dir: Optional[str] = None 67 validate_input_ids: Optional[bool] = True 68 shared_embeddings: bool = False 69 70 # network architecture configuration 71 encoder_tokenizer: Any = MISSING 72 encoder: Any = MISSING 73 74 decoder_tokenizer: Any = MISSING 75 decoder: Any = MISSING 76 77 head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) 78 79 # dataset configurations 80 train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 81 src_file_name=MISSING, 82 tgt_file_name=MISSING, 83 tokens_in_batch=512, 84 clean=True, 85 shuffle=True, 86 cache_ids=False, 87 use_cache=False, 88 ) 89 validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 90 src_file_name=MISSING, 91 tgt_file_name=MISSING, 92 tokens_in_batch=512, 93 clean=False, 94 shuffle=False, 95 cache_ids=False, 96 use_cache=False, 97 ) 98 test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( 99 src_file_name=MISSING, 100 tgt_file_name=MISSING, 101 tokens_in_batch=512, 102 clean=False, 103 shuffle=False, 104 cache_ids=False, 105 use_cache=False, 106 ) 107 optim: Optional[OptimConfig] = MTOptimConfig() 108 109 110 @dataclass 111 class AAYNBaseConfig(MTEncDecModelConfig): 112 113 # Attention is All You Need Base Configuration 114 encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 115 decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') 116 117 encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( 118 library='nemo', 119 model_name=None, 120 pretrained=False, 121 hidden_size=512, 122 inner_size=2048, 123 num_layers=6, 124 num_attention_heads=8, 125 ffn_dropout=0.1, 126 attn_score_dropout=0.1, 127 attn_layer_dropout=0.1, 128 ) 129 130 decoder: NeMoTransformerConfig = NeMoTransformerConfig( 131 library='nemo', 132 model_name=None, 133 pretrained=False, 134 hidden_size=512, 135 inner_size=2048, 136 num_layers=6, 137 num_attention_heads=8, 138 ffn_dropout=0.1, 139 attn_score_dropout=0.1, 140 attn_layer_dropout=0.1, 141 ) 142 143 144 @dataclass 145 class MTBottleneckModelConfig(AAYNBaseConfig): 146 model_type: str = 'nll' 147 min_logv: float = -6 148 latent_size: int = -1 # -1 will take value of encoder hidden 149 non_recon_warmup_batches: int = 200000 150 recon_per_token: bool = True 151 log_timing: bool = True 152 153 encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( 154 library='nemo', 155 model_name=None, 156 pretrained=False, 157 hidden_size=512, 158 inner_size=2048, 159 num_layers=6, 160 num_attention_heads=8, 161 ffn_dropout=0.1, 162 attn_score_dropout=0.1, 163 attn_layer_dropout=0.1, 164 arch='seq2seq', 165 hidden_steps=32, 166 hidden_blocks=1, 167 hidden_init_method='params', 168 ) 169 170 decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( 171 library='nemo', 172 model_name=None, 173 pretrained=False, 174 inner_size=2048, 175 num_layers=6, 176 num_attention_heads=8, 177 ffn_dropout=0.1, 178 attn_score_dropout=0.1, 179 attn_layer_dropout=0.1, 180 arch='seq2seq', 181 ) 182 [end of nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py] [start of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass 16 from typing import Any, Dict, Optional 17 18 from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict 19 20 from nemo.collections.common.parts.adapter_modules import LinearAdapterConfig 21 from nemo.collections.nlp.data.token_classification.punctuation_capitalization_dataset import ( 22 PunctuationCapitalizationEvalDataConfig, 23 PunctuationCapitalizationTrainDataConfig, 24 legacy_data_config_to_new_data_config, 25 ) 26 from nemo.core.config import TrainerConfig 27 from nemo.core.config.modelPT import NemoConfig 28 from nemo.utils.exp_manager import ExpManagerConfig 29 30 31 @dataclass 32 class FreezeConfig: 33 is_enabled: bool = False 34 """Freeze audio encoder weight and add Conformer Layers on top of it""" 35 d_model: Optional[int] = 256 36 """`d_model` parameter of ``ConformerLayer``""" 37 d_ff: Optional[int] = 1024 38 """``d_ff`` parameter of ``ConformerLayer``""" 39 num_layers: Optional[int] = 8 40 """``num_layers`` number of ``ConformerLayer`` modules to add on top of audio encoder""" 41 42 43 @dataclass 44 class AdapterConfig: 45 config: Optional[LinearAdapterConfig] = None 46 """Linear adapter config see ``collections.common.parts.LinearAdapterConfig``""" 47 enable: bool = False 48 """Use adapters for audio encoder""" 49 50 51 @dataclass 52 class FusionConfig: 53 num_layers: Optional[int] = 4 54 """"Number of layers to use in fusion""" 55 num_attention_heads: Optional[int] = 4 56 """Number of attention heads to use in fusion""" 57 inner_size: Optional[int] = 2048 58 """Fusion inner size""" 59 60 61 @dataclass 62 class AudioEncoderConfig: 63 pretrained_model: str = MISSING 64 """A configuration for restoring pretrained audio encoder""" 65 freeze: Optional[FreezeConfig] = None 66 adapter: Optional[AdapterConfig] = None 67 fusion: Optional[FusionConfig] = None 68 69 70 @dataclass 71 class TokenizerConfig: 72 """A structure and default values of source text tokenizer.""" 73 74 vocab_file: Optional[str] = None 75 """A path to vocabulary file which is used in ``'word'``, ``'char'``, and HuggingFace tokenizers""" 76 77 tokenizer_name: str = MISSING 78 """A name of the tokenizer used for tokenization of source sequences. Possible options are ``'sentencepiece'``, 79 ``'word'``, ``'char'``, HuggingFace tokenizers (e.g. ``'bert-base-uncased'``). For more options see function 80 ``nemo.collections.nlp.modules.common.get_tokenizer``. The tokenizer must have properties ``cls_id``, ``pad_id``, 81 ``sep_id``, ``unk_id``.""" 82 83 special_tokens: Optional[Dict[str, str]] = None 84 """A dictionary with special tokens passed to constructors of ``'char'``, ``'word'``, ``'sentencepiece'``, and 85 various HuggingFace tokenizers.""" 86 87 tokenizer_model: Optional[str] = None 88 """A path to a tokenizer model required for ``'sentencepiece'`` tokenizer.""" 89 90 91 @dataclass 92 class LanguageModelConfig: 93 """ 94 A structure and default values of language model configuration of punctuation and capitalization model. BERT like 95 HuggingFace models are supported. Provide a valid ``pretrained_model_name`` and, optionally, you may 96 reinitialize model via ``config_file`` or ``config``. 97 98 Alternatively you can initialize the language model using ``lm_checkpoint``. 99 100 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 101 """ 102 103 pretrained_model_name: str = MISSING 104 """A mandatory parameter containing name of HuggingFace pretrained model. For example, ``'bert-base-uncased'``.""" 105 106 config_file: Optional[str] = None 107 """A path to a file with HuggingFace model config which is used to reinitialize language model.""" 108 109 config: Optional[Dict] = None 110 """A HuggingFace config which is used to reinitialize language model.""" 111 112 lm_checkpoint: Optional[str] = None 113 """A path to a ``torch`` checkpoint of a language model.""" 114 115 116 @dataclass 117 class HeadConfig: 118 """ 119 A structure and default values of configuration of capitalization or punctuation model head. This config defines a 120 multilayer perceptron which is applied to output of a language model. Number of units in the hidden layer is equal 121 to the dimension of the language model. 122 123 This config is a part of :class:`PunctuationCapitalizationModelConfig` config. 124 """ 125 126 num_fc_layers: int = 1 127 """A number of hidden layers in a multilayer perceptron.""" 128 129 fc_dropout: float = 0.1 130 """A dropout used in an MLP.""" 131 132 activation: str = 'relu' 133 """An activation used in hidden layers.""" 134 135 use_transformer_init: bool = True 136 """Whether to initialize the weights of the classifier head with the approach that was used for language model 137 initialization.""" 138 139 140 @dataclass 141 class ClassLabelsConfig: 142 """ 143 A structure and default values of a mandatory part of config which contains names of files which are saved in .nemo 144 checkpoint. These files can also be used for passing label vocabulary to the model. For using them as label 145 vocabularies you will need to provide path these files in parameter 146 ``model.common_dataset_parameters.label_vocab_dir``. Each line in labels files 147 contains 1 label. The values are sorted, ``<line number>==<label id>``, starting from ``0``. A label with ``0`` id 148 must contain neutral label which must be equal to ``model.common_dataset_parameters.pad_label``. 149 150 This config is a part of :class:`~CommonDatasetParametersConfig`. 151 """ 152 153 punct_labels_file: str = MISSING 154 """A name of punctuation labels file.""" 155 156 capit_labels_file: str = MISSING 157 """A name of capitalization labels file.""" 158 159 160 @dataclass 161 class CommonDatasetParametersConfig: 162 """ 163 A structure and default values of common dataset parameters config which includes label and loss mask information. 164 If you omit parameters ``punct_label_ids``, ``capit_label_ids``, ``label_vocab_dir``, then labels will be inferred 165 from a training dataset or loaded from a checkpoint. 166 167 Parameters ``ignore_extra_tokens`` and ``ignore_start_end`` are responsible for forming loss mask. A loss mask 168 defines on which tokens loss is computed. 169 170 This parameter is a part of config :class:`~PunctuationCapitalizationModelConfig`. 171 """ 172 173 pad_label: str = MISSING 174 """A mandatory parameter which should contain label used for punctuation and capitalization label padding. It 175 also serves as a neutral label for both punctuation and capitalization. If any of ``punct_label_ids``, 176 ``capit_label_ids`` parameters is provided, then ``pad_label`` must have ``0`` id in them. In addition, if ``label_vocab_dir`` 177 is provided, then ``pad_label`` must be on the first lines in files ``class_labels.punct_labels_file`` and 178 ``class_labels.capit_labels_file``.""" 179 180 ignore_extra_tokens: bool = False 181 """Whether to compute loss on not first tokens in words. If this parameter is ``True``, then loss mask is ``False`` 182 for all tokens in a word except the first.""" 183 184 ignore_start_end: bool = True 185 """If ``False``, then loss is computed on [CLS] and [SEP] tokens.""" 186 187 punct_label_ids: Optional[Dict[str, int]] = None 188 """A dictionary with punctuation label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit this 189 parameter and pass label ids through ``class_labels.punct_labels_file`` or let the model to infer label ids from 190 dataset or load them from checkpoint.""" 191 192 capit_label_ids: Optional[Dict[str, int]] = None 193 """A dictionary with capitalization label ids. ``pad_label`` must have ``0`` id in this dictionary. You can omit 194 this parameter and pass label ids through ``class_labels.capit_labels_file`` or let model to infer label ids from 195 dataset or load them from checkpoint.""" 196 197 label_vocab_dir: Optional[str] = None 198 """A path to directory which contains class labels files. See :class:`ClassLabelsConfig`. If this parameter is 199 provided, then labels will be loaded from files which are located in ``label_vocab_dir`` and have names specified 200 in ``model.class_labels`` configuration section. A label specified in ``pad_label`` has to be on the first lines 201 of ``model.class_labels`` files.""" 202 203 204 @dataclass 205 class PunctuationCapitalizationModelConfig: 206 """ 207 A configuration of 208 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 209 model. 210 211 See an example of model config in 212 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 213 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 214 215 This config is a part of :class:`~PunctuationCapitalizationConfig`. 216 """ 217 218 class_labels: ClassLabelsConfig = ClassLabelsConfig() 219 """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. 220 These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` 221 for passing vocabularies, please provide path to vocabulary files in 222 ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" 223 224 common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() 225 """Label ids and loss mask information information.""" 226 227 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 228 """A configuration for creating training dataset and data loader.""" 229 230 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 231 """A configuration for creating validation datasets and data loaders.""" 232 233 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 234 """A configuration for creating test datasets and data loaders.""" 235 236 punct_head: HeadConfig = HeadConfig() 237 """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" 238 239 capit_head: HeadConfig = HeadConfig() 240 """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" 241 242 tokenizer: Any = TokenizerConfig() 243 """A configuration for source text tokenizer.""" 244 245 language_model: LanguageModelConfig = LanguageModelConfig() 246 """A configuration of a BERT-like language model which serves as a model body.""" 247 248 optim: Optional[Any] = None 249 """A configuration of optimizer and learning rate scheduler. There is much variability in such config. For 250 description see `Optimizers 251 <https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/core/core.html#optimizers>`_ section in 252 documentation and `primer <https://github.com/NVIDIA/NeMo/blob/main/tutorials/00_NeMo_Primer.ipynb>_ tutorial.""" 253 254 255 @dataclass 256 class PunctuationCapitalizationLexicalAudioModelConfig(PunctuationCapitalizationModelConfig): 257 """ 258 A configuration of 259 :class:`~nemo.collections.nlp.models.token_classification.punctuation_lexical_audio_capitalization_model.PunctuationCapitalizationLexicalAudioModel` 260 model. 261 262 See an example of model config in 263 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 264 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_lexical_audio_config.yaml>`_ 265 266 Audio encoder can be frozen during training with ``freeze_audio_encoder`` parameter. 267 Adapter can be added to audio encoder with ``use_adapters`` and ``adapter_config`` parameters. 268 More conformer layers can be added on top of pretrained audio encoder with ``frozen_conf_d_model``, ``frozen_conf_d_ff`` and ``frozen_conf_num_layers`` parameters. 269 """ 270 271 train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None 272 """A configuration for creating training dataset and data loader.""" 273 274 validation_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 275 """A configuration for creating validation datasets and data loaders.""" 276 277 test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None 278 """A configuration for creating test datasets and data loaders.""" 279 280 audio_encoder: Optional[AudioEncoderConfig] = None 281 282 restore_lexical_encoder_from: Optional[str] = None 283 """"Path to .nemo checkpoint to load weights from""" # add more comments 284 285 use_weighted_loss: Optional[bool] = False 286 """If set to ``True`` CrossEntropyLoss will be weighted""" 287 288 289 @dataclass 290 class PunctuationCapitalizationConfig(NemoConfig): 291 """ 292 A config for punctuation model training and testing. 293 294 See an example of full config in 295 `nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml 296 <https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_ 297 """ 298 299 pretrained_model: Optional[str] = None 300 """Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options 301 by calling method 302 :func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`. 303 """ 304 305 name: Optional[str] = 'Punctuation_and_Capitalization' 306 """A name of the model. Used for naming output directories and ``.nemo`` checkpoints.""" 307 308 do_training: bool = True 309 """Whether to perform training of the model.""" 310 311 do_testing: bool = False 312 """Whether ot perform testing of the model.""" 313 314 model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() 315 """A configuration for the 316 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` 317 model.""" 318 319 trainer: Optional[TrainerConfig] = TrainerConfig() 320 """Contains ``Trainer`` Lightning class constructor parameters.""" 321 322 exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) 323 """A configuration with various NeMo training options such as output directories, resuming from checkpoint, 324 tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" 325 326 327 @dataclass 328 class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): 329 model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() 330 331 332 def is_legacy_model_config(model_cfg: DictConfig) -> bool: 333 """ 334 Test if model config is old style config. Old style configs are configs which were used before 335 ``common_dataset_parameters`` item was added. Old style datasets use ``dataset`` instead of 336 ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. Old style configs do not support 337 tarred datasets. 338 339 Args: 340 model_cfg: model configuration 341 342 Returns: 343 whether ``model_config`` is legacy 344 """ 345 return 'common_dataset_parameters' not in model_cfg 346 347 348 def legacy_model_config_to_new_model_config(model_cfg: DictConfig) -> DictConfig: 349 """ 350 Transform old style config into 351 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig`. 352 Old style configs are configs which were used before ``common_dataset_parameters`` item was added. Old style 353 datasets use ``dataset`` instead of ``common_dataset_parameters``, ``batch_size`` instead of ``tokens_in_batch``. 354 Old style configs do not support tarred datasets. 355 356 Args: 357 model_cfg: old style config 358 359 Returns: 360 model config which follows dataclass 361 :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_config.PunctuationCapitalizationModelConfig` 362 """ 363 train_ds = model_cfg.get('train_ds') 364 validation_ds = model_cfg.get('validation_ds') 365 test_ds = model_cfg.get('test_ds') 366 dataset = model_cfg.dataset 367 punct_head_config = model_cfg.get('punct_head', {}) 368 capit_head_config = model_cfg.get('capit_head', {}) 369 omega_conf = OmegaConf.structured( 370 PunctuationCapitalizationModelConfig( 371 class_labels=model_cfg.class_labels, 372 common_dataset_parameters=CommonDatasetParametersConfig( 373 pad_label=dataset.pad_label, 374 ignore_extra_tokens=dataset.get( 375 'ignore_extra_tokens', CommonDatasetParametersConfig.ignore_extra_tokens 376 ), 377 ignore_start_end=dataset.get('ignore_start_end', CommonDatasetParametersConfig.ignore_start_end), 378 punct_label_ids=model_cfg.punct_label_ids, 379 capit_label_ids=model_cfg.capit_label_ids, 380 ), 381 train_ds=None 382 if train_ds is None 383 else legacy_data_config_to_new_data_config(train_ds, dataset, train=True), 384 validation_ds=None 385 if validation_ds is None 386 else legacy_data_config_to_new_data_config(validation_ds, dataset, train=False), 387 test_ds=None if test_ds is None else legacy_data_config_to_new_data_config(test_ds, dataset, train=False), 388 punct_head=HeadConfig( 389 num_fc_layers=punct_head_config.get('punct_num_fc_layers', HeadConfig.num_fc_layers), 390 fc_dropout=punct_head_config.get('fc_dropout', HeadConfig.fc_dropout), 391 activation=punct_head_config.get('activation', HeadConfig.activation), 392 use_transformer_init=punct_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 393 ), 394 capit_head=HeadConfig( 395 num_fc_layers=capit_head_config.get('capit_num_fc_layers', HeadConfig.num_fc_layers), 396 fc_dropout=capit_head_config.get('fc_dropout', HeadConfig.fc_dropout), 397 activation=capit_head_config.get('activation', HeadConfig.activation), 398 use_transformer_init=capit_head_config.get('use_transformer_init', HeadConfig.use_transformer_init), 399 ), 400 tokenizer=model_cfg.tokenizer, 401 language_model=model_cfg.language_model, 402 optim=model_cfg.optim, 403 ) 404 ) 405 with open_dict(omega_conf): 406 retain_during_legacy_conversion = model_cfg.get('retain_during_legacy_conversion', {}) 407 for key in retain_during_legacy_conversion.keys(): 408 omega_conf[key] = retain_during_legacy_conversion[key] 409 return omega_conf 410 [end of nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py] [start of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """Transformer based language model.""" 16 from MeCab import Model 17 from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule 18 from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule 19 from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( 20 MegatronRetrievalTransformerEncoderModule, 21 ) 22 from nemo.collections.nlp.modules.common.megatron.utils import ( 23 ApexGuardDefaults, 24 init_method_normal, 25 scaled_init_method_normal, 26 ) 27 28 try: 29 from apex.transformer.enums import AttnMaskType, ModelType 30 31 HAVE_APEX = True 32 except (ImportError, ModuleNotFoundError): 33 HAVE_APEX = False 34 # fake missing classes with None attributes 35 AttnMaskType = ApexGuardDefaults() 36 ModelType = ApexGuardDefaults() 37 38 try: 39 from megatron.core import ModelParallelConfig 40 41 HAVE_MEGATRON_CORE = True 42 43 except (ImportError, ModuleNotFoundError): 44 45 ModelParallelConfig = ApexGuardDefaults 46 47 HAVE_MEGATRON_CORE = False 48 49 __all__ = [] 50 51 AVAILABLE_ENCODERS = ["transformer", "perceiver", "retro"] 52 53 54 def get_encoder_model( 55 config: ModelParallelConfig, 56 arch, 57 hidden_size, 58 ffn_hidden_size, 59 num_layers, 60 num_attention_heads, 61 apply_query_key_layer_scaling=False, 62 kv_channels=None, 63 init_method=None, 64 scaled_init_method=None, 65 encoder_attn_mask_type=AttnMaskType.padding, 66 pre_process=True, 67 post_process=True, 68 init_method_std=0.02, 69 megatron_amp_O2=False, 70 hidden_dropout=0.1, 71 attention_dropout=0.1, 72 ffn_dropout=0.0, 73 precision=16, 74 fp32_residual_connection=False, 75 activations_checkpoint_method=None, 76 activations_checkpoint_num_layers=1, 77 activations_checkpoint_granularity=None, 78 layernorm_epsilon=1e-5, 79 bias_activation_fusion=True, 80 bias_dropout_add_fusion=True, 81 masked_softmax_fusion=True, 82 persist_layer_norm=False, 83 openai_gelu=False, 84 activation="gelu", 85 onnx_safe=False, 86 bias=True, 87 normalization="layernorm", 88 headscale=False, 89 transformer_block_type="pre_ln", 90 hidden_steps=32, 91 parent_model_type=ModelType.encoder_or_decoder, 92 layer_type=None, 93 chunk_size=64, 94 num_self_attention_per_cross_attention=1, 95 layer_number_offset=0, # this is use only for attention norm_factor scaling 96 megatron_legacy=False, 97 normalize_attention_scores=True, 98 sequence_parallel=False, 99 num_moe_experts=1, 100 moe_frequency=1, 101 moe_dropout=0.0, 102 turn_off_rop=False, # turn off the RoP positional embedding 103 version=1, # model version 104 position_embedding_type='learned_absolute', 105 use_flash_attention=False, 106 ): 107 """Build language model and return along with the key to save.""" 108 109 if kv_channels is None: 110 assert ( 111 hidden_size % num_attention_heads == 0 112 ), 'hidden_size must be divisible by num_attention_heads if kv_channels is None' 113 kv_channels = hidden_size // num_attention_heads 114 115 if init_method is None: 116 init_method = init_method_normal(init_method_std) 117 118 if scaled_init_method is None: 119 scaled_init_method = scaled_init_method_normal(init_method_std, num_layers) 120 121 if arch == "transformer": 122 # Language encoder. 123 encoder = MegatronTransformerEncoderModule( 124 config=config, 125 init_method=init_method, 126 output_layer_init_method=scaled_init_method, 127 hidden_size=hidden_size, 128 num_layers=num_layers, 129 num_attention_heads=num_attention_heads, 130 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 131 kv_channels=kv_channels, 132 ffn_hidden_size=ffn_hidden_size, 133 encoder_attn_mask_type=encoder_attn_mask_type, 134 pre_process=pre_process, 135 post_process=post_process, 136 megatron_amp_O2=megatron_amp_O2, 137 hidden_dropout=hidden_dropout, 138 attention_dropout=attention_dropout, 139 ffn_dropout=ffn_dropout, 140 precision=precision, 141 fp32_residual_connection=fp32_residual_connection, 142 activations_checkpoint_method=activations_checkpoint_method, 143 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 144 activations_checkpoint_granularity=activations_checkpoint_granularity, 145 layernorm_epsilon=layernorm_epsilon, 146 bias_activation_fusion=bias_activation_fusion, 147 bias_dropout_add_fusion=bias_dropout_add_fusion, 148 masked_softmax_fusion=masked_softmax_fusion, 149 persist_layer_norm=persist_layer_norm, 150 openai_gelu=openai_gelu, 151 onnx_safe=onnx_safe, 152 activation=activation, 153 bias=bias, 154 normalization=normalization, 155 transformer_block_type=transformer_block_type, 156 headscale=headscale, 157 parent_model_type=parent_model_type, 158 megatron_legacy=megatron_legacy, 159 normalize_attention_scores=normalize_attention_scores, 160 num_moe_experts=num_moe_experts, 161 moe_frequency=moe_frequency, 162 moe_dropout=moe_dropout, 163 position_embedding_type=position_embedding_type, 164 use_flash_attention=use_flash_attention, 165 ) 166 elif arch == "retro": 167 encoder = MegatronRetrievalTransformerEncoderModule( 168 config=config, 169 init_method=init_method, 170 output_layer_init_method=scaled_init_method, 171 hidden_size=hidden_size, 172 num_layers=num_layers, 173 num_attention_heads=num_attention_heads, 174 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 175 kv_channels=kv_channels, 176 layer_type=layer_type, 177 ffn_hidden_size=ffn_hidden_size, 178 pre_process=pre_process, 179 post_process=post_process, 180 megatron_amp_O2=megatron_amp_O2, 181 hidden_dropout=hidden_dropout, 182 attention_dropout=attention_dropout, 183 precision=precision, 184 fp32_residual_connection=fp32_residual_connection, 185 activations_checkpoint_method=activations_checkpoint_method, 186 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 187 activations_checkpoint_granularity=activations_checkpoint_granularity, 188 layernorm_epsilon=layernorm_epsilon, 189 bias_activation_fusion=bias_activation_fusion, 190 bias_dropout_add_fusion=bias_dropout_add_fusion, 191 masked_softmax_fusion=masked_softmax_fusion, 192 persist_layer_norm=persist_layer_norm, 193 openai_gelu=openai_gelu, 194 onnx_safe=onnx_safe, 195 activation=activation, 196 bias=bias, 197 normalization=normalization, 198 transformer_block_type=transformer_block_type, 199 parent_model_type=parent_model_type, 200 chunk_size=chunk_size, 201 layer_number_offset=layer_number_offset, 202 megatron_legacy=megatron_legacy, 203 normalize_attention_scores=normalize_attention_scores, 204 turn_off_rop=turn_off_rop, 205 version=version, 206 ) 207 elif arch == "perceiver": 208 encoder = MegatronPerceiverEncoderModule( 209 config=config, 210 init_method=init_method, 211 output_layer_init_method=scaled_init_method, 212 hidden_size=hidden_size, 213 num_layers=num_layers, 214 num_attention_heads=num_attention_heads, 215 apply_query_key_layer_scaling=apply_query_key_layer_scaling, 216 kv_channels=kv_channels, 217 ffn_hidden_size=ffn_hidden_size, 218 encoder_attn_mask_type=encoder_attn_mask_type, 219 pre_process=pre_process, 220 post_process=post_process, 221 megatron_amp_O2=megatron_amp_O2, 222 hidden_dropout=hidden_dropout, 223 attention_dropout=attention_dropout, 224 ffn_dropout=ffn_dropout, 225 precision=precision, 226 fp32_residual_connection=fp32_residual_connection, 227 activations_checkpoint_method=activations_checkpoint_method, 228 activations_checkpoint_num_layers=activations_checkpoint_num_layers, 229 activations_checkpoint_granularity=activations_checkpoint_granularity, 230 layernorm_epsilon=layernorm_epsilon, 231 bias_activation_fusion=bias_activation_fusion, 232 bias_dropout_add_fusion=bias_dropout_add_fusion, 233 masked_softmax_fusion=masked_softmax_fusion, 234 persist_layer_norm=persist_layer_norm, 235 openai_gelu=openai_gelu, 236 onnx_safe=onnx_safe, 237 activation=activation, 238 bias=bias, 239 normalization=normalization, 240 transformer_block_type=transformer_block_type, 241 headscale=headscale, 242 parent_model_type=parent_model_type, 243 hidden_steps=hidden_steps, 244 num_self_attention_per_cross_attention=num_self_attention_per_cross_attention, 245 megatron_legacy=megatron_legacy, 246 normalize_attention_scores=normalize_attention_scores, 247 ) 248 else: 249 raise ValueError(f"Unknown encoder arch = {arch}. Available encoder arch = {AVAILABLE_ENCODERS}") 250 251 return encoder 252 [end of nemo/collections/nlp/modules/common/megatron/megatron_encoders.py] [start of nemo/collections/tts/models/fastpitch.py] 1 # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 import contextlib 15 from dataclasses import dataclass 16 from pathlib import Path 17 from typing import List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import DictConfig, OmegaConf, open_dict 22 from pytorch_lightning import Trainer 23 from pytorch_lightning.loggers import TensorBoardLogger 24 25 from nemo.collections.common.parts.preprocessing import parsers 26 from nemo.collections.tts.losses.aligner_loss import BinLoss, ForwardSumLoss 27 from nemo.collections.tts.losses.fastpitchloss import DurationLoss, EnergyLoss, MelLoss, PitchLoss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.modules.fastpitch import FastPitchModule 30 from nemo.collections.tts.parts.mixins import FastPitchAdapterModelMixin 31 from nemo.collections.tts.parts.utils.callbacks import LoggingCallback 32 from nemo.collections.tts.parts.utils.helpers import ( 33 batch_from_ragged, 34 g2p_backward_compatible_support, 35 plot_alignment_to_numpy, 36 plot_spectrogram_to_numpy, 37 process_batch, 38 sample_tts_input, 39 ) 40 from nemo.core.classes import Exportable 41 from nemo.core.classes.common import PretrainedModelInfo, typecheck 42 from nemo.core.neural_types.elements import ( 43 Index, 44 LengthsType, 45 MelSpectrogramType, 46 ProbsType, 47 RegressionValuesType, 48 TokenDurationType, 49 TokenIndex, 50 TokenLogDurationType, 51 ) 52 from nemo.core.neural_types.neural_type import NeuralType 53 from nemo.utils import logging, model_utils 54 55 56 @dataclass 57 class G2PConfig: 58 _target_: str = "nemo.collections.tts.g2p.models.en_us_arpabet.EnglishG2p" 59 phoneme_dict: str = "scripts/tts_dataset_files/cmudict-0.7b_nv22.10" 60 heteronyms: str = "scripts/tts_dataset_files/heteronyms-052722" 61 phoneme_probability: float = 0.5 62 63 64 @dataclass 65 class TextTokenizer: 66 _target_: str = "nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers.EnglishPhonemesTokenizer" 67 punct: bool = True 68 stresses: bool = True 69 chars: bool = True 70 apostrophe: bool = True 71 pad_with_space: bool = True 72 add_blank_at: bool = True 73 g2p: G2PConfig = G2PConfig() 74 75 76 @dataclass 77 class TextTokenizerConfig: 78 text_tokenizer: TextTokenizer = TextTokenizer() 79 80 81 class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): 82 """FastPitch model (https://arxiv.org/abs/2006.06873) that is used to generate mel spectrogram from text.""" 83 84 def __init__(self, cfg: DictConfig, trainer: Trainer = None): 85 # Convert to Hydra 1.0 compatible DictConfig 86 cfg = model_utils.convert_model_config_to_dict_config(cfg) 87 cfg = model_utils.maybe_update_config_version(cfg) 88 89 # Setup normalizer 90 self.normalizer = None 91 self.text_normalizer_call = None 92 self.text_normalizer_call_kwargs = {} 93 self._setup_normalizer(cfg) 94 95 self.learn_alignment = cfg.get("learn_alignment", False) 96 97 # Setup vocabulary (=tokenizer) and input_fft_kwargs (supported only with self.learn_alignment=True) 98 input_fft_kwargs = {} 99 if self.learn_alignment: 100 self.vocab = None 101 102 self.ds_class = cfg.train_ds.dataset._target_ 103 self.ds_class_name = self.ds_class.split(".")[-1] 104 if not self.ds_class in [ 105 "nemo.collections.tts.data.dataset.TTSDataset", 106 "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset", 107 "nemo.collections.tts.torch.data.TTSDataset", 108 ]: 109 raise ValueError(f"Unknown dataset class: {self.ds_class}.") 110 111 self._setup_tokenizer(cfg) 112 assert self.vocab is not None 113 input_fft_kwargs["n_embed"] = len(self.vocab.tokens) 114 input_fft_kwargs["padding_idx"] = self.vocab.pad 115 116 self._parser = None 117 self._tb_logger = None 118 super().__init__(cfg=cfg, trainer=trainer) 119 120 self.bin_loss_warmup_epochs = cfg.get("bin_loss_warmup_epochs", 100) 121 self.log_images = cfg.get("log_images", False) 122 self.log_train_images = False 123 124 default_prosody_loss_scale = 0.1 if self.learn_alignment else 1.0 125 dur_loss_scale = cfg.get("dur_loss_scale", default_prosody_loss_scale) 126 pitch_loss_scale = cfg.get("pitch_loss_scale", default_prosody_loss_scale) 127 energy_loss_scale = cfg.get("energy_loss_scale", default_prosody_loss_scale) 128 129 self.mel_loss_fn = MelLoss() 130 self.pitch_loss_fn = PitchLoss(loss_scale=pitch_loss_scale) 131 self.duration_loss_fn = DurationLoss(loss_scale=dur_loss_scale) 132 self.energy_loss_fn = EnergyLoss(loss_scale=energy_loss_scale) 133 134 self.aligner = None 135 if self.learn_alignment: 136 aligner_loss_scale = cfg.get("aligner_loss_scale", 1.0) 137 self.aligner = instantiate(self._cfg.alignment_module) 138 self.forward_sum_loss_fn = ForwardSumLoss(loss_scale=aligner_loss_scale) 139 self.bin_loss_fn = BinLoss(loss_scale=aligner_loss_scale) 140 141 self.preprocessor = instantiate(self._cfg.preprocessor) 142 input_fft = instantiate(self._cfg.input_fft, **input_fft_kwargs) 143 output_fft = instantiate(self._cfg.output_fft) 144 duration_predictor = instantiate(self._cfg.duration_predictor) 145 pitch_predictor = instantiate(self._cfg.pitch_predictor) 146 speaker_encoder = instantiate(self._cfg.get("speaker_encoder", None)) 147 energy_embedding_kernel_size = cfg.get("energy_embedding_kernel_size", 0) 148 energy_predictor = instantiate(self._cfg.get("energy_predictor", None)) 149 150 # [TODO] may remove if we change the pre-trained config 151 # cfg: condition_types = [ "add" ] 152 n_speakers = cfg.get("n_speakers", 0) 153 speaker_emb_condition_prosody = cfg.get("speaker_emb_condition_prosody", False) 154 speaker_emb_condition_decoder = cfg.get("speaker_emb_condition_decoder", False) 155 speaker_emb_condition_aligner = cfg.get("speaker_emb_condition_aligner", False) 156 min_token_duration = cfg.get("min_token_duration", 0) 157 use_log_energy = cfg.get("use_log_energy", True) 158 if n_speakers > 1 and "add" not in input_fft.cond_input.condition_types: 159 input_fft.cond_input.condition_types.append("add") 160 if speaker_emb_condition_prosody: 161 duration_predictor.cond_input.condition_types.append("add") 162 pitch_predictor.cond_input.condition_types.append("add") 163 if speaker_emb_condition_decoder: 164 output_fft.cond_input.condition_types.append("add") 165 if speaker_emb_condition_aligner and self.aligner is not None: 166 self.aligner.cond_input.condition_types.append("add") 167 168 self.fastpitch = FastPitchModule( 169 input_fft, 170 output_fft, 171 duration_predictor, 172 pitch_predictor, 173 energy_predictor, 174 self.aligner, 175 speaker_encoder, 176 n_speakers, 177 cfg.symbols_embedding_dim, 178 cfg.pitch_embedding_kernel_size, 179 energy_embedding_kernel_size, 180 cfg.n_mel_channels, 181 min_token_duration, 182 cfg.max_token_duration, 183 use_log_energy, 184 ) 185 self._input_types = self._output_types = None 186 self.export_config = { 187 "emb_range": (0, self.fastpitch.encoder.word_emb.num_embeddings), 188 "enable_volume": False, 189 "enable_ragged_batches": False, 190 } 191 if self.fastpitch.speaker_emb is not None: 192 self.export_config["num_speakers"] = cfg.n_speakers 193 194 self.log_config = cfg.get("log_config", None) 195 196 # Adapter modules setup (from FastPitchAdapterModelMixin) 197 self.setup_adapters() 198 199 def _get_default_text_tokenizer_conf(self): 200 text_tokenizer: TextTokenizerConfig = TextTokenizerConfig() 201 return OmegaConf.create(OmegaConf.to_yaml(text_tokenizer)) 202 203 def _setup_normalizer(self, cfg): 204 if "text_normalizer" in cfg: 205 normalizer_kwargs = {} 206 207 if "whitelist" in cfg.text_normalizer: 208 normalizer_kwargs["whitelist"] = self.register_artifact( 209 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 210 ) 211 try: 212 import nemo_text_processing 213 214 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 215 except Exception as e: 216 logging.error(e) 217 raise ImportError( 218 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 219 ) 220 221 self.text_normalizer_call = self.normalizer.normalize 222 if "text_normalizer_call_kwargs" in cfg: 223 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 224 225 def _setup_tokenizer(self, cfg): 226 text_tokenizer_kwargs = {} 227 228 if "g2p" in cfg.text_tokenizer: 229 # for backward compatibility 230 if ( 231 self._is_model_being_restored() 232 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 233 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 234 ): 235 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 236 cfg.text_tokenizer.g2p["_target_"] 237 ) 238 239 g2p_kwargs = {} 240 241 if "phoneme_dict" in cfg.text_tokenizer.g2p: 242 g2p_kwargs["phoneme_dict"] = self.register_artifact( 243 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 244 ) 245 246 if "heteronyms" in cfg.text_tokenizer.g2p: 247 g2p_kwargs["heteronyms"] = self.register_artifact( 248 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 249 ) 250 251 # for backward compatability 252 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 253 254 # TODO @xueyang: rename the instance of tokenizer because vocab is misleading. 255 self.vocab = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 256 257 @property 258 def tb_logger(self): 259 if self._tb_logger is None: 260 if self.logger is None and self.logger.experiment is None: 261 return None 262 tb_logger = self.logger.experiment 263 for logger in self.trainer.loggers: 264 if isinstance(logger, TensorBoardLogger): 265 tb_logger = logger.experiment 266 break 267 self._tb_logger = tb_logger 268 return self._tb_logger 269 270 @property 271 def parser(self): 272 if self._parser is not None: 273 return self._parser 274 275 if self.learn_alignment: 276 self._parser = self.vocab.encode 277 else: 278 self._parser = parsers.make_parser( 279 labels=self._cfg.labels, 280 name='en', 281 unk_id=-1, 282 blank_id=-1, 283 do_normalize=True, 284 abbreviation_version="fastpitch", 285 make_table=False, 286 ) 287 return self._parser 288 289 def parse(self, str_input: str, normalize=True) -> torch.tensor: 290 if self.training: 291 logging.warning("parse() is meant to be called in eval mode.") 292 293 if normalize and self.text_normalizer_call is not None: 294 str_input = self.text_normalizer_call(str_input, **self.text_normalizer_call_kwargs) 295 296 if self.learn_alignment: 297 eval_phon_mode = contextlib.nullcontext() 298 if hasattr(self.vocab, "set_phone_prob"): 299 eval_phon_mode = self.vocab.set_phone_prob(prob=1.0) 300 301 # Disable mixed g2p representation if necessary 302 with eval_phon_mode: 303 tokens = self.parser(str_input) 304 else: 305 tokens = self.parser(str_input) 306 307 x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device) 308 return x 309 310 @typecheck( 311 input_types={ 312 "text": NeuralType(('B', 'T_text'), TokenIndex()), 313 "durs": NeuralType(('B', 'T_text'), TokenDurationType()), 314 "pitch": NeuralType(('B', 'T_audio'), RegressionValuesType()), 315 "energy": NeuralType(('B', 'T_audio'), RegressionValuesType(), optional=True), 316 "speaker": NeuralType(('B'), Index(), optional=True), 317 "pace": NeuralType(optional=True), 318 "spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 319 "attn_prior": NeuralType(('B', 'T_spec', 'T_text'), ProbsType(), optional=True), 320 "mel_lens": NeuralType(('B'), LengthsType(), optional=True), 321 "input_lens": NeuralType(('B'), LengthsType(), optional=True), 322 # reference_* data is used for multi-speaker FastPitch training 323 "reference_spec": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType(), optional=True), 324 "reference_spec_lens": NeuralType(('B'), LengthsType(), optional=True), 325 } 326 ) 327 def forward( 328 self, 329 *, 330 text, 331 durs=None, 332 pitch=None, 333 energy=None, 334 speaker=None, 335 pace=1.0, 336 spec=None, 337 attn_prior=None, 338 mel_lens=None, 339 input_lens=None, 340 reference_spec=None, 341 reference_spec_lens=None, 342 ): 343 return self.fastpitch( 344 text=text, 345 durs=durs, 346 pitch=pitch, 347 energy=energy, 348 speaker=speaker, 349 pace=pace, 350 spec=spec, 351 attn_prior=attn_prior, 352 mel_lens=mel_lens, 353 input_lens=input_lens, 354 reference_spec=reference_spec, 355 reference_spec_lens=reference_spec_lens, 356 ) 357 358 @typecheck(output_types={"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType())}) 359 def generate_spectrogram( 360 self, 361 tokens: 'torch.tensor', 362 speaker: Optional[int] = None, 363 pace: float = 1.0, 364 reference_spec: Optional['torch.tensor'] = None, 365 reference_spec_lens: Optional['torch.tensor'] = None, 366 ) -> torch.tensor: 367 if self.training: 368 logging.warning("generate_spectrogram() is meant to be called in eval mode.") 369 if isinstance(speaker, int): 370 speaker = torch.tensor([speaker]).to(self.device) 371 spect, *_ = self( 372 text=tokens, 373 durs=None, 374 pitch=None, 375 speaker=speaker, 376 pace=pace, 377 reference_spec=reference_spec, 378 reference_spec_lens=reference_spec_lens, 379 ) 380 return spect 381 382 def training_step(self, batch, batch_idx): 383 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 384 None, 385 None, 386 None, 387 None, 388 None, 389 None, 390 ) 391 if self.learn_alignment: 392 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 393 batch_dict = batch 394 else: 395 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 396 audio = batch_dict.get("audio") 397 audio_lens = batch_dict.get("audio_lens") 398 text = batch_dict.get("text") 399 text_lens = batch_dict.get("text_lens") 400 attn_prior = batch_dict.get("align_prior_matrix", None) 401 pitch = batch_dict.get("pitch", None) 402 energy = batch_dict.get("energy", None) 403 speaker = batch_dict.get("speaker_id", None) 404 reference_audio = batch_dict.get("reference_audio", None) 405 reference_audio_len = batch_dict.get("reference_audio_lens", None) 406 else: 407 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 408 409 mels, spec_len = self.preprocessor(input_signal=audio, length=audio_lens) 410 reference_spec, reference_spec_len = None, None 411 if reference_audio is not None: 412 reference_spec, reference_spec_len = self.preprocessor( 413 input_signal=reference_audio, length=reference_audio_len 414 ) 415 416 ( 417 mels_pred, 418 _, 419 _, 420 log_durs_pred, 421 pitch_pred, 422 attn_soft, 423 attn_logprob, 424 attn_hard, 425 attn_hard_dur, 426 pitch, 427 energy_pred, 428 energy_tgt, 429 ) = self( 430 text=text, 431 durs=durs, 432 pitch=pitch, 433 energy=energy, 434 speaker=speaker, 435 pace=1.0, 436 spec=mels if self.learn_alignment else None, 437 reference_spec=reference_spec, 438 reference_spec_lens=reference_spec_len, 439 attn_prior=attn_prior, 440 mel_lens=spec_len, 441 input_lens=text_lens, 442 ) 443 if durs is None: 444 durs = attn_hard_dur 445 446 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 447 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 448 loss = mel_loss + dur_loss 449 if self.learn_alignment: 450 ctc_loss = self.forward_sum_loss_fn(attn_logprob=attn_logprob, in_lens=text_lens, out_lens=spec_len) 451 bin_loss_weight = min(self.current_epoch / self.bin_loss_warmup_epochs, 1.0) * 1.0 452 bin_loss = self.bin_loss_fn(hard_attention=attn_hard, soft_attention=attn_soft) * bin_loss_weight 453 loss += ctc_loss + bin_loss 454 455 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 456 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 457 loss += pitch_loss + energy_loss 458 459 self.log("t_loss", loss) 460 self.log("t_mel_loss", mel_loss) 461 self.log("t_dur_loss", dur_loss) 462 self.log("t_pitch_loss", pitch_loss) 463 if energy_tgt is not None: 464 self.log("t_energy_loss", energy_loss) 465 if self.learn_alignment: 466 self.log("t_ctc_loss", ctc_loss) 467 self.log("t_bin_loss", bin_loss) 468 469 # Log images to tensorboard 470 if self.log_images and self.log_train_images and isinstance(self.logger, TensorBoardLogger): 471 self.log_train_images = False 472 473 self.tb_logger.add_image( 474 "train_mel_target", 475 plot_spectrogram_to_numpy(mels[0].data.cpu().float().numpy()), 476 self.global_step, 477 dataformats="HWC", 478 ) 479 spec_predict = mels_pred[0].data.cpu().float().numpy() 480 self.tb_logger.add_image( 481 "train_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 482 ) 483 if self.learn_alignment: 484 attn = attn_hard[0].data.cpu().float().numpy().squeeze() 485 self.tb_logger.add_image( 486 "train_attn", plot_alignment_to_numpy(attn.T), self.global_step, dataformats="HWC", 487 ) 488 soft_attn = attn_soft[0].data.cpu().float().numpy().squeeze() 489 self.tb_logger.add_image( 490 "train_soft_attn", plot_alignment_to_numpy(soft_attn.T), self.global_step, dataformats="HWC", 491 ) 492 493 return loss 494 495 def validation_step(self, batch, batch_idx): 496 attn_prior, durs, speaker, energy, reference_audio, reference_audio_len = ( 497 None, 498 None, 499 None, 500 None, 501 None, 502 None, 503 ) 504 if self.learn_alignment: 505 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 506 batch_dict = batch 507 else: 508 batch_dict = process_batch(batch, self._train_dl.dataset.sup_data_types_set) 509 audio = batch_dict.get("audio") 510 audio_lens = batch_dict.get("audio_lens") 511 text = batch_dict.get("text") 512 text_lens = batch_dict.get("text_lens") 513 attn_prior = batch_dict.get("align_prior_matrix", None) 514 pitch = batch_dict.get("pitch", None) 515 energy = batch_dict.get("energy", None) 516 speaker = batch_dict.get("speaker_id", None) 517 reference_audio = batch_dict.get("reference_audio", None) 518 reference_audio_len = batch_dict.get("reference_audio_lens", None) 519 else: 520 audio, audio_lens, text, text_lens, durs, pitch, speaker = batch 521 522 mels, mel_lens = self.preprocessor(input_signal=audio, length=audio_lens) 523 reference_spec, reference_spec_len = None, None 524 if reference_audio is not None: 525 reference_spec, reference_spec_len = self.preprocessor( 526 input_signal=reference_audio, length=reference_audio_len 527 ) 528 529 # Calculate val loss on ground truth durations to better align L2 loss in time 530 (mels_pred, _, _, log_durs_pred, pitch_pred, _, _, _, attn_hard_dur, pitch, energy_pred, energy_tgt,) = self( 531 text=text, 532 durs=durs, 533 pitch=pitch, 534 energy=energy, 535 speaker=speaker, 536 pace=1.0, 537 spec=mels if self.learn_alignment else None, 538 reference_spec=reference_spec, 539 reference_spec_lens=reference_spec_len, 540 attn_prior=attn_prior, 541 mel_lens=mel_lens, 542 input_lens=text_lens, 543 ) 544 if durs is None: 545 durs = attn_hard_dur 546 547 mel_loss = self.mel_loss_fn(spect_predicted=mels_pred, spect_tgt=mels) 548 dur_loss = self.duration_loss_fn(log_durs_predicted=log_durs_pred, durs_tgt=durs, len=text_lens) 549 pitch_loss = self.pitch_loss_fn(pitch_predicted=pitch_pred, pitch_tgt=pitch, len=text_lens) 550 energy_loss = self.energy_loss_fn(energy_predicted=energy_pred, energy_tgt=energy_tgt, length=text_lens) 551 loss = mel_loss + dur_loss + pitch_loss + energy_loss 552 553 val_outputs = { 554 "val_loss": loss, 555 "mel_loss": mel_loss, 556 "dur_loss": dur_loss, 557 "pitch_loss": pitch_loss, 558 "energy_loss": energy_loss if energy_tgt is not None else None, 559 "mel_target": mels if batch_idx == 0 else None, 560 "mel_pred": mels_pred if batch_idx == 0 else None, 561 } 562 self.validation_step_outputs.append(val_outputs) 563 return val_outputs 564 565 def on_validation_epoch_end(self): 566 collect = lambda key: torch.stack([x[key] for x in self.validation_step_outputs]).mean() 567 val_loss = collect("val_loss") 568 mel_loss = collect("mel_loss") 569 dur_loss = collect("dur_loss") 570 pitch_loss = collect("pitch_loss") 571 self.log("val_loss", val_loss, sync_dist=True) 572 self.log("val_mel_loss", mel_loss, sync_dist=True) 573 self.log("val_dur_loss", dur_loss, sync_dist=True) 574 self.log("val_pitch_loss", pitch_loss, sync_dist=True) 575 if self.validation_step_outputs[0]["energy_loss"] is not None: 576 energy_loss = collect("energy_loss") 577 self.log("val_energy_loss", energy_loss, sync_dist=True) 578 579 _, _, _, _, _, spec_target, spec_predict = self.validation_step_outputs[0].values() 580 581 if self.log_images and isinstance(self.logger, TensorBoardLogger): 582 self.tb_logger.add_image( 583 "val_mel_target", 584 plot_spectrogram_to_numpy(spec_target[0].data.cpu().float().numpy()), 585 self.global_step, 586 dataformats="HWC", 587 ) 588 spec_predict = spec_predict[0].data.cpu().float().numpy() 589 self.tb_logger.add_image( 590 "val_mel_predicted", plot_spectrogram_to_numpy(spec_predict), self.global_step, dataformats="HWC", 591 ) 592 self.log_train_images = True 593 self.validation_step_outputs.clear() # free memory) 594 595 def _setup_train_dataloader(self, cfg): 596 phon_mode = contextlib.nullcontext() 597 if hasattr(self.vocab, "set_phone_prob"): 598 phon_mode = self.vocab.set_phone_prob(self.vocab.phoneme_probability) 599 600 with phon_mode: 601 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 602 603 sampler = dataset.get_sampler(cfg.dataloader_params.batch_size) 604 return torch.utils.data.DataLoader( 605 dataset, collate_fn=dataset.collate_fn, sampler=sampler, **cfg.dataloader_params 606 ) 607 608 def _setup_test_dataloader(self, cfg): 609 phon_mode = contextlib.nullcontext() 610 if hasattr(self.vocab, "set_phone_prob"): 611 phon_mode = self.vocab.set_phone_prob(0.0) 612 613 with phon_mode: 614 dataset = instantiate(cfg.dataset, text_tokenizer=self.vocab,) 615 616 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 617 618 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 619 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 620 raise ValueError(f"No dataset for {name}") 621 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 622 raise ValueError(f"No dataloader_params for {name}") 623 if shuffle_should_be: 624 if 'shuffle' not in cfg.dataloader_params: 625 logging.warning( 626 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 627 "config. Manually setting to True" 628 ) 629 with open_dict(cfg.dataloader_params): 630 cfg.dataloader_params.shuffle = True 631 elif not cfg.dataloader_params.shuffle: 632 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 633 elif cfg.dataloader_params.shuffle: 634 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 635 636 if self.ds_class == "nemo.collections.tts.data.dataset.TTSDataset": 637 phon_mode = contextlib.nullcontext() 638 if hasattr(self.vocab, "set_phone_prob"): 639 phon_mode = self.vocab.set_phone_prob(prob=None if name == "val" else self.vocab.phoneme_probability) 640 641 with phon_mode: 642 dataset = instantiate( 643 cfg.dataset, 644 text_normalizer=self.normalizer, 645 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 646 text_tokenizer=self.vocab, 647 ) 648 else: 649 dataset = instantiate(cfg.dataset) 650 651 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 652 653 def setup_training_data(self, cfg): 654 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 655 self._train_dl = self._setup_train_dataloader(cfg) 656 else: 657 self._train_dl = self.__setup_dataloader_from_config(cfg) 658 659 def setup_validation_data(self, cfg): 660 if self.ds_class == "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 661 self._validation_dl = self._setup_test_dataloader(cfg) 662 else: 663 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="val") 664 665 def setup_test_data(self, cfg): 666 """Omitted.""" 667 pass 668 669 def configure_callbacks(self): 670 if not self.log_config: 671 return [] 672 673 sample_ds_class = self.log_config.dataset._target_ 674 if sample_ds_class != "nemo.collections.tts.data.text_to_speech_dataset.TextToSpeechDataset": 675 raise ValueError(f"Logging callback only supported for TextToSpeechDataset, got {sample_ds_class}") 676 677 data_loader = self._setup_test_dataloader(self.log_config) 678 679 generators = instantiate(self.log_config.generators) 680 log_dir = Path(self.log_config.log_dir) if self.log_config.log_dir else None 681 log_callback = LoggingCallback( 682 generators=generators, 683 data_loader=data_loader, 684 log_epochs=self.log_config.log_epochs, 685 epoch_frequency=self.log_config.epoch_frequency, 686 output_dir=log_dir, 687 loggers=self.trainer.loggers, 688 log_tensorboard=self.log_config.log_tensorboard, 689 log_wandb=self.log_config.log_wandb, 690 ) 691 692 return [log_callback] 693 694 @classmethod 695 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 696 """ 697 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 698 Returns: 699 List of available pre-trained models. 700 """ 701 list_of_models = [] 702 703 # en-US, single speaker, 22050Hz, LJSpeech (ARPABET). 704 model = PretrainedModelInfo( 705 pretrained_model_name="tts_en_fastpitch", 706 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/1.8.1/files/tts_en_fastpitch_align.nemo", 707 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is ARPABET-based.", 708 class_=cls, 709 ) 710 list_of_models.append(model) 711 712 # en-US, single speaker, 22050Hz, LJSpeech (IPA). 713 model = PretrainedModelInfo( 714 pretrained_model_name="tts_en_fastpitch_ipa", 715 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch/versions/IPA_1.13.0/files/tts_en_fastpitch_align_ipa.nemo", 716 description="This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent. It is IPA-based.", 717 class_=cls, 718 ) 719 list_of_models.append(model) 720 721 # en-US, multi-speaker, 44100Hz, HiFiTTS. 722 model = PretrainedModelInfo( 723 pretrained_model_name="tts_en_fastpitch_multispeaker", 724 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_multispeaker_fastpitchhifigan/versions/1.10.0/files/tts_en_fastpitch_multispeaker.nemo", 725 description="This model is trained on HiFITTS sampled at 44100Hz with and can be used to generate male and female English voices with an American accent.", 726 class_=cls, 727 ) 728 list_of_models.append(model) 729 730 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 21.02 731 model = PretrainedModelInfo( 732 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2102", 733 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2102.nemo", 734 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 21.02 Dataset sampled at 22050Hz and can be used to generate male German voices.", 735 class_=cls, 736 ) 737 list_of_models.append(model) 738 739 # de-DE, single male speaker, grapheme-based tokenizer, 22050 Hz, Thorsten Mรผllerโ€™s German Neutral-TTS Dataset, 22.10 740 model = PretrainedModelInfo( 741 pretrained_model_name="tts_de_fastpitch_singleSpeaker_thorstenNeutral_2210", 742 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitchhifigan/versions/1.15.0/files/tts_de_fastpitch_thorstens2210.nemo", 743 description="This model is trained on a single male speaker data in Thorsten Mรผllerโ€™s German Neutral 22.10 Dataset sampled at 22050Hz and can be used to generate male German voices.", 744 class_=cls, 745 ) 746 list_of_models.append(model) 747 748 # de-DE, multi-speaker, 5 speakers, 44100 Hz, HUI-Audio-Corpus-German Clean. 749 model = PretrainedModelInfo( 750 pretrained_model_name="tts_de_fastpitch_multispeaker_5", 751 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_de_fastpitch_multispeaker_5/versions/1.11.0/files/tts_de_fastpitch_multispeaker_5.nemo", 752 description="This model is trained on 5 speakers in HUI-Audio-Corpus-German clean subset sampled at 44100Hz with and can be used to generate male and female German voices.", 753 class_=cls, 754 ) 755 list_of_models.append(model) 756 757 # es, 174 speakers, 44100Hz, OpenSLR (IPA) 758 model = PretrainedModelInfo( 759 pretrained_model_name="tts_es_fastpitch_multispeaker", 760 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_es_multispeaker_fastpitchhifigan/versions/1.15.0/files/tts_es_fastpitch_multispeaker.nemo", 761 description="This model is trained on 174 speakers in 6 crowdsourced Latin American Spanish OpenSLR datasets sampled at 44100Hz and can be used to generate male and female Spanish voices with Latin American accents.", 762 class_=cls, 763 ) 764 list_of_models.append(model) 765 766 # zh, single female speaker, 22050Hz, SFSpeech Bilingual Chinese/English dataset, improved model using richer 767 # dict and jieba word segmenter for polyphone disambiguation. 768 model = PretrainedModelInfo( 769 pretrained_model_name="tts_zh_fastpitch_sfspeech", 770 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_zh_fastpitch_hifigan_sfspeech/versions/1.15.0/files/tts_zh_fastpitch_sfspeech.nemo", 771 description="This model is trained on a single female speaker in SFSpeech Bilingual Chinese/English dataset" 772 " sampled at 22050Hz and can be used to generate female Mandarin Chinese voices. It is improved" 773 " using richer dict and jieba word segmenter for polyphone disambiguation.", 774 class_=cls, 775 ) 776 list_of_models.append(model) 777 778 # en, multi speaker, LibriTTS, 16000 Hz 779 # stft 25ms 10ms matching ASR params 780 # for use during Enhlish ASR training/adaptation 781 model = PretrainedModelInfo( 782 pretrained_model_name="tts_en_fastpitch_for_asr_finetuning", 783 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastpitch_spectrogram_enhancer_for_asr_finetuning/versions/1.20.0/files/tts_en_fastpitch_for_asr_finetuning.nemo", 784 description="This model is trained on LibriSpeech, train-960 subset." 785 " STFT parameters follow those commonly used in ASR: 25 ms window, 10 ms hop." 786 " This model is supposed to be used with its companion SpetrogramEnhancer for " 787 " ASR fine-tuning. Usage for regular TTS tasks is not advised.", 788 class_=cls, 789 ) 790 list_of_models.append(model) 791 792 return list_of_models 793 794 # Methods for model exportability 795 def _prepare_for_export(self, **kwargs): 796 super()._prepare_for_export(**kwargs) 797 798 tensor_shape = ('T') if self.export_config["enable_ragged_batches"] else ('B', 'T') 799 800 # Define input_types and output_types as required by export() 801 self._input_types = { 802 "text": NeuralType(tensor_shape, TokenIndex()), 803 "pitch": NeuralType(tensor_shape, RegressionValuesType()), 804 "pace": NeuralType(tensor_shape), 805 "volume": NeuralType(tensor_shape, optional=True), 806 "batch_lengths": NeuralType(('B'), optional=True), 807 "speaker": NeuralType(('B'), Index(), optional=True), 808 } 809 self._output_types = { 810 "spect": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 811 "num_frames": NeuralType(('B'), TokenDurationType()), 812 "durs_predicted": NeuralType(('B', 'T'), TokenDurationType()), 813 "log_durs_predicted": NeuralType(('B', 'T'), TokenLogDurationType()), 814 "pitch_predicted": NeuralType(('B', 'T'), RegressionValuesType()), 815 } 816 if self.export_config["enable_volume"]: 817 self._output_types["volume_aligned"] = NeuralType(('B', 'T'), RegressionValuesType()) 818 819 def _export_teardown(self): 820 self._input_types = self._output_types = None 821 822 @property 823 def disabled_deployment_input_names(self): 824 """Implement this method to return a set of input names disabled for export""" 825 disabled_inputs = set() 826 if self.fastpitch.speaker_emb is None: 827 disabled_inputs.add("speaker") 828 if not self.export_config["enable_ragged_batches"]: 829 disabled_inputs.add("batch_lengths") 830 if not self.export_config["enable_volume"]: 831 disabled_inputs.add("volume") 832 return disabled_inputs 833 834 @property 835 def input_types(self): 836 return self._input_types 837 838 @property 839 def output_types(self): 840 return self._output_types 841 842 def input_example(self, max_batch=1, max_dim=44): 843 """ 844 Generates input examples for tracing etc. 845 Returns: 846 A tuple of input examples. 847 """ 848 par = next(self.fastpitch.parameters()) 849 inputs = sample_tts_input(self.export_config, par.device, max_batch=max_batch, max_dim=max_dim) 850 if 'enable_ragged_batches' not in self.export_config: 851 inputs.pop('batch_lengths', None) 852 return (inputs,) 853 854 def forward_for_export(self, text, pitch, pace, volume=None, batch_lengths=None, speaker=None): 855 if self.export_config["enable_ragged_batches"]: 856 text, pitch, pace, volume_tensor, lens = batch_from_ragged( 857 text, pitch, pace, batch_lengths, padding_idx=self.fastpitch.encoder.padding_idx, volume=volume 858 ) 859 if volume is not None: 860 volume = volume_tensor 861 return self.fastpitch.infer(text=text, pitch=pitch, pace=pace, volume=volume, speaker=speaker) 862 863 def interpolate_speaker( 864 self, original_speaker_1, original_speaker_2, weight_speaker_1, weight_speaker_2, new_speaker_id 865 ): 866 """ 867 This method performs speaker interpolation between two original speakers the model is trained on. 868 869 Inputs: 870 original_speaker_1: Integer speaker ID of first existing speaker in the model 871 original_speaker_2: Integer speaker ID of second existing speaker in the model 872 weight_speaker_1: Floating point weight associated in to first speaker during weight combination 873 weight_speaker_2: Floating point weight associated in to second speaker during weight combination 874 new_speaker_id: Integer speaker ID of new interpolated speaker in the model 875 """ 876 if self.fastpitch.speaker_emb is None: 877 raise Exception( 878 "Current FastPitch model is not a multi-speaker FastPitch model. Speaker interpolation can only \ 879 be performed with a multi-speaker model" 880 ) 881 n_speakers = self.fastpitch.speaker_emb.weight.data.size()[0] 882 if original_speaker_1 >= n_speakers or original_speaker_2 >= n_speakers or new_speaker_id >= n_speakers: 883 raise Exception( 884 f"Parameters original_speaker_1, original_speaker_2, new_speaker_id should be less than the total \ 885 total number of speakers FastPitch was trained on (n_speakers = {n_speakers})." 886 ) 887 speaker_emb_1 = ( 888 self.fastpitch.speaker_emb(torch.tensor(original_speaker_1, dtype=torch.int32).cuda()).clone().detach() 889 ) 890 speaker_emb_2 = ( 891 self.fastpitch.speaker_emb(torch.tensor(original_speaker_2, dtype=torch.int32).cuda()).clone().detach() 892 ) 893 new_speaker_emb = weight_speaker_1 * speaker_emb_1 + weight_speaker_2 * speaker_emb_2 894 self.fastpitch.speaker_emb.weight.data[new_speaker_id] = new_speaker_emb 895 [end of nemo/collections/tts/models/fastpitch.py] [start of nemo/collections/tts/models/tacotron2.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import contextlib 16 from dataclasses import dataclass 17 from typing import Any, Dict, List, Optional 18 19 import torch 20 from hydra.utils import instantiate 21 from omegaconf import MISSING, DictConfig, OmegaConf, open_dict 22 from omegaconf.errors import ConfigAttributeError 23 from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger 24 from torch import nn 25 26 from nemo.collections.common.parts.preprocessing import parsers 27 from nemo.collections.tts.losses.tacotron2loss import Tacotron2Loss 28 from nemo.collections.tts.models.base import SpectrogramGenerator 29 from nemo.collections.tts.parts.utils.helpers import ( 30 g2p_backward_compatible_support, 31 get_mask_from_lengths, 32 tacotron2_log_to_tb_func, 33 tacotron2_log_to_wandb_func, 34 ) 35 from nemo.core.classes.common import PretrainedModelInfo, typecheck 36 from nemo.core.neural_types.elements import ( 37 AudioSignal, 38 EmbeddedTextType, 39 LengthsType, 40 LogitsType, 41 MelSpectrogramType, 42 SequenceToSequenceAlignmentType, 43 ) 44 from nemo.core.neural_types.neural_type import NeuralType 45 from nemo.utils import logging, model_utils 46 47 48 @dataclass 49 class Preprocessor: 50 _target_: str = MISSING 51 pad_value: float = MISSING 52 53 54 @dataclass 55 class Tacotron2Config: 56 preprocessor: Preprocessor = Preprocessor() 57 encoder: Dict[Any, Any] = MISSING 58 decoder: Dict[Any, Any] = MISSING 59 postnet: Dict[Any, Any] = MISSING 60 labels: List = MISSING 61 train_ds: Optional[Dict[Any, Any]] = None 62 validation_ds: Optional[Dict[Any, Any]] = None 63 64 65 class Tacotron2Model(SpectrogramGenerator): 66 """Tacotron 2 Model that is used to generate mel spectrograms from text""" 67 68 def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None): 69 # Convert to Hydra 1.0 compatible DictConfig 70 cfg = model_utils.convert_model_config_to_dict_config(cfg) 71 cfg = model_utils.maybe_update_config_version(cfg) 72 73 # setup normalizer 74 self.normalizer = None 75 self.text_normalizer_call = None 76 self.text_normalizer_call_kwargs = {} 77 self._setup_normalizer(cfg) 78 79 # setup tokenizer 80 self.tokenizer = None 81 if hasattr(cfg, 'text_tokenizer'): 82 self._setup_tokenizer(cfg) 83 84 self.num_tokens = len(self.tokenizer.tokens) 85 self.tokenizer_pad = self.tokenizer.pad 86 self.tokenizer_unk = self.tokenizer.oov 87 # assert self.tokenizer is not None 88 else: 89 self.num_tokens = len(cfg.labels) + 3 90 91 super().__init__(cfg=cfg, trainer=trainer) 92 93 schema = OmegaConf.structured(Tacotron2Config) 94 # ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes 95 if isinstance(cfg, dict): 96 cfg = OmegaConf.create(cfg) 97 elif not isinstance(cfg, DictConfig): 98 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 99 # Ensure passed cfg is compliant with schema 100 try: 101 OmegaConf.merge(cfg, schema) 102 self.pad_value = cfg.preprocessor.pad_value 103 except ConfigAttributeError: 104 self.pad_value = cfg.preprocessor.params.pad_value 105 logging.warning( 106 "Your config is using an old NeMo yaml configuration. Please ensure that the yaml matches the " 107 "current version in the main branch for future compatibility." 108 ) 109 110 self._parser = None 111 self.audio_to_melspec_precessor = instantiate(cfg.preprocessor) 112 self.text_embedding = nn.Embedding(self.num_tokens, 512) 113 self.encoder = instantiate(self._cfg.encoder) 114 self.decoder = instantiate(self._cfg.decoder) 115 self.postnet = instantiate(self._cfg.postnet) 116 self.loss = Tacotron2Loss() 117 self.calculate_loss = True 118 119 @property 120 def parser(self): 121 if self._parser is not None: 122 return self._parser 123 124 ds_class_name = self._cfg.train_ds.dataset._target_.split(".")[-1] 125 if ds_class_name == "TTSDataset": 126 self._parser = None 127 elif hasattr(self._cfg, "labels"): 128 self._parser = parsers.make_parser( 129 labels=self._cfg.labels, 130 name='en', 131 unk_id=-1, 132 blank_id=-1, 133 do_normalize=True, 134 abbreviation_version="fastpitch", 135 make_table=False, 136 ) 137 else: 138 raise ValueError("Wanted to setup parser, but model does not have necessary paramaters") 139 140 return self._parser 141 142 def parse(self, text: str, normalize=True) -> torch.Tensor: 143 if self.training: 144 logging.warning("parse() is meant to be called in eval mode.") 145 if normalize and self.text_normalizer_call is not None: 146 text = self.text_normalizer_call(text, **self.text_normalizer_call_kwargs) 147 148 eval_phon_mode = contextlib.nullcontext() 149 if hasattr(self.tokenizer, "set_phone_prob"): 150 eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0) 151 152 with eval_phon_mode: 153 if self.tokenizer is not None: 154 tokens = self.tokenizer.encode(text) 155 else: 156 tokens = self.parser(text) 157 # Old parser doesn't add bos and eos ids, so maunally add it 158 tokens = [len(self._cfg.labels)] + tokens + [len(self._cfg.labels) + 1] 159 tokens_tensor = torch.tensor(tokens).unsqueeze_(0).to(self.device) 160 return tokens_tensor 161 162 @property 163 def input_types(self): 164 if self.training: 165 return { 166 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 167 "token_len": NeuralType(('B'), LengthsType()), 168 "audio": NeuralType(('B', 'T'), AudioSignal()), 169 "audio_len": NeuralType(('B'), LengthsType()), 170 } 171 else: 172 return { 173 "tokens": NeuralType(('B', 'T'), EmbeddedTextType()), 174 "token_len": NeuralType(('B'), LengthsType()), 175 "audio": NeuralType(('B', 'T'), AudioSignal(), optional=True), 176 "audio_len": NeuralType(('B'), LengthsType(), optional=True), 177 } 178 179 @property 180 def output_types(self): 181 if not self.calculate_loss and not self.training: 182 return { 183 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 184 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 185 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 186 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 187 "pred_length": NeuralType(('B'), LengthsType()), 188 } 189 return { 190 "spec_pred_dec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 191 "spec_pred_postnet": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 192 "gate_pred": NeuralType(('B', 'T'), LogitsType()), 193 "spec_target": NeuralType(('B', 'D', 'T'), MelSpectrogramType()), 194 "spec_target_len": NeuralType(('B'), LengthsType()), 195 "alignments": NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()), 196 } 197 198 @typecheck() 199 def forward(self, *, tokens, token_len, audio=None, audio_len=None): 200 if audio is not None and audio_len is not None: 201 spec_target, spec_target_len = self.audio_to_melspec_precessor(audio, audio_len) 202 else: 203 if self.training or self.calculate_loss: 204 raise ValueError( 205 f"'audio' and 'audio_len' can not be None when either 'self.training' or 'self.calculate_loss' is True." 206 ) 207 208 token_embedding = self.text_embedding(tokens).transpose(1, 2) 209 encoder_embedding = self.encoder(token_embedding=token_embedding, token_len=token_len) 210 211 if self.training: 212 spec_pred_dec, gate_pred, alignments = self.decoder( 213 memory=encoder_embedding, decoder_inputs=spec_target, memory_lengths=token_len 214 ) 215 else: 216 spec_pred_dec, gate_pred, alignments, pred_length = self.decoder( 217 memory=encoder_embedding, memory_lengths=token_len 218 ) 219 220 spec_pred_postnet = self.postnet(mel_spec=spec_pred_dec) 221 222 if not self.calculate_loss and not self.training: 223 return spec_pred_dec, spec_pred_postnet, gate_pred, alignments, pred_length 224 225 return spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments 226 227 @typecheck( 228 input_types={"tokens": NeuralType(('B', 'T'), EmbeddedTextType())}, 229 output_types={"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType())}, 230 ) 231 def generate_spectrogram(self, *, tokens): 232 self.eval() 233 self.calculate_loss = False 234 token_len = torch.tensor([len(i) for i in tokens]).to(self.device) 235 tensors = self(tokens=tokens, token_len=token_len) 236 spectrogram_pred = tensors[1] 237 238 if spectrogram_pred.shape[0] > 1: 239 # Silence all frames past the predicted end 240 mask = ~get_mask_from_lengths(tensors[-1]) 241 mask = mask.expand(spectrogram_pred.shape[1], mask.size(0), mask.size(1)) 242 mask = mask.permute(1, 0, 2) 243 spectrogram_pred.data.masked_fill_(mask, self.pad_value) 244 245 return spectrogram_pred 246 247 def training_step(self, batch, batch_idx): 248 audio, audio_len, tokens, token_len = batch 249 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, _ = self.forward( 250 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 251 ) 252 253 loss, _ = self.loss( 254 spec_pred_dec=spec_pred_dec, 255 spec_pred_postnet=spec_pred_postnet, 256 gate_pred=gate_pred, 257 spec_target=spec_target, 258 spec_target_len=spec_target_len, 259 pad_value=self.pad_value, 260 ) 261 262 output = { 263 'loss': loss, 264 'progress_bar': {'training_loss': loss}, 265 'log': {'loss': loss}, 266 } 267 return output 268 269 def validation_step(self, batch, batch_idx): 270 audio, audio_len, tokens, token_len = batch 271 spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments = self.forward( 272 audio=audio, audio_len=audio_len, tokens=tokens, token_len=token_len 273 ) 274 275 loss, gate_target = self.loss( 276 spec_pred_dec=spec_pred_dec, 277 spec_pred_postnet=spec_pred_postnet, 278 gate_pred=gate_pred, 279 spec_target=spec_target, 280 spec_target_len=spec_target_len, 281 pad_value=self.pad_value, 282 ) 283 loss = { 284 "val_loss": loss, 285 "mel_target": spec_target, 286 "mel_postnet": spec_pred_postnet, 287 "gate": gate_pred, 288 "gate_target": gate_target, 289 "alignments": alignments, 290 } 291 self.validation_step_outputs.append(loss) 292 return loss 293 294 def on_validation_epoch_end(self): 295 if self.logger is not None and self.logger.experiment is not None: 296 logger = self.logger.experiment 297 for logger in self.trainer.loggers: 298 if isinstance(logger, TensorBoardLogger): 299 logger = logger.experiment 300 break 301 if isinstance(logger, TensorBoardLogger): 302 tacotron2_log_to_tb_func( 303 logger, 304 self.validation_step_outputs[0].values(), 305 self.global_step, 306 tag="val", 307 log_images=True, 308 add_audio=False, 309 ) 310 elif isinstance(logger, WandbLogger): 311 tacotron2_log_to_wandb_func( 312 logger, 313 self.validation_step_outputs[0].values(), 314 self.global_step, 315 tag="val", 316 log_images=True, 317 add_audio=False, 318 ) 319 avg_loss = torch.stack( 320 [x['val_loss'] for x in self.validation_step_outputs] 321 ).mean() # This reduces across batches, not workers! 322 self.log('val_loss', avg_loss) 323 self.validation_step_outputs.clear() # free memory 324 325 def _setup_normalizer(self, cfg): 326 if "text_normalizer" in cfg: 327 normalizer_kwargs = {} 328 329 if "whitelist" in cfg.text_normalizer: 330 normalizer_kwargs["whitelist"] = self.register_artifact( 331 'text_normalizer.whitelist', cfg.text_normalizer.whitelist 332 ) 333 334 try: 335 import nemo_text_processing 336 337 self.normalizer = instantiate(cfg.text_normalizer, **normalizer_kwargs) 338 except Exception as e: 339 logging.error(e) 340 raise ImportError( 341 "`nemo_text_processing` not installed, see https://github.com/NVIDIA/NeMo-text-processing for more details" 342 ) 343 344 self.text_normalizer_call = self.normalizer.normalize 345 if "text_normalizer_call_kwargs" in cfg: 346 self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs 347 348 def _setup_tokenizer(self, cfg): 349 text_tokenizer_kwargs = {} 350 if "g2p" in cfg.text_tokenizer and cfg.text_tokenizer.g2p is not None: 351 # for backward compatibility 352 if ( 353 self._is_model_being_restored() 354 and (cfg.text_tokenizer.g2p.get('_target_', None) is not None) 355 and cfg.text_tokenizer.g2p["_target_"].startswith("nemo_text_processing.g2p") 356 ): 357 cfg.text_tokenizer.g2p["_target_"] = g2p_backward_compatible_support( 358 cfg.text_tokenizer.g2p["_target_"] 359 ) 360 361 g2p_kwargs = {} 362 363 if "phoneme_dict" in cfg.text_tokenizer.g2p: 364 g2p_kwargs["phoneme_dict"] = self.register_artifact( 365 'text_tokenizer.g2p.phoneme_dict', cfg.text_tokenizer.g2p.phoneme_dict, 366 ) 367 368 if "heteronyms" in cfg.text_tokenizer.g2p: 369 g2p_kwargs["heteronyms"] = self.register_artifact( 370 'text_tokenizer.g2p.heteronyms', cfg.text_tokenizer.g2p.heteronyms, 371 ) 372 373 text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p, **g2p_kwargs) 374 375 self.tokenizer = instantiate(cfg.text_tokenizer, **text_tokenizer_kwargs) 376 377 def __setup_dataloader_from_config(self, cfg, shuffle_should_be: bool = True, name: str = "train"): 378 if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig): 379 raise ValueError(f"No dataset for {name}") 380 if "dataloader_params" not in cfg or not isinstance(cfg.dataloader_params, DictConfig): 381 raise ValueError(f"No dataloder_params for {name}") 382 if shuffle_should_be: 383 if 'shuffle' not in cfg.dataloader_params: 384 logging.warning( 385 f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its " 386 "config. Manually setting to True" 387 ) 388 with open_dict(cfg.dataloader_params): 389 cfg.dataloader_params.shuffle = True 390 elif not cfg.dataloader_params.shuffle: 391 logging.error(f"The {name} dataloader for {self} has shuffle set to False!!!") 392 elif not shuffle_should_be and cfg.dataloader_params.shuffle: 393 logging.error(f"The {name} dataloader for {self} has shuffle set to True!!!") 394 395 dataset = instantiate( 396 cfg.dataset, 397 text_normalizer=self.normalizer, 398 text_normalizer_call_kwargs=self.text_normalizer_call_kwargs, 399 text_tokenizer=self.tokenizer, 400 ) 401 402 return torch.utils.data.DataLoader(dataset, collate_fn=dataset.collate_fn, **cfg.dataloader_params) 403 404 def setup_training_data(self, cfg): 405 self._train_dl = self.__setup_dataloader_from_config(cfg) 406 407 def setup_validation_data(self, cfg): 408 self._validation_dl = self.__setup_dataloader_from_config(cfg, shuffle_should_be=False, name="validation") 409 410 @classmethod 411 def list_available_models(cls) -> 'List[PretrainedModelInfo]': 412 """ 413 This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. 414 Returns: 415 List of available pre-trained models. 416 """ 417 list_of_models = [] 418 model = PretrainedModelInfo( 419 pretrained_model_name="tts_en_tacotron2", 420 location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_tacotron2/versions/1.10.0/files/tts_en_tacotron2.nemo", 421 description="This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.", 422 class_=cls, 423 aliases=["Tacotron2-22050Hz"], 424 ) 425 list_of_models.append(model) 426 return list_of_models 427 [end of nemo/collections/tts/models/tacotron2.py] [start of nemo/core/config/modelPT.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 from dataclasses import dataclass, field 16 from typing import Any, Dict, Optional 17 18 from omegaconf import MISSING 19 20 from nemo.core import config 21 from nemo.core.classes.dataset import DatasetConfig 22 from nemo.utils import exp_manager 23 24 25 @dataclass 26 class SchedConfig: 27 name: str = MISSING 28 min_lr: float = 0.0 29 last_epoch: int = -1 30 31 32 @dataclass 33 class OptimConfig: 34 name: str = MISSING 35 sched: Optional[SchedConfig] = None 36 37 38 @dataclass 39 class ModelConfig: 40 """ 41 Model component inside ModelPT 42 """ 43 44 # ... 45 train_ds: Optional[DatasetConfig] = None 46 validation_ds: Optional[DatasetConfig] = None 47 test_ds: Optional[DatasetConfig] = None 48 optim: Optional[OptimConfig] = None 49 50 51 @dataclass 52 class HydraConfig: 53 run: Dict[str, Any] = field(default_factory=lambda: {"dir": "."}) 54 job_logging: Dict[str, Any] = field(default_factory=lambda: {"root": {"handlers": None}}) 55 56 57 @dataclass 58 class NemoConfig: 59 name: str = MISSING 60 model: ModelConfig = MISSING 61 trainer: config.TrainerConfig = config.TrainerConfig( 62 strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' 63 ) 64 exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() 65 hydra: HydraConfig = HydraConfig() 66 67 68 class ModelConfigBuilder: 69 def __init__(self, model_cfg: ModelConfig): 70 """ 71 Base class for any Model Config Builder. 72 73 A Model Config Builder is a utility class that accepts a ModelConfig dataclass, 74 and via a set of utility methods (that are implemented by the subclassed ModelConfigBuilder), 75 builds a finalized ModelConfig that can be supplied to a NemoModel dataclass as 76 the `model` component. 77 78 Subclasses *must* implement the private method `_finalize_cfg`. 79 Inside this method, they must update `self.model_cfg` with all interdependent config 80 options that need to be set (either updated by user explicitly or with their default value). 81 82 The updated model config must then be preserved in `self.model_cfg`. 83 84 Example: 85 # Create the config builder 86 config_builder = <subclass>ModelConfigBuilder() 87 88 # Update the components of the config that are modifiable 89 config_builder.set_X(X) 90 config_builder.set_Y(Y) 91 92 # Create a "finalized" config dataclass that will contain all the updates 93 # that were specified by the builder 94 model_config = config_builder.build() 95 96 # Use model config as is (or further update values), then create a new Model 97 model = nemo.<domain>.models.<ModelName>Model(cfg=model_config, trainer=Trainer()) 98 99 Supported build methods: 100 - set_train_ds: All model configs can accept a subclass of `DatasetConfig` as their 101 training config. Subclasses can override this method to enable auto-complete 102 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 103 104 - set_validation_ds: All model configs can accept a subclass of `DatasetConfig` as their 105 validation config. Subclasses can override this method to enable auto-complete 106 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 107 108 - set_test_ds: All model configs can accept a subclass of `DatasetConfig` as their 109 test config. Subclasses can override this method to enable auto-complete 110 by replacing `Optional[DatasetConfig]` with `Optional[<subclass of DatasetConfig>]`. 111 112 - set_optim: A build method that supports changes to the Optimizer (and optionally, 113 the Scheduler) used for training the model. The function accepts two inputs - 114 115 `cfg`: A subclass of `OptimizerParams` - any OptimizerParams subclass can be used, 116 in order to select an appropriate Optimizer. Examples: AdamParams. 117 118 `sched_cfg`: A subclass of `SchedulerParams` - any SchedulerParams subclass can be used, 119 in order to select an appropriate Scheduler. Examples: CosineAnnealingParams. 120 Note that this argument is optional. 121 122 - build(): The method which should return a "finalized" ModelConfig dataclass. 123 Subclasses *should* always override this method, and update the signature 124 of this method with the return type of the Dataclass, so that it enables 125 autocomplete for the user. 126 127 Example: 128 def build(self) -> EncDecCTCConfig: 129 return super().build() 130 131 Any additional build methods must be added by subclasses of ModelConfigBuilder. 132 133 Args: 134 model_cfg: 135 """ 136 self.model_cfg = model_cfg 137 self.train_ds_cfg = None 138 self.validation_ds_cfg = None 139 self.test_ds_cfg = None 140 self.optim_cfg = None 141 142 def set_train_ds(self, cfg: Optional[DatasetConfig] = None): 143 self.model_cfg.train_ds = cfg 144 145 def set_validation_ds(self, cfg: Optional[DatasetConfig] = None): 146 self.model_cfg.validation_ds = cfg 147 148 def set_test_ds(self, cfg: Optional[DatasetConfig] = None): 149 self.model_cfg.test_ds = cfg 150 151 def set_optim(self, cfg: config.OptimizerParams, sched_cfg: Optional[config.SchedulerParams] = None): 152 @dataclass 153 class WrappedOptimConfig(OptimConfig, cfg.__class__): 154 pass 155 156 # Setup optim 157 optim_name = cfg.__class__.__name__.replace("Params", "").lower() 158 wrapped_cfg = WrappedOptimConfig(name=optim_name, sched=None, **vars(cfg)) 159 160 if sched_cfg is not None: 161 162 @dataclass 163 class WrappedSchedConfig(SchedConfig, sched_cfg.__class__): 164 pass 165 166 # Setup scheduler 167 sched_name = sched_cfg.__class__.__name__.replace("Params", "") 168 wrapped_sched_cfg = WrappedSchedConfig(name=sched_name, **vars(sched_cfg)) 169 170 wrapped_cfg.sched = wrapped_sched_cfg 171 172 self.model_cfg.optim = wrapped_cfg 173 174 def _finalize_cfg(self): 175 raise NotImplementedError() 176 177 def build(self) -> ModelConfig: 178 # validate config 179 self._finalize_cfg() 180 181 return self.model_cfg 182 [end of nemo/core/config/modelPT.py] [start of nemo/utils/exp_manager.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import glob 16 import os 17 import subprocess 18 import sys 19 import time 20 import warnings 21 from dataclasses import dataclass 22 from datetime import timedelta 23 from pathlib import Path 24 from shutil import copy, move 25 from typing import Any, Dict, List, Optional, Tuple, Union 26 27 import pytorch_lightning 28 import torch 29 from hydra.core.hydra_config import HydraConfig 30 from hydra.utils import get_original_cwd 31 from omegaconf import DictConfig, OmegaConf, open_dict 32 from pytorch_lightning.callbacks import Callback, ModelCheckpoint 33 from pytorch_lightning.callbacks.early_stopping import EarlyStopping 34 from pytorch_lightning.callbacks.timer import Interval, Timer 35 from pytorch_lightning.loggers import MLFlowLogger, TensorBoardLogger, WandbLogger 36 from pytorch_lightning.loops import _TrainingEpochLoop 37 from pytorch_lightning.strategies.ddp import DDPStrategy 38 39 from nemo.collections.common.callbacks import EMA 40 from nemo.constants import NEMO_ENV_VARNAME_TESTING, NEMO_ENV_VARNAME_VERSION 41 from nemo.utils import logging, timers 42 from nemo.utils.app_state import AppState 43 from nemo.utils.callbacks import NeMoModelCheckpoint, PreemptionCallback 44 from nemo.utils.env_var_parsing import get_envbool 45 from nemo.utils.exceptions import NeMoBaseException 46 from nemo.utils.get_rank import is_global_rank_zero 47 from nemo.utils.lightning_logger_patch import add_filehandlers_to_pl_logger 48 from nemo.utils.loggers import ClearMLLogger, ClearMLParams, DLLogger, DLLoggerParams, MLFlowParams 49 from nemo.utils.model_utils import uninject_model_parallel_rank 50 51 52 class NotFoundError(NeMoBaseException): 53 """ Raised when a file or folder is not found""" 54 55 56 class LoggerMisconfigurationError(NeMoBaseException): 57 """ Raised when a mismatch between trainer.logger and exp_manager occurs""" 58 59 def __init__(self, message): 60 message = ( 61 message 62 + " You can disable lighning's trainer from creating a logger by passing logger=False to its constructor." 63 ) 64 super().__init__(message) 65 66 67 class CheckpointMisconfigurationError(NeMoBaseException): 68 """ Raised when a mismatch between trainer.callbacks and exp_manager occurs""" 69 70 71 @dataclass 72 class EarlyStoppingParams: 73 monitor: str = "val_loss" # The metric that early stopping should consider. 74 mode: str = "min" # inform early stopping whether to look for increase or decrease in monitor. 75 min_delta: float = 0.001 # smallest change to consider as improvement. 76 patience: int = 10 # how many (continuous) validation cycles to wait with no improvement and stopping training. 77 verbose: bool = True 78 strict: bool = True 79 check_finite: bool = True 80 stopping_threshold: Optional[float] = None 81 divergence_threshold: Optional[float] = None 82 check_on_train_epoch_end: Optional[bool] = None 83 log_rank_zero_only: bool = False 84 85 86 @dataclass 87 class CallbackParams: 88 filepath: Optional[str] = None # Deprecated 89 dirpath: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 90 filename: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 91 monitor: Optional[str] = "val_loss" 92 verbose: Optional[bool] = True 93 save_last: Optional[bool] = True 94 save_top_k: Optional[int] = 3 95 save_weights_only: Optional[bool] = False 96 mode: Optional[str] = "min" 97 auto_insert_metric_name: bool = True 98 every_n_epochs: Optional[int] = 1 99 every_n_train_steps: Optional[int] = None 100 train_time_interval: Optional[str] = None 101 prefix: Optional[str] = None # If None, exp_manager will attempt to handle the filepath 102 postfix: str = ".nemo" 103 save_best_model: bool = False 104 always_save_nemo: bool = False 105 save_nemo_on_train_end: Optional[bool] = True # Whether to automatically save .nemo file durin on_train_end hook 106 model_parallel_size: Optional[int] = None # tensor parallel size * pipeline parallel size 107 save_on_train_epoch_end: Optional[bool] = False # Save after training, not after validation 108 109 110 @dataclass 111 class StepTimingParams: 112 reduction: Optional[str] = "mean" 113 # if True torch.cuda.synchronize() is called on start/stop 114 sync_cuda: Optional[bool] = False 115 # if positive, defines the size of a sliding window for computing mean 116 buffer_size: Optional[int] = 1 117 118 119 @dataclass 120 class EMAParams: 121 enable: Optional[bool] = False 122 decay: Optional[float] = 0.999 123 cpu_offload: Optional[bool] = False 124 validate_original_weights: Optional[bool] = False 125 every_n_steps: int = 1 126 127 128 @dataclass 129 class ExpManagerConfig: 130 """Experiment Manager config for validation of passed arguments. 131 """ 132 133 # Log dir creation parameters 134 explicit_log_dir: Optional[str] = None 135 exp_dir: Optional[str] = None 136 name: Optional[str] = None 137 version: Optional[str] = None 138 use_datetime_version: Optional[bool] = True 139 resume_if_exists: Optional[bool] = False 140 resume_past_end: Optional[bool] = False 141 resume_ignore_no_checkpoint: Optional[bool] = False 142 resume_from_checkpoint: Optional[str] = None 143 # Logging parameters 144 create_tensorboard_logger: Optional[bool] = True 145 summary_writer_kwargs: Optional[Dict[Any, Any]] = None 146 create_wandb_logger: Optional[bool] = False 147 wandb_logger_kwargs: Optional[Dict[Any, Any]] = None 148 create_mlflow_logger: Optional[bool] = False 149 mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() 150 create_dllogger_logger: Optional[bool] = False 151 dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() 152 create_clearml_logger: Optional[bool] = False 153 clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() 154 # Checkpointing parameters 155 create_checkpoint_callback: Optional[bool] = True 156 checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() 157 create_early_stopping_callback: Optional[bool] = False 158 early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() 159 create_preemption_callback: Optional[bool] = True 160 # Additional exp_manager arguments 161 files_to_copy: Optional[List[str]] = None 162 # logs timing of train/val/test steps 163 log_step_timing: Optional[bool] = True 164 step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() 165 # Configures creation of log files for different ranks 166 log_local_rank_0_only: Optional[bool] = False 167 log_global_rank_0_only: Optional[bool] = False 168 # disable initial validation when resuming from a checkpoint saved during validation 169 disable_validation_on_resume: Optional[bool] = True 170 ema: Optional[EMAParams] = EMAParams() 171 # Wall clock time limit 172 max_time_per_run: Optional[str] = None 173 # time to sleep non 0 ranks during initialization 174 seconds_to_sleep: float = 5 175 176 177 class TimingCallback(Callback): 178 """ 179 Logs execution time of train/val/test steps 180 """ 181 182 def __init__(self, timer_kwargs={}): 183 self.timer = timers.NamedTimer(**timer_kwargs) 184 185 def _on_batch_start(self, name): 186 # reset only if we do not return mean of a sliding window 187 if self.timer.buffer_size <= 0: 188 self.timer.reset(name) 189 190 self.timer.start(name) 191 192 def _on_batch_end(self, name, pl_module): 193 self.timer.stop(name) 194 # Set the `batch_size=1` as WAR for `dataloader_iter`, which is not used for any metric 195 pl_module.log( 196 name + ' in s', 197 self.timer[name], 198 on_step=True, 199 on_epoch=False, 200 batch_size=1, 201 prog_bar=(name == "train_step_timing"), 202 ) 203 204 def on_train_batch_start(self, trainer, pl_module, batch, batch_idx): 205 self._on_batch_start("train_step_timing") 206 207 def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): 208 self._on_batch_end("train_step_timing", pl_module) 209 210 def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 211 self._on_batch_start("validation_step_timing") 212 213 def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 214 self._on_batch_end("validation_step_timing", pl_module) 215 216 def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0): 217 self._on_batch_start("test_step_timing") 218 219 def on_test_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0): 220 self._on_batch_end("test_step_timing", pl_module) 221 222 def on_before_backward(self, trainer, pl_module, loss): 223 self._on_batch_start("train_backward_timing") 224 225 def on_after_backward(self, trainer, pl_module): 226 self._on_batch_end("train_backward_timing", pl_module) 227 228 229 def exp_manager(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None) -> Optional[Path]: 230 """ 231 exp_manager is a helper function used to manage folders for experiments. It follows the pytorch lightning paradigm 232 of exp_dir/model_or_experiment_name/version. If the lightning trainer has a logger, exp_manager will get exp_dir, 233 name, and version from the logger. Otherwise it will use the exp_dir and name arguments to create the logging 234 directory. exp_manager also allows for explicit folder creation via explicit_log_dir. 235 236 The version can be a datetime string or an integer. Datestime version can be disabled if use_datetime_version is set 237 to False. It optionally creates TensorBoardLogger, WandBLogger, DLLogger, MLFlowLogger, ClearMLLogger, 238 ModelCheckpoint objects from pytorch lightning. 239 It copies sys.argv, and git information if available to the logging directory. It creates a log file for each 240 process to log their output into. 241 242 exp_manager additionally has a resume feature (resume_if_exists) which can be used to continuing training from 243 the constructed log_dir. When you need to continue the training repeatedly (like on a cluster which you need 244 multiple consecutive jobs), you need to avoid creating the version folders. Therefore from v1.0.0, when 245 resume_if_exists is set to True, creating the version folders is ignored. 246 247 Args: 248 trainer (pytorch_lightning.Trainer): The lightning trainer. 249 cfg (DictConfig, dict): Can have the following keys: 250 251 - explicit_log_dir (str, Path): Can be used to override exp_dir/name/version folder creation. Defaults to 252 None, which will use exp_dir, name, and version to construct the logging directory. 253 - exp_dir (str, Path): The base directory to create the logging directory. Defaults to None, which logs to 254 ./nemo_experiments. 255 - name (str): The name of the experiment. Defaults to None which turns into "default" via name = name or 256 "default". 257 - version (str): The version of the experiment. Defaults to None which uses either a datetime string or 258 lightning's TensorboardLogger system of using version_{int}. 259 - use_datetime_version (bool): Whether to use a datetime string for version. Defaults to True. 260 - resume_if_exists (bool): Whether this experiment is resuming from a previous run. If True, it sets 261 trainer._checkpoint_connector._ckpt_path so that the trainer should auto-resume. exp_manager will move files 262 under log_dir to log_dir/run_{int}. Defaults to False. From v1.0.0, when resume_if_exists is True, 263 we would not create version folders to make it easier to find the log folder for next runs. 264 - resume_past_end (bool): exp_manager errors out if resume_if_exists is True and a checkpoint matching 265 ``*end.ckpt`` indicating a previous training run fully completed. This behaviour can be disabled, in which 266 case the ``*end.ckpt`` will be loaded by setting resume_past_end to True. Defaults to False. 267 - resume_ignore_no_checkpoint (bool): exp_manager errors out if resume_if_exists is True and no checkpoint 268 could be found. This behaviour can be disabled, in which case exp_manager will print a message and 269 continue without restoring, by setting resume_ignore_no_checkpoint to True. Defaults to False. 270 - resume_from_checkpoint (str): Can be used to specify a path to a specific checkpoint file to load from. This will 271 override any checkpoint found when resume_if_exists is True. Defaults to None. 272 - create_tensorboard_logger (bool): Whether to create a tensorboard logger and attach it to the pytorch 273 lightning trainer. Defaults to True. 274 - summary_writer_kwargs (dict): A dictionary of kwargs that can be passed to lightning's TensorboardLogger 275 class. Note that log_dir is passed by exp_manager and cannot exist in this dict. Defaults to None. 276 - create_wandb_logger (bool): Whether to create a Weights and Baises logger and attach it to the pytorch 277 lightning trainer. Defaults to False. 278 - wandb_logger_kwargs (dict): A dictionary of kwargs that can be passed to lightning's WandBLogger 279 class. Note that name and project are required parameters if create_wandb_logger is True. 280 Defaults to None. 281 - create_mlflow_logger (bool): Whether to create an MLFlow logger and attach it to the pytorch lightning 282 training. Defaults to False 283 - mlflow_logger_kwargs (dict): optional parameters for the MLFlow logger 284 - create_dllogger_logger (bool): Whether to create an DLLogger logger and attach it to the pytorch lightning 285 training. Defaults to False 286 - dllogger_logger_kwargs (dict): optional parameters for the DLLogger logger 287 - create_clearml_logger (bool): Whether to create an ClearML logger and attach it to the pytorch lightning 288 training. Defaults to False 289 - clearml_logger_kwargs (dict): optional parameters for the ClearML logger 290 - create_checkpoint_callback (bool): Whether to create a ModelCheckpoint callback and attach it to the 291 pytorch lightning trainer. The ModelCheckpoint saves the top 3 models with the best "val_loss", the most 292 recent checkpoint under ``*last.ckpt``, and the final checkpoint after training completes under ``*end.ckpt``. 293 Defaults to True. 294 - create_early_stopping_callback (bool): Flag to decide if early stopping should be used to stop training. Default is False. 295 See EarlyStoppingParams dataclass above. 296 - create_preemption_callback (bool): Flag to decide whether to enable preemption callback to save checkpoints and exit training 297 immediately upon preemption. Default is True. 298 - files_to_copy (list): A list of files to copy to the experiment logging directory. Defaults to None which 299 copies no files. 300 - log_local_rank_0_only (bool): Whether to only create log files for local rank 0. Defaults to False. 301 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 302 - log_global_rank_0_only (bool): Whether to only create log files for global rank 0. Defaults to False. 303 Set this to True if you are using DDP with many GPUs and do not want many log files in your exp dir. 304 - max_time (str): The maximum wall clock time *per run*. This is intended to be used on clusters where you want 305 a checkpoint to be saved after this specified time and be able to resume from that checkpoint. Defaults to None. 306 - seconds_to_sleep (float): seconds to sleep non rank 0 processes for. Used to give enough time for rank 0 to initialize 307 308 returns: 309 log_dir (Path): The final logging directory where logging files are saved. Usually the concatenation of 310 exp_dir, name, and version. 311 """ 312 # Add rank information to logger 313 # Note: trainer.global_rank and trainer.is_global_zero are not set until trainer.fit, so have to hack around it 314 local_rank = int(os.environ.get("LOCAL_RANK", 0)) 315 global_rank = trainer.node_rank * trainer.num_devices + local_rank 316 logging.rank = global_rank 317 318 if cfg is None: 319 logging.error("exp_manager did not receive a cfg argument. It will be disabled.") 320 return 321 if trainer.fast_dev_run: 322 logging.info("Trainer was called with fast_dev_run. exp_manager will return without any functionality.") 323 return 324 325 # Ensure passed cfg is compliant with ExpManagerConfig 326 schema = OmegaConf.structured(ExpManagerConfig) 327 if isinstance(cfg, dict): 328 cfg = OmegaConf.create(cfg) 329 elif not isinstance(cfg, DictConfig): 330 raise ValueError(f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig") 331 cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True)) 332 cfg = OmegaConf.merge(schema, cfg) 333 334 error_checks(trainer, cfg) # Ensures that trainer options are compliant with NeMo and exp_manager arguments 335 336 log_dir, exp_dir, name, version = get_log_dir( 337 trainer=trainer, 338 exp_dir=cfg.exp_dir, 339 name=cfg.name, 340 version=cfg.version, 341 explicit_log_dir=cfg.explicit_log_dir, 342 use_datetime_version=cfg.use_datetime_version, 343 resume_if_exists=cfg.resume_if_exists, 344 ) 345 346 check_resume( 347 trainer, 348 log_dir, 349 cfg.resume_if_exists, 350 cfg.resume_past_end, 351 cfg.resume_ignore_no_checkpoint, 352 cfg.checkpoint_callback_params.dirpath, 353 cfg.resume_from_checkpoint, 354 ) 355 356 checkpoint_name = name 357 # If name returned from get_log_dir is "", use cfg.name for checkpointing 358 if checkpoint_name is None or checkpoint_name == '': 359 checkpoint_name = cfg.name or "default" 360 361 # Set mlflow name if it's not set, before the main name is erased 362 if cfg.create_mlflow_logger and (not cfg.mlflow_logger_kwargs.get("experiment_name", None)): 363 cfg.mlflow_logger_kwargs.experiment_name = cfg.name 364 logging.warning( 365 'mlflow logger specified but no experiment name set. Using the same as Tensorboard: %s', 366 cfg.mlflow_logger_kwargs.experiment_name, 367 ) 368 369 cfg.name = name # Used for configure_loggers so that the log_dir is properly set even if name is "" 370 cfg.version = version 371 372 # update app_state with log_dir, exp_dir, etc 373 app_state = AppState() 374 app_state.log_dir = log_dir 375 app_state.exp_dir = exp_dir 376 app_state.name = name 377 app_state.version = version 378 app_state.checkpoint_name = checkpoint_name 379 app_state.create_checkpoint_callback = cfg.create_checkpoint_callback 380 app_state.checkpoint_callback_params = cfg.checkpoint_callback_params 381 382 # Create the logging directory if it does not exist 383 os.makedirs(log_dir, exist_ok=True) # Cannot limit creation to global zero as all ranks write to own log file 384 logging.info(f'Experiments will be logged at {log_dir}') 385 trainer._default_root_dir = log_dir 386 387 if cfg.log_local_rank_0_only is True and cfg.log_global_rank_0_only is True: 388 raise ValueError( 389 f"Cannot set both log_local_rank_0_only and log_global_rank_0_only to True. Please set either one or neither." 390 ) 391 392 # This is set if the env var NEMO_TESTING is set to True. 393 nemo_testing = get_envbool(NEMO_ENV_VARNAME_TESTING, False) 394 395 # Handle logging to file 396 log_file = log_dir / f'nemo_log_globalrank-{global_rank}_localrank-{local_rank}.txt' 397 if cfg.log_local_rank_0_only is True and not nemo_testing: 398 if local_rank == 0: 399 logging.add_file_handler(log_file) 400 elif cfg.log_global_rank_0_only is True and not nemo_testing: 401 if global_rank == 0: 402 logging.add_file_handler(log_file) 403 else: 404 # Logs on all ranks. 405 logging.add_file_handler(log_file) 406 407 # For some reason, LearningRateLogger requires trainer to have a logger. Safer to create logger on all ranks 408 # not just global rank 0. 409 if ( 410 cfg.create_tensorboard_logger 411 or cfg.create_wandb_logger 412 or cfg.create_mlflow_logger 413 or cfg.create_dllogger_logger 414 or cfg.create_clearml_logger 415 ): 416 configure_loggers( 417 trainer, 418 exp_dir, 419 log_dir, 420 cfg.name, 421 cfg.version, 422 cfg.checkpoint_callback_params, 423 cfg.create_tensorboard_logger, 424 cfg.summary_writer_kwargs, 425 cfg.create_wandb_logger, 426 cfg.wandb_logger_kwargs, 427 cfg.create_mlflow_logger, 428 cfg.mlflow_logger_kwargs, 429 cfg.create_dllogger_logger, 430 cfg.dllogger_logger_kwargs, 431 cfg.create_clearml_logger, 432 cfg.clearml_logger_kwargs, 433 ) 434 435 # add loggers timing callbacks 436 if cfg.log_step_timing: 437 timing_callback = TimingCallback(timer_kwargs=cfg.step_timing_kwargs or {}) 438 trainer.callbacks.insert(0, timing_callback) 439 440 if cfg.ema.enable: 441 ema_callback = EMA( 442 decay=cfg.ema.decay, 443 validate_original_weights=cfg.ema.validate_original_weights, 444 cpu_offload=cfg.ema.cpu_offload, 445 every_n_steps=cfg.ema.every_n_steps, 446 ) 447 trainer.callbacks.append(ema_callback) 448 449 if cfg.create_early_stopping_callback: 450 early_stop_callback = EarlyStopping(**cfg.early_stopping_callback_params) 451 trainer.callbacks.append(early_stop_callback) 452 453 if cfg.create_checkpoint_callback: 454 configure_checkpointing( 455 trainer, 456 log_dir, 457 checkpoint_name, 458 cfg.resume_if_exists, 459 cfg.checkpoint_callback_params, 460 cfg.create_preemption_callback, 461 ) 462 463 if cfg.disable_validation_on_resume: 464 # extend training loop to skip initial validation when resuming from checkpoint 465 configure_no_restart_validation_training_loop(trainer) 466 # Setup a stateless timer for use on clusters. 467 if cfg.max_time_per_run is not None: 468 found_ptl_timer = False 469 for idx, callback in enumerate(trainer.callbacks): 470 if isinstance(callback, Timer): 471 # NOTE: PTL does not expose a `trainer.max_time`. By the time we are in this function, PTL has already setup a timer if the user specifies `trainer.max_time` so best we can do is replace that. 472 # Working: If only `trainer.max_time` is set - it behaves as a normal PTL timer. If only `exp_manager.max_time_per_run` is set - it behaves as a StateLessTimer. If both are set, it also behaves as a StateLessTimer. 473 logging.warning( 474 f'Found a PTL Timer callback, replacing with a StatelessTimer callback. This will happen if you set trainer.max_time as well as exp_manager.max_time_per_run.' 475 ) 476 trainer.callbacks[idx] = StatelessTimer(cfg.max_time_per_run) 477 found_ptl_timer = True 478 break 479 480 if not found_ptl_timer: 481 trainer.max_time = cfg.max_time_per_run 482 trainer.callbacks.append(StatelessTimer(cfg.max_time_per_run)) 483 484 if is_global_rank_zero(): 485 # Move files_to_copy to folder and add git information if present 486 if cfg.files_to_copy: 487 for _file in cfg.files_to_copy: 488 copy(Path(_file), log_dir) 489 490 # Create files for cmd args and git info 491 with open(log_dir / 'cmd-args.log', 'w', encoding='utf-8') as _file: 492 _file.write(" ".join(sys.argv)) 493 494 # Try to get git hash 495 git_repo, git_hash = get_git_hash() 496 if git_repo: 497 with open(log_dir / 'git-info.log', 'w', encoding='utf-8') as _file: 498 _file.write(f'commit hash: {git_hash}') 499 _file.write(get_git_diff()) 500 501 # Add err_file logging to global_rank zero 502 logging.add_err_file_handler(log_dir / 'nemo_error_log.txt') 503 504 # Add lightning file logging to global_rank zero 505 add_filehandlers_to_pl_logger(log_dir / 'lightning_logs.txt', log_dir / 'nemo_error_log.txt') 506 507 elif trainer.num_nodes * trainer.num_devices > 1: 508 # sleep other ranks so rank 0 can finish 509 # doing the initialization such as moving files 510 time.sleep(cfg.seconds_to_sleep) 511 512 return log_dir 513 514 515 def error_checks(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictConfig, Dict]] = None): 516 """ 517 Checks that the passed trainer is compliant with NeMo and exp_manager's passed configuration. Checks that: 518 - Throws error when hydra has changed the working directory. This causes issues with lightning's DDP 519 - Throws error when trainer has loggers defined but create_tensorboard_logger or create_wandB_logger 520 or create_mlflow_logger or create_dllogger_logger is True 521 - Prints error messages when 1) run on multi-node and not Slurm, and 2) run on multi-gpu without DDP 522 """ 523 if HydraConfig.initialized() and get_original_cwd() != os.getcwd(): 524 raise ValueError( 525 "Hydra changed the working directory. This interferes with ExpManger's functionality. Please pass " 526 "hydra.run.dir=. to your python script." 527 ) 528 if trainer.logger is not None and ( 529 cfg.create_tensorboard_logger or cfg.create_wandb_logger or cfg.create_mlflow_logger 530 ): 531 raise LoggerMisconfigurationError( 532 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and either " 533 f"create_tensorboard_logger: {cfg.create_tensorboard_logger} or create_wandb_logger: " 534 f"{cfg.create_wandb_logger} or create_mlflow_logger: {cfg.create_mlflow_logger}" 535 f"or create_dllogger_logger: {cfg.create_mlflow_logger} was set to True. " 536 "These can only be used if trainer does not already have a logger." 537 ) 538 if trainer.num_nodes > 1 and not check_slurm(trainer): 539 logging.error( 540 "You are running multi-node training without SLURM handling the processes." 541 " Please note that this is not tested in NeMo and could result in errors." 542 ) 543 if trainer.num_devices > 1 and not isinstance(trainer.strategy, DDPStrategy): 544 logging.error( 545 "You are running multi-gpu without ddp.Please note that this is not tested in NeMo and could result in " 546 "errors." 547 ) 548 549 550 def check_resume( 551 trainer: 'pytorch_lightning.Trainer', 552 log_dir: str, 553 resume_if_exists: bool = False, 554 resume_past_end: bool = False, 555 resume_ignore_no_checkpoint: bool = False, 556 dirpath: str = None, 557 resume_from_checkpoint: str = None, 558 ): 559 """Checks that resume=True was used correctly with the arguments pass to exp_manager. Sets 560 trainer._checkpoint_connector._ckpt_path as necessary. 561 562 Returns: 563 log_dir (Path): The log_dir 564 exp_dir (str): The base exp_dir without name nor version 565 name (str): The name of the experiment 566 version (str): The version of the experiment 567 568 Raises: 569 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 570 ValueError: If resume is True, and there were more than 1 checkpoint could found. 571 """ 572 573 if not log_dir: 574 raise ValueError(f"Resuming requires the log_dir {log_dir} to be passed to exp_manager") 575 576 checkpoint = None 577 if resume_from_checkpoint: 578 checkpoint = resume_from_checkpoint 579 if resume_if_exists: 580 # Use <log_dir>/checkpoints/ unless `dirpath` is set 581 checkpoint_dir = Path(dirpath) if dirpath else Path(Path(log_dir) / "checkpoints") 582 583 # when using distributed checkpointing, checkpoint_dir is a directory of directories 584 # we check for this here 585 dist_checkpoints = [d for d in list(checkpoint_dir.glob("*")) if d.is_dir()] 586 end_dist_checkpoints = [d for d in dist_checkpoints if d.match("*end")] 587 last_dist_checkpoints = [d for d in dist_checkpoints if d.match("*last")] 588 589 end_checkpoints = end_dist_checkpoints if end_dist_checkpoints else list(checkpoint_dir.rglob("*end.ckpt")) 590 last_checkpoints = last_dist_checkpoints if last_dist_checkpoints else list(checkpoint_dir.rglob("*last.ckpt")) 591 592 if not checkpoint_dir.exists() or (not len(end_checkpoints) > 0 and not len(last_checkpoints) > 0): 593 if resume_ignore_no_checkpoint: 594 warn = f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. " 595 if checkpoint is None: 596 warn += "Training from scratch." 597 elif checkpoint == resume_from_checkpoint: 598 warn += f"Training from {resume_from_checkpoint}." 599 logging.warning(warn) 600 else: 601 raise NotFoundError( 602 f"There were no checkpoints found in checkpoint_dir or no checkpoint folder at checkpoint_dir :{checkpoint_dir}. Cannot resume." 603 ) 604 elif len(end_checkpoints) > 0: 605 if resume_past_end: 606 if len(end_checkpoints) > 1: 607 if 'mp_rank' in str(end_checkpoints[0]): 608 checkpoint = end_checkpoints[0] 609 else: 610 raise ValueError(f"Multiple checkpoints {end_checkpoints} that matches *end.ckpt.") 611 else: 612 raise ValueError( 613 f"Found {end_checkpoints[0]} indicating that the last training run has already completed." 614 ) 615 elif len(last_checkpoints) > 1: 616 if 'mp_rank' in str(last_checkpoints[0]) or 'tp_rank' in str(last_checkpoints[0]): 617 checkpoint = last_checkpoints[0] 618 checkpoint = uninject_model_parallel_rank(checkpoint) 619 else: 620 raise ValueError(f"Multiple checkpoints {last_checkpoints} that matches *last.ckpt.") 621 else: 622 checkpoint = last_checkpoints[0] 623 624 # PTL 2.0 supports ckpt_path instead of resume_from_checkpoint as the trainer flag 625 if checkpoint is not None: 626 trainer.ckpt_path = str(checkpoint) 627 logging.info(f'Resuming training from checkpoint: {trainer.ckpt_path}') 628 629 if is_global_rank_zero(): 630 # Check to see if any files exist that need to be moved 631 files_to_move = [] 632 if Path(log_dir).exists(): 633 for child in Path(log_dir).iterdir(): 634 if child.is_file(): 635 files_to_move.append(child) 636 637 if len(files_to_move) > 0: 638 # Move old files to a new folder 639 other_run_dirs = Path(log_dir).glob("run_*") 640 run_count = 0 641 for fold in other_run_dirs: 642 if fold.is_dir(): 643 run_count += 1 644 new_run_dir = Path(Path(log_dir) / f"run_{run_count}") 645 new_run_dir.mkdir() 646 for _file in files_to_move: 647 move(str(_file), str(new_run_dir)) 648 649 650 def check_explicit_log_dir( 651 trainer: 'pytorch_lightning.Trainer', explicit_log_dir: Union[Path, str], exp_dir: str, name: str, version: str 652 ) -> Tuple[Path, str, str, str]: 653 """ Checks that the passed arguments are compatible with explicit_log_dir. 654 655 Returns: 656 log_dir (Path): the log_dir 657 exp_dir (str): the base exp_dir without name nor version 658 name (str): The name of the experiment 659 version (str): The version of the experiment 660 661 Raise: 662 LoggerMisconfigurationError 663 """ 664 if trainer.logger is not None: 665 raise LoggerMisconfigurationError( 666 "The pytorch lightning trainer that was passed to exp_manager contained a logger and explicit_log_dir: " 667 f"{explicit_log_dir} was pass to exp_manager. Please remove the logger from the lightning trainer." 668 ) 669 # Checking only (explicit_log_dir) vs (exp_dir and version). 670 # The `name` will be used as the actual name of checkpoint/archive. 671 if exp_dir or version: 672 logging.error( 673 f"exp_manager received explicit_log_dir: {explicit_log_dir} and at least one of exp_dir: {exp_dir}, " 674 f"or version: {version}. Please note that exp_dir, name, and version will be ignored." 675 ) 676 if is_global_rank_zero() and Path(explicit_log_dir).exists(): 677 logging.warning(f"Exp_manager is logging to {explicit_log_dir}, but it already exists.") 678 return Path(explicit_log_dir), str(explicit_log_dir), "", "" 679 680 681 def get_log_dir( 682 trainer: 'pytorch_lightning.Trainer', 683 exp_dir: str = None, 684 name: str = None, 685 version: str = None, 686 explicit_log_dir: str = None, 687 use_datetime_version: bool = True, 688 resume_if_exists: bool = False, 689 ) -> Tuple[Path, str, str, str]: 690 """ 691 Obtains the log_dir used for exp_manager. 692 693 Returns: 694 log_dir (Path): the log_dir 695 exp_dir (str): the base exp_dir without name nor version 696 name (str): The name of the experiment 697 version (str): The version of the experiment 698 explicit_log_dir (str): The explicit path to the log folder. Defaults to False. 699 use_datetime_version (bool): Uses date and time as the version of the log folder. Defaults to True. 700 resume_if_exists (bool): if resume_if_exists of the exp_manager's config is enabled or not. When enabled, the 701 version folders would not get created. 702 703 Raise: 704 LoggerMisconfigurationError: If trainer is incompatible with arguments 705 NotFoundError: If resume is True, resume_ignore_no_checkpoint is False, and checkpoints could not be found. 706 ValueError: If resume is True, and there were more than 1 checkpoint could found. 707 """ 708 if explicit_log_dir: # If explicit log_dir was passed, short circuit 709 return check_explicit_log_dir(trainer, explicit_log_dir, exp_dir, name, version) 710 711 # Default exp_dir to ./nemo_experiments if None was passed 712 _exp_dir = exp_dir 713 if exp_dir is None: 714 _exp_dir = str(Path.cwd() / 'nemo_experiments') 715 716 # If the user has already defined a logger for the trainer, use the logger defaults for logging directory 717 if trainer.logger is not None: 718 if trainer.logger.save_dir: 719 if exp_dir: 720 raise LoggerMisconfigurationError( 721 "The pytorch lightning trainer that was passed to exp_manager contained a logger, the logger's " 722 f"save_dir was not None, and exp_dir ({exp_dir}) was not None. If trainer.logger.save_dir " 723 "exists, exp_manager will use trainer.logger.save_dir as the logging directory and exp_dir " 724 "must be None." 725 ) 726 _exp_dir = trainer.logger.save_dir 727 if name: 728 raise LoggerMisconfigurationError( 729 "The pytorch lightning trainer that was passed to exp_manager contained a logger, and name: " 730 f"{name} was also passed to exp_manager. If the trainer contains a " 731 "logger, exp_manager will use trainer.logger.name, and name passed to exp_manager must be None." 732 ) 733 name = trainer.logger.name 734 version = f"version_{trainer.logger.version}" 735 # Use user-defined exp_dir, project_name, exp_name, and versioning options 736 else: 737 name = name or "default" 738 version = version or os.environ.get(NEMO_ENV_VARNAME_VERSION, None) 739 740 if not version: 741 if resume_if_exists: 742 logging.warning( 743 "No version folders would be created under the log folder as 'resume_if_exists' is enabled." 744 ) 745 version = None 746 elif is_global_rank_zero(): 747 if use_datetime_version: 748 version = time.strftime('%Y-%m-%d_%H-%M-%S') 749 else: 750 tensorboard_logger = TensorBoardLogger(save_dir=Path(_exp_dir), name=name, version=version) 751 version = f"version_{tensorboard_logger.version}" 752 os.environ[NEMO_ENV_VARNAME_VERSION] = "" if version is None else version 753 754 log_dir = Path(_exp_dir) / Path(str(name)) / Path("" if version is None else str(version)) 755 return log_dir, str(_exp_dir), name, version 756 757 758 def get_git_hash(): 759 """ 760 Helper function that tries to get the commit hash if running inside a git folder 761 762 returns: 763 Bool: Whether the git subprocess ran without error 764 str: git subprocess output or error message 765 """ 766 try: 767 return ( 768 True, 769 subprocess.check_output(['git', 'rev-parse', 'HEAD'], stderr=subprocess.STDOUT).decode(), 770 ) 771 except subprocess.CalledProcessError as err: 772 return False, "{}\n".format(err.output.decode("utf-8")) 773 774 775 def get_git_diff(): 776 """ 777 Helper function that tries to get the git diff if running inside a git folder 778 779 returns: 780 Bool: Whether the git subprocess ran without error 781 str: git subprocess output or error message 782 """ 783 try: 784 return subprocess.check_output(['git', 'diff'], stderr=subprocess.STDOUT).decode() 785 except subprocess.CalledProcessError as err: 786 return "{}\n".format(err.output.decode("utf-8")) 787 788 789 def configure_loggers( 790 trainer: 'pytorch_lightning.Trainer', 791 exp_dir: [Path, str], 792 log_dir: [Path, str], 793 name: str, 794 version: str, 795 checkpoint_callback_params: dict, 796 create_tensorboard_logger: bool, 797 summary_writer_kwargs: dict, 798 create_wandb_logger: bool, 799 wandb_kwargs: dict, 800 create_mlflow_logger: bool, 801 mlflow_kwargs: dict, 802 create_dllogger_logger: bool, 803 dllogger_kwargs: dict, 804 create_clearml_logger: bool, 805 clearml_kwargs: dict, 806 ): 807 """ 808 Creates TensorboardLogger and/or WandBLogger / MLFlowLogger / DLlogger / ClearMLLogger and attach them to trainer. 809 Raises ValueError if summary_writer_kwargs or wandb_kwargs are misconfigured. 810 """ 811 # Potentially create tensorboard logger and/or WandBLogger / MLFlowLogger / DLLogger 812 logger_list = [] 813 if create_tensorboard_logger: 814 if summary_writer_kwargs is None: 815 summary_writer_kwargs = {} 816 elif "log_dir" in summary_writer_kwargs: 817 raise ValueError( 818 "You cannot pass `log_dir` as part of `summary_writer_kwargs`. `log_dir` is handled by lightning's " 819 "TensorBoardLogger logger." 820 ) 821 tensorboard_logger = TensorBoardLogger(save_dir=exp_dir, name=name, version=version, **summary_writer_kwargs) 822 logger_list.append(tensorboard_logger) 823 logging.info("TensorboardLogger has been set up") 824 825 if create_wandb_logger: 826 if wandb_kwargs is None: 827 wandb_kwargs = {} 828 if "name" not in wandb_kwargs and "project" not in wandb_kwargs: 829 raise ValueError("name and project are required for wandb_logger") 830 831 # Update the wandb save_dir 832 if wandb_kwargs.get('save_dir', None) is None: 833 wandb_kwargs['save_dir'] = exp_dir 834 os.makedirs(wandb_kwargs['save_dir'], exist_ok=True) 835 wandb_logger = WandbLogger(version=version, **wandb_kwargs) 836 837 logger_list.append(wandb_logger) 838 logging.info("WandBLogger has been set up") 839 840 if create_mlflow_logger: 841 mlflow_logger = MLFlowLogger(run_name=version, **mlflow_kwargs) 842 843 logger_list.append(mlflow_logger) 844 logging.info("MLFlowLogger has been set up") 845 846 if create_dllogger_logger: 847 dllogger_logger = DLLogger(**dllogger_kwargs) 848 849 logger_list.append(dllogger_logger) 850 logging.info("DLLogger has been set up") 851 852 if create_clearml_logger: 853 clearml_logger = ClearMLLogger( 854 clearml_cfg=clearml_kwargs, 855 log_dir=log_dir, 856 prefix=name, 857 save_best_model=checkpoint_callback_params.save_best_model, 858 ) 859 860 logger_list.append(clearml_logger) 861 logging.info("ClearMLLogger has been set up") 862 863 trainer._logger_connector.configure_logger(logger_list) 864 865 866 def configure_checkpointing( 867 trainer: 'pytorch_lightning.Trainer', 868 log_dir: Path, 869 name: str, 870 resume: bool, 871 params: 'DictConfig', 872 create_preemption_callback: bool, 873 ): 874 """ Adds ModelCheckpoint to trainer. Raises CheckpointMisconfigurationError if trainer already has a ModelCheckpoint 875 callback 876 """ 877 for callback in trainer.callbacks: 878 if isinstance(callback, ModelCheckpoint): 879 raise CheckpointMisconfigurationError( 880 "The pytorch lightning trainer that was passed to exp_manager contained a ModelCheckpoint " 881 "and create_checkpoint_callback was set to True. Please either set create_checkpoint_callback " 882 "to False, or remove ModelCheckpoint from the lightning trainer" 883 ) 884 # Create the callback and attach it to trainer 885 if "filepath" in params: 886 if params.filepath is not None: 887 logging.warning("filepath is deprecated. Please switch to dirpath and filename instead") 888 if params.dirpath is None: 889 params.dirpath = Path(params.filepath).parent 890 if params.filename is None: 891 params.filename = Path(params.filepath).name 892 with open_dict(params): 893 del params["filepath"] 894 if params.dirpath is None: 895 params.dirpath = Path(log_dir / 'checkpoints') 896 if params.filename is None: 897 params.filename = f'{name}--{{{params.monitor}:.4f}}-{{epoch}}' 898 if params.prefix is None: 899 params.prefix = name 900 NeMoModelCheckpoint.CHECKPOINT_NAME_LAST = params.filename + '-last' 901 902 logging.debug(params.dirpath) 903 logging.debug(params.filename) 904 logging.debug(params.prefix) 905 906 if "val" in params.monitor: 907 if ( 908 trainer.max_epochs is not None 909 and trainer.max_epochs != -1 910 and trainer.max_epochs < trainer.check_val_every_n_epoch 911 ): 912 logging.error( 913 "The checkpoint callback was told to monitor a validation value but trainer.max_epochs(" 914 f"{trainer.max_epochs}) was less than trainer.check_val_every_n_epoch({trainer.check_val_every_n_epoch}" 915 f"). It is very likely this run will fail with ModelCheckpoint(monitor='{params.monitor}') not found " 916 "in the returned metrics. Please ensure that validation is run within trainer.max_epochs." 917 ) 918 elif trainer.max_steps is not None and trainer.max_steps != -1: 919 logging.warning( 920 "The checkpoint callback was told to monitor a validation value and trainer's max_steps was set to " 921 f"{trainer.max_steps}. Please ensure that max_steps will run for at least " 922 f"{trainer.check_val_every_n_epoch} epochs to ensure that checkpointing will not error out." 923 ) 924 925 checkpoint_callback = NeMoModelCheckpoint(n_resume=resume, **params) 926 checkpoint_callback.last_model_path = trainer.ckpt_path or "" 927 if 'mp_rank' in checkpoint_callback.last_model_path or 'tp_rank' in checkpoint_callback.last_model_path: 928 checkpoint_callback.last_model_path = uninject_model_parallel_rank(checkpoint_callback.last_model_path) 929 trainer.callbacks.append(checkpoint_callback) 930 if create_preemption_callback: 931 # Check if cuda is avialable as preemption is supported only on GPUs 932 if torch.cuda.is_available(): 933 ## By default PreemptionCallback handles SIGTERM. To handle other signals pass the signal in the call as below: 934 ## PreemptionCallback(checkpoint_callback, signal.SIGCHLD) 935 preemption_callback = PreemptionCallback(checkpoint_callback) 936 trainer.callbacks.append(preemption_callback) 937 else: 938 logging.info("Preemption is supported only on GPUs, disabling preemption") 939 940 941 def check_slurm(trainer): 942 try: 943 return trainer.accelerator_connector.is_slurm_managing_tasks 944 except AttributeError: 945 return False 946 947 948 class StatelessTimer(Timer): 949 """Extension of PTL timers to be per run.""" 950 951 def __init__(self, duration: timedelta = None, interval: str = Interval.step, verbose: bool = True,) -> None: 952 super().__init__(duration, interval, verbose) 953 954 # Override PTL Timer's state dict to not store elapsed time information so that we can restore and continue training. 955 def state_dict(self) -> Dict[str, Any]: 956 return {} 957 958 def load_state_dict(self, state_dict: Dict[str, Any]) -> None: 959 return 960 961 962 def configure_no_restart_validation_training_loop(trainer: pytorch_lightning.Trainer) -> None: 963 if type(trainer.fit_loop.epoch_loop) != _TrainingEpochLoop: 964 warnings.warn("Detected custom epoch loop. Skipping no validation on restart support.", UserWarning) 965 return 966 ## Pass trainer object to avoid trainer getting overwritten as None 967 loop = SkipResumeTrainingValidationLoop(trainer, trainer.min_steps, trainer.max_steps) 968 trainer.fit_loop.epoch_loop = loop 969 970 971 class SkipResumeTrainingValidationLoop(_TrainingEpochLoop): 972 """ 973 Extend the PTL Epoch loop to skip validating when resuming. 974 This happens when resuming a checkpoint that has already run validation, but loading restores 975 the training state before validation has run. 976 """ 977 978 def _should_check_val_fx(self) -> bool: 979 if self.restarting and self.global_step % self.trainer.val_check_batch == 0: 980 return False 981 return super()._should_check_val_fx() 982 983 984 def clean_exp_ckpt(exp_log_dir: Union[str, Path], remove_ckpt: bool = True, remove_nemo: bool = False): 985 """ 986 Helper method that removes Pytorch Lightning .ckpt files or NeMo .nemo files from the checkpoint directory 987 988 Args: 989 exp_log_dir: str path to the root directory of the current experiment. 990 remove_ckpt: bool, whether to remove all *.ckpt files in the checkpoints directory. 991 remove_nemo: bool, whether to remove all *.nemo files in the checkpoints directory. 992 """ 993 exp_log_dir = str(exp_log_dir) 994 995 if remove_ckpt: 996 logging.info("Deleting *.ckpt files ...") 997 ckpt_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.ckpt")) 998 for filepath in ckpt_files: 999 os.remove(filepath) 1000 logging.info(f"Deleted file : {filepath}") 1001 1002 if remove_nemo: 1003 logging.info("Deleting *.nemo files ...") 1004 nemo_files = glob.glob(os.path.join(exp_log_dir, "checkpoints", "*.nemo")) 1005 for filepath in nemo_files: 1006 os.remove(filepath) 1007 logging.info(f"Deleted file : {filepath}") 1008 [end of nemo/utils/exp_manager.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR model with CTC decoder. To evaluate a model with 19 # Transducer (RNN-T) decoder use another script 'scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py'. 20 # NeMo's beam search decoders are capable of using the KenLM's N-gram models 21 # to find the best candidates. This script supports both character level and BPE level 22 # encodings and models which is detected automatically from the type of the model. 23 # You may train the LM model with 'scripts/asr_language_modeling/ngram_lm/train_kenlm.py'. 24 25 # Config Help 26 27 To discover all arguments of the script, please run : 28 python eval_beamsearch_ngram.py --help 29 python eval_beamsearch_ngram.py --cfg job 30 31 # USAGE 32 33 python eval_beamsearch_ngram.py nemo_model_file=<path to the .nemo file of the model> \ 34 input_manifest=<path to the evaluation JSON manifest file> \ 35 kenlm_model_file=<path to the binary KenLM model> \ 36 beam_width=[<list of the beam widths, separated with commas>] \ 37 beam_alpha=[<list of the beam alphas, separated with commas>] \ 38 beam_beta=[<list of the beam betas, separated with commas>] \ 39 preds_output_folder=<optional folder to store the predictions> \ 40 probs_cache_file=null \ 41 decoding_mode=beamsearch_ngram 42 ... 43 44 45 # Grid Search for Hyper parameters 46 47 For grid search, you can provide a list of arguments as follows - 48 49 beam_width=[4,8,16,....] \ 50 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 51 beam_beta=[-1.0,-0.5,0.0,...,1.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 from dataclasses import dataclass, field, is_dataclass 64 from pathlib import Path 65 from typing import List, Optional 66 67 import editdistance 68 import numpy as np 69 import torch 70 from omegaconf import MISSING, OmegaConf 71 from sklearn.model_selection import ParameterGrid 72 from tqdm.auto import tqdm 73 74 import nemo.collections.asr as nemo_asr 75 from nemo.collections.asr.models import EncDecHybridRNNTCTCModel 76 from nemo.collections.asr.parts.submodules import ctc_beam_decoding 77 from nemo.collections.asr.parts.utils.transcribe_utils import PunctuationCapitalization, TextProcessingConfig 78 from nemo.core.config import hydra_runner 79 from nemo.utils import logging 80 81 # fmt: off 82 83 84 @dataclass 85 class EvalBeamSearchNGramConfig: 86 """ 87 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 88 """ 89 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 90 nemo_model_file: str = MISSING 91 92 # File paths 93 input_manifest: str = MISSING # The manifest file of the evaluation set 94 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 95 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 96 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 97 98 # Parameters for inference 99 acoustic_batch_size: int = 16 # The batch size to calculate log probabilities 100 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 101 device: str = "cuda" # The device to load the model onto to calculate log probabilities 102 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 103 104 # Beam Search hyperparameters 105 106 # The decoding scheme to be used for evaluation. 107 # Can be one of ["greedy", "beamsearch", "beamsearch_ngram"] 108 decoding_mode: str = "beamsearch_ngram" 109 110 beam_width: List[int] = field(default_factory=lambda: [128]) # The width or list of the widths for the beam search decoding 111 beam_alpha: List[float] = field(default_factory=lambda: [1.0]) # The alpha parameter or list of the alphas for the beam search decoding 112 beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding 113 114 decoding_strategy: str = "beam" 115 decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) 116 117 text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( 118 punctuation_marks = ".,?", 119 separate_punctuation = False, 120 do_lowercase = False, 121 rm_punctuation = False, 122 ) 123 # fmt: on 124 125 126 def beam_search_eval( 127 model: nemo_asr.models.ASRModel, 128 cfg: EvalBeamSearchNGramConfig, 129 all_probs: List[torch.Tensor], 130 target_transcripts: List[str], 131 preds_output_file: str = None, 132 lm_path: str = None, 133 beam_alpha: float = 1.0, 134 beam_beta: float = 0.0, 135 beam_width: int = 128, 136 beam_batch_size: int = 128, 137 progress_bar: bool = True, 138 punctuation_capitalization: PunctuationCapitalization = None, 139 ): 140 level = logging.getEffectiveLevel() 141 logging.setLevel(logging.CRITICAL) 142 # Reset config 143 model.change_decoding_strategy(None) 144 145 # Override the beam search config with current search candidate configuration 146 cfg.decoding.beam_size = beam_width 147 cfg.decoding.beam_alpha = beam_alpha 148 cfg.decoding.beam_beta = beam_beta 149 cfg.decoding.return_best_hypothesis = False 150 cfg.decoding.kenlm_path = cfg.kenlm_model_file 151 152 # Update model's decoding strategy config 153 model.cfg.decoding.strategy = cfg.decoding_strategy 154 model.cfg.decoding.beam = cfg.decoding 155 156 # Update model's decoding strategy 157 if isinstance(model, EncDecHybridRNNTCTCModel): 158 model.change_decoding_strategy(model.cfg.decoding, decoder_type='ctc') 159 decoding = model.ctc_decoding 160 else: 161 model.change_decoding_strategy(model.cfg.decoding) 162 decoding = model.decoding 163 logging.setLevel(level) 164 165 wer_dist_first = cer_dist_first = 0 166 wer_dist_best = cer_dist_best = 0 167 words_count = 0 168 chars_count = 0 169 sample_idx = 0 170 if preds_output_file: 171 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 172 173 if progress_bar: 174 it = tqdm( 175 range(int(np.ceil(len(all_probs) / beam_batch_size))), 176 desc=f"Beam search decoding with width={beam_width}, alpha={beam_alpha}, beta={beam_beta}", 177 ncols=120, 178 ) 179 else: 180 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 181 for batch_idx in it: 182 # disabling type checking 183 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 184 probs_lens = torch.tensor([prob.shape[0] for prob in probs_batch]) 185 with torch.no_grad(): 186 packed_batch = torch.zeros(len(probs_batch), max(probs_lens), probs_batch[0].shape[-1], device='cpu') 187 188 for prob_index in range(len(probs_batch)): 189 packed_batch[prob_index, : probs_lens[prob_index], :] = torch.tensor( 190 probs_batch[prob_index], device=packed_batch.device, dtype=packed_batch.dtype 191 ) 192 193 _, beams_batch = decoding.ctc_decoder_predictions_tensor( 194 packed_batch, decoder_lengths=probs_lens, return_hypotheses=True, 195 ) 196 197 for beams_idx, beams in enumerate(beams_batch): 198 target = target_transcripts[sample_idx + beams_idx] 199 target_split_w = target.split() 200 target_split_c = list(target) 201 words_count += len(target_split_w) 202 chars_count += len(target_split_c) 203 wer_dist_min = cer_dist_min = 10000 204 for candidate_idx, candidate in enumerate(beams): # type: (int, ctc_beam_decoding.rnnt_utils.Hypothesis) 205 pred_text = candidate.text 206 if cfg.text_processing.do_lowercase: 207 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 208 if cfg.text_processing.rm_punctuation: 209 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 210 if cfg.text_processing.separate_punctuation: 211 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 212 pred_split_w = pred_text.split() 213 wer_dist = editdistance.eval(target_split_w, pred_split_w) 214 pred_split_c = list(pred_text) 215 cer_dist = editdistance.eval(target_split_c, pred_split_c) 216 217 wer_dist_min = min(wer_dist_min, wer_dist) 218 cer_dist_min = min(cer_dist_min, cer_dist) 219 220 if candidate_idx == 0: 221 # first candidate 222 wer_dist_first += wer_dist 223 cer_dist_first += cer_dist 224 225 score = candidate.score 226 if preds_output_file: 227 out_file.write('{}\t{}\n'.format(pred_text, score)) 228 wer_dist_best += wer_dist_min 229 cer_dist_best += cer_dist_min 230 sample_idx += len(probs_batch) 231 232 if preds_output_file: 233 out_file.close() 234 logging.info(f"Stored the predictions of beam search decoding at '{preds_output_file}'.") 235 236 if lm_path: 237 logging.info( 238 'WER/CER with beam search decoding and N-gram model = {:.2%}/{:.2%}'.format( 239 wer_dist_first / words_count, cer_dist_first / chars_count 240 ) 241 ) 242 else: 243 logging.info( 244 'WER/CER with beam search decoding = {:.2%}/{:.2%}'.format( 245 wer_dist_first / words_count, cer_dist_first / chars_count 246 ) 247 ) 248 logging.info( 249 'Oracle WER/CER in candidates with perfect LM= {:.2%}/{:.2%}'.format( 250 wer_dist_best / words_count, cer_dist_best / chars_count 251 ) 252 ) 253 logging.info(f"=================================================================================") 254 255 return wer_dist_first / words_count, cer_dist_first / chars_count 256 257 258 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 259 def main(cfg: EvalBeamSearchNGramConfig): 260 logging.warning("This file will be renamed to eval_beamsearch_ngram_ctc.py in the future NeMo (1.21) release.") 261 if is_dataclass(cfg): 262 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 263 264 valid_decoding_modes = ["greedy", "beamsearch", "beamsearch_ngram"] 265 if cfg.decoding_mode not in valid_decoding_modes: 266 raise ValueError( 267 f"Given decoding_mode={cfg.decoding_mode} is invalid. Available options are :\n" f"{valid_decoding_modes}" 268 ) 269 270 if cfg.nemo_model_file.endswith('.nemo'): 271 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 272 else: 273 logging.warning( 274 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 275 ) 276 asr_model = nemo_asr.models.ASRModel.from_pretrained( 277 cfg.nemo_model_file, map_location=torch.device(cfg.device) 278 ) 279 280 target_transcripts = [] 281 manifest_dir = Path(cfg.input_manifest).parent 282 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 283 audio_file_paths = [] 284 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 285 data = json.loads(line) 286 audio_file = Path(data['audio_filepath']) 287 if not audio_file.is_file() and not audio_file.is_absolute(): 288 audio_file = manifest_dir / audio_file 289 target_transcripts.append(data['text']) 290 audio_file_paths.append(str(audio_file.absolute())) 291 292 punctuation_capitalization = PunctuationCapitalization(cfg.text_processing.punctuation_marks) 293 if cfg.text_processing.do_lowercase: 294 target_transcripts = punctuation_capitalization.do_lowercase(target_transcripts) 295 if cfg.text_processing.rm_punctuation: 296 target_transcripts = punctuation_capitalization.rm_punctuation(target_transcripts) 297 if cfg.text_processing.separate_punctuation: 298 target_transcripts = punctuation_capitalization.separate_punctuation(target_transcripts) 299 300 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 301 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 302 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 303 with open(cfg.probs_cache_file, 'rb') as probs_file: 304 all_probs = pickle.load(probs_file) 305 306 if len(all_probs) != len(audio_file_paths): 307 raise ValueError( 308 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 309 f"match the manifest file. You may need to delete the probabilities cached file." 310 ) 311 else: 312 313 @contextlib.contextmanager 314 def default_autocast(): 315 yield 316 317 if cfg.use_amp: 318 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 319 logging.info("AMP is enabled!\n") 320 autocast = torch.cuda.amp.autocast 321 322 else: 323 autocast = default_autocast 324 else: 325 326 autocast = default_autocast 327 328 with autocast(): 329 with torch.no_grad(): 330 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 331 asr_model.cur_decoder = 'ctc' 332 all_logits = asr_model.transcribe(audio_file_paths, batch_size=cfg.acoustic_batch_size, logprobs=True) 333 334 all_probs = all_logits 335 if cfg.probs_cache_file: 336 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 337 with open(cfg.probs_cache_file, 'wb') as f_dump: 338 pickle.dump(all_probs, f_dump) 339 340 wer_dist_greedy = 0 341 cer_dist_greedy = 0 342 words_count = 0 343 chars_count = 0 344 for batch_idx, probs in enumerate(all_probs): 345 preds = np.argmax(probs, axis=1) 346 preds_tensor = torch.tensor(preds, device='cpu').unsqueeze(0) 347 if isinstance(asr_model, EncDecHybridRNNTCTCModel): 348 pred_text = asr_model.ctc_decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 349 else: 350 pred_text = asr_model._wer.decoding.ctc_decoder_predictions_tensor(preds_tensor)[0][0] 351 352 if cfg.text_processing.do_lowercase: 353 pred_text = punctuation_capitalization.do_lowercase([pred_text])[0] 354 if cfg.text_processing.rm_punctuation: 355 pred_text = punctuation_capitalization.rm_punctuation([pred_text])[0] 356 if cfg.text_processing.separate_punctuation: 357 pred_text = punctuation_capitalization.separate_punctuation([pred_text])[0] 358 359 pred_split_w = pred_text.split() 360 target_split_w = target_transcripts[batch_idx].split() 361 pred_split_c = list(pred_text) 362 target_split_c = list(target_transcripts[batch_idx]) 363 364 wer_dist = editdistance.eval(target_split_w, pred_split_w) 365 cer_dist = editdistance.eval(target_split_c, pred_split_c) 366 367 wer_dist_greedy += wer_dist 368 cer_dist_greedy += cer_dist 369 words_count += len(target_split_w) 370 chars_count += len(target_split_c) 371 372 logging.info('Greedy WER/CER = {:.2%}/{:.2%}'.format(wer_dist_greedy / words_count, cer_dist_greedy / chars_count)) 373 374 asr_model = asr_model.to('cpu') 375 376 if cfg.decoding_mode == "beamsearch_ngram": 377 if not os.path.exists(cfg.kenlm_model_file): 378 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 379 lm_path = cfg.kenlm_model_file 380 else: 381 lm_path = None 382 383 # 'greedy' decoding_mode would skip the beam search decoding 384 if cfg.decoding_mode in ["beamsearch_ngram", "beamsearch"]: 385 if cfg.beam_width is None or cfg.beam_alpha is None or cfg.beam_beta is None: 386 raise ValueError("beam_width, beam_alpha and beam_beta are needed to perform beam search decoding.") 387 params = {'beam_width': cfg.beam_width, 'beam_alpha': cfg.beam_alpha, 'beam_beta': cfg.beam_beta} 388 hp_grid = ParameterGrid(params) 389 hp_grid = list(hp_grid) 390 391 best_wer_beam_size, best_cer_beam_size = None, None 392 best_wer_alpha, best_cer_alpha = None, None 393 best_wer_beta, best_cer_beta = None, None 394 best_wer, best_cer = 1e6, 1e6 395 396 logging.info(f"==============================Starting the beam search decoding===============================") 397 logging.info(f"Grid search size: {len(hp_grid)}") 398 logging.info(f"It may take some time...") 399 logging.info(f"==============================================================================================") 400 401 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 402 os.mkdir(cfg.preds_output_folder) 403 for hp in hp_grid: 404 if cfg.preds_output_folder: 405 preds_output_file = os.path.join( 406 cfg.preds_output_folder, 407 f"preds_out_width{hp['beam_width']}_alpha{hp['beam_alpha']}_beta{hp['beam_beta']}.tsv", 408 ) 409 else: 410 preds_output_file = None 411 412 candidate_wer, candidate_cer = beam_search_eval( 413 asr_model, 414 cfg, 415 all_probs=all_probs, 416 target_transcripts=target_transcripts, 417 preds_output_file=preds_output_file, 418 lm_path=lm_path, 419 beam_width=hp["beam_width"], 420 beam_alpha=hp["beam_alpha"], 421 beam_beta=hp["beam_beta"], 422 beam_batch_size=cfg.beam_batch_size, 423 progress_bar=True, 424 punctuation_capitalization=punctuation_capitalization, 425 ) 426 427 if candidate_cer < best_cer: 428 best_cer_beam_size = hp["beam_width"] 429 best_cer_alpha = hp["beam_alpha"] 430 best_cer_beta = hp["beam_beta"] 431 best_cer = candidate_cer 432 433 if candidate_wer < best_wer: 434 best_wer_beam_size = hp["beam_width"] 435 best_wer_alpha = hp["beam_alpha"] 436 best_wer_beta = hp["beam_beta"] 437 best_wer = candidate_wer 438 439 logging.info( 440 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 441 f'Beam alpha = {best_wer_alpha}, Beam beta = {best_wer_beta}' 442 ) 443 444 logging.info( 445 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 446 f'Beam alpha = {best_cer_alpha}, Beam beta = {best_cer_beta}' 447 ) 448 logging.info(f"=================================================================================") 449 450 451 if __name__ == '__main__': 452 main() 453 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py] [start of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # 15 16 """ 17 # This script would evaluate an N-gram language model trained with KenLM library (https://github.com/kpu/kenlm) in 18 # fusion with beam search decoders on top of a trained ASR Transducer model. NeMo's beam search decoders are capable of using the 19 # KenLM's N-gram models to find the best candidates. This script supports both character level and BPE level 20 # encodings and models which is detected automatically from the type of the model. 21 # You may train the LM model with 'scripts/ngram_lm/train_kenlm.py'. 22 23 # Config Help 24 25 To discover all arguments of the script, please run : 26 python eval_beamsearch_ngram.py --help 27 python eval_beamsearch_ngram.py --cfg job 28 29 # USAGE 30 31 python eval_beamsearch_ngram_transducer.py nemo_model_file=<path to the .nemo file of the model> \ 32 input_manifest=<path to the evaluation JSON manifest file \ 33 kenlm_model_file=<path to the binary KenLM model> \ 34 beam_width=[<list of the beam widths, separated with commas>] \ 35 beam_alpha=[<list of the beam alphas, separated with commas>] \ 36 preds_output_folder=<optional folder to store the predictions> \ 37 probs_cache_file=null \ 38 decoding_strategy=<greedy_batch or maes decoding> 39 maes_prefix_alpha=[<list of the maes prefix alphas, separated with commas>] \ 40 maes_expansion_gamma=[<list of the maes expansion gammas, separated with commas>] \ 41 hat_subtract_ilm=<in case of HAT model: subtract internal LM or not> \ 42 hat_ilm_weight=[<in case of HAT model: list of the HAT internal LM weights, separated with commas>] \ 43 ... 44 45 46 # Grid Search for Hyper parameters 47 48 For grid search, you can provide a list of arguments as follows - 49 50 beam_width=[4,8,16,....] \ 51 beam_alpha=[-2.0,-1.0,...,1.0,2.0] \ 52 53 # You may find more info on how to use this script at: 54 # https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html 55 56 """ 57 58 59 import contextlib 60 import json 61 import os 62 import pickle 63 import tempfile 64 from dataclasses import dataclass, field, is_dataclass 65 from pathlib import Path 66 from typing import List, Optional 67 68 import editdistance 69 import numpy as np 70 import torch 71 from omegaconf import MISSING, OmegaConf 72 from sklearn.model_selection import ParameterGrid 73 from tqdm.auto import tqdm 74 75 import nemo.collections.asr as nemo_asr 76 from nemo.collections.asr.parts.submodules import rnnt_beam_decoding 77 from nemo.core.config import hydra_runner 78 from nemo.utils import logging 79 80 # fmt: off 81 82 83 @dataclass 84 class EvalBeamSearchNGramConfig: 85 """ 86 Evaluate an ASR model with beam search decoding and n-gram KenLM language model. 87 """ 88 # # The path of the '.nemo' file of the ASR model or the name of a pretrained model (ngc / huggingface) 89 nemo_model_file: str = MISSING 90 91 # File paths 92 input_manifest: str = MISSING # The manifest file of the evaluation set 93 kenlm_model_file: Optional[str] = None # The path of the KenLM binary model file 94 preds_output_folder: Optional[str] = None # The optional folder where the predictions are stored 95 probs_cache_file: Optional[str] = None # The cache file for storing the logprobs of the model 96 97 # Parameters for inference 98 acoustic_batch_size: int = 128 # The batch size to calculate log probabilities 99 beam_batch_size: int = 128 # The batch size to be used for beam search decoding 100 device: str = "cuda" # The device to load the model onto to calculate log probabilities 101 use_amp: bool = False # Whether to use AMP if available to calculate log probabilities 102 num_workers: int = 1 # Number of workers for DataLoader 103 104 # The decoding scheme to be used for evaluation 105 decoding_strategy: str = "greedy_batch" # ["greedy_batch", "beam", "tsd", "alsd", "maes"] 106 107 # Beam Search hyperparameters 108 beam_width: List[int] = field(default_factory=lambda: [8]) # The width or list of the widths for the beam search decoding 109 beam_alpha: List[float] = field(default_factory=lambda: [0.2]) # The alpha parameter or list of the alphas for the beam search decoding 110 111 maes_prefix_alpha: List[int] = field(default_factory=lambda: [2]) # The maes_prefix_alpha or list of the maes_prefix_alpha for the maes decoding 112 maes_expansion_gamma: List[float] = field(default_factory=lambda: [2.3]) # The maes_expansion_gamma or list of the maes_expansion_gamma for the maes decoding 113 114 # HAT related parameters (only for internal lm subtraction) 115 hat_subtract_ilm: bool = False 116 hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) 117 118 decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) 119 120 121 # fmt: on 122 123 124 def decoding_step( 125 model: nemo_asr.models.ASRModel, 126 cfg: EvalBeamSearchNGramConfig, 127 all_probs: List[torch.Tensor], 128 target_transcripts: List[str], 129 preds_output_file: str = None, 130 beam_batch_size: int = 128, 131 progress_bar: bool = True, 132 ): 133 level = logging.getEffectiveLevel() 134 logging.setLevel(logging.CRITICAL) 135 # Reset config 136 model.change_decoding_strategy(None) 137 138 cfg.decoding.hat_ilm_weight = cfg.decoding.hat_ilm_weight * cfg.hat_subtract_ilm 139 # Override the beam search config with current search candidate configuration 140 cfg.decoding.return_best_hypothesis = False 141 cfg.decoding.ngram_lm_model = cfg.kenlm_model_file 142 cfg.decoding.hat_subtract_ilm = cfg.hat_subtract_ilm 143 144 # Update model's decoding strategy config 145 model.cfg.decoding.strategy = cfg.decoding_strategy 146 model.cfg.decoding.beam = cfg.decoding 147 148 # Update model's decoding strategy 149 model.change_decoding_strategy(model.cfg.decoding) 150 logging.setLevel(level) 151 152 wer_dist_first = cer_dist_first = 0 153 wer_dist_best = cer_dist_best = 0 154 words_count = 0 155 chars_count = 0 156 sample_idx = 0 157 if preds_output_file: 158 out_file = open(preds_output_file, 'w', encoding='utf_8', newline='\n') 159 160 if progress_bar: 161 if cfg.decoding_strategy == "greedy_batch": 162 description = "Greedy_batch decoding.." 163 else: 164 description = f"{cfg.decoding_strategy} decoding with bw={cfg.decoding.beam_size}, ba={cfg.decoding.ngram_lm_alpha}, ma={cfg.decoding.maes_prefix_alpha}, mg={cfg.decoding.maes_expansion_gamma}, hat_ilmw={cfg.decoding.hat_ilm_weight}" 165 it = tqdm(range(int(np.ceil(len(all_probs) / beam_batch_size))), desc=description, ncols=120) 166 else: 167 it = range(int(np.ceil(len(all_probs) / beam_batch_size))) 168 for batch_idx in it: 169 # disabling type checking 170 probs_batch = all_probs[batch_idx * beam_batch_size : (batch_idx + 1) * beam_batch_size] 171 probs_lens = torch.tensor([prob.shape[-1] for prob in probs_batch]) 172 with torch.no_grad(): 173 packed_batch = torch.zeros(len(probs_batch), probs_batch[0].shape[0], max(probs_lens), device='cpu') 174 175 for prob_index in range(len(probs_batch)): 176 packed_batch[prob_index, :, : probs_lens[prob_index]] = torch.tensor( 177 probs_batch[prob_index].unsqueeze(0), device=packed_batch.device, dtype=packed_batch.dtype 178 ) 179 best_hyp_batch, beams_batch = model.decoding.rnnt_decoder_predictions_tensor( 180 packed_batch, probs_lens, return_hypotheses=True, 181 ) 182 if cfg.decoding_strategy == "greedy_batch": 183 beams_batch = [[x] for x in best_hyp_batch] 184 185 for beams_idx, beams in enumerate(beams_batch): 186 target = target_transcripts[sample_idx + beams_idx] 187 target_split_w = target.split() 188 target_split_c = list(target) 189 words_count += len(target_split_w) 190 chars_count += len(target_split_c) 191 wer_dist_min = cer_dist_min = 10000 192 for candidate_idx, candidate in enumerate(beams): # type: (int, rnnt_beam_decoding.rnnt_utils.Hypothesis) 193 pred_text = candidate.text 194 pred_split_w = pred_text.split() 195 wer_dist = editdistance.eval(target_split_w, pred_split_w) 196 pred_split_c = list(pred_text) 197 cer_dist = editdistance.eval(target_split_c, pred_split_c) 198 199 wer_dist_min = min(wer_dist_min, wer_dist) 200 cer_dist_min = min(cer_dist_min, cer_dist) 201 202 if candidate_idx == 0: 203 # first candidate 204 wer_dist_first += wer_dist 205 cer_dist_first += cer_dist 206 207 score = candidate.score 208 if preds_output_file: 209 out_file.write('{}\t{}\n'.format(pred_text, score)) 210 wer_dist_best += wer_dist_min 211 cer_dist_best += cer_dist_min 212 sample_idx += len(probs_batch) 213 214 if cfg.decoding_strategy == "greedy_batch": 215 return wer_dist_first / words_count, cer_dist_first / chars_count 216 217 if preds_output_file: 218 out_file.close() 219 logging.info(f"Stored the predictions of {cfg.decoding_strategy} decoding at '{preds_output_file}'.") 220 221 if cfg.decoding.ngram_lm_model: 222 logging.info( 223 f"WER/CER with {cfg.decoding_strategy} decoding and N-gram model = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 224 ) 225 else: 226 logging.info( 227 f"WER/CER with {cfg.decoding_strategy} decoding = {wer_dist_first / words_count:.2%}/{cer_dist_first / chars_count:.2%}" 228 ) 229 logging.info( 230 f"Oracle WER/CER in candidates with perfect LM= {wer_dist_best / words_count:.2%}/{cer_dist_best / chars_count:.2%}" 231 ) 232 logging.info(f"=================================================================================") 233 234 return wer_dist_first / words_count, cer_dist_first / chars_count 235 236 237 @hydra_runner(config_path=None, config_name='EvalBeamSearchNGramConfig', schema=EvalBeamSearchNGramConfig) 238 def main(cfg: EvalBeamSearchNGramConfig): 239 if is_dataclass(cfg): 240 cfg = OmegaConf.structured(cfg) # type: EvalBeamSearchNGramConfig 241 242 valid_decoding_strategis = ["greedy_batch", "beam", "tsd", "alsd", "maes"] 243 if cfg.decoding_strategy not in valid_decoding_strategis: 244 raise ValueError( 245 f"Given decoding_strategy={cfg.decoding_strategy} is invalid. Available options are :\n" 246 f"{valid_decoding_strategis}" 247 ) 248 249 if cfg.nemo_model_file.endswith('.nemo'): 250 asr_model = nemo_asr.models.ASRModel.restore_from(cfg.nemo_model_file, map_location=torch.device(cfg.device)) 251 else: 252 logging.warning( 253 "nemo_model_file does not end with .nemo, therefore trying to load a pretrained model with this name." 254 ) 255 asr_model = nemo_asr.models.ASRModel.from_pretrained( 256 cfg.nemo_model_file, map_location=torch.device(cfg.device) 257 ) 258 259 if cfg.kenlm_model_file: 260 if not os.path.exists(cfg.kenlm_model_file): 261 raise FileNotFoundError(f"Could not find the KenLM model file '{cfg.kenlm_model_file}'.") 262 if cfg.decoding_strategy != "maes": 263 raise ValueError(f"Decoding with kenlm model is supported only for maes decoding algorithm.") 264 lm_path = cfg.kenlm_model_file 265 else: 266 lm_path = None 267 cfg.beam_alpha = [0.0] 268 if cfg.hat_subtract_ilm: 269 assert lm_path, "kenlm must be set for hat internal lm subtraction" 270 271 if cfg.decoding_strategy != "maes": 272 cfg.maes_prefix_alpha, cfg.maes_expansion_gamma, cfg.hat_ilm_weight = [0], [0], [0] 273 274 target_transcripts = [] 275 manifest_dir = Path(cfg.input_manifest).parent 276 with open(cfg.input_manifest, 'r', encoding='utf_8') as manifest_file: 277 audio_file_paths = [] 278 for line in tqdm(manifest_file, desc=f"Reading Manifest {cfg.input_manifest} ...", ncols=120): 279 data = json.loads(line) 280 audio_file = Path(data['audio_filepath']) 281 if not audio_file.is_file() and not audio_file.is_absolute(): 282 audio_file = manifest_dir / audio_file 283 target_transcripts.append(data['text']) 284 audio_file_paths.append(str(audio_file.absolute())) 285 286 if cfg.probs_cache_file and os.path.exists(cfg.probs_cache_file): 287 logging.info(f"Found a pickle file of probabilities at '{cfg.probs_cache_file}'.") 288 logging.info(f"Loading the cached pickle file of probabilities from '{cfg.probs_cache_file}' ...") 289 with open(cfg.probs_cache_file, 'rb') as probs_file: 290 all_probs = pickle.load(probs_file) 291 292 if len(all_probs) != len(audio_file_paths): 293 raise ValueError( 294 f"The number of samples in the probabilities file '{cfg.probs_cache_file}' does not " 295 f"match the manifest file. You may need to delete the probabilities cached file." 296 ) 297 else: 298 299 @contextlib.contextmanager 300 def default_autocast(): 301 yield 302 303 if cfg.use_amp: 304 if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 305 logging.info("AMP is enabled!\n") 306 autocast = torch.cuda.amp.autocast 307 308 else: 309 autocast = default_autocast 310 else: 311 312 autocast = default_autocast 313 314 # manual calculation of encoder_embeddings 315 with autocast(): 316 with torch.no_grad(): 317 asr_model.eval() 318 asr_model.encoder.freeze() 319 device = next(asr_model.parameters()).device 320 all_probs = [] 321 with tempfile.TemporaryDirectory() as tmpdir: 322 with open(os.path.join(tmpdir, 'manifest.json'), 'w', encoding='utf-8') as fp: 323 for audio_file in audio_file_paths: 324 entry = {'audio_filepath': audio_file, 'duration': 100000, 'text': ''} 325 fp.write(json.dumps(entry) + '\n') 326 config = { 327 'paths2audio_files': audio_file_paths, 328 'batch_size': cfg.acoustic_batch_size, 329 'temp_dir': tmpdir, 330 'num_workers': cfg.num_workers, 331 'channel_selector': None, 332 'augmentor': None, 333 } 334 temporary_datalayer = asr_model._setup_transcribe_dataloader(config) 335 for test_batch in tqdm(temporary_datalayer, desc="Transcribing", disable=True): 336 encoded, encoded_len = asr_model.forward( 337 input_signal=test_batch[0].to(device), input_signal_length=test_batch[1].to(device) 338 ) 339 # dump encoder embeddings per file 340 for idx in range(encoded.shape[0]): 341 encoded_no_pad = encoded[idx, :, : encoded_len[idx]] 342 all_probs.append(encoded_no_pad) 343 344 if cfg.probs_cache_file: 345 logging.info(f"Writing pickle files of probabilities at '{cfg.probs_cache_file}'...") 346 with open(cfg.probs_cache_file, 'wb') as f_dump: 347 pickle.dump(all_probs, f_dump) 348 349 if cfg.decoding_strategy == "greedy_batch": 350 asr_model = asr_model.to('cpu') 351 candidate_wer, candidate_cer = decoding_step( 352 asr_model, 353 cfg, 354 all_probs=all_probs, 355 target_transcripts=target_transcripts, 356 beam_batch_size=cfg.beam_batch_size, 357 progress_bar=True, 358 ) 359 logging.info(f"Greedy batch WER/CER = {candidate_wer:.2%}/{candidate_cer:.2%}") 360 361 asr_model = asr_model.to('cpu') 362 363 # 'greedy_batch' decoding_strategy would skip the beam search decoding 364 if cfg.decoding_strategy in ["beam", "tsd", "alsd", "maes"]: 365 if cfg.beam_width is None or cfg.beam_alpha is None: 366 raise ValueError("beam_width and beam_alpha are needed to perform beam search decoding.") 367 params = { 368 'beam_width': cfg.beam_width, 369 'beam_alpha': cfg.beam_alpha, 370 'maes_prefix_alpha': cfg.maes_prefix_alpha, 371 'maes_expansion_gamma': cfg.maes_expansion_gamma, 372 'hat_ilm_weight': cfg.hat_ilm_weight, 373 } 374 hp_grid = ParameterGrid(params) 375 hp_grid = list(hp_grid) 376 377 best_wer_beam_size, best_cer_beam_size = None, None 378 best_wer_alpha, best_cer_alpha = None, None 379 best_wer, best_cer = 1e6, 1e6 380 381 logging.info( 382 f"==============================Starting the {cfg.decoding_strategy} decoding===============================" 383 ) 384 logging.info(f"Grid search size: {len(hp_grid)}") 385 logging.info(f"It may take some time...") 386 logging.info(f"==============================================================================================") 387 388 if cfg.preds_output_folder and not os.path.exists(cfg.preds_output_folder): 389 os.mkdir(cfg.preds_output_folder) 390 for hp in hp_grid: 391 if cfg.preds_output_folder: 392 results_file = f"preds_out_{cfg.decoding_strategy}_bw{hp['beam_width']}" 393 if cfg.decoding_strategy == "maes": 394 results_file = f"{results_file}_ma{hp['maes_prefix_alpha']}_mg{hp['maes_expansion_gamma']}" 395 if cfg.kenlm_model_file: 396 results_file = f"{results_file}_ba{hp['beam_alpha']}" 397 if cfg.hat_subtract_ilm: 398 results_file = f"{results_file}_hat_ilmw{hp['hat_ilm_weight']}" 399 preds_output_file = os.path.join(cfg.preds_output_folder, f"{results_file}.tsv") 400 else: 401 preds_output_file = None 402 403 cfg.decoding.beam_size = hp["beam_width"] 404 cfg.decoding.ngram_lm_alpha = hp["beam_alpha"] 405 cfg.decoding.maes_prefix_alpha = hp["maes_prefix_alpha"] 406 cfg.decoding.maes_expansion_gamma = hp["maes_expansion_gamma"] 407 cfg.decoding.hat_ilm_weight = hp["hat_ilm_weight"] 408 409 candidate_wer, candidate_cer = decoding_step( 410 asr_model, 411 cfg, 412 all_probs=all_probs, 413 target_transcripts=target_transcripts, 414 preds_output_file=preds_output_file, 415 beam_batch_size=cfg.beam_batch_size, 416 progress_bar=True, 417 ) 418 419 if candidate_cer < best_cer: 420 best_cer_beam_size = hp["beam_width"] 421 best_cer_alpha = hp["beam_alpha"] 422 best_cer_ma = hp["maes_prefix_alpha"] 423 best_cer_mg = hp["maes_expansion_gamma"] 424 best_cer_hat_ilm_weight = hp["hat_ilm_weight"] 425 best_cer = candidate_cer 426 427 if candidate_wer < best_wer: 428 best_wer_beam_size = hp["beam_width"] 429 best_wer_alpha = hp["beam_alpha"] 430 best_wer_ma = hp["maes_prefix_alpha"] 431 best_wer_ga = hp["maes_expansion_gamma"] 432 best_wer_hat_ilm_weight = hp["hat_ilm_weight"] 433 best_wer = candidate_wer 434 435 wer_hat_parameter = "" 436 if cfg.hat_subtract_ilm: 437 wer_hat_parameter = f"HAT ilm weight = {best_wer_hat_ilm_weight}, " 438 logging.info( 439 f'Best WER Candidate = {best_wer:.2%} :: Beam size = {best_wer_beam_size}, ' 440 f'Beam alpha = {best_wer_alpha}, {wer_hat_parameter}' 441 f'maes_prefix_alpha = {best_wer_ma}, maes_expansion_gamma = {best_wer_ga} ' 442 ) 443 444 cer_hat_parameter = "" 445 if cfg.hat_subtract_ilm: 446 cer_hat_parameter = f"HAT ilm weight = {best_cer_hat_ilm_weight}" 447 logging.info( 448 f'Best CER Candidate = {best_cer:.2%} :: Beam size = {best_cer_beam_size}, ' 449 f'Beam alpha = {best_cer_alpha}, {cer_hat_parameter} ' 450 f'maes_prefix_alpha = {best_cer_ma}, maes_expansion_gamma = {best_cer_mg}' 451 ) 452 logging.info(f"=================================================================================") 453 454 455 if __name__ == '__main__': 456 main() 457 [end of scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py] [start of scripts/confidence_ensembles/build_ensemble.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 """ 16 This script provides a functionality to create confidence-based ensembles 17 from a collection of pretrained models. 18 19 For more details see the paper https://arxiv.org/abs/2306.15824 20 or tutorial in tutorials/asr/Confidence_Ensembles.ipynb 21 22 You would typically use this script by providing a yaml config file or overriding 23 default options from command line. 24 25 Usage examples: 26 27 1. Building an ensemble of two monolingual models with default settings (no confidence tuning). 28 29 python build_ensemble.py --config-path=. --config-name=ensemble_config.yaml 30 ensemble.0.model=stt_it_conformer_ctc_large 31 ensemble.0.training_manifest=<path to the Italian data of 100+ utterances (no transcription required)> 32 ensemble.1.model=stt_es_conformer_ctc_large 33 ensemble.1.training_manifest=<path to the Spanish data of 100+ utterances (no transcription required)> 34 output_path=<path to the desired location of the .nemo checkpoint> 35 36 You can have more than 2 models and can control transcription settings (e.g., batch size) 37 with ``transcription.<any argument of examples/asr/transcribe_speech.py>`` parameters. 38 39 2. If you want to get improved results, you can enable tuning of the confidence and logistic regression (LR) parameters. 40 E.g. 41 42 python build_ensemble.py 43 <all arguments like in the previous example> 44 ensemble.0.dev_manifest=<path to the dev data that's required for tuning> 45 ... 46 # IMPORTANT: see the note below if you use > 2 models! 47 ensemble.N.dev_manifest=<path to the dev data that's required for tuning> 48 tune_confidence=True # to allow confidence tuning. LR is tuned by default 49 50 As with any tuning, it is recommended to have reasonably large validation set for each model, 51 otherwise you might overfit to the validation data. 52 53 Note that if you add additional models (> 2) you will need to modify ensemble_config.yaml 54 or create a new one with added models in there. While it's theoretically possible to 55 fully override such parameters from commandline, hydra is very unfriendly for such 56 use-cases, so it's strongly recommended to be creating new configs. 57 58 3. If you want to precisely control tuning grid search, you can do that with 59 60 python build_ensemble.py 61 <all arguments as in the previous examples> 62 tune_confidence_config.confidence_type='[entropy_renyi_exp,entropy_tsallis_exp]' # only tune over this set 63 tune_confidence_config.alpha='[0.1,0.5,1.0]' # only tune over this set 64 65 You can check the dataclasses in this file for the full list of supported 66 arguments and their default values. 67 """ 68 69 import atexit 70 71 # using default logging to be able to silence unnecessary messages from nemo 72 import logging 73 import os 74 import random 75 import sys 76 import tempfile 77 from copy import deepcopy 78 from dataclasses import dataclass 79 from pathlib import Path 80 from typing import Dict, List, Optional, Tuple 81 82 import joblib 83 import numpy as np 84 import pytorch_lightning as pl 85 from omegaconf import MISSING, DictConfig, OmegaConf 86 from sklearn.linear_model import LogisticRegression 87 from sklearn.metrics import confusion_matrix 88 from sklearn.pipeline import Pipeline, make_pipeline 89 from sklearn.preprocessing import StandardScaler 90 from tqdm import tqdm 91 92 from nemo.collections.asr.models.confidence_ensemble import ( 93 ConfidenceEnsembleModel, 94 ConfidenceSpec, 95 compute_confidence, 96 get_filtered_logprobs, 97 ) 98 from nemo.collections.asr.parts.utils.asr_confidence_utils import ( 99 ConfidenceConfig, 100 ConfidenceMethodConfig, 101 get_confidence_aggregation_bank, 102 get_confidence_measure_bank, 103 ) 104 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis 105 from nemo.core.config import hydra_runner 106 107 LOG = logging.getLogger(__file__) 108 109 # adding Python path. If not found, asking user to get the file 110 try: 111 sys.path.append(str(Path(__file__).parents[2] / "examples" / "asr")) 112 import transcribe_speech 113 except ImportError: 114 # if users run script normally from nemo repo, this shouldn't be triggered as 115 # we modify the path above. But if they downloaded the build_ensemble.py as 116 # an isolated script, we'd ask them to also download corresponding version 117 # of the transcribe_speech.py 118 print( 119 "Current script depends on 'examples/asr/transcribe_speech.py', but can't find it. " 120 "If it's not present, download it from the NeMo github manually and put inside this folder." 121 ) 122 123 124 @dataclass 125 class EnsembleConfig: 126 # .nemo path or pretrained name 127 model: str = MISSING 128 # path to the training data manifest (non-tarred) 129 training_manifest: str = MISSING 130 # specify to limit the number of training samples 131 # 100 is most likely enough, but setting higher default just in case 132 max_training_samples: int = 1000 133 # specify to provide dev data manifest for HP tuning 134 dev_manifest: Optional[str] = None 135 136 137 @dataclass 138 class TuneConfidenceConfig: 139 # important parameter, so should always be tuned 140 exclude_blank: Tuple[bool] = (True, False) 141 # prod is pretty much always worse, so not including by default 142 aggregation: Tuple[str] = ("mean", "min", "max") 143 # not including max prob, as there is always an entropy-based metric 144 # that's better but otherwise including everything 145 confidence_type: Tuple[str] = ( 146 "entropy_renyi_exp", 147 "entropy_renyi_lin", 148 "entropy_tsallis_exp", 149 "entropy_tsallis_lin", 150 "entropy_gibbs_lin", 151 "entropy_gibbs_exp", 152 ) 153 154 # TODO: currently it's not possible to efficiently tune temperature, as we always 155 # apply log-softmax in the decoder, so to try different values it will be required 156 # to rerun the decoding, which is very slow. To support this for one-off experiments 157 # it's possible to modify the code of CTC decoder / Transducer joint to 158 # remove log-softmax and then apply it directly in this script with the temperature 159 # 160 # Alternatively, one can run this script multiple times with different values of 161 # temperature and pick the best performing ensemble. Note that this will increase 162 # tuning time by the number of temperature values tried. On the other hand, 163 # the above approach is a lot more efficient and will only slightly increase 164 # the total tuning runtime. 165 166 # very important to tune for max prob, but for entropy metrics 1.0 is almost always best 167 # temperature: Tuple[float] = (1.0,) 168 169 # not that important, but can sometimes make a small difference 170 alpha: Tuple[float] = (0.25, 0.33, 0.5, 1.0) 171 172 def get_grid_size(self) -> int: 173 """Returns the total number of points in the search space.""" 174 if "max_prob" in self.confidence_type: 175 return ( 176 len(self.exclude_blank) 177 * len(self.aggregation) 178 * ((len(self.confidence_type) - 1) * len(self.alpha) + 1) 179 ) 180 return len(self.exclude_blank) * len(self.aggregation) * len(self.confidence_type) * len(self.alpha) 181 182 183 @dataclass 184 class TuneLogisticRegressionConfig: 185 # will have log-uniform grid over this range with that many points 186 # note that a value of 10000.0 (not regularization) is always added 187 C_num_points: int = 10 188 C_min: float = 0.0001 189 C_max: float = 10.0 190 191 # not too important 192 multi_class: Tuple[str] = ("ovr", "multinomial") 193 194 # should try to include weights directly if the data is too imbalanced 195 class_weight: Tuple = (None, "balanced") 196 197 # increase if getting many warnings that algorithm didn't converge 198 max_iter: int = 1000 199 200 201 @dataclass 202 class BuildEnsembleConfig: 203 # where to save the resulting ensemble model 204 output_path: str = MISSING 205 206 # each model specification 207 ensemble: List[EnsembleConfig] = MISSING 208 209 random_seed: int = 0 # for reproducibility 210 211 # default confidence, can override 212 confidence: ConfidenceConfig = ConfidenceConfig( 213 # we keep frame confidences and apply aggregation manually to get full-utterance confidence 214 preserve_frame_confidence=True, 215 exclude_blank=True, 216 aggregation="mean", 217 method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), 218 ) 219 temperature: float = 1.0 220 221 # this is optional, but can be used to change any aspect of the transcription 222 # config, such as batch size or amp usage. Note that model, data and confidence 223 # will be overriden by this script 224 transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() 225 226 # set to True to tune the confidence. 227 # requires dev manifests to be specified for each model 228 tune_confidence: bool = False 229 # used to specify what to tune over. By default runs tuning over some 230 # reasonalbe grid, so that it does not take forever. 231 # Can be changed as needed 232 tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() 233 234 # very fast to tune and can be important in case of imbalanced datasets 235 # will automatically set to False if dev data is not available 236 tune_logistic_regression: bool = True 237 tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() 238 239 def __post_init__(self): 240 """Checking that if any dev data is provided, all are provided. 241 242 Will also auto-set tune_logistic_regression to False if no dev data 243 is available. 244 245 If tune_confidence is set to True (user choice) and no dev data is 246 provided, will raise an error. 247 """ 248 num_dev_data = 0 249 for ensemble_cfg in self.ensemble: 250 num_dev_data += ensemble_cfg.dev_manifest is not None 251 if num_dev_data == 0: 252 if self.tune_confidence: 253 raise ValueError("tune_confidence is set to True, but no dev data is provided") 254 LOG.info("Setting tune_logistic_regression = False since no dev data is provided") 255 self.tune_logistic_regression = False 256 return 257 258 if num_dev_data < len(self.ensemble): 259 raise ValueError( 260 "Some ensemble configs specify dev data, but some don't. Either all have to specify it or none!" 261 ) 262 263 264 def calculate_score(features: np.ndarray, labels: np.ndarray, pipe: Pipeline) -> Tuple[float, np.ndarray]: 265 """Score is always calculated as mean of the per-class scores. 266 267 This is done to account for possible class imbalances. 268 269 Args: 270 features: numpy array of features of shape [N x D], where N is the 271 number of objects (typically a total number of utterances in 272 all datasets) and D is the total number of confidence scores 273 used to train the model (typically = number of models). 274 labels: numpy array of shape [N] contatining ground-truth model indices. 275 pipe: classification pipeline (currently, standardization + logistic 276 regression). 277 278 Returns: 279 tuple: score value in [0, 1] and full classification confusion matrix. 280 """ 281 predictions = pipe.predict(features) 282 conf_m = confusion_matrix(labels, predictions) 283 score = np.diag(conf_m).sum() / conf_m.sum() 284 return score, conf_m 285 286 287 def train_model_selection( 288 training_features: np.ndarray, 289 training_labels: np.ndarray, 290 dev_features: Optional[np.ndarray] = None, 291 dev_labels: Optional[np.ndarray] = None, 292 tune_lr: bool = False, 293 tune_lr_cfg: Optional[TuneLogisticRegressionConfig] = None, 294 verbose: bool = False, 295 ) -> Tuple[Pipeline, float]: 296 """Trains model selection block with an (optional) tuning of the parameters. 297 298 Returns a pipeline consisting of feature standardization and logistic 299 regression. If tune_lr is set to True, dev features/labels will be used 300 to tune the hyperparameters of the logistic regression with the grid 301 search that's defined via ``tune_lr_cfg``. 302 303 If no tuning is requested, uses the following parameters:: 304 305 best_pipe = make_pipeline( 306 StandardScaler(), 307 LogisticRegression( 308 multi_class="multinomial", 309 C=10000.0, 310 max_iter=1000, 311 class_weight="balanced", 312 ), 313 ) 314 315 Args: 316 training_features: numpy array of features of shape [N x D], where N is 317 the number of objects (typically a total number of utterances in 318 all training datasets) and D is the total number of confidence 319 scores used to train the model (typically = number of models). 320 training_labels: numpy array of shape [N] contatining ground-truth 321 model indices. 322 dev_features: same as training, but for the validation subset. 323 dev_labels: same as training, but for the validation subset. 324 tune_lr: controls whether tuning of LR hyperparameters is performed. 325 If set to True, it's required to also provide dev features/labels. 326 tune_lr_cfg: specifies what values of LR hyperparameters to try. 327 verbose: if True, will output final training/dev scores. 328 329 Returns: 330 tuple: trained model selection pipeline, best score (or -1 if no tuning 331 was done). 332 """ 333 if not tune_lr: 334 # default parameters: C=10000.0 disables regularization 335 best_pipe = make_pipeline( 336 StandardScaler(), 337 LogisticRegression(multi_class="multinomial", C=10000.0, max_iter=1000, class_weight="balanced"), 338 ) 339 max_score = -1 340 else: 341 C_pms = np.append( 342 np.exp(np.linspace(np.log(tune_lr_cfg.C_min), np.log(tune_lr_cfg.C_max), tune_lr_cfg.C_num_points)), 343 10000.0, 344 ) 345 max_score = 0 346 best_pipe = None 347 for class_weight in tune_lr_cfg.class_weight: 348 for multi_class in tune_lr_cfg.multi_class: 349 for C in C_pms: 350 pipe = make_pipeline( 351 StandardScaler(), 352 LogisticRegression( 353 multi_class=multi_class, C=C, max_iter=tune_lr_cfg.max_iter, class_weight=class_weight 354 ), 355 ) 356 pipe.fit(training_features, training_labels) 357 score, confusion = calculate_score(dev_features, dev_labels, pipe) 358 if score > max_score: 359 max_score = score 360 best_pipe = pipe 361 362 best_pipe.fit(training_features, training_labels) 363 if verbose: 364 accuracy, confusion = calculate_score(training_features, training_labels, best_pipe) 365 LOG.info("Training fit accuracy: %.4f", accuracy * 100.0) 366 LOG.info("Training confusion matrix:\n%s", str(confusion)) 367 if dev_features is not None and verbose: 368 accuracy, confusion = calculate_score(dev_features, dev_labels, best_pipe) 369 LOG.info("Dev fit accuracy: %.4f", accuracy * 100.0) 370 LOG.info("Dev confusion matrix:\n%s", str(confusion)) 371 372 return best_pipe, max_score 373 374 375 def subsample_manifest(manifest_file: str, max_samples: int) -> str: 376 """Will save a subsampled version of the manifest to the same folder. 377 378 Have to save to the same folder to support relative paths. 379 380 Args: 381 manifest_file: path to the manifest file that needs subsampling. 382 max_samples: how many samples to retain. Will randomly select that 383 many lines from the manifest. 384 385 Returns: 386 str: the path to the subsampled manifest file. 387 """ 388 with open(manifest_file, "rt", encoding="utf-8") as fin: 389 lines = fin.readlines() 390 if max_samples < len(lines): 391 lines = random.sample(lines, max_samples) 392 output_file = manifest_file + "-subsampled" 393 with open(output_file, "wt", encoding="utf-8") as fout: 394 fout.write("".join(lines)) 395 return output_file 396 397 398 def cleanup_subsampled_manifests(subsampled_manifests: List[str]): 399 """Removes all generated subsamples manifests.""" 400 for manifest in subsampled_manifests: 401 os.remove(manifest) 402 403 404 def compute_all_confidences( 405 hypothesis: Hypothesis, tune_confidence_cfg: TuneConfidenceConfig 406 ) -> Dict[ConfidenceSpec, float]: 407 """Computes a set of confidence scores from a given hypothesis. 408 409 Works with the output of both CTC and Transducer decoding. 410 411 Args: 412 hypothesis: generated hypothesis as returned from the transcribe 413 method of the ASR model. 414 tune_confidence_cfg: config specifying what confidence scores to 415 compute. 416 417 Returns: 418 dict: dictionary with confidenct spec -> confidence score mapping. 419 """ 420 conf_values = {} 421 422 for exclude_blank in tune_confidence_cfg.exclude_blank: 423 filtered_logprobs = get_filtered_logprobs(hypothesis, exclude_blank) 424 vocab_size = filtered_logprobs.shape[1] 425 for aggregation in tune_confidence_cfg.aggregation: 426 aggr_func = get_confidence_aggregation_bank()[aggregation] 427 for conf_type in tune_confidence_cfg.confidence_type: 428 conf_func = get_confidence_measure_bank()[conf_type] 429 if conf_type == "max_prob": # skipping alpha in this case 430 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=1.0)).cpu().item() 431 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, 1.0)] = conf_value 432 else: 433 for alpha in tune_confidence_cfg.alpha: 434 conf_value = aggr_func(conf_func(filtered_logprobs, v=vocab_size, t=alpha)).cpu().item() 435 conf_values[ConfidenceSpec(exclude_blank, aggregation, conf_type, alpha)] = conf_value 436 437 return conf_values 438 439 440 def find_best_confidence( 441 train_confidences: List[List[Dict[ConfidenceSpec, float]]], 442 train_labels: List[int], 443 dev_confidences: List[List[Dict[ConfidenceSpec, float]]], 444 dev_labels: List[int], 445 tune_lr: bool, 446 tune_lr_config: TuneConfidenceConfig, 447 ) -> Tuple[ConfidenceConfig, Pipeline]: 448 """Finds the best confidence configuration for model selection. 449 450 Will loop over all values in the confidence dictionary and fit the LR 451 model (optionally tuning its HPs). The best performing confidence (on the 452 dev set) will be used for the final LR model. 453 454 Args: 455 train_confidences: this is an object of type 456 ``List[List[Dict[ConfidenceSpec, float]]]``. The shape of this 457 object is [M, N, S], where 458 M: number of models 459 N: number of utterances in all training sets 460 S: number of confidence scores to try 461 462 This argument will be used to construct np.array objects for each 463 of the confidence scores with the shape [M, N] 464 465 train_labels: ground-truth labels of the correct model for each data 466 points. This is a list of size [N] 467 dev_confidences: same as training, but for the validation subset. 468 dev_labels: same as training, but for the validation subset. 469 tune_lr: controls whether tuning of LR hyperparameters is performed. 470 tune_lr_cfg: specifies what values of LR hyperparameters to try. 471 472 Returns: 473 tuple: best confidence config, best model selection pipeline 474 """ 475 max_score = 0 476 best_pipe = None 477 best_conf_spec = None 478 LOG.info("Evaluation all confidences. Total grid size: %d", len(train_confidences[0][0].keys())) 479 for conf_spec in tqdm(train_confidences[0][0].keys()): 480 cur_train_confidences = [] 481 for model_confs in train_confidences: 482 cur_train_confidences.append([]) 483 for model_conf in model_confs: 484 cur_train_confidences[-1].append(model_conf[conf_spec]) 485 cur_dev_confidences = [] 486 for model_confs in dev_confidences: 487 cur_dev_confidences.append([]) 488 for model_conf in model_confs: 489 cur_dev_confidences[-1].append(model_conf[conf_spec]) 490 # transposing with zip(*list) 491 training_features = np.array(list(zip(*cur_train_confidences))) 492 training_labels = np.array(train_labels) 493 dev_features = np.array(list(zip(*cur_dev_confidences))) 494 dev_labels = np.array(dev_labels) 495 pipe, score = train_model_selection( 496 training_features, training_labels, dev_features, dev_labels, tune_lr, tune_lr_config, 497 ) 498 if max_score < score: 499 max_score = score 500 best_pipe = pipe 501 best_conf_spec = conf_spec 502 LOG.info("Found better parameters: %s. New score: %.4f", str(conf_spec), max_score) 503 504 return best_conf_spec.to_confidence_config(), best_pipe 505 506 507 @hydra_runner(config_name="BuildEnsembleConfig", schema=BuildEnsembleConfig) 508 def main(cfg: BuildEnsembleConfig): 509 # silencing all messages from nemo/ptl to avoid dumping tons of configs to the stdout 510 logging.getLogger('pytorch_lightning').setLevel(logging.CRITICAL) 511 logging.getLogger('nemo_logger').setLevel(logging.CRITICAL) 512 LOG.info(f'Build ensemble config:\n{OmegaConf.to_yaml(cfg)}') 513 514 # to ensure post init is called 515 cfg = BuildEnsembleConfig(**cfg) 516 517 pl.seed_everything(cfg.random_seed) 518 cfg.transcription.random_seed = None # seed is already applied 519 cfg.transcription.return_transcriptions = True 520 cfg.transcription.preserve_alignment = True 521 cfg.transcription.ctc_decoding.temperature = cfg.temperature 522 cfg.transcription.rnnt_decoding.temperature = cfg.temperature 523 # this ensures that generated output is after log-softmax for consistency with CTC 524 525 train_confidences = [] 526 dev_confidences = [] 527 train_labels = [] 528 dev_labels = [] 529 530 # registering clean-up function that will hold on to this list and 531 # should clean up even if there is partial error in some of the transcribe 532 # calls 533 subsampled_manifests = [] 534 atexit.register(cleanup_subsampled_manifests, subsampled_manifests) 535 536 # note that we loop over the same config. 537 # This is intentional, as we need to run all models on all datasets 538 # this loop will do the following things: 539 # 1. Goes through each model X each training dataset 540 # 2. Computes predictions by directly calling transcribe_speech.main 541 # 3. Converts transcription to the confidence score(s) as specified in the config 542 # 4. If dev sets are provided, computes the same for them 543 # 5. Creates a list of ground-truth model indices by mapping each model 544 # to its own training dataset as specified in the config. 545 # 6. After the loop, we either run tuning over all confidence scores or 546 # directly use a single score to fit logistic regression and save the 547 # final ensemble model. 548 for model_idx, model_cfg in enumerate(cfg.ensemble): 549 train_model_confidences = [] 550 dev_model_confidences = [] 551 for data_idx, data_cfg in enumerate(cfg.ensemble): 552 if model_idx == 0: # generating subsampled manifests only one time 553 subsampled_manifests.append( 554 subsample_manifest(data_cfg.training_manifest, data_cfg.max_training_samples) 555 ) 556 subsampled_manifest = subsampled_manifests[data_idx] 557 558 if model_cfg.model.endswith(".nemo"): 559 cfg.transcription.model_path = model_cfg.model 560 else: # assuming pretrained model 561 cfg.transcription.pretrained_name = model_cfg.model 562 563 cfg.transcription.dataset_manifest = subsampled_manifest 564 565 # training 566 with tempfile.NamedTemporaryFile() as output_file: 567 cfg.transcription.output_filename = output_file.name 568 LOG.info("Transcribing training dataset %d with model %d", data_idx, model_idx) 569 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 570 LOG.info("Generating confidence scores") 571 # TODO: parallelize this loop? 572 for transcription in tqdm(transcriptions): 573 if cfg.tune_confidence: 574 train_model_confidences.append( 575 compute_all_confidences(transcription, cfg.tune_confidence_config) 576 ) 577 else: 578 train_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 579 if model_idx == 0: # labels are the same for all models 580 train_labels.append(data_idx) 581 582 # optional dev 583 if data_cfg.dev_manifest is not None: 584 cfg.transcription.dataset_manifest = data_cfg.dev_manifest 585 with tempfile.NamedTemporaryFile() as output_file: 586 cfg.transcription.output_filename = output_file.name 587 LOG.info("Transcribing dev dataset %d with model %d", data_idx, model_idx) 588 transcriptions = transcribe_speech.main(deepcopy(cfg.transcription)) 589 LOG.info("Generating confidence scores") 590 for transcription in tqdm(transcriptions): 591 if cfg.tune_confidence: 592 dev_model_confidences.append( 593 compute_all_confidences(transcription, cfg.tune_confidence_config) 594 ) 595 else: 596 dev_model_confidences.append(compute_confidence(transcription, cfg.confidence)) 597 if model_idx == 0: # labels are the same for all models 598 dev_labels.append(data_idx) 599 600 train_confidences.append(train_model_confidences) 601 if dev_model_confidences: 602 dev_confidences.append(dev_model_confidences) 603 604 if cfg.tune_confidence: 605 best_confidence, model_selection_block = find_best_confidence( 606 train_confidences, 607 train_labels, 608 dev_confidences, 609 dev_labels, 610 cfg.tune_logistic_regression, 611 cfg.tune_logistic_regression_config, 612 ) 613 else: 614 best_confidence = cfg.confidence 615 # transposing with zip(*list) 616 training_features = np.array(list(zip(*train_confidences))) 617 training_labels = np.array(train_labels) 618 if dev_confidences: 619 dev_features = np.array(list(zip(*dev_confidences))) 620 dev_labels = np.array(dev_labels) 621 else: 622 dev_features = None 623 dev_labels = None 624 model_selection_block, _ = train_model_selection( 625 training_features, 626 training_labels, 627 dev_features, 628 dev_labels, 629 cfg.tune_logistic_regression, 630 cfg.tune_logistic_regression_config, 631 verbose=True, 632 ) 633 634 with tempfile.TemporaryDirectory() as tmpdir: 635 model_selection_block_path = os.path.join(tmpdir, 'model_selection_block.pkl') 636 joblib.dump(model_selection_block, model_selection_block_path) 637 638 # creating ensemble checkpoint 639 ensemble_model = ConfidenceEnsembleModel( 640 cfg=DictConfig( 641 { 642 'model_selection_block': model_selection_block_path, 643 'confidence': best_confidence, 644 'temperature': cfg.temperature, 645 'load_models': [model_cfg.model for model_cfg in cfg.ensemble], 646 } 647 ), 648 trainer=None, 649 ) 650 ensemble_model.save_to(cfg.output_path) 651 652 653 if __name__ == '__main__': 654 main() 655 [end of scripts/confidence_ensembles/build_ensemble.py] [start of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] 1 # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import json 16 import os 17 from dataclasses import dataclass, is_dataclass 18 from pathlib import Path 19 from typing import Optional 20 21 import pytorch_lightning as pl 22 import torch 23 from omegaconf import MISSING, OmegaConf 24 from sklearn.model_selection import ParameterGrid 25 26 from nemo.collections.asr.metrics.rnnt_wer import RNNTDecodingConfig 27 from nemo.collections.asr.metrics.wer import CTCDecodingConfig 28 from nemo.collections.asr.models import ASRModel, EncDecRNNTModel 29 from nemo.collections.asr.parts.utils.asr_confidence_benchmarking_utils import ( 30 apply_confidence_parameters, 31 run_confidence_benchmark, 32 ) 33 from nemo.collections.asr.parts.utils.asr_confidence_utils import ConfidenceConfig 34 from nemo.core.config import hydra_runner 35 from nemo.utils import logging, model_utils 36 37 """ 38 Get confidence metrics and curve plots for a given model, dataset, and confidence parameters. 39 40 # Arguments 41 model_path: Path to .nemo ASR checkpoint 42 pretrained_name: Name of pretrained ASR model (from NGC registry) 43 dataset_manifest: Path to dataset JSON manifest file (in NeMo format) 44 output_dir: Output directory to store a report and curve plot directories 45 46 batch_size: batch size during inference 47 num_workers: number of workers during inference 48 49 cuda: Optional int to enable or disable execution of model on certain CUDA device 50 amp: Bool to decide if Automatic Mixed Precision should be used during inference 51 audio_type: Str filetype of the audio. Supported = wav, flac, mp3 52 53 target_level: Word- or token-level confidence. Supported = word, token, auto (for computing both word and token) 54 confidence_cfg: Config with confidence parameters 55 grid_params: Dictionary with lists of parameters to iteratively benchmark on 56 57 # Usage 58 ASR model can be specified by either "model_path" or "pretrained_name". 59 Data for transcription are defined with "dataset_manifest". 60 Results are returned as a benchmark report and curve plots. 61 62 python benchmark_asr_confidence.py \ 63 model_path=null \ 64 pretrained_name=null \ 65 dataset_manifest="" \ 66 output_dir="" \ 67 batch_size=64 \ 68 num_workers=8 \ 69 cuda=0 \ 70 amp=True \ 71 target_level="word" \ 72 confidence_cfg.exclude_blank=False \ 73 'grid_params="{\"aggregation\": [\"min\", \"prod\"], \"alpha\": [0.33, 0.5]}"' 74 """ 75 76 77 def get_experiment_params(cfg): 78 """Get experiment parameters from a confidence config and generate the experiment name. 79 80 Returns: 81 List of experiment parameters. 82 String with the experiment name. 83 """ 84 blank = "no_blank" if cfg.exclude_blank else "blank" 85 aggregation = cfg.aggregation 86 method_name = cfg.method_cfg.name 87 alpha = cfg.method_cfg.alpha 88 if method_name == "entropy": 89 entropy_type = cfg.method_cfg.entropy_type 90 entropy_norm = cfg.method_cfg.entropy_norm 91 experiment_param_list = [ 92 aggregation, 93 str(cfg.exclude_blank), 94 method_name, 95 entropy_type, 96 entropy_norm, 97 str(alpha), 98 ] 99 experiment_str = "-".join([aggregation, blank, method_name, entropy_type, entropy_norm, str(alpha)]) 100 else: 101 experiment_param_list = [aggregation, str(cfg.exclude_blank), method_name, "-", "-", str(alpha)] 102 experiment_str = "-".join([aggregation, blank, method_name, str(alpha)]) 103 return experiment_param_list, experiment_str 104 105 106 @dataclass 107 class ConfidenceBenchmarkingConfig: 108 # Required configs 109 model_path: Optional[str] = None # Path to a .nemo file 110 pretrained_name: Optional[str] = None # Name of a pretrained model 111 dataset_manifest: str = MISSING 112 output_dir: str = MISSING 113 114 # General configs 115 batch_size: int = 32 116 num_workers: int = 4 117 118 # Set `cuda` to int to define CUDA device. If 'None', will look for CUDA 119 # device anyway, and do inference on CPU only if CUDA device is not found. 120 # If `cuda` is a negative number, inference will be on CPU only. 121 cuda: Optional[int] = None 122 amp: bool = False 123 audio_type: str = "wav" 124 125 # Confidence configs 126 target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) 127 confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) 128 grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on 129 130 131 @hydra_runner(config_name="ConfidenceBenchmarkingConfig", schema=ConfidenceBenchmarkingConfig) 132 def main(cfg: ConfidenceBenchmarkingConfig): 133 torch.set_grad_enabled(False) 134 135 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 136 137 if is_dataclass(cfg): 138 cfg = OmegaConf.structured(cfg) 139 140 if cfg.model_path is None and cfg.pretrained_name is None: 141 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None!") 142 143 # setup GPU 144 if cfg.cuda is None: 145 if torch.cuda.is_available(): 146 device = [0] # use 0th CUDA device 147 accelerator = 'gpu' 148 else: 149 device = 1 150 accelerator = 'cpu' 151 else: 152 device = [cfg.cuda] 153 accelerator = 'gpu' 154 155 map_location = torch.device('cuda:{}'.format(device[0]) if accelerator == 'gpu' else 'cpu') 156 157 # setup model 158 if cfg.model_path is not None: 159 # restore model from .nemo file path 160 model_cfg = ASRModel.restore_from(restore_path=cfg.model_path, return_config=True) 161 classpath = model_cfg.target # original class path 162 imported_class = model_utils.import_class_by_path(classpath) # type: ASRModel 163 logging.info(f"Restoring model : {imported_class.__name__}") 164 asr_model = imported_class.restore_from( 165 restore_path=cfg.model_path, map_location=map_location 166 ) # type: ASRModel 167 else: 168 # restore model by name 169 asr_model = ASRModel.from_pretrained( 170 model_name=cfg.pretrained_name, map_location=map_location 171 ) # type: ASRModel 172 173 trainer = pl.Trainer(devices=device, accelerator=accelerator) 174 asr_model.set_trainer(trainer) 175 asr_model = asr_model.eval() 176 177 # Check if ctc or rnnt model 178 is_rnnt = isinstance(asr_model, EncDecRNNTModel) 179 180 # Check that the model has the `change_decoding_strategy` method 181 if not hasattr(asr_model, 'change_decoding_strategy'): 182 raise RuntimeError("The asr_model you are using must have the `change_decoding_strategy` method.") 183 184 # get filenames and reference texts from manifest 185 filepaths = [] 186 reference_texts = [] 187 if os.stat(cfg.dataset_manifest).st_size == 0: 188 logging.error(f"The input dataset_manifest {cfg.dataset_manifest} is empty. Exiting!") 189 return None 190 manifest_dir = Path(cfg.dataset_manifest).parent 191 with open(cfg.dataset_manifest, 'r') as f: 192 for line in f: 193 item = json.loads(line) 194 audio_file = Path(item['audio_filepath']) 195 if not audio_file.is_file() and not audio_file.is_absolute(): 196 audio_file = manifest_dir / audio_file 197 filepaths.append(str(audio_file.absolute())) 198 reference_texts.append(item['text']) 199 200 # setup AMP (optional) 201 autocast = None 202 if cfg.amp and torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'): 203 logging.info("AMP enabled!\n") 204 autocast = torch.cuda.amp.autocast 205 206 # do grid-based benchmarking if grid_params is provided, otherwise a regular one 207 work_dir = Path(cfg.output_dir) 208 os.makedirs(work_dir, exist_ok=True) 209 report_legend = ( 210 ",".join( 211 [ 212 "model_type", 213 "aggregation", 214 "blank", 215 "method_name", 216 "entropy_type", 217 "entropy_norm", 218 "alpha", 219 "target_level", 220 "auc_roc", 221 "auc_pr", 222 "auc_nt", 223 "nce", 224 "ece", 225 "auc_yc", 226 "std_yc", 227 "max_yc", 228 ] 229 ) 230 + "\n" 231 ) 232 model_typename = "RNNT" if is_rnnt else "CTC" 233 report_file = work_dir / Path("report.csv") 234 if cfg.grid_params: 235 asr_model.change_decoding_strategy( 236 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 237 if is_rnnt 238 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 239 ) 240 params = json.loads(cfg.grid_params) 241 hp_grid = ParameterGrid(params) 242 hp_grid = list(hp_grid) 243 244 logging.info(f"==============================Running a benchmarking with grid search=========================") 245 logging.info(f"Grid search size: {len(hp_grid)}") 246 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directories near the file") 247 logging.info(f"==============================================================================================") 248 249 with open(report_file, "tw", encoding="utf-8") as f: 250 f.write(report_legend) 251 f.flush() 252 for i, hp in enumerate(hp_grid): 253 logging.info(f"Run # {i + 1}, grid: `{hp}`") 254 asr_model.change_decoding_strategy(apply_confidence_parameters(asr_model.cfg.decoding, hp)) 255 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 256 plot_dir = work_dir / Path(experiment_name) 257 results = run_confidence_benchmark( 258 asr_model, 259 cfg.target_level, 260 filepaths, 261 reference_texts, 262 cfg.batch_size, 263 cfg.num_workers, 264 plot_dir, 265 autocast, 266 ) 267 for level, result in results.items(): 268 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 269 f.flush() 270 else: 271 asr_model.change_decoding_strategy( 272 RNNTDecodingConfig(fused_batch_size=-1, strategy="greedy_batch", confidence_cfg=cfg.confidence_cfg) 273 if is_rnnt 274 else CTCDecodingConfig(confidence_cfg=cfg.confidence_cfg) 275 ) 276 param_list, experiment_name = get_experiment_params(asr_model.cfg.decoding.confidence_cfg) 277 plot_dir = work_dir / Path(experiment_name) 278 279 logging.info(f"==============================Running a single benchmarking===================================") 280 logging.info(f"Results will be written to:\nreport file `{report_file}`\nand plot directory `{plot_dir}`") 281 282 with open(report_file, "tw", encoding="utf-8") as f: 283 f.write(report_legend) 284 f.flush() 285 results = run_confidence_benchmark( 286 asr_model, 287 cfg.batch_size, 288 cfg.num_workers, 289 cfg.target_level, 290 filepaths, 291 reference_texts, 292 plot_dir, 293 autocast, 294 ) 295 for level, result in results.items(): 296 f.write(f"{model_typename},{','.join(param_list)},{level},{','.join([str(r) for r in result])}\n") 297 logging.info(f"===========================================Done===============================================") 298 299 300 if __name__ == '__main__': 301 main() 302 [end of scripts/speech_recognition/confidence/benchmark_asr_confidence.py] [start of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] 1 # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 """ 15 # This script converts an existing audio dataset with a manifest to 16 # a tarred and sharded audio dataset that can be read by the 17 # TarredAudioToTextDataLayer. 18 19 # Please make sure your audio_filepath DOES NOT CONTAIN '-sub'! 20 # Because we will use it to handle files which have duplicate filenames but with different offsets 21 # (see function create_shard for details) 22 23 24 # Bucketing can help to improve the training speed. You may use --buckets_num to specify the number of buckets. 25 # It creates multiple tarred datasets, one per bucket, based on the audio durations. 26 # The range of [min_duration, max_duration) is split into equal sized buckets. 27 # Recommend to use --sort_in_shards to speedup the training by reducing the paddings in the batches 28 # More info on how to use bucketing feature: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/datasets.html 29 30 # If valid NVIDIA DALI version is installed, will also generate the corresponding DALI index files that need to be 31 # supplied to the config in order to utilize webdataset for efficient large dataset handling. 32 # NOTE: DALI + Webdataset is NOT compatible with Bucketing support ! 33 34 # Usage: 35 1) Creating a new tarfile dataset 36 37 python convert_to_tarred_audio_dataset.py \ 38 --manifest_path=<path to the manifest file> \ 39 --target_dir=<path to output directory> \ 40 --num_shards=<number of tarfiles that will contain the audio> \ 41 --max_duration=<float representing maximum duration of audio samples> \ 42 --min_duration=<float representing minimum duration of audio samples> \ 43 --shuffle --shuffle_seed=1 \ 44 --sort_in_shards \ 45 --workers=-1 46 47 48 2) Concatenating more tarfiles to a pre-existing tarred dataset 49 50 python convert_to_tarred_audio_dataset.py \ 51 --manifest_path=<path to the tarred manifest file> \ 52 --metadata_path=<path to the metadata.yaml (or metadata_version_{X}.yaml) file> \ 53 --target_dir=<path to output directory where the original tarfiles are contained> \ 54 --max_duration=<float representing maximum duration of audio samples> \ 55 --min_duration=<float representing minimum duration of audio samples> \ 56 --shuffle --shuffle_seed=1 \ 57 --sort_in_shards \ 58 --workers=-1 \ 59 --concat_manifest_paths \ 60 <space separated paths to 1 or more manifest files to concatenate into the original tarred dataset> 61 62 3) Writing an empty metadata file 63 64 python convert_to_tarred_audio_dataset.py \ 65 --target_dir=<path to output directory> \ 66 # any other optional argument 67 --num_shards=8 \ 68 --max_duration=16.7 \ 69 --min_duration=0.01 \ 70 --shuffle \ 71 --workers=-1 \ 72 --sort_in_shards \ 73 --shuffle_seed=1 \ 74 --write_metadata 75 76 """ 77 import argparse 78 import copy 79 import json 80 import os 81 import random 82 import tarfile 83 from collections import defaultdict 84 from dataclasses import dataclass, field 85 from datetime import datetime 86 from typing import Any, List, Optional 87 88 from joblib import Parallel, delayed 89 from omegaconf import DictConfig, OmegaConf, open_dict 90 91 try: 92 import create_dali_tarred_dataset_index as dali_index 93 94 DALI_INDEX_SCRIPT_AVAILABLE = True 95 except (ImportError, ModuleNotFoundError, FileNotFoundError): 96 DALI_INDEX_SCRIPT_AVAILABLE = False 97 98 parser = argparse.ArgumentParser( 99 description="Convert an existing ASR dataset to tarballs compatible with TarredAudioToTextDataLayer." 100 ) 101 parser.add_argument( 102 "--manifest_path", default=None, type=str, required=False, help="Path to the existing dataset's manifest." 103 ) 104 105 parser.add_argument( 106 '--concat_manifest_paths', 107 nargs='+', 108 default=None, 109 type=str, 110 required=False, 111 help="Path to the additional dataset's manifests that will be concatenated with base dataset.", 112 ) 113 114 # Optional arguments 115 parser.add_argument( 116 "--target_dir", 117 default='./tarred', 118 type=str, 119 help="Target directory for resulting tarballs and manifest. Defaults to `./tarred`. Creates the path if necessary.", 120 ) 121 122 parser.add_argument( 123 "--metadata_path", required=False, default=None, type=str, help="Path to metadata file for the dataset.", 124 ) 125 126 parser.add_argument( 127 "--num_shards", 128 default=-1, 129 type=int, 130 help="Number of shards (tarballs) to create. Used for partitioning data among workers.", 131 ) 132 parser.add_argument( 133 '--max_duration', 134 default=None, 135 required=True, 136 type=float, 137 help='Maximum duration of audio clip in the dataset. By default, it is None and is required to be set.', 138 ) 139 parser.add_argument( 140 '--min_duration', 141 default=None, 142 type=float, 143 help='Minimum duration of audio clip in the dataset. By default, it is None and will not filter files.', 144 ) 145 parser.add_argument( 146 "--shuffle", 147 action='store_true', 148 help="Whether or not to randomly shuffle the samples in the manifest before tarring/sharding.", 149 ) 150 151 parser.add_argument( 152 "--keep_files_together", 153 action='store_true', 154 help="Whether or not to keep entries from the same file (but different offsets) together when sorting before tarring/sharding.", 155 ) 156 157 parser.add_argument( 158 "--sort_in_shards", 159 action='store_true', 160 help="Whether or not to sort samples inside the shards based on their duration.", 161 ) 162 163 parser.add_argument( 164 "--buckets_num", type=int, default=1, help="Number of buckets to create based on duration.", 165 ) 166 167 parser.add_argument("--shuffle_seed", type=int, default=None, help="Random seed for use if shuffling is enabled.") 168 parser.add_argument( 169 '--write_metadata', 170 action='store_true', 171 help=( 172 "Flag to write a blank metadata with the current call config. " 173 "Note that the metadata will not contain the number of shards, " 174 "and it must be filled out by the user." 175 ), 176 ) 177 parser.add_argument( 178 "--no_shard_manifests", 179 action='store_true', 180 help="Do not write sharded manifests along with the aggregated manifest.", 181 ) 182 parser.add_argument('--workers', type=int, default=1, help='Number of worker processes') 183 args = parser.parse_args() 184 185 186 @dataclass 187 class ASRTarredDatasetConfig: 188 num_shards: int = -1 189 shuffle: bool = False 190 max_duration: Optional[float] = None 191 min_duration: Optional[float] = None 192 shuffle_seed: Optional[int] = None 193 sort_in_shards: bool = True 194 shard_manifests: bool = True 195 keep_files_together: bool = False 196 197 198 @dataclass 199 class ASRTarredDatasetMetadata: 200 created_datetime: Optional[str] = None 201 version: int = 0 202 num_samples_per_shard: Optional[int] = None 203 is_concatenated_manifest: bool = False 204 205 dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() 206 history: Optional[List[Any]] = field(default_factory=lambda: []) 207 208 def __post_init__(self): 209 self.created_datetime = self.get_current_datetime() 210 211 def get_current_datetime(self): 212 return datetime.now().strftime("%m-%d-%Y %H-%M-%S") 213 214 @classmethod 215 def from_config(cls, config: DictConfig): 216 obj = cls() 217 obj.__dict__.update(**config) 218 return obj 219 220 @classmethod 221 def from_file(cls, filepath: str): 222 config = OmegaConf.load(filepath) 223 return ASRTarredDatasetMetadata.from_config(config=config) 224 225 226 class ASRTarredDatasetBuilder: 227 """ 228 Helper class that constructs a tarred dataset from scratch, or concatenates tarred datasets 229 together and constructs manifests for them. 230 """ 231 232 def __init__(self): 233 self.config = None 234 235 def configure(self, config: ASRTarredDatasetConfig): 236 """ 237 Sets the config generated from command line overrides. 238 239 Args: 240 config: ASRTarredDatasetConfig dataclass object. 241 """ 242 self.config = config # type: ASRTarredDatasetConfig 243 244 if self.config.num_shards < 0: 245 raise ValueError("`num_shards` must be > 0. Please fill in the metadata information correctly.") 246 247 def create_new_dataset(self, manifest_path: str, target_dir: str = "./tarred/", num_workers: int = 0): 248 """ 249 Creates a new tarred dataset from a given manifest file. 250 251 Args: 252 manifest_path: Path to the original ASR manifest. 253 target_dir: Output directory. 254 num_workers: Integer denoting number of parallel worker processes which will write tarfiles. 255 Defaults to 1 - which denotes sequential worker process. 256 257 Output: 258 Writes tarfiles, along with the tarred dataset compatible manifest file. 259 Also preserves a record of the metadata used to construct this tarred dataset. 260 """ 261 if self.config is None: 262 raise ValueError("Config has not been set. Please call `configure(config: ASRTarredDatasetConfig)`") 263 264 if manifest_path is None: 265 raise FileNotFoundError("Manifest filepath cannot be None !") 266 267 config = self.config # type: ASRTarredDatasetConfig 268 269 if not os.path.exists(target_dir): 270 os.makedirs(target_dir) 271 272 # Read the existing manifest 273 entries, total_duration, filtered_entries, filtered_duration = self._read_manifest(manifest_path, config) 274 275 if len(filtered_entries) > 0: 276 print(f"Filtered {len(filtered_entries)} files which amounts to {filtered_duration} seconds of audio.") 277 print( 278 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 279 ) 280 281 if len(entries) == 0: 282 print("No tarred dataset was created as there were 0 valid samples after filtering!") 283 return 284 if config.shuffle: 285 random.seed(config.shuffle_seed) 286 print("Shuffling...") 287 if config.keep_files_together: 288 filename_entries = defaultdict(list) 289 for ent in entries: 290 filename_entries[ent["audio_filepath"]].append(ent) 291 filenames = list(filename_entries.keys()) 292 random.shuffle(filenames) 293 shuffled_entries = [] 294 for filename in filenames: 295 shuffled_entries += filename_entries[filename] 296 entries = shuffled_entries 297 else: 298 random.shuffle(entries) 299 300 # Create shards and updated manifest entries 301 print(f"Number of samples added : {len(entries)}") 302 print(f"Remainder: {len(entries) % config.num_shards}") 303 304 start_indices = [] 305 end_indices = [] 306 # Build indices 307 for i in range(config.num_shards): 308 start_idx = (len(entries) // config.num_shards) * i 309 end_idx = start_idx + (len(entries) // config.num_shards) 310 print(f"Shard {i} has entries {start_idx} ~ {end_idx}") 311 files = set() 312 for ent_id in range(start_idx, end_idx): 313 files.add(entries[ent_id]["audio_filepath"]) 314 print(f"Shard {i} contains {len(files)} files") 315 if i == config.num_shards - 1: 316 # We discard in order to have the same number of entries per shard. 317 print(f"Have {len(entries) - end_idx} entries left over that will be discarded.") 318 319 start_indices.append(start_idx) 320 end_indices.append(end_idx) 321 322 manifest_folder, _ = os.path.split(manifest_path) 323 324 with Parallel(n_jobs=num_workers, verbose=config.num_shards) as parallel: 325 # Call parallel tarfile construction 326 new_entries_list = parallel( 327 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, i, manifest_folder) 328 for i, (start_idx, end_idx) in enumerate(zip(start_indices, end_indices)) 329 ) 330 331 if config.shard_manifests: 332 sharded_manifests_dir = target_dir + '/sharded_manifests' 333 if not os.path.exists(sharded_manifests_dir): 334 os.makedirs(sharded_manifests_dir) 335 336 for manifest in new_entries_list: 337 shard_id = manifest[0]['shard_id'] 338 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 339 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 340 for entry in manifest: 341 json.dump(entry, m2) 342 m2.write('\n') 343 344 # Flatten the list of list of entries to a list of entries 345 new_entries = [sample for manifest in new_entries_list for sample in manifest] 346 del new_entries_list 347 348 print("Total number of entries in manifest :", len(new_entries)) 349 350 # Write manifest 351 new_manifest_path = os.path.join(target_dir, 'tarred_audio_manifest.json') 352 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 353 for entry in new_entries: 354 json.dump(entry, m2) 355 m2.write('\n') 356 357 # Write metadata (default metadata for new datasets) 358 new_metadata_path = os.path.join(target_dir, 'metadata.yaml') 359 metadata = ASRTarredDatasetMetadata() 360 361 # Update metadata 362 metadata.dataset_config = config 363 metadata.num_samples_per_shard = len(new_entries) // config.num_shards 364 365 # Write metadata 366 metadata_yaml = OmegaConf.structured(metadata) 367 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 368 369 def create_concatenated_dataset( 370 self, 371 base_manifest_path: str, 372 manifest_paths: List[str], 373 metadata: ASRTarredDatasetMetadata, 374 target_dir: str = "./tarred_concatenated/", 375 num_workers: int = 1, 376 ): 377 """ 378 Creates new tarfiles in order to create a concatenated dataset, whose manifest contains the data for 379 both the original dataset as well as the new data submitted in manifest paths. 380 381 Args: 382 base_manifest_path: Path to the manifest file which contains the information for the original 383 tarred dataset (with flattened paths). 384 manifest_paths: List of one or more paths to manifest files that will be concatenated with above 385 base tarred dataset. 386 metadata: ASRTarredDatasetMetadata dataclass instance with overrides from command line. 387 target_dir: Output directory 388 389 Output: 390 Writes tarfiles which with indices mapping to a "concatenated" tarred dataset, 391 along with the tarred dataset compatible manifest file which includes information 392 about all the datasets that comprise the concatenated dataset. 393 394 Also preserves a record of the metadata used to construct this tarred dataset. 395 """ 396 if not os.path.exists(target_dir): 397 os.makedirs(target_dir) 398 399 if base_manifest_path is None: 400 raise FileNotFoundError("Base manifest filepath cannot be None !") 401 402 if manifest_paths is None or len(manifest_paths) == 0: 403 raise FileNotFoundError("List of additional manifest filepaths cannot be None !") 404 405 config = ASRTarredDatasetConfig(**(metadata.dataset_config)) 406 407 # Read the existing manifest (no filtering here) 408 base_entries, _, _, _ = self._read_manifest(base_manifest_path, config) 409 print(f"Read base manifest containing {len(base_entries)} samples.") 410 411 # Precompute number of samples per shard 412 if metadata.num_samples_per_shard is None: 413 num_samples_per_shard = len(base_entries) // config.num_shards 414 else: 415 num_samples_per_shard = metadata.num_samples_per_shard 416 417 print("Number of samples per shard :", num_samples_per_shard) 418 419 # Compute min and max duration and update config (if no metadata passed) 420 print(f"Selected max duration : {config.max_duration}") 421 print(f"Selected min duration : {config.min_duration}") 422 423 entries = [] 424 for new_manifest_idx in range(len(manifest_paths)): 425 new_entries, total_duration, filtered_new_entries, filtered_duration = self._read_manifest( 426 manifest_paths[new_manifest_idx], config 427 ) 428 429 if len(filtered_new_entries) > 0: 430 print( 431 f"Filtered {len(filtered_new_entries)} files which amounts to {filtered_duration:0.2f}" 432 f" seconds of audio from manifest {manifest_paths[new_manifest_idx]}." 433 ) 434 print( 435 f"After filtering, manifest has {len(entries)} files which amounts to {total_duration} seconds of audio." 436 ) 437 438 entries.extend(new_entries) 439 440 if len(entries) == 0: 441 print("No tarred dataset was created as there were 0 valid samples after filtering!") 442 return 443 444 if config.shuffle: 445 random.seed(config.shuffle_seed) 446 print("Shuffling...") 447 random.shuffle(entries) 448 449 # Drop last section of samples that cannot be added onto a chunk 450 drop_count = len(entries) % num_samples_per_shard 451 total_new_entries = len(entries) 452 entries = entries[:-drop_count] 453 454 print( 455 f"Dropping {drop_count} samples from total new samples {total_new_entries} since they cannot " 456 f"be added into a uniformly sized chunk." 457 ) 458 459 # Create shards and updated manifest entries 460 num_added_shards = len(entries) // num_samples_per_shard 461 462 print(f"Number of samples in base dataset : {len(base_entries)}") 463 print(f"Number of samples in additional datasets : {len(entries)}") 464 print(f"Number of added shards : {num_added_shards}") 465 print(f"Remainder: {len(entries) % num_samples_per_shard}") 466 467 start_indices = [] 468 end_indices = [] 469 shard_indices = [] 470 for i in range(num_added_shards): 471 start_idx = (len(entries) // num_added_shards) * i 472 end_idx = start_idx + (len(entries) // num_added_shards) 473 shard_idx = i + config.num_shards 474 print(f"Shard {shard_idx} has entries {start_idx + len(base_entries)} ~ {end_idx + len(base_entries)}") 475 476 start_indices.append(start_idx) 477 end_indices.append(end_idx) 478 shard_indices.append(shard_idx) 479 480 manifest_folder, _ = os.path.split(base_manifest_path) 481 482 with Parallel(n_jobs=num_workers, verbose=num_added_shards) as parallel: 483 # Call parallel tarfile construction 484 new_entries_list = parallel( 485 delayed(self._create_shard)(entries[start_idx:end_idx], target_dir, shard_idx, manifest_folder) 486 for i, (start_idx, end_idx, shard_idx) in enumerate(zip(start_indices, end_indices, shard_indices)) 487 ) 488 489 if config.shard_manifests: 490 sharded_manifests_dir = target_dir + '/sharded_manifests' 491 if not os.path.exists(sharded_manifests_dir): 492 os.makedirs(sharded_manifests_dir) 493 494 for manifest in new_entries_list: 495 shard_id = manifest[0]['shard_id'] 496 new_manifest_shard_path = os.path.join(sharded_manifests_dir, f'manifest_{shard_id}.json') 497 with open(new_manifest_shard_path, 'w', encoding='utf-8') as m2: 498 for entry in manifest: 499 json.dump(entry, m2) 500 m2.write('\n') 501 502 # Flatten the list of list of entries to a list of entries 503 new_entries = [sample for manifest in new_entries_list for sample in manifest] 504 del new_entries_list 505 506 # Write manifest 507 if metadata is None: 508 new_version = 1 # start with `1`, where `0` indicates the base manifest + dataset 509 else: 510 new_version = metadata.version + 1 511 512 print("Total number of entries in manifest :", len(base_entries) + len(new_entries)) 513 514 new_manifest_path = os.path.join(target_dir, f'tarred_audio_manifest_version_{new_version}.json') 515 with open(new_manifest_path, 'w', encoding='utf-8') as m2: 516 # First write all the entries of base manifest 517 for entry in base_entries: 518 json.dump(entry, m2) 519 m2.write('\n') 520 521 # Finally write the new entries 522 for entry in new_entries: 523 json.dump(entry, m2) 524 m2.write('\n') 525 526 # Preserve historical metadata 527 base_metadata = metadata 528 529 # Write metadata (updated metadata for concatenated datasets) 530 new_metadata_path = os.path.join(target_dir, f'metadata_version_{new_version}.yaml') 531 metadata = ASRTarredDatasetMetadata() 532 533 # Update config 534 config.num_shards = config.num_shards + num_added_shards 535 536 # Update metadata 537 metadata.version = new_version 538 metadata.dataset_config = config 539 metadata.num_samples_per_shard = num_samples_per_shard 540 metadata.is_concatenated_manifest = True 541 metadata.created_datetime = metadata.get_current_datetime() 542 543 # Attach history 544 current_metadata = OmegaConf.structured(base_metadata.history) 545 metadata.history = current_metadata 546 547 # Write metadata 548 metadata_yaml = OmegaConf.structured(metadata) 549 OmegaConf.save(metadata_yaml, new_metadata_path, resolve=True) 550 551 def _read_manifest(self, manifest_path: str, config: ASRTarredDatasetConfig): 552 """Read and filters data from the manifest""" 553 # Read the existing manifest 554 entries = [] 555 total_duration = 0.0 556 filtered_entries = [] 557 filtered_duration = 0.0 558 with open(manifest_path, 'r', encoding='utf-8') as m: 559 for line in m: 560 entry = json.loads(line) 561 if (config.max_duration is None or entry['duration'] < config.max_duration) and ( 562 config.min_duration is None or entry['duration'] >= config.min_duration 563 ): 564 entries.append(entry) 565 total_duration += entry["duration"] 566 else: 567 filtered_entries.append(entry) 568 filtered_duration += entry['duration'] 569 570 return entries, total_duration, filtered_entries, filtered_duration 571 572 def _create_shard(self, entries, target_dir, shard_id, manifest_folder): 573 """Creates a tarball containing the audio files from `entries`. 574 """ 575 if self.config.sort_in_shards: 576 entries.sort(key=lambda x: x["duration"], reverse=False) 577 578 new_entries = [] 579 tar = tarfile.open(os.path.join(target_dir, f'audio_{shard_id}.tar'), mode='w', dereference=True) 580 581 count = dict() 582 for entry in entries: 583 # We squash the filename since we do not preserve directory structure of audio files in the tarball. 584 if os.path.exists(entry["audio_filepath"]): 585 audio_filepath = entry["audio_filepath"] 586 else: 587 audio_filepath = os.path.join(manifest_folder, entry["audio_filepath"]) 588 if not os.path.exists(audio_filepath): 589 raise FileNotFoundError(f"Could not find {entry['audio_filepath']}!") 590 591 base, ext = os.path.splitext(audio_filepath) 592 base = base.replace('/', '_') 593 # Need the following replacement as long as WebDataset splits on first period 594 base = base.replace('.', '_') 595 squashed_filename = f'{base}{ext}' 596 if squashed_filename not in count: 597 tar.add(audio_filepath, arcname=squashed_filename) 598 to_write = squashed_filename 599 count[squashed_filename] = 1 600 else: 601 to_write = base + "-sub" + str(count[squashed_filename]) + ext 602 count[squashed_filename] += 1 603 604 new_entry = { 605 'audio_filepath': to_write, 606 'duration': entry['duration'], 607 'shard_id': shard_id, # Keep shard ID for recordkeeping 608 } 609 610 if 'label' in entry: 611 new_entry['label'] = entry['label'] 612 613 if 'text' in entry: 614 new_entry['text'] = entry['text'] 615 616 if 'offset' in entry: 617 new_entry['offset'] = entry['offset'] 618 619 if 'lang' in entry: 620 new_entry['lang'] = entry['lang'] 621 622 new_entries.append(new_entry) 623 624 tar.close() 625 return new_entries 626 627 @classmethod 628 def setup_history(cls, base_metadata: ASRTarredDatasetMetadata, history: List[Any]): 629 if 'history' in base_metadata.keys(): 630 for history_val in base_metadata.history: 631 cls.setup_history(history_val, history) 632 633 if base_metadata is not None: 634 metadata_copy = copy.deepcopy(base_metadata) 635 with open_dict(metadata_copy): 636 metadata_copy.pop('history', None) 637 history.append(metadata_copy) 638 639 640 def main(): 641 if args.buckets_num > 1: 642 bucket_length = (args.max_duration - args.min_duration) / float(args.buckets_num) 643 for i in range(args.buckets_num): 644 min_duration = args.min_duration + i * bucket_length 645 max_duration = min_duration + bucket_length 646 if i == args.buckets_num - 1: 647 # add a small number to cover the samples with exactly duration of max_duration in the last bucket. 648 max_duration += 1e-5 649 target_dir = os.path.join(args.target_dir, f"bucket{i+1}") 650 print(f"Creating bucket {i+1} with min_duration={min_duration} and max_duration={max_duration} ...") 651 print(f"Results are being saved at: {target_dir}.") 652 create_tar_datasets(min_duration=min_duration, max_duration=max_duration, target_dir=target_dir) 653 print(f"Bucket {i+1} is created.") 654 else: 655 create_tar_datasets(min_duration=args.min_duration, max_duration=args.max_duration, target_dir=args.target_dir) 656 657 658 def create_tar_datasets(min_duration: float, max_duration: float, target_dir: str): 659 builder = ASRTarredDatasetBuilder() 660 661 shard_manifests = False if args.no_shard_manifests else True 662 663 if args.write_metadata: 664 metadata = ASRTarredDatasetMetadata() 665 dataset_cfg = ASRTarredDatasetConfig( 666 num_shards=args.num_shards, 667 shuffle=args.shuffle, 668 max_duration=max_duration, 669 min_duration=min_duration, 670 shuffle_seed=args.shuffle_seed, 671 sort_in_shards=args.sort_in_shards, 672 shard_manifests=shard_manifests, 673 keep_files_together=args.keep_files_together, 674 ) 675 metadata.dataset_config = dataset_cfg 676 677 output_path = os.path.join(target_dir, 'default_metadata.yaml') 678 OmegaConf.save(metadata, output_path, resolve=True) 679 print(f"Default metadata written to {output_path}") 680 exit(0) 681 682 if args.concat_manifest_paths is None or len(args.concat_manifest_paths) == 0: 683 print("Creating new tarred dataset ...") 684 685 # Create a tarred dataset from scratch 686 config = ASRTarredDatasetConfig( 687 num_shards=args.num_shards, 688 shuffle=args.shuffle, 689 max_duration=max_duration, 690 min_duration=min_duration, 691 shuffle_seed=args.shuffle_seed, 692 sort_in_shards=args.sort_in_shards, 693 shard_manifests=shard_manifests, 694 keep_files_together=args.keep_files_together, 695 ) 696 builder.configure(config) 697 builder.create_new_dataset(manifest_path=args.manifest_path, target_dir=target_dir, num_workers=args.workers) 698 699 else: 700 if args.buckets_num > 1: 701 raise ValueError("Concatenation feature does not support buckets_num > 1.") 702 print("Concatenating multiple tarred datasets ...") 703 704 # Implicitly update config from base details 705 if args.metadata_path is not None: 706 metadata = ASRTarredDatasetMetadata.from_file(args.metadata_path) 707 else: 708 raise ValueError("`metadata` yaml file path must be provided!") 709 710 # Preserve history 711 history = [] 712 builder.setup_history(OmegaConf.structured(metadata), history) 713 metadata.history = history 714 715 # Add command line overrides (everything other than num_shards) 716 metadata.dataset_config.max_duration = max_duration 717 metadata.dataset_config.min_duration = min_duration 718 metadata.dataset_config.shuffle = args.shuffle 719 metadata.dataset_config.shuffle_seed = args.shuffle_seed 720 metadata.dataset_config.sort_in_shards = args.sort_in_shards 721 metadata.dataset_config.shard_manifests = shard_manifests 722 723 builder.configure(metadata.dataset_config) 724 725 # Concatenate a tarred dataset onto a previous one 726 builder.create_concatenated_dataset( 727 base_manifest_path=args.manifest_path, 728 manifest_paths=args.concat_manifest_paths, 729 metadata=metadata, 730 target_dir=target_dir, 731 num_workers=args.workers, 732 ) 733 734 if DALI_INDEX_SCRIPT_AVAILABLE and dali_index.INDEX_CREATOR_AVAILABLE: 735 print("Constructing DALI Tarfile Index - ", target_dir) 736 index_config = dali_index.DALITarredIndexConfig(tar_dir=target_dir, workers=args.workers) 737 dali_index.main(index_config) 738 739 740 if __name__ == "__main__": 741 main() 742 [end of scripts/speech_recognition/convert_to_tarred_audio_dataset.py] [start of tools/nemo_forced_aligner/align.py] 1 # Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 15 import copy 16 import math 17 import os 18 from dataclasses import dataclass, field, is_dataclass 19 from pathlib import Path 20 from typing import List, Optional 21 22 import torch 23 from omegaconf import OmegaConf 24 from utils.data_prep import ( 25 add_t_start_end_to_utt_obj, 26 get_batch_starts_ends, 27 get_batch_variables, 28 get_manifest_lines_batch, 29 is_entry_in_all_lines, 30 is_entry_in_any_lines, 31 ) 32 from utils.make_ass_files import make_ass_files 33 from utils.make_ctm_files import make_ctm_files 34 from utils.make_output_manifest import write_manifest_out_line 35 from utils.viterbi_decoding import viterbi_decoding 36 37 from nemo.collections.asr.models.ctc_models import EncDecCTCModel 38 from nemo.collections.asr.models.hybrid_rnnt_ctc_models import EncDecHybridRNNTCTCModel 39 from nemo.collections.asr.parts.utils.streaming_utils import FrameBatchASR 40 from nemo.collections.asr.parts.utils.transcribe_utils import setup_model 41 from nemo.core.config import hydra_runner 42 from nemo.utils import logging 43 44 """ 45 Align the utterances in manifest_filepath. 46 Results are saved in ctm files in output_dir. 47 48 Arguments: 49 pretrained_name: string specifying the name of a CTC NeMo ASR model which will be automatically downloaded 50 from NGC and used for generating the log-probs which we will use to do alignment. 51 Note: NFA can only use CTC models (not Transducer models) at the moment. 52 model_path: string specifying the local filepath to a CTC NeMo ASR model which will be used to generate the 53 log-probs which we will use to do alignment. 54 Note: NFA can only use CTC models (not Transducer models) at the moment. 55 Note: if a model_path is provided, it will override the pretrained_name. 56 manifest_filepath: filepath to the manifest of the data you want to align, 57 containing 'audio_filepath' and 'text' fields. 58 output_dir: the folder where output CTM files and new JSON manifest will be saved. 59 align_using_pred_text: if True, will transcribe the audio using the specified model and then use that transcription 60 as the reference text for the forced alignment. 61 transcribe_device: None, or a string specifying the device that will be used for generating log-probs (i.e. "transcribing"). 62 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 63 (otherwise will set it to 'cpu'). 64 viterbi_device: None, or string specifying the device that will be used for doing Viterbi decoding. 65 The string needs to be in a format recognized by torch.device(). If None, NFA will set it to 'cuda' if it is available 66 (otherwise will set it to 'cpu'). 67 batch_size: int specifying batch size that will be used for generating log-probs and doing Viterbi decoding. 68 use_local_attention: boolean flag specifying whether to try to use local attention for the ASR Model (will only 69 work if the ASR Model is a Conformer model). If local attention is used, we will set the local attention context 70 size to [64,64]. 71 additional_segment_grouping_separator: an optional string used to separate the text into smaller segments. 72 If this is not specified, then the whole text will be treated as a single segment. 73 remove_blank_tokens_from_ctm: a boolean denoting whether to remove <blank> tokens from token-level output CTMs. 74 audio_filepath_parts_in_utt_id: int specifying how many of the 'parts' of the audio_filepath 75 we will use (starting from the final part of the audio_filepath) to determine the 76 utt_id that will be used in the CTM files. Note also that any spaces that are present in the audio_filepath 77 will be replaced with dashes, so as not to change the number of space-separated elements in the 78 CTM files. 79 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 1 => utt_id will be "e1" 80 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 2 => utt_id will be "d_e1" 81 e.g. if audio_filepath is "/a/b/c/d/e 1.wav" and audio_filepath_parts_in_utt_id is 3 => utt_id will be "c_d_e1" 82 use_buffered_infer: False, if set True, using streaming to do get the logits for alignment 83 This flag is useful when aligning large audio file. 84 However, currently the chunk streaming inference does not support batch inference, 85 which means even you set batch_size > 1, it will only infer one by one instead of doing 86 the whole batch inference together. 87 chunk_len_in_secs: float chunk length in seconds 88 total_buffer_in_secs: float Length of buffer (chunk + left and right padding) in seconds 89 chunk_batch_size: int batch size for buffered chunk inference, 90 which will cut one audio into segments and do inference on chunk_batch_size segments at a time 91 92 simulate_cache_aware_streaming: False, if set True, using cache aware streaming to do get the logits for alignment 93 94 save_output_file_formats: List of strings specifying what type of output files to save (default: ["ctm", "ass"]) 95 ctm_file_config: CTMFileConfig to specify the configuration of the output CTM files 96 ass_file_config: ASSFileConfig to specify the configuration of the output ASS files 97 """ 98 99 100 @dataclass 101 class CTMFileConfig: 102 remove_blank_tokens: bool = False 103 # minimum duration (in seconds) for timestamps in the CTM.If any line in the CTM has a 104 # duration lower than this, it will be enlarged from the middle outwards until it 105 # meets the minimum_timestamp_duration, or reaches the beginning or end of the audio file. 106 # Note that this may cause timestamps to overlap. 107 minimum_timestamp_duration: float = 0 108 109 110 @dataclass 111 class ASSFileConfig: 112 fontsize: int = 20 113 vertical_alignment: str = "center" 114 # if resegment_text_to_fill_space is True, the ASS files will use new segments 115 # such that each segment will not take up more than (approximately) max_lines_per_segment 116 # when the ASS file is applied to a video 117 resegment_text_to_fill_space: bool = False 118 max_lines_per_segment: int = 2 119 text_already_spoken_rgb: List[int] = field(default_factory=lambda: [49, 46, 61]) # dark gray 120 text_being_spoken_rgb: List[int] = field(default_factory=lambda: [57, 171, 9]) # dark green 121 text_not_yet_spoken_rgb: List[int] = field(default_factory=lambda: [194, 193, 199]) # light gray 122 123 124 @dataclass 125 class AlignmentConfig: 126 # Required configs 127 pretrained_name: Optional[str] = None 128 model_path: Optional[str] = None 129 manifest_filepath: Optional[str] = None 130 output_dir: Optional[str] = None 131 132 # General configs 133 align_using_pred_text: bool = False 134 transcribe_device: Optional[str] = None 135 viterbi_device: Optional[str] = None 136 batch_size: int = 1 137 use_local_attention: bool = True 138 additional_segment_grouping_separator: Optional[str] = None 139 audio_filepath_parts_in_utt_id: int = 1 140 141 # Buffered chunked streaming configs 142 use_buffered_chunked_streaming: bool = False 143 chunk_len_in_secs: float = 1.6 144 total_buffer_in_secs: float = 4.0 145 chunk_batch_size: int = 32 146 147 # Cache aware streaming configs 148 simulate_cache_aware_streaming: Optional[bool] = False 149 150 # Output file configs 151 save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) 152 ctm_file_config: CTMFileConfig = CTMFileConfig() 153 ass_file_config: ASSFileConfig = ASSFileConfig() 154 155 156 @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) 157 def main(cfg: AlignmentConfig): 158 159 logging.info(f'Hydra config: {OmegaConf.to_yaml(cfg)}') 160 161 if is_dataclass(cfg): 162 cfg = OmegaConf.structured(cfg) 163 164 # Validate config 165 if cfg.model_path is None and cfg.pretrained_name is None: 166 raise ValueError("Both cfg.model_path and cfg.pretrained_name cannot be None") 167 168 if cfg.model_path is not None and cfg.pretrained_name is not None: 169 raise ValueError("One of cfg.model_path and cfg.pretrained_name must be None") 170 171 if cfg.manifest_filepath is None: 172 raise ValueError("cfg.manifest_filepath must be specified") 173 174 if cfg.output_dir is None: 175 raise ValueError("cfg.output_dir must be specified") 176 177 if cfg.batch_size < 1: 178 raise ValueError("cfg.batch_size cannot be zero or a negative number") 179 180 if cfg.additional_segment_grouping_separator == "" or cfg.additional_segment_grouping_separator == " ": 181 raise ValueError("cfg.additional_grouping_separator cannot be empty string or space character") 182 183 if cfg.ctm_file_config.minimum_timestamp_duration < 0: 184 raise ValueError("cfg.minimum_timestamp_duration cannot be a negative number") 185 186 if cfg.ass_file_config.vertical_alignment not in ["top", "center", "bottom"]: 187 raise ValueError("cfg.ass_file_config.vertical_alignment must be one of 'top', 'center' or 'bottom'") 188 189 for rgb_list in [ 190 cfg.ass_file_config.text_already_spoken_rgb, 191 cfg.ass_file_config.text_already_spoken_rgb, 192 cfg.ass_file_config.text_already_spoken_rgb, 193 ]: 194 if len(rgb_list) != 3: 195 raise ValueError( 196 "cfg.ass_file_config.text_already_spoken_rgb," 197 " cfg.ass_file_config.text_being_spoken_rgb," 198 " and cfg.ass_file_config.text_already_spoken_rgb all need to contain" 199 " exactly 3 elements." 200 ) 201 202 # Validate manifest contents 203 if not is_entry_in_all_lines(cfg.manifest_filepath, "audio_filepath"): 204 raise RuntimeError( 205 "At least one line in cfg.manifest_filepath does not contain an 'audio_filepath' entry. " 206 "All lines must contain an 'audio_filepath' entry." 207 ) 208 209 if cfg.align_using_pred_text: 210 if is_entry_in_any_lines(cfg.manifest_filepath, "pred_text"): 211 raise RuntimeError( 212 "Cannot specify cfg.align_using_pred_text=True when the manifest at cfg.manifest_filepath " 213 "contains 'pred_text' entries. This is because the audio will be transcribed and may produce " 214 "a different 'pred_text'. This may cause confusion." 215 ) 216 else: 217 if not is_entry_in_all_lines(cfg.manifest_filepath, "text"): 218 raise RuntimeError( 219 "At least one line in cfg.manifest_filepath does not contain a 'text' entry. " 220 "NFA requires all lines to contain a 'text' entry when cfg.align_using_pred_text=False." 221 ) 222 223 # init devices 224 if cfg.transcribe_device is None: 225 transcribe_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 226 else: 227 transcribe_device = torch.device(cfg.transcribe_device) 228 logging.info(f"Device to be used for transcription step (`transcribe_device`) is {transcribe_device}") 229 230 if cfg.viterbi_device is None: 231 viterbi_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 232 else: 233 viterbi_device = torch.device(cfg.viterbi_device) 234 logging.info(f"Device to be used for viterbi step (`viterbi_device`) is {viterbi_device}") 235 236 if transcribe_device.type == 'cuda' or viterbi_device.type == 'cuda': 237 logging.warning( 238 'One or both of transcribe_device and viterbi_device are GPUs. If you run into OOM errors ' 239 'it may help to change both devices to be the CPU.' 240 ) 241 242 # load model 243 model, _ = setup_model(cfg, transcribe_device) 244 model.eval() 245 246 if isinstance(model, EncDecHybridRNNTCTCModel): 247 model.change_decoding_strategy(decoder_type="ctc") 248 249 if cfg.use_local_attention: 250 logging.info( 251 "Flag use_local_attention is set to True => will try to use local attention for model if it allows it" 252 ) 253 model.change_attention_model(self_attention_model="rel_pos_local_attn", att_context_size=[64, 64]) 254 255 if not (isinstance(model, EncDecCTCModel) or isinstance(model, EncDecHybridRNNTCTCModel)): 256 raise NotImplementedError( 257 f"Model is not an instance of NeMo EncDecCTCModel or ENCDecHybridRNNTCTCModel." 258 " Currently only instances of these models are supported" 259 ) 260 261 if cfg.ctm_file_config.minimum_timestamp_duration > 0: 262 logging.warning( 263 f"cfg.ctm_file_config.minimum_timestamp_duration has been set to {cfg.ctm_file_config.minimum_timestamp_duration} seconds. " 264 "This may cause the alignments for some tokens/words/additional segments to be overlapping." 265 ) 266 267 buffered_chunk_params = {} 268 if cfg.use_buffered_chunked_streaming: 269 model_cfg = copy.deepcopy(model._cfg) 270 271 OmegaConf.set_struct(model_cfg.preprocessor, False) 272 # some changes for streaming scenario 273 model_cfg.preprocessor.dither = 0.0 274 model_cfg.preprocessor.pad_to = 0 275 276 if model_cfg.preprocessor.normalize != "per_feature": 277 logging.error( 278 "Only EncDecCTCModelBPE models trained with per_feature normalization are supported currently" 279 ) 280 # Disable config overwriting 281 OmegaConf.set_struct(model_cfg.preprocessor, True) 282 283 feature_stride = model_cfg.preprocessor['window_stride'] 284 model_stride_in_secs = feature_stride * cfg.model_downsample_factor 285 total_buffer = cfg.total_buffer_in_secs 286 chunk_len = float(cfg.chunk_len_in_secs) 287 tokens_per_chunk = math.ceil(chunk_len / model_stride_in_secs) 288 mid_delay = math.ceil((chunk_len + (total_buffer - chunk_len) / 2) / model_stride_in_secs) 289 logging.info(f"tokens_per_chunk is {tokens_per_chunk}, mid_delay is {mid_delay}") 290 291 model = FrameBatchASR( 292 asr_model=model, 293 frame_len=chunk_len, 294 total_buffer=cfg.total_buffer_in_secs, 295 batch_size=cfg.chunk_batch_size, 296 ) 297 buffered_chunk_params = { 298 "delay": mid_delay, 299 "model_stride_in_secs": model_stride_in_secs, 300 "tokens_per_chunk": tokens_per_chunk, 301 } 302 # get start and end line IDs of batches 303 starts, ends = get_batch_starts_ends(cfg.manifest_filepath, cfg.batch_size) 304 305 # init output_timestep_duration = None and we will calculate and update it during the first batch 306 output_timestep_duration = None 307 308 # init f_manifest_out 309 os.makedirs(cfg.output_dir, exist_ok=True) 310 tgt_manifest_name = str(Path(cfg.manifest_filepath).stem) + "_with_output_file_paths.json" 311 tgt_manifest_filepath = str(Path(cfg.output_dir) / tgt_manifest_name) 312 f_manifest_out = open(tgt_manifest_filepath, 'w') 313 314 # get alignment and save in CTM batch-by-batch 315 for start, end in zip(starts, ends): 316 manifest_lines_batch = get_manifest_lines_batch(cfg.manifest_filepath, start, end) 317 318 (log_probs_batch, y_batch, T_batch, U_batch, utt_obj_batch, output_timestep_duration,) = get_batch_variables( 319 manifest_lines_batch, 320 model, 321 cfg.additional_segment_grouping_separator, 322 cfg.align_using_pred_text, 323 cfg.audio_filepath_parts_in_utt_id, 324 output_timestep_duration, 325 cfg.simulate_cache_aware_streaming, 326 cfg.use_buffered_chunked_streaming, 327 buffered_chunk_params, 328 ) 329 330 alignments_batch = viterbi_decoding(log_probs_batch, y_batch, T_batch, U_batch, viterbi_device) 331 332 for utt_obj, alignment_utt in zip(utt_obj_batch, alignments_batch): 333 334 utt_obj = add_t_start_end_to_utt_obj(utt_obj, alignment_utt, output_timestep_duration) 335 336 if "ctm" in cfg.save_output_file_formats: 337 utt_obj = make_ctm_files(utt_obj, cfg.output_dir, cfg.ctm_file_config,) 338 339 if "ass" in cfg.save_output_file_formats: 340 utt_obj = make_ass_files(utt_obj, cfg.output_dir, cfg.ass_file_config) 341 342 write_manifest_out_line( 343 f_manifest_out, utt_obj, 344 ) 345 346 f_manifest_out.close() 347 348 return None 349 350 351 if __name__ == "__main__": 352 main() 353 [end of tools/nemo_forced_aligner/align.py] </code> Here is an example of a patch file. It consists of changes to the code base. It specifies the file names, the line numbers of each change, and the removed and added lines. A single patch file can contain changes to multiple files. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch> I need you to solve the provided issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the format shown above. Respond below:
NVIDIA/NeMo
15db83ec4a65e649d83b61d7a4a58d911586e853
Ubuntu 22.04 Python 3.11 [asr]: multiple errors `dataclasses ValueError: mutable default * for field * is not allowed: use default_factory` **Describe the bug** After installing latest stable `1.19.1` from pipI, or the latest current commit with `[asr]` extras, I'm getting this error from `hydra-core==1.2.0` when trying to import `nemo.collections.asr`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 15, in <module> from nemo.collections.asr.losses.angularloss import AngularSoftmaxLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/angularloss.py", line 18, in <module> from nemo.core.classes import Loss, Typing, typecheck File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/__init__.py", line 16, in <module> from nemo.core.classes import * File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/core/classes/__init__.py", line 16, in <module> import hydra File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/__init__.py", line 5, in <module> from hydra import utils File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/utils.py", line 8, in <module> import hydra._internal.instantiate._instantiate2 File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 12, in <module> from hydra._internal.utils import _locate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/_internal/utils.py", line 18, in <module> from hydra.core.utils import get_valid_filename, validate_config_path File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/utils.py", line 20, in <module> from hydra.core.hydra_config import HydraConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/core/hydra_config.py", line 6, in <module> from hydra.conf import HydraConf File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 46, in <module> class JobConf: File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/hydra/conf/__init__.py", line 75, in JobConf @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'hydra.conf.JobConf.JobConfig.OverrideDirname'> for field override_dirname is not allowed: use default_factory ``` If I then manually upgrade `hydra-core` to the current latest (`1.3.2`), I get similar errors from `nemo-toolkit`: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` It's easy to fix with a patch like this: ``` --- nemo/collections/common/parts/adapter_modules.py.orig 2023-08-04 13:55:53.464534800 +0200 +++ nemo/collections/common/parts/adapter_modules.py 2023-08-04 14:05:45.579537700 +0200 @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, is_dataclass, field from typing import Any, Optional from hydra.utils import instantiate @@ -151,5 +151,5 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=adapter_mixin_strategies.ResidualAddAdapterStrategyConfig) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) ``` However then another error of the kind comes up: ``` Traceback (most recent call last): File "/home/alex/T7/src/speaker_verification/nemo_speaker/test.py", line 5, in <module> import nemo.collections.asr as nemo_asr File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/__init__.py", line 18, in <module> from nemo.collections.asr.models.clustering_diarizer import ClusteringDiarizer File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/models/clustering_diarizer.py", line 29, in <module> from nemo.collections.asr.metrics.der import score_labels File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/der.py", line 24, in <module> from nemo.collections.asr.metrics.wer import word_error_rate File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/metrics/wer.py", line 27, in <module> from nemo.collections.asr.parts.submodules import ctc_beam_decoding, ctc_greedy_decoding File "/home/alex/T7/src/speaker_verification/nemo_speaker/venv11/lib/python3.11/site-packages/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py", line 593, in <module> @dataclass ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.collections.asr.parts.submodules.ctc_beam_decoding.FlashlightConfig'> for field flashlight_cfg is not allowed: use default_factory ``` This can also be fixed accordingly, but all in all, it appears these issues are pretty common in the code base. Looks like NeMo isn't Python 3.11 ready, at least the `asr` collection. **Steps/Code to reproduce bug** With Python 3.11 1. Install `nemo-toolkit[asr]` either 1.19.1, or current HEAD from git (b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50) 2. Run `import nemo.collections.asr` **Expected behavior** Import `nemo.collections.asr` without errors **Environment overview (please complete the following information)** - Environment location: Bare-metal - Method of NeMo install: `pip install nemo-toolkit==1.19.1` or from source b498d438fc4c35ebf364a9a1c5cd3e29a2c0fe50 **Environment details** - Ubuntu 22.04 - PyTorch version 2.0.1 - Python version 3.11.4
Seems to be a similar to #7002 Interesting. The fix is easy but needs to be applied to basically every single place that has this constructor for our adapter configs. Let me see if I can update it. But no guarantees on how soon fixes will come in main. Looking forward to it @titu1994 ! Thanks ๐Ÿ˜ƒ @titu1994 I was looking to use NeMo speaker diarization with Python 3.11 and hit this dataclass issue. I patched everything involved in the specific code paths I needed: https://github.com/lmnt-com/NeMo/commit/d89acf9f0152e43dee29d7d1c4667ee34c26ffd7 I was using the neural diarizer as described in https://github.com/NVIDIA/NeMo/tree/main/examples/speaker_tasks/diarization I'd be happy to upstream this if it's helpful. I haven't checked whether this is backwards compatible for earlier python/dataclass versions, do you know? For reference, what led me to this issue, though it's duplicative to the above discussion: [Similar error](https://github.com/huggingface/datasets/issues/5230) [StackOverflow solution](https://stackoverflow.com/questions/53632152/why-cant-dataclasses-have-mutable-defaults-in-their-class-attributes-declaratio) @shaper Thanks for sharing. For brevity, you don't really need a `lambda` when you don't pass any init parameters, like this: ``` field(default_factory=lambda: ConfidenceConfig()) ``` You can just do ``` field(default_factory=ConfidenceConfig) ``` It's only needed when you do pass parameter(s), like ``` field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) ``` This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days. I have the same issue. @tango4j suggested using one of the models from https://huggingface.co/spaces/hf-audio/open_asr_leaderboard, but I cannot import nemo.collections.asr: ``` Traceback (most recent call last): File "/opt/pycharm-2022.3.3/plugins/python/helpers/pycharm/docrunner.py", line 138, in __run exec(compile(example.source, filename, "single", File "<doctest NeMoASR[2]>", line 1, in <module> NeMoASR().apply_asr(file) ^^^^^^^^^ File "/home/cbj/python/cbj/cbj/transcribe/pretrained.py", line 504, in __init__ import nemo.collections.asr as nemo_asr File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` For documentation (I had to search in the provided links): Mutable defaults were never allowed in dataclasses (by convention), but in python 3.11 they improved the check: Instead of checking some types (dict, list, set) they now use hashable as indicator for mutable. An alternative to default_factory would be to use frozen dataclasses, but I don't know whether in this code base the configs are used as mutable objects or not. You need to update to NeMo 1.20, omegaconf did a fix that should resolve this I have NeMo 1.20.0. With `pip install nemo_toolkit` and `pip install pytorch_lightning` I installed yesterday nemo. So it should be the newest PyPI version. ``` $ pip show nemo_toolkit Name: nemo-toolkit Version: 1.20.0 Summary: NeMo - a toolkit for Conversational AI Home-page: https://github.com/nvidia/nemo Author: NVIDIA Author-email: nemo-toolkit@nvidia.com License: Apache2 Location: /opt/py/2023/lib/python3.11/site-packages Requires: huggingface-hub, numba, numpy, onnx, python-dateutil, ruamel.yaml, scikit-learn, setuptools, tensorboard, text-unidecode, torch, tqdm, wget, wrapt Required-by: $ pip show omegaconf Name: omegaconf Version: 2.3.0 Summary: A flexible configuration library Home-page: https://github.com/omry/omegaconf Author: Omry Yadan Author-email: omry@yadan.net License: Location: /home/cbj/.local/lib/python3.11/site-packages Requires: antlr4-python3-runtime, PyYAML Required-by: hydra-core $ python -c "import nemo.collections.asr as nemo_asr" Traceback (most recent call last): File "<string>", line 1, in <module> File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/__init__.py", line 15, in <module> from nemo.collections.asr import data, losses, models, modules File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/__init__.py", line 16, in <module> from nemo.collections.asr.losses.audio_losses import SDRLoss File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/losses/audio_losses.py", line 21, in <module> from nemo.collections.asr.parts.preprocessing.features import make_seq_mask_like File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/__init__.py", line 16, in <module> from nemo.collections.asr.parts.preprocessing.features import FeaturizerFactory, FilterbankFeatures, WaveformFeaturizer File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 44, in <module> from nemo.collections.asr.parts.preprocessing.perturb import AudioAugmentor File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/asr/parts/preprocessing/perturb.py", line 50, in <module> from nemo.collections.common.parts.preprocessing import collections, parsers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/__init__.py", line 16, in <module> from nemo.collections.common import data, losses, parts, tokenizers File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/__init__.py", line 15, in <module> from nemo.collections.common.parts.adapter_modules import LinearAdapter, LinearAdapterConfig File "/opt/py/2023/lib/python3.11/site-packages/nemo/collections/common/parts/adapter_modules.py", line 147, in <module> @dataclass ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1230, in dataclass return wrap(cls) ^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 1220, in wrap return _process_class(cls, init, repr, eq, order, unsafe_hash, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 958, in _process_class cls_fields.append(_get_field(cls, name, type, kw_only)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/py/2023/lib/python3.11/dataclasses.py", line 815, in _get_field raise ValueError(f'mutable default {type(f.default)} for field ' ValueError: mutable default <class 'nemo.core.classes.mixins.adapter_mixin_strategies.ResidualAddAdapterStrategyConfig'> for field adapter_strategy is not allowed: use default_factory ``` Hmm ok I'll take a look
2023-10-03T19:14:38Z
<patch> <patch> diff --git a/examples/asr/experimental/k2/align_speech_parallel.py b/examples/asr/experimental/k2/align_speech_parallel.py --- a/examples/asr/experimental/k2/align_speech_parallel.py +++ b/examples/asr/experimental/k2/align_speech_parallel.py @@ -74,7 +74,7 @@ import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Optional import pytorch_lightning as ptl @@ -94,12 +94,14 @@ @dataclass class ParallelAlignmentConfig: model: Optional[str] = None # name - predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4) - aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig() + predict_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(return_sample_id=True, num_workers=4) + ) + aligner_args: K2AlignerWrapperModelConfig = field(default_factory=lambda: K2AlignerWrapperModelConfig()) output_path: str = MISSING model_stride: int = 8 - trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp") + trainer: TrainerConfig = field(default_factory=lambda: TrainerConfig(gpus=-1, accelerator="ddp")) # there arguments will be ignored return_predictions: bool = False diff --git a/nemo/collections/asr/metrics/rnnt_wer.py b/nemo/collections/asr/metrics/rnnt_wer.py --- a/nemo/collections/asr/metrics/rnnt_wer.py +++ b/nemo/collections/asr/metrics/rnnt_wer.py @@ -15,7 +15,7 @@ import copy import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1299,7 +1299,7 @@ class RNNTDecodingConfig: preserve_alignments: Optional[bool] = None # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # RNNT Joint fused batch size fused_batch_size: Optional[int] = None @@ -1317,10 +1317,10 @@ class RNNTDecodingConfig: rnnt_timestamp_type: str = "all" # can be char, word or all for both # greedy decoding config - greedy: greedy_decode.GreedyRNNTInferConfig = greedy_decode.GreedyRNNTInferConfig() + greedy: greedy_decode.GreedyRNNTInferConfig = field(default_factory=lambda: greedy_decode.GreedyRNNTInferConfig()) # beam decoding config - beam: beam_decode.BeamRNNTInferConfig = beam_decode.BeamRNNTInferConfig(beam_size=4) + beam: beam_decode.BeamRNNTInferConfig = field(default_factory=lambda: beam_decode.BeamRNNTInferConfig(beam_size=4)) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/metrics/wer.py b/nemo/collections/asr/metrics/wer.py --- a/nemo/collections/asr/metrics/wer.py +++ b/nemo/collections/asr/metrics/wer.py @@ -14,7 +14,7 @@ import re from abc import abstractmethod -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Callable, Dict, List, Optional, Tuple, Union import editdistance @@ -1297,13 +1297,17 @@ class CTCDecodingConfig: batch_dim_index: int = 0 # greedy decoding config - greedy: ctc_greedy_decoding.GreedyCTCInferConfig = ctc_greedy_decoding.GreedyCTCInferConfig() + greedy: ctc_greedy_decoding.GreedyCTCInferConfig = field( + default_factory=lambda: ctc_greedy_decoding.GreedyCTCInferConfig() + ) # beam decoding config - beam: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + beam: ctc_beam_decoding.BeamCTCInferConfig = field( + default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=4) + ) # confidence config - confidence_cfg: ConfidenceConfig = ConfidenceConfig() + confidence_cfg: ConfidenceConfig = field(default_factory=lambda: ConfidenceConfig()) # can be used to change temperature for decoding temperature: float = 1.0 diff --git a/nemo/collections/asr/models/configs/aligner_config.py b/nemo/collections/asr/models/configs/aligner_config.py --- a/nemo/collections/asr/models/configs/aligner_config.py +++ b/nemo/collections/asr/models/configs/aligner_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.parts.k2.classes import GraphModuleConfig @@ -35,10 +35,10 @@ class AlignerWrapperModelConfig: word_output: bool = True cpu_decoding: bool = False decode_batch_size: int = 0 - ctc_cfg: AlignerCTCConfig = AlignerCTCConfig() - rnnt_cfg: AlignerRNNTConfig = AlignerRNNTConfig() + ctc_cfg: AlignerCTCConfig = field(default_factory=lambda: AlignerCTCConfig()) + rnnt_cfg: AlignerRNNTConfig = field(default_factory=lambda: AlignerRNNTConfig()) @dataclass class K2AlignerWrapperModelConfig(AlignerWrapperModelConfig): - decoder_module_cfg: GraphModuleConfig = GraphModuleConfig() + decoder_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) diff --git a/nemo/collections/asr/models/configs/asr_models_config.py b/nemo/collections/asr/models/configs/asr_models_config.py --- a/nemo/collections/asr/models/configs/asr_models_config.py +++ b/nemo/collections/asr/models/configs/asr_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -74,24 +74,32 @@ class EncDecCTCConfig(model_cfg.ModelConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=True) - validation_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ASRDatasetConfig = ASRDatasetConfig(manifest_filepath=None, shuffle=False) + train_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=True)) + validation_ds: ASRDatasetConfig = field( + default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ASRDatasetConfig = field(default_factory=lambda: ASRDatasetConfig(manifest_filepath=None, shuffle=False)) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() - decoding: CTCDecodingConfig = CTCDecodingConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) + decoding: CTCDecodingConfig = field(default_factory=lambda: CTCDecodingConfig()) @dataclass class EncDecCTCModelConfig(model_cfg.NemoConfig): - model: EncDecCTCConfig = EncDecCTCConfig() + model: EncDecCTCConfig = field(default_factory=lambda: EncDecCTCConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/classification_models_config.py b/nemo/collections/asr/models/configs/classification_models_config.py --- a/nemo/collections/asr/models/configs/classification_models_config.py +++ b/nemo/collections/asr/models/configs/classification_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional from omegaconf import MISSING @@ -72,30 +72,40 @@ class EncDecClassificationConfig(model_cfg.ModelConfig): timesteps: int = MISSING # Dataset configs - train_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: EncDecClassificationDatasetConfig = EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: EncDecClassificationDatasetConfig = field( + default_factory=lambda: EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=timesteps + preprocessor: AudioToMFCCPreprocessorConfig = field(default_factory=lambda: AudioToMFCCPreprocessorConfig()) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=-1) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig() - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig()) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) + + def __post_init__(self): + if self.crop_or_pad_augment is not None: + self.crop_or_pad_augment.audio_length = self.timesteps @dataclass class EncDecClassificationModelConfig(model_cfg.NemoConfig): - model: EncDecClassificationConfig = EncDecClassificationConfig() + model: EncDecClassificationConfig = field(default_factory=lambda: EncDecClassificationConfig()) diff --git a/nemo/collections/asr/models/configs/diarizer_config.py b/nemo/collections/asr/models/configs/diarizer_config.py --- a/nemo/collections/asr/models/configs/diarizer_config.py +++ b/nemo/collections/asr/models/configs/diarizer_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from typing import Any, Dict, Optional, Tuple, Union @@ -78,9 +78,9 @@ class ASRDiarizerParams(DiarizerComponentConfig): @dataclass class ASRDiarizerConfig(DiarizerComponentConfig): model_path: Optional[str] = "stt_en_conformer_ctc_large" - parameters: ASRDiarizerParams = ASRDiarizerParams() - ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = ASRDiarizerCTCDecoderParams() - realigning_lm_parameters: ASRRealigningLMParams = ASRRealigningLMParams() + parameters: ASRDiarizerParams = field(default_factory=lambda: ASRDiarizerParams()) + ctc_decoder_parameters: ASRDiarizerCTCDecoderParams = field(default_factory=lambda: ASRDiarizerCTCDecoderParams()) + realigning_lm_parameters: ASRRealigningLMParams = field(default_factory=lambda: ASRRealigningLMParams()) @dataclass @@ -102,7 +102,7 @@ class VADParams(DiarizerComponentConfig): class VADConfig(DiarizerComponentConfig): model_path: str = "vad_multilingual_marblenet" # .nemo local model path or pretrained VAD model name external_vad_manifest: Optional[str] = None - parameters: VADParams = VADParams() + parameters: VADParams = field(default_factory=lambda: VADParams()) @dataclass @@ -121,7 +121,7 @@ class SpeakerEmbeddingsParams(DiarizerComponentConfig): class SpeakerEmbeddingsConfig(DiarizerComponentConfig): # .nemo local model path or pretrained model name (titanet_large, ecapa_tdnn or speakerverification_speakernet) model_path: Optional[str] = None - parameters: SpeakerEmbeddingsParams = SpeakerEmbeddingsParams() + parameters: SpeakerEmbeddingsParams = field(default_factory=lambda: SpeakerEmbeddingsParams()) @dataclass @@ -142,7 +142,7 @@ class ClusteringParams(DiarizerComponentConfig): @dataclass class ClusteringConfig(DiarizerComponentConfig): - parameters: ClusteringParams = ClusteringParams() + parameters: ClusteringParams = field(default_factory=lambda: ClusteringParams()) @dataclass @@ -166,7 +166,7 @@ class MSDDParams(DiarizerComponentConfig): @dataclass class MSDDConfig(DiarizerComponentConfig): model_path: Optional[str] = "diar_msdd_telephonic" - parameters: MSDDParams = MSDDParams() + parameters: MSDDParams = field(default_factory=lambda: MSDDParams()) @dataclass @@ -176,16 +176,16 @@ class DiarizerConfig(DiarizerComponentConfig): oracle_vad: bool = False # If True, uses RTTM files provided in the manifest file to get VAD timestamps collar: float = 0.25 # Collar value for scoring ignore_overlap: bool = True # Consider or ignore overlap segments while scoring - vad: VADConfig = VADConfig() - speaker_embeddings: SpeakerEmbeddingsConfig = SpeakerEmbeddingsConfig() - clustering: ClusteringConfig = ClusteringConfig() - msdd_model: MSDDConfig = MSDDConfig() - asr: ASRDiarizerConfig = ASRDiarizerConfig() + vad: VADConfig = field(default_factory=lambda: VADConfig()) + speaker_embeddings: SpeakerEmbeddingsConfig = field(default_factory=lambda: SpeakerEmbeddingsConfig()) + clustering: ClusteringConfig = field(default_factory=lambda: ClusteringConfig()) + msdd_model: MSDDConfig = field(default_factory=lambda: MSDDConfig()) + asr: ASRDiarizerConfig = field(default_factory=lambda: ASRDiarizerConfig()) @dataclass class NeuralDiarizerInferenceConfig(DiarizerComponentConfig): - diarizer: DiarizerConfig = DiarizerConfig() + diarizer: DiarizerConfig = field(default_factory=lambda: DiarizerConfig()) device: str = "cpu" verbose: bool = False batch_size: int = 64 diff --git a/nemo/collections/asr/models/configs/k2_sequence_models_config.py b/nemo/collections/asr/models/configs/k2_sequence_models_config.py --- a/nemo/collections/asr/models/configs/k2_sequence_models_config.py +++ b/nemo/collections/asr/models/configs/k2_sequence_models_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from nemo.collections.asr.models.configs.asr_models_config import EncDecCTCConfig from nemo.collections.asr.parts.k2.classes import GraphModuleConfig as BackendConfig @@ -26,14 +26,14 @@ class GraphModuleConfig: split_batch_size: int = 0 dec_type: str = "topo" transcribe_training: bool = True - backend_cfg: BackendConfig = BackendConfig() + backend_cfg: BackendConfig = field(default_factory=lambda: BackendConfig()) @dataclass class EncDecK2SeqConfig(EncDecCTCConfig): - graph_module_cfg: GraphModuleConfig = GraphModuleConfig() + graph_module_cfg: GraphModuleConfig = field(default_factory=lambda: GraphModuleConfig()) @dataclass class EncDecK2SeqModelConfig(NemoConfig): - model: EncDecK2SeqConfig = EncDecK2SeqConfig() + model: EncDecK2SeqConfig = field(default_factory=lambda: EncDecK2SeqConfig()) diff --git a/nemo/collections/asr/models/configs/matchboxnet_config.py b/nemo/collections/asr/models/configs/matchboxnet_config.py --- a/nemo/collections/asr/models/configs/matchboxnet_config.py +++ b/nemo/collections/asr/models/configs/matchboxnet_config.py @@ -107,30 +107,38 @@ class MatchboxNetModelConfig(clf_cfg.EncDecClassificationConfig): labels: List[str] = MISSING # Dataset configs - train_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=False + train_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig( + manifest_filepath=None, shuffle=True, trim_silence=False + ) ) - validation_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + validation_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) - test_ds: clf_cfg.EncDecClassificationDatasetConfig = clf_cfg.EncDecClassificationDatasetConfig( - manifest_filepath=None, shuffle=False + test_ds: clf_cfg.EncDecClassificationDatasetConfig = field( + default_factory=lambda: clf_cfg.EncDecClassificationDatasetConfig(manifest_filepath=None, shuffle=False) ) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMFCCPreprocessorConfig = AudioToMFCCPreprocessorConfig(window_size=0.025) - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig( - freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + preprocessor: AudioToMFCCPreprocessorConfig = field( + default_factory=lambda: AudioToMFCCPreprocessorConfig(window_size=0.025) + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig( + freq_masks=2, time_masks=2, freq_width=15, time_width=25, rect_masks=5, rect_time=25, rect_freq=15 + ) ) - crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = CropOrPadSpectrogramAugmentationConfig( - audio_length=128 + crop_or_pad_augment: Optional[CropOrPadSpectrogramAugmentationConfig] = field( + default_factory=lambda: CropOrPadSpectrogramAugmentationConfig(audio_length=128) ) - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderClassificationConfig = ConvASRDecoderClassificationConfig() + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderClassificationConfig = field(default_factory=lambda: ConvASRDecoderClassificationConfig()) @dataclass diff --git a/nemo/collections/asr/models/configs/quartznet_config.py b/nemo/collections/asr/models/configs/quartznet_config.py --- a/nemo/collections/asr/models/configs/quartznet_config.py +++ b/nemo/collections/asr/models/configs/quartznet_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Callable, List, Optional from omegaconf import MISSING @@ -174,20 +174,30 @@ class JasperModelConfig(ctc_cfg.EncDecCTCConfig): labels: List[str] = MISSING # Dataset configs - train_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig( - manifest_filepath=None, shuffle=True, trim_silence=True + train_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=True, trim_silence=True) + ) + validation_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) + ) + test_ds: ctc_cfg.ASRDatasetConfig = field( + default_factory=lambda: ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) ) - validation_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) - test_ds: ctc_cfg.ASRDatasetConfig = ctc_cfg.ASRDatasetConfig(manifest_filepath=None, shuffle=False) # Optimizer / Scheduler config - optim: Optional[model_cfg.OptimConfig] = model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + optim: Optional[model_cfg.OptimConfig] = field( + default_factory=lambda: model_cfg.OptimConfig(sched=model_cfg.SchedConfig()) + ) # Model general component configs - preprocessor: AudioToMelSpectrogramPreprocessorConfig = AudioToMelSpectrogramPreprocessorConfig() - spec_augment: Optional[SpectrogramAugmentationConfig] = SpectrogramAugmentationConfig() - encoder: ConvASREncoderConfig = ConvASREncoderConfig(activation="relu") - decoder: ConvASRDecoderConfig = ConvASRDecoderConfig() + preprocessor: AudioToMelSpectrogramPreprocessorConfig = field( + default_factory=lambda: AudioToMelSpectrogramPreprocessorConfig() + ) + spec_augment: Optional[SpectrogramAugmentationConfig] = field( + default_factory=lambda: SpectrogramAugmentationConfig() + ) + encoder: ConvASREncoderConfig = field(default_factory=lambda: ConvASREncoderConfig(activation="relu")) + decoder: ConvASRDecoderConfig = field(default_factory=lambda: ConvASRDecoderConfig()) @dataclass diff --git a/nemo/collections/asr/modules/audio_preprocessing.py b/nemo/collections/asr/modules/audio_preprocessing.py --- a/nemo/collections/asr/modules/audio_preprocessing.py +++ b/nemo/collections/asr/modules/audio_preprocessing.py @@ -634,6 +634,12 @@ def __init__(self, audio_length): super(CropOrPadSpectrogramAugmentation, self).__init__() self.audio_length = audio_length + if self.audio_length < 0: + raise ValueError( + 'audio_length must be non-negative. If using a dataclass with OmegaConf, ' + 'please call OmegaConf.to_object(cfg) to call appropriate __post_init__ methods.' + ) + @typecheck() @torch.no_grad() def forward(self, input_signal, length): diff --git a/nemo/collections/asr/parts/k2/classes.py b/nemo/collections/asr/parts/k2/classes.py --- a/nemo/collections/asr/parts/k2/classes.py +++ b/nemo/collections/asr/parts/k2/classes.py @@ -13,7 +13,7 @@ # limitations under the License. from abc import ABC -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple import torch @@ -43,7 +43,7 @@ class GraphModuleConfig: topo_with_self_loops: bool = True token_lm: Optional[Any] = None intersect_pruned: bool = False - intersect_conf: GraphIntersectDenseConfig = GraphIntersectDenseConfig() + intersect_conf: GraphIntersectDenseConfig = field(default_factory=lambda: GraphIntersectDenseConfig()) boost_coeff: float = 0.0 predictor_window_size: int = 0 predictor_step_size: int = 1 diff --git a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py --- a/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py +++ b/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py @@ -13,7 +13,7 @@ # limitations under the License. import math -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional import torch @@ -183,7 +183,7 @@ class MultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format(MultiHeadAttentionAdapter.__module__, MultiHeadAttentionAdapter.__name__) @@ -287,7 +287,7 @@ class RelPositionMultiHeadAttentionAdapterConfig: n_feat: int dropout_rate: float = 0.0 proj_dim: Optional[int] = None - adapter_strategy: Optional[Any] = MHAResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field(default_factory=lambda: MHAResidualAddAdapterStrategyConfig()) _target_: str = "{0}.{1}".format( RelPositionMultiHeadAttentionAdapter.__module__, RelPositionMultiHeadAttentionAdapter.__name__ ) @@ -336,7 +336,9 @@ class PositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(PositionalEncodingAdapter.__module__, PositionalEncodingAdapter.__name__) @@ -378,5 +380,7 @@ class RelPositionalEncodingAdapterConfig: d_model: int max_len: int = 5000 xscale: float = 1.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(RelPositionalEncodingAdapter.__module__, RelPositionalEncodingAdapter.__name__) diff --git a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_beam_decoding.py @@ -14,7 +14,7 @@ import math import os -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import torch @@ -602,5 +602,5 @@ class BeamCTCInferConfig: beam_beta: float = 0.0 kenlm_path: Optional[str] = None - flashlight_cfg: Optional[FlashlightConfig] = FlashlightConfig() - pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = PyCTCDecodeConfig() + flashlight_cfg: Optional[FlashlightConfig] = field(default_factory=lambda: FlashlightConfig()) + pyctcdecode_cfg: Optional[PyCTCDecodeConfig] = field(default_factory=lambda: PyCTCDecodeConfig()) diff --git a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/ctc_greedy_decoding.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional import torch @@ -253,7 +253,7 @@ class GreedyCTCInferConfig: preserve_alignments: bool = False compute_timestamps: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py --- a/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py +++ b/nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py @@ -26,7 +26,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import List, Optional, Tuple, Union import numpy as np @@ -2185,7 +2185,7 @@ class GreedyRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed @@ -2201,7 +2201,7 @@ class GreedyBatchedRNNTInferConfig: max_symbols_per_step: Optional[int] = 10 preserve_alignments: bool = False preserve_frame_confidence: bool = False - confidence_method_cfg: Optional[ConfidenceMethodConfig] = ConfidenceMethodConfig() + confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/asr/parts/utils/asr_confidence_utils.py b/nemo/collections/asr/parts/utils/asr_confidence_utils.py --- a/nemo/collections/asr/parts/utils/asr_confidence_utils.py +++ b/nemo/collections/asr/parts/utils/asr_confidence_utils.py @@ -14,7 +14,7 @@ import math from abc import ABC, abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, field from functools import partial from typing import List, Optional @@ -175,7 +175,7 @@ class ConfidenceConfig: preserve_word_confidence: bool = False exclude_blank: bool = True aggregation: str = "min" - method_cfg: ConfidenceMethodConfig = ConfidenceMethodConfig() + method_cfg: ConfidenceMethodConfig = field(default_factory=lambda: ConfidenceMethodConfig()) def __post_init__(self): # OmegaConf.structured ensures that post_init check is always executed diff --git a/nemo/collections/common/parts/adapter_modules.py b/nemo/collections/common/parts/adapter_modules.py --- a/nemo/collections/common/parts/adapter_modules.py +++ b/nemo/collections/common/parts/adapter_modules.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from typing import Any, Optional from hydra.utils import instantiate @@ -160,5 +160,7 @@ class LinearAdapterConfig: activation: str = 'swish' norm_position: str = 'pre' dropout: float = 0.0 - adapter_strategy: Optional[Any] = adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + adapter_strategy: Optional[Any] = field( + default_factory=lambda: adapter_mixin_strategies.ResidualAddAdapterStrategyConfig() + ) _target_: str = "{0}.{1}".format(LinearAdapter.__module__, LinearAdapter.__name__) diff --git a/nemo/collections/common/tokenizers/en_ja_tokenizers.py b/nemo/collections/common/tokenizers/en_ja_tokenizers.py --- a/nemo/collections/common/tokenizers/en_ja_tokenizers.py +++ b/nemo/collections/common/tokenizers/en_ja_tokenizers.py @@ -14,11 +14,19 @@ import re from typing import List -import ipadic -import MeCab from pangu import spacing from sacremoses import MosesDetokenizer, MosesPunctNormalizer, MosesTokenizer +try: + import ipadic + import MeCab + + HAVE_MECAB = True + HAVE_IPADIC = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + HAVE_IPADIC = False + class EnJaProcessor: """ @@ -67,6 +75,9 @@ class JaMecabProcessor: """ def __init__(self): + if not HAVE_MECAB or not HAVE_IPADIC: + raise ImportError("Please ensure that you have installed `MeCab` and `ipadic` to use JaMecabProcessor") + self.mecab_tokenizer = MeCab.Tagger(ipadic.MECAB_ARGS + " -Owakati") def detokenize(self, text: List[str]) -> str: diff --git a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py --- a/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py +++ b/nemo/collections/nlp/models/machine_translation/mt_enc_dec_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Optional, Tuple from omegaconf.omegaconf import MISSING @@ -46,7 +46,7 @@ class MTOptimConfig(OptimConfig): lr: float = 1e-3 betas: Tuple[float, float] = (0.9, 0.98) weight_decay: float = 0.0 - sched: Optional[MTSchedConfig] = MTSchedConfig() + sched: Optional[MTSchedConfig] = field(default_factory=lambda: MTSchedConfig()) @dataclass @@ -74,70 +74,80 @@ class MTEncDecModelConfig(EncDecNLPModelConfig): decoder_tokenizer: Any = MISSING decoder: Any = MISSING - head: TokenClassifierConfig = TokenClassifierConfig(log_softmax=True) + head: TokenClassifierConfig = field(default_factory=lambda: TokenClassifierConfig(log_softmax=True)) # dataset configurations - train_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=True, - shuffle=True, - cache_ids=False, - use_cache=False, + train_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=True, + shuffle=True, + cache_ids=False, + use_cache=False, + ) ) - validation_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + validation_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - test_ds: Optional[TranslationDataConfig] = TranslationDataConfig( - src_file_name=MISSING, - tgt_file_name=MISSING, - tokens_in_batch=512, - clean=False, - shuffle=False, - cache_ids=False, - use_cache=False, + test_ds: Optional[TranslationDataConfig] = field( + default_factory=lambda: TranslationDataConfig( + src_file_name=MISSING, + tgt_file_name=MISSING, + tokens_in_batch=512, + clean=False, + shuffle=False, + cache_ids=False, + use_cache=False, + ) ) - optim: Optional[OptimConfig] = MTOptimConfig() + optim: Optional[OptimConfig] = field(default_factory=lambda: MTOptimConfig()) @dataclass class AAYNBaseConfig(MTEncDecModelConfig): # Attention is All You Need Base Configuration - encoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - decoder_tokenizer: TokenizerConfig = TokenizerConfig(library='yttm') - - encoder: NeMoTransformerEncoderConfig = NeMoTransformerEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + encoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + decoder_tokenizer: TokenizerConfig = field(default_factory=lambda: TokenizerConfig(library='yttm')) + + encoder: NeMoTransformerEncoderConfig = field( + default_factory=lambda: NeMoTransformerEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) - decoder: NeMoTransformerConfig = NeMoTransformerConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, + decoder: NeMoTransformerConfig = field( + default_factory=lambda: NeMoTransformerConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + ) ) @@ -150,32 +160,36 @@ class MTBottleneckModelConfig(AAYNBaseConfig): recon_per_token: bool = True log_timing: bool = True - encoder: NeMoTransformerBottleneckEncoderConfig = NeMoTransformerBottleneckEncoderConfig( - library='nemo', - model_name=None, - pretrained=False, - hidden_size=512, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', - hidden_steps=32, - hidden_blocks=1, - hidden_init_method='params', + encoder: NeMoTransformerBottleneckEncoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckEncoderConfig( + library='nemo', + model_name=None, + pretrained=False, + hidden_size=512, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + hidden_steps=32, + hidden_blocks=1, + hidden_init_method='params', + ) ) - decoder: NeMoTransformerBottleneckDecoderConfig = NeMoTransformerBottleneckDecoderConfig( - library='nemo', - model_name=None, - pretrained=False, - inner_size=2048, - num_layers=6, - num_attention_heads=8, - ffn_dropout=0.1, - attn_score_dropout=0.1, - attn_layer_dropout=0.1, - arch='seq2seq', + decoder: NeMoTransformerBottleneckDecoderConfig = field( + default_factory=lambda: NeMoTransformerBottleneckDecoderConfig( + library='nemo', + model_name=None, + pretrained=False, + inner_size=2048, + num_layers=6, + num_attention_heads=8, + ffn_dropout=0.1, + attn_score_dropout=0.1, + attn_layer_dropout=0.1, + arch='seq2seq', + ) ) diff --git a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py --- a/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py +++ b/nemo/collections/nlp/models/token_classification/punctuation_capitalization_config.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, Optional from omegaconf.omegaconf import MISSING, DictConfig, OmegaConf, open_dict @@ -215,13 +215,15 @@ class PunctuationCapitalizationModelConfig: This config is a part of :class:`~PunctuationCapitalizationConfig`. """ - class_labels: ClassLabelsConfig = ClassLabelsConfig() + class_labels: ClassLabelsConfig = field(default_factory=lambda: ClassLabelsConfig()) """A mandatory parameter containing a dictionary with names of label id files used in .nemo checkpoints. These file names can also be used for passing label vocabularies to the model. If you wish to use ``class_labels`` for passing vocabularies, please provide path to vocabulary files in ``model.common_dataset_parameters.label_vocab_dir`` parameter.""" - common_dataset_parameters: Optional[CommonDatasetParametersConfig] = CommonDatasetParametersConfig() + common_dataset_parameters: Optional[CommonDatasetParametersConfig] = field( + default_factory=lambda: CommonDatasetParametersConfig() + ) """Label ids and loss mask information information.""" train_ds: Optional[PunctuationCapitalizationTrainDataConfig] = None @@ -233,16 +235,16 @@ class PunctuationCapitalizationModelConfig: test_ds: Optional[PunctuationCapitalizationEvalDataConfig] = None """A configuration for creating test datasets and data loaders.""" - punct_head: HeadConfig = HeadConfig() + punct_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating punctuation MLP head that is applied to a language model outputs.""" - capit_head: HeadConfig = HeadConfig() + capit_head: HeadConfig = field(default_factory=lambda: HeadConfig()) """A configuration for creating capitalization MLP head that is applied to a language model outputs.""" - tokenizer: Any = TokenizerConfig() + tokenizer: Any = field(default_factory=lambda: TokenizerConfig()) """A configuration for source text tokenizer.""" - language_model: LanguageModelConfig = LanguageModelConfig() + language_model: LanguageModelConfig = field(default_factory=lambda: LanguageModelConfig()) """A configuration of a BERT-like language model which serves as a model body.""" optim: Optional[Any] = None @@ -311,22 +313,30 @@ class PunctuationCapitalizationConfig(NemoConfig): do_testing: bool = False """Whether ot perform testing of the model.""" - model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig() + model: PunctuationCapitalizationModelConfig = field(default_factory=lambda: PunctuationCapitalizationModelConfig()) """A configuration for the :class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel` model.""" - trainer: Optional[TrainerConfig] = TrainerConfig() + trainer: Optional[TrainerConfig] = field(default_factory=lambda: TrainerConfig()) """Contains ``Trainer`` Lightning class constructor parameters.""" - exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[]) + exp_manager: Optional[ExpManagerConfig] = field( + default_factory=lambda: ExpManagerConfig(name=None, files_to_copy=[]) + ) """A configuration with various NeMo training options such as output directories, resuming from checkpoint, tensorboard and W&B logging, and so on. For possible options see :ref:`exp-manager-label`.""" + def __post_init__(self): + if self.exp_manager is not None: + self.exp_manager.name = self.name + @dataclass class PunctuationCapitalizationLexicalAudioConfig(PunctuationCapitalizationConfig): - model: PunctuationCapitalizationLexicalAudioModelConfig = PunctuationCapitalizationLexicalAudioModelConfig() + model: PunctuationCapitalizationLexicalAudioModelConfig = field( + default_factory=lambda: PunctuationCapitalizationLexicalAudioModelConfig() + ) def is_legacy_model_config(model_cfg: DictConfig) -> bool: diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py --- a/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py +++ b/nemo/collections/nlp/modules/common/megatron/megatron_encoders.py @@ -13,7 +13,6 @@ # limitations under the License. """Transformer based language model.""" -from MeCab import Model from nemo.collections.nlp.modules.common.megatron.megatron_perceiver_encoders import MegatronPerceiverEncoderModule from nemo.collections.nlp.modules.common.megatron.megatron_transformer_encoder import MegatronTransformerEncoderModule from nemo.collections.nlp.modules.common.megatron.retrieval_transformer import ( @@ -25,6 +24,13 @@ scaled_init_method_normal, ) +try: + from MeCab import Model + + HAVE_MECAB = True +except (ImportError, ModuleNotFoundError): + HAVE_MECAB = False + try: from apex.transformer.enums import AttnMaskType, ModelType diff --git a/nemo/collections/tts/models/fastpitch.py b/nemo/collections/tts/models/fastpitch.py --- a/nemo/collections/tts/models/fastpitch.py +++ b/nemo/collections/tts/models/fastpitch.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import List, Optional @@ -70,12 +70,12 @@ class TextTokenizer: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) @dataclass class TextTokenizerConfig: - text_tokenizer: TextTokenizer = TextTokenizer() + text_tokenizer: TextTokenizer = field(default_factory=lambda: TextTokenizer()) class FastPitchModel(SpectrogramGenerator, Exportable, FastPitchAdapterModelMixin): diff --git a/nemo/collections/tts/models/tacotron2.py b/nemo/collections/tts/models/tacotron2.py --- a/nemo/collections/tts/models/tacotron2.py +++ b/nemo/collections/tts/models/tacotron2.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch @@ -53,7 +53,7 @@ class Preprocessor: @dataclass class Tacotron2Config: - preprocessor: Preprocessor = Preprocessor() + preprocessor: Preprocessor = field(default_factory=lambda: Preprocessor()) encoder: Dict[Any, Any] = MISSING decoder: Dict[Any, Any] = MISSING postnet: Dict[Any, Any] = MISSING diff --git a/nemo/core/config/modelPT.py b/nemo/core/config/modelPT.py --- a/nemo/core/config/modelPT.py +++ b/nemo/core/config/modelPT.py @@ -58,11 +58,13 @@ class HydraConfig: class NemoConfig: name: str = MISSING model: ModelConfig = MISSING - trainer: config.TrainerConfig = config.TrainerConfig( - strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + trainer: config.TrainerConfig = field( + default_factory=lambda: config.TrainerConfig( + strategy="ddp", enable_checkpointing=False, logger=False, log_every_n_steps=1, accelerator='gpu' + ) ) - exp_manager: Optional[Any] = exp_manager.ExpManagerConfig() - hydra: HydraConfig = HydraConfig() + exp_manager: Optional[Any] = field(default_factory=lambda: exp_manager.ExpManagerConfig()) + hydra: HydraConfig = field(default_factory=lambda: HydraConfig()) class ModelConfigBuilder: diff --git a/nemo/utils/exp_manager.py b/nemo/utils/exp_manager.py --- a/nemo/utils/exp_manager.py +++ b/nemo/utils/exp_manager.py @@ -18,7 +18,7 @@ import sys import time import warnings -from dataclasses import dataclass +from dataclasses import dataclass, field from datetime import timedelta from pathlib import Path from shutil import copy, move @@ -146,28 +146,30 @@ class ExpManagerConfig: create_wandb_logger: Optional[bool] = False wandb_logger_kwargs: Optional[Dict[Any, Any]] = None create_mlflow_logger: Optional[bool] = False - mlflow_logger_kwargs: Optional[MLFlowParams] = MLFlowParams() + mlflow_logger_kwargs: Optional[MLFlowParams] = field(default_factory=lambda: MLFlowParams()) create_dllogger_logger: Optional[bool] = False - dllogger_logger_kwargs: Optional[DLLoggerParams] = DLLoggerParams() + dllogger_logger_kwargs: Optional[DLLoggerParams] = field(default_factory=lambda: DLLoggerParams()) create_clearml_logger: Optional[bool] = False - clearml_logger_kwargs: Optional[ClearMLParams] = ClearMLParams() + clearml_logger_kwargs: Optional[ClearMLParams] = field(default_factory=lambda: ClearMLParams()) # Checkpointing parameters create_checkpoint_callback: Optional[bool] = True - checkpoint_callback_params: Optional[CallbackParams] = CallbackParams() + checkpoint_callback_params: Optional[CallbackParams] = field(default_factory=lambda: CallbackParams()) create_early_stopping_callback: Optional[bool] = False - early_stopping_callback_params: Optional[EarlyStoppingParams] = EarlyStoppingParams() + early_stopping_callback_params: Optional[EarlyStoppingParams] = field( + default_factory=lambda: EarlyStoppingParams() + ) create_preemption_callback: Optional[bool] = True # Additional exp_manager arguments files_to_copy: Optional[List[str]] = None # logs timing of train/val/test steps log_step_timing: Optional[bool] = True - step_timing_kwargs: Optional[StepTimingParams] = StepTimingParams() + step_timing_kwargs: Optional[StepTimingParams] = field(default_factory=lambda: StepTimingParams()) # Configures creation of log files for different ranks log_local_rank_0_only: Optional[bool] = False log_global_rank_0_only: Optional[bool] = False # disable initial validation when resuming from a checkpoint saved during validation disable_validation_on_resume: Optional[bool] = True - ema: Optional[EMAParams] = EMAParams() + ema: Optional[EMAParams] = field(default_factory=lambda: EMAParams()) # Wall clock time limit max_time_per_run: Optional[str] = None # time to sleep non 0 ranks during initialization diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py @@ -112,14 +112,14 @@ class EvalBeamSearchNGramConfig: beam_beta: List[float] = field(default_factory=lambda: [0.0]) # The beta parameter or list of the betas for the beam search decoding decoding_strategy: str = "beam" - decoding: ctc_beam_decoding.BeamCTCInferConfig = ctc_beam_decoding.BeamCTCInferConfig(beam_size=128) + decoding: ctc_beam_decoding.BeamCTCInferConfig = field(default_factory=lambda: ctc_beam_decoding.BeamCTCInferConfig(beam_size=128)) - text_processing: Optional[TextProcessingConfig] = TextProcessingConfig( + text_processing: Optional[TextProcessingConfig] = field(default_factory=lambda: TextProcessingConfig( punctuation_marks = ".,?", separate_punctuation = False, do_lowercase = False, rm_punctuation = False, - ) + )) # fmt: on diff --git a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py --- a/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py +++ b/scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram_transducer.py @@ -115,7 +115,7 @@ class EvalBeamSearchNGramConfig: hat_subtract_ilm: bool = False hat_ilm_weight: List[float] = field(default_factory=lambda: [0.0]) - decoding: rnnt_beam_decoding.BeamRNNTInferConfig = rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128) + decoding: rnnt_beam_decoding.BeamRNNTInferConfig = field(default_factory=lambda: rnnt_beam_decoding.BeamRNNTInferConfig(beam_size=128)) # fmt: on diff --git a/scripts/confidence_ensembles/build_ensemble.py b/scripts/confidence_ensembles/build_ensemble.py --- a/scripts/confidence_ensembles/build_ensemble.py +++ b/scripts/confidence_ensembles/build_ensemble.py @@ -75,7 +75,7 @@ import sys import tempfile from copy import deepcopy -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Tuple @@ -209,19 +209,23 @@ class BuildEnsembleConfig: random_seed: int = 0 # for reproducibility # default confidence, can override - confidence: ConfidenceConfig = ConfidenceConfig( - # we keep frame confidences and apply aggregation manually to get full-utterance confidence - preserve_frame_confidence=True, - exclude_blank=True, - aggregation="mean", - method_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + confidence: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig( + # we keep frame confidences and apply aggregation manually to get full-utterance confidence + preserve_frame_confidence=True, + exclude_blank=True, + aggregation="mean", + measure_cfg=ConfidenceMethodConfig(name="entropy", entropy_type="renyi", alpha=0.25, entropy_norm="lin",), + ) ) temperature: float = 1.0 # this is optional, but can be used to change any aspect of the transcription # config, such as batch size or amp usage. Note that model, data and confidence # will be overriden by this script - transcription: transcribe_speech.TranscriptionConfig = transcribe_speech.TranscriptionConfig() + transcription: transcribe_speech.TranscriptionConfig = field( + default_factory=lambda: transcribe_speech.TranscriptionConfig() + ) # set to True to tune the confidence. # requires dev manifests to be specified for each model @@ -229,12 +233,14 @@ class BuildEnsembleConfig: # used to specify what to tune over. By default runs tuning over some # reasonalbe grid, so that it does not take forever. # Can be changed as needed - tune_confidence_config: TuneConfidenceConfig = TuneConfidenceConfig() + tune_confidence_config: TuneConfidenceConfig = field(default_factory=lambda: TuneConfidenceConfig()) # very fast to tune and can be important in case of imbalanced datasets # will automatically set to False if dev data is not available tune_logistic_regression: bool = True - tune_logistic_regression_config: TuneLogisticRegressionConfig = TuneLogisticRegressionConfig() + tune_logistic_regression_config: TuneLogisticRegressionConfig = field( + default_factory=lambda: TuneLogisticRegressionConfig() + ) def __post_init__(self): """Checking that if any dev data is provided, all are provided. diff --git a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py --- a/scripts/speech_recognition/confidence/benchmark_asr_confidence.py +++ b/scripts/speech_recognition/confidence/benchmark_asr_confidence.py @@ -14,7 +14,7 @@ import json import os -from dataclasses import dataclass, is_dataclass +from dataclasses import dataclass, field, is_dataclass from pathlib import Path from typing import Optional @@ -124,7 +124,9 @@ class ConfidenceBenchmarkingConfig: # Confidence configs target_level: str = "auto" # Choices: "word", "token", "auto" (for both word- and token-level confidence) - confidence_cfg: ConfidenceConfig = ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + confidence_cfg: ConfidenceConfig = field( + default_factory=lambda: ConfidenceConfig(preserve_word_confidence=True, preserve_token_confidence=True) + ) grid_params: Optional[str] = None # a dictionary with lists of parameters to iteratively benchmark on diff --git a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py --- a/scripts/speech_recognition/convert_to_tarred_audio_dataset.py +++ b/scripts/speech_recognition/convert_to_tarred_audio_dataset.py @@ -202,7 +202,7 @@ class ASRTarredDatasetMetadata: num_samples_per_shard: Optional[int] = None is_concatenated_manifest: bool = False - dataset_config: Optional[ASRTarredDatasetConfig] = ASRTarredDatasetConfig() + dataset_config: Optional[ASRTarredDatasetConfig] = field(default_factory=lambda: ASRTarredDatasetConfig()) history: Optional[List[Any]] = field(default_factory=lambda: []) def __post_init__(self): diff --git a/tools/nemo_forced_aligner/align.py b/tools/nemo_forced_aligner/align.py --- a/tools/nemo_forced_aligner/align.py +++ b/tools/nemo_forced_aligner/align.py @@ -149,8 +149,8 @@ class AlignmentConfig: # Output file configs save_output_file_formats: List[str] = field(default_factory=lambda: ["ctm", "ass"]) - ctm_file_config: CTMFileConfig = CTMFileConfig() - ass_file_config: ASSFileConfig = ASSFileConfig() + ctm_file_config: CTMFileConfig = field(default_factory=lambda: CTMFileConfig()) + ass_file_config: ASSFileConfig = field(default_factory=lambda: ASSFileConfig()) @hydra_runner(config_name="AlignmentConfig", schema=AlignmentConfig) </patch> </s> </patch>
diff --git a/tests/collections/asr/test_text_to_text_dataset.py b/tests/collections/asr/test_text_to_text_dataset.py --- a/tests/collections/asr/test_text_to_text_dataset.py +++ b/tests/collections/asr/test_text_to_text_dataset.py @@ -15,7 +15,7 @@ import json import multiprocessing import os -from dataclasses import dataclass +from dataclasses import dataclass, field from pathlib import Path import pytest @@ -118,7 +118,7 @@ class TextTokenizerCfg: apostrophe: bool = True pad_with_space: bool = True add_blank_at: bool = True - g2p: G2PConfig = G2PConfig() + g2p: G2PConfig = field(default_factory=lambda: G2PConfig()) config = OmegaConf.create(OmegaConf.to_yaml(TextTokenizerCfg())) return instantiate(config)
1.0
slackapi__python-slack-events-api-71
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): </issue> <code> [start of README.rst] 1 Slack Events API adapter for Python 2 =================================== 3 4 .. image:: https://badge.fury.io/py/slackeventsapi.svg 5 :target: https://pypi.org/project/slackeventsapi/ 6 .. image:: https://travis-ci.org/slackapi/python-slack-events-api.svg?branch=master 7 :target: https://travis-ci.org/slackapi/python-slack-events-api 8 .. image:: https://codecov.io/gh/slackapi/python-slack-events-api/branch/master/graph/badge.svg 9 :target: https://codecov.io/gh/slackapi/python-slack-events-api 10 11 12 The Slack Events Adapter is a Python-based solution to receive and parse events 13 from Slackโ€™s Events API. This library uses an event emitter framework to allow 14 you to easily process Slack events by simply attaching functions 15 to event listeners. 16 17 This adapter enhances and simplifies Slack's Events API by incorporating useful best practices, patterns, and opportunities to abstract out common tasks. 18 19 ๐Ÿ’ก We wrote a `blog post which explains how`_ the Events API can help you, why we built these tools, and how you can use them to build production-ready Slack apps. 20 21 .. _blog post which explains how: https://medium.com/@SlackAPI/enhancing-slacks-events-api-7535827829ab 22 23 24 ๐Ÿค– Installation 25 ---------------- 26 27 .. code:: shell 28 29 pip install slackeventsapi 30 31 ๐Ÿค– App Setup 32 -------------------- 33 34 Before you can use the `Events API`_ you must 35 `create a Slack App`_, and turn on 36 `Event Subscriptions`_. 37 38 ๐Ÿ’ก When you add the Request URL to your app's Event Subscription settings, 39 Slack will send a request containing a `challenge` code to verify that your 40 server is alive. This package handles that URL Verification event for you, so 41 all you need to do is start the example app, start ngrok and configure your 42 URL accordingly. 43 44 โœ… Once you have your `Request URL` verified, your app is ready to start 45 receiving Team Events. 46 47 ๐Ÿ”‘ Your server will begin receiving Events from Slack's Events API as soon as a 48 user has authorized your app. 49 50 ๐Ÿค– Development workflow: 51 =========================== 52 53 (1) Create a Slack app on https://api.slack.com/apps 54 (2) Add a `bot user` for your app 55 (3) Start the example app on your **Request URL** endpoint 56 (4) Start ngrok and copy the **HTTPS** URL 57 (5) Add your **Request URL** and subscribe your app to events 58 (6) Go to your ngrok URL (e.g. https://myapp12.ngrok.com/) and auth your app 59 60 **๐ŸŽ‰ Once your app has been authorized, you will begin receiving Slack Events** 61 62 โš ๏ธ Ngrok is a great tool for developing Slack apps, but we don't recommend using ngrok 63 for production apps. 64 65 ๐Ÿค– Usage 66 ---------- 67 **โš ๏ธ Keep your app's credentials safe!** 68 69 - For development, keep them in virtualenv variables. 70 71 - For production, use a secure data store. 72 73 - Never post your app's credentials to github. 74 75 .. code:: python 76 77 SLACK_SIGNING_SECRET = os.environ["SLACK_SIGNING_SECRET"] 78 79 Create a Slack Event Adapter for receiving actions via the Events API 80 ----------------------------------------------------------------------- 81 **Using the built-in Flask server:** 82 83 .. code:: python 84 85 from slackeventsapi import SlackEventAdapter 86 87 88 slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, endpoint="/slack/events") 89 90 91 # Create an event listener for "reaction_added" events and print the emoji name 92 @slack_events_adapter.on("reaction_added") 93 def reaction_added(event_data): 94 emoji = event_data["event"]["reaction"] 95 print(emoji) 96 97 98 # Start the server on port 3000 99 slack_events_adapter.start(port=3000) 100 101 102 **Using your existing Flask instance:** 103 104 105 .. code:: python 106 107 from flask import Flask 108 from slackeventsapi import SlackEventAdapter 109 110 111 # This `app` represents your existing Flask app 112 app = Flask(__name__) 113 114 115 # An example of one of your Flask app's routes 116 @app.route("/") 117 def hello(): 118 return "Hello there!" 119 120 121 # Bind the Events API route to your existing Flask app by passing the server 122 # instance as the last param, or with `server=app`. 123 slack_events_adapter = SlackEventAdapter(SLACK_SIGNING_SECRET, "/slack/events", app) 124 125 126 # Create an event listener for "reaction_added" events and print the emoji name 127 @slack_events_adapter.on("reaction_added") 128 def reaction_added(event_data): 129 emoji = event_data["event"]["reaction"] 130 print(emoji) 131 132 133 # Start the server on port 3000 134 if __name__ == "__main__": 135 app.run(port=3000) 136 137 For a comprehensive list of available Slack `Events` and more information on 138 `Scopes`, see https://api.slack.com/events-api 139 140 ๐Ÿค– Example event listeners 141 ----------------------------- 142 143 See `example.py`_ for usage examples. This example also utilizes the 144 SlackClient Web API client. 145 146 .. _example.py: /example/ 147 148 ๐Ÿค” Support 149 ----------- 150 151 Need help? Join `Slack Community`_ and talk to us in `#slack-api`_. 152 153 You can also `create an Issue`_ right here on GitHub. 154 155 .. _Events API: https://api.slack.com/events-api 156 .. _create a Slack App: https://api.slack.com/apps/new 157 .. _Event Subscriptions: https://api.slack.com/events-api#subscriptions 158 .. _Slack Community: http://slackcommunity.com/ 159 .. _#slack-api: https://dev4slack.slack.com/messages/slack-api/ 160 .. _create an Issue: https://github.com/slackapi/python-slack-events-api/issues/new 161 [end of README.rst] [start of /dev/null] 1 [end of /dev/null] [start of slackeventsapi/server.py] 1 from flask import Flask, request, make_response, Blueprint 2 import json 3 import platform 4 import sys 5 import hmac 6 import hashlib 7 from time import time 8 from .version import __version__ 9 10 11 class SlackServer(Flask): 12 def __init__(self, signing_secret, endpoint, emitter, server): 13 self.signing_secret = signing_secret 14 self.emitter = emitter 15 self.endpoint = endpoint 16 self.package_info = self.get_package_info() 17 18 # If a server is passed in, bind the event handler routes to it, 19 # otherwise create a new Flask instance. 20 if server: 21 if isinstance(server, Flask) or isinstance(server, Blueprint): 22 self.bind_route(server) 23 else: 24 raise TypeError("Server must be an instance of Flask or Blueprint") 25 else: 26 Flask.__init__(self, __name__) 27 self.bind_route(self) 28 29 def get_package_info(self): 30 client_name = __name__.split('.')[0] 31 client_version = __version__ # Version is returned from version.py 32 33 # Collect the package info, Python version and OS version. 34 package_info = { 35 "client": "{0}/{1}".format(client_name, client_version), 36 "python": "Python/{v.major}.{v.minor}.{v.micro}".format(v=sys.version_info), 37 "system": "{0}/{1}".format(platform.system(), platform.release()) 38 } 39 40 # Concatenate and format the user-agent string to be passed into request headers 41 ua_string = [] 42 for key, val in package_info.items(): 43 ua_string.append(val) 44 45 return " ".join(ua_string) 46 47 def verify_signature(self, timestamp, signature): 48 # Verify the request signature of the request sent from Slack 49 # Generate a new hash using the app's signing secret and request data 50 51 # Compare the generated hash and incoming request signature 52 # Python 2.7.6 doesn't support compare_digest 53 # It's recommended to use Python 2.7.7+ 54 # noqa See https://docs.python.org/2/whatsnew/2.7.html#pep-466-network-security-enhancements-for-python-2-7 55 req = str.encode('v0:' + str(timestamp) + ':') + request.get_data() 56 request_hash = 'v0=' + hmac.new( 57 str.encode(self.signing_secret), 58 req, hashlib.sha256 59 ).hexdigest() 60 61 if hasattr(hmac, "compare_digest"): 62 # Compare byte strings for Python 2 63 if (sys.version_info[0] == 2): 64 return hmac.compare_digest(bytes(request_hash), bytes(signature)) 65 else: 66 return hmac.compare_digest(request_hash, signature) 67 else: 68 if len(request_hash) != len(signature): 69 return False 70 result = 0 71 if isinstance(request_hash, bytes) and isinstance(signature, bytes): 72 for x, y in zip(request_hash, signature): 73 result |= x ^ y 74 else: 75 for x, y in zip(request_hash, signature): 76 result |= ord(x) ^ ord(y) 77 return result == 0 78 79 def bind_route(self, server): 80 @server.route(self.endpoint, methods=['GET', 'POST']) 81 def event(): 82 # If a GET request is made, return 404. 83 if request.method == 'GET': 84 return make_response("These are not the slackbots you're looking for.", 404) 85 86 # Each request comes with request timestamp and request signature 87 # emit an error if the timestamp is out of range 88 req_timestamp = request.headers.get('X-Slack-Request-Timestamp') 89 if abs(time() - int(req_timestamp)) > 60 * 5: 90 slack_exception = SlackEventAdapterException('Invalid request timestamp') 91 self.emitter.emit('error', slack_exception) 92 return make_response("", 403) 93 94 # Verify the request signature using the app's signing secret 95 # emit an error if the signature can't be verified 96 req_signature = request.headers.get('X-Slack-Signature') 97 if not self.verify_signature(req_timestamp, req_signature): 98 slack_exception = SlackEventAdapterException('Invalid request signature') 99 self.emitter.emit('error', slack_exception) 100 return make_response("", 403) 101 102 # Parse the request payload into JSON 103 event_data = json.loads(request.data.decode('utf-8')) 104 105 # Echo the URL verification challenge code back to Slack 106 if "challenge" in event_data: 107 return make_response( 108 event_data.get("challenge"), 200, {"content_type": "application/json"} 109 ) 110 111 # Parse the Event payload and emit the event to the event listener 112 if "event" in event_data: 113 event_type = event_data["event"]["type"] 114 self.emitter.emit(event_type, event_data) 115 response = make_response("", 200) 116 response.headers['X-Slack-Powered-By'] = self.package_info 117 return response 118 119 120 class SlackEventAdapterException(Exception): 121 """ 122 Base exception for all errors raised by the SlackClient library 123 """ 124 125 def __init__(self, msg=None): 126 if msg is None: 127 # default error message 128 msg = "An error occurred in the SlackEventsApiAdapter library" 129 super(SlackEventAdapterException, self).__init__(msg) 130 [end of slackeventsapi/server.py] </code> Here is an example of a patch file. It consists of changes to the code base. It specifies the file names, the line numbers of each change, and the removed and added lines. A single patch file can contain changes to multiple files. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch> I need you to solve the provided issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the format shown above. Respond below:
slackapi/python-slack-events-api
0c0ce604b502508622fb14c278a0d64841fa32e3
Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s): Passing Flask app proxy as server Hi Guys, I have an app factory on my setup and the app object usually it is invoked as : `from flask import current_app as app` However, the slackeventsapi complains about the app object : `TypeError("Server must be an instance of Flask")` I have fixed adding the following to server.py : `from werkzeug.local import LocalProxy # Importing the localproxy class` Line 25 Changed from : ` if isinstance(server, Flask):` to : `if isinstance(server, Flask) or isinstance(server, LocalProxy):` Basically, if a Flask app proxy is passed the api will carry on without complaining since it has the same methods as the Flask app object. I hope this help other people and it is considered as a solution if more information is needed I am help to provide. Thanks for the good work with the API. ### What type of issue is this? (place an `x` in one of the `[ ]`) - [X] bug ? - [X] enhancement (feature request) - [ ] question - [ ] documentation related - [ ] testing related - [ ] discussion ### Requirements * [X] I've read and understood the [Contributing guidelines](https://github.com/slackapi/python-slack-events-api/blob/master/.github/contributing.md) and have done my best effort to follow them. * [X] I've read and agree to the [Code of Conduct](https://slackhq.github.io/code-of-conduct). * [X] I've searched for any related issues and avoided creating a duplicate issue. #### Reproducible in: slackeventsapi version: slackeventsapi==2.1.0 python version: Python 3.7.3 OS version(s):
2020-06-12T06:58:10Z
<patch> <patch> diff --git a/example/current_app/main.py b/example/current_app/main.py new file mode 100644 --- /dev/null +++ b/example/current_app/main.py @@ -0,0 +1,49 @@ +# ------------------ +# Only for running this script here +import sys +from os.path import dirname +sys.path.insert(1, f"{dirname(__file__)}/../..") +# ------------------ + +import os +from slack import WebClient +import logging +logging.basicConfig(level=logging.DEBUG) + +from flask import Flask + +app = Flask(__name__) + +with app.app_context(): + from test_module.slack_app import slack_events_adapter + + slack_bot_token = os.environ["SLACK_BOT_TOKEN"] + slack_client = WebClient(slack_bot_token) + + + @slack_events_adapter.on("message") + def handle_message(event_data): + message = event_data["event"] + if message.get("subtype") is None and "hi" in message.get('text'): + channel = message["channel"] + message = "Hi <@%s>! :tada:" % message["user"] + slack_client.chat_postMessage(channel=channel, text=message) + + + @slack_events_adapter.on("error") + def error_handler(err): + print("ERROR: " + str(err)) + +# (Terminal A) +# source env/bin/activate +# (env) $ export SLACK_BOT_TOKEN=xoxb-*** +# (env) $ export SLACK_SIGNING_SECRET=** +# (env) $ cd example/current_app +# (env) $ FLASK_APP=main.py FLASK_ENV=development flask run --port 3000 + +# (Terminal B) +# ngrok http 3000 + +# in Slack +# /invite @{your app's bot user} +# post a message "hi" in the channel diff --git a/slackeventsapi/server.py b/slackeventsapi/server.py --- a/slackeventsapi/server.py +++ b/slackeventsapi/server.py @@ -1,10 +1,13 @@ -from flask import Flask, request, make_response, Blueprint +import hashlib +import hmac import json import platform import sys -import hmac -import hashlib from time import time + +from flask import Flask, request, make_response, Blueprint +from werkzeug.local import LocalProxy + from .version import __version__ @@ -18,10 +21,10 @@ def __init__(self, signing_secret, endpoint, emitter, server): # If a server is passed in, bind the event handler routes to it, # otherwise create a new Flask instance. if server: - if isinstance(server, Flask) or isinstance(server, Blueprint): + if isinstance(server, (Flask, Blueprint, LocalProxy)): self.bind_route(server) else: - raise TypeError("Server must be an instance of Flask or Blueprint") + raise TypeError("Server must be an instance of Flask, Blueprint, or LocalProxy") else: Flask.__init__(self, __name__) self.bind_route(self) </patch> </s> </patch>
diff --git a/example/current_app/test_module/__init__.py b/example/current_app/test_module/__init__.py new file mode 100644 diff --git a/example/current_app/test_module/slack_app.py b/example/current_app/test_module/slack_app.py new file mode 100644 --- /dev/null +++ b/example/current_app/test_module/slack_app.py @@ -0,0 +1,16 @@ +# ------------------ +# Only for running this script here +import logging +import sys +from os.path import dirname + +sys.path.insert(1, f"{dirname(__file__)}/../../..") +logging.basicConfig(level=logging.DEBUG) +# ------------------ + +from flask import current_app as app +from slackeventsapi import SlackEventAdapter +import os + +slack_signing_secret = os.environ["SLACK_SIGNING_SECRET"] +slack_events_adapter = SlackEventAdapter(slack_signing_secret, "/slack/events", app) diff --git a/tests/test_server.py b/tests/test_server.py --- a/tests/test_server.py +++ b/tests/test_server.py @@ -18,7 +18,7 @@ def test_server_not_flask(): with pytest.raises(TypeError) as e: invalid_flask = "I am not a Flask" SlackEventAdapter("SIGNING_SECRET", "/slack/events", invalid_flask) - assert e.value.args[0] == 'Server must be an instance of Flask or Blueprint' + assert e.value.args[0] == 'Server must be an instance of Flask, Blueprint, or LocalProxy' def test_blueprint_server():
1.0
celery__celery-2598
You will be provided with a partial code base and an issue statement explaining a problem to resolve. <issue> CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ``` </issue> <code> [start of README.rst] 1 ================================= 2 celery - Distributed Task Queue 3 ================================= 4 5 .. image:: http://cloud.github.com/downloads/celery/celery/celery_128.png 6 7 |build-status| |coverage-status| 8 9 :Version: 3.2.0a1 (Cipater) 10 :Web: http://celeryproject.org/ 11 :Download: http://pypi.python.org/pypi/celery/ 12 :Source: http://github.com/celery/celery/ 13 :Keywords: task queue, job queue, asynchronous, async, rabbitmq, amqp, redis, 14 python, webhooks, queue, distributed 15 16 -- 17 18 What is a Task Queue? 19 ===================== 20 21 Task queues are used as a mechanism to distribute work across threads or 22 machines. 23 24 A task queue's input is a unit of work, called a task, dedicated worker 25 processes then constantly monitor the queue for new work to perform. 26 27 Celery communicates via messages, usually using a broker 28 to mediate between clients and workers. To initiate a task a client puts a 29 message on the queue, the broker then delivers the message to a worker. 30 31 A Celery system can consist of multiple workers and brokers, giving way 32 to high availability and horizontal scaling. 33 34 Celery is a library written in Python, but the protocol can be implemented in 35 any language. So far there's RCelery_ for the Ruby programming language, and a 36 `PHP client`, but language interoperability can also be achieved 37 by using webhooks. 38 39 .. _RCelery: https://github.com/leapfrogonline/rcelery 40 .. _`PHP client`: https://github.com/gjedeer/celery-php 41 .. _`using webhooks`: 42 http://docs.celeryproject.org/en/latest/userguide/remote-tasks.html 43 44 What do I need? 45 =============== 46 47 Celery version 3.0 runs on, 48 49 - Python (2.6, 2.7, 3.3, 3.4) 50 - PyPy (1.8, 1.9) 51 - Jython (2.5, 2.7). 52 53 This is the last version to support Python 2.5, 54 and from Celery 3.1, Python 2.6 or later is required. 55 The last version to support Python 2.4 was Celery series 2.2. 56 57 *Celery* is usually used with a message broker to send and receive messages. 58 The RabbitMQ, Redis transports are feature complete, 59 but there's also experimental support for a myriad of other solutions, including 60 using SQLite for local development. 61 62 *Celery* can run on a single machine, on multiple machines, or even 63 across datacenters. 64 65 Get Started 66 =========== 67 68 If this is the first time you're trying to use Celery, or you are 69 new to Celery 3.0 coming from previous versions then you should read our 70 getting started tutorials: 71 72 - `First steps with Celery`_ 73 74 Tutorial teaching you the bare minimum needed to get started with Celery. 75 76 - `Next steps`_ 77 78 A more complete overview, showing more features. 79 80 .. _`First steps with Celery`: 81 http://docs.celeryproject.org/en/latest/getting-started/first-steps-with-celery.html 82 83 .. _`Next steps`: 84 http://docs.celeryproject.org/en/latest/getting-started/next-steps.html 85 86 Celery is... 87 ========== 88 89 - **Simple** 90 91 Celery is easy to use and maintain, and does *not need configuration files*. 92 93 It has an active, friendly community you can talk to for support, 94 including a `mailing-list`_ and and an IRC channel. 95 96 Here's one of the simplest applications you can make:: 97 98 from celery import Celery 99 100 app = Celery('hello', broker='amqp://guest@localhost//') 101 102 @app.task 103 def hello(): 104 return 'hello world' 105 106 - **Highly Available** 107 108 Workers and clients will automatically retry in the event 109 of connection loss or failure, and some brokers support 110 HA in way of *Master/Master* or *Master/Slave* replication. 111 112 - **Fast** 113 114 A single Celery process can process millions of tasks a minute, 115 with sub-millisecond round-trip latency (using RabbitMQ, 116 py-librabbitmq, and optimized settings). 117 118 - **Flexible** 119 120 Almost every part of *Celery* can be extended or used on its own, 121 Custom pool implementations, serializers, compression schemes, logging, 122 schedulers, consumers, producers, autoscalers, broker transports and much more. 123 124 It supports... 125 ============ 126 127 - **Message Transports** 128 129 - RabbitMQ_, Redis_, 130 - MongoDB_ (experimental), Amazon SQS (experimental), 131 - CouchDB_ (experimental), SQLAlchemy_ (experimental), 132 - Django ORM (experimental), `IronMQ`_ 133 - and more... 134 135 - **Concurrency** 136 137 - Prefork, Eventlet_, gevent_, threads/single threaded 138 139 - **Result Stores** 140 141 - AMQP, Redis 142 - memcached, MongoDB 143 - SQLAlchemy, Django ORM 144 - Apache Cassandra, IronCache 145 146 - **Serialization** 147 148 - *pickle*, *json*, *yaml*, *msgpack*. 149 - *zlib*, *bzip2* compression. 150 - Cryptographic message signing. 151 152 .. _`Eventlet`: http://eventlet.net/ 153 .. _`gevent`: http://gevent.org/ 154 155 .. _RabbitMQ: http://rabbitmq.com 156 .. _Redis: http://redis.io 157 .. _MongoDB: http://mongodb.org 158 .. _Beanstalk: http://kr.github.com/beanstalkd 159 .. _CouchDB: http://couchdb.apache.org 160 .. _SQLAlchemy: http://sqlalchemy.org 161 .. _`IronMQ`: http://iron.io 162 163 Framework Integration 164 ===================== 165 166 Celery is easy to integrate with web frameworks, some of which even have 167 integration packages: 168 169 +--------------------+------------------------+ 170 | `Django`_ | not needed | 171 +--------------------+------------------------+ 172 | `Pyramid`_ | `pyramid_celery`_ | 173 +--------------------+------------------------+ 174 | `Pylons`_ | `celery-pylons`_ | 175 +--------------------+------------------------+ 176 | `Flask`_ | not needed | 177 +--------------------+------------------------+ 178 | `web2py`_ | `web2py-celery`_ | 179 +--------------------+------------------------+ 180 | `Tornado`_ | `tornado-celery`_ | 181 +--------------------+------------------------+ 182 183 The integration packages are not strictly necessary, but they can make 184 development easier, and sometimes they add important hooks like closing 185 database connections at ``fork``. 186 187 .. _`Django`: http://djangoproject.com/ 188 .. _`Pylons`: http://www.pylonsproject.org/ 189 .. _`Flask`: http://flask.pocoo.org/ 190 .. _`web2py`: http://web2py.com/ 191 .. _`Bottle`: http://bottlepy.org/ 192 .. _`Pyramid`: http://docs.pylonsproject.org/en/latest/docs/pyramid.html 193 .. _`pyramid_celery`: http://pypi.python.org/pypi/pyramid_celery/ 194 .. _`django-celery`: http://pypi.python.org/pypi/django-celery 195 .. _`celery-pylons`: http://pypi.python.org/pypi/celery-pylons 196 .. _`web2py-celery`: http://code.google.com/p/web2py-celery/ 197 .. _`Tornado`: http://www.tornadoweb.org/ 198 .. _`tornado-celery`: http://github.com/mher/tornado-celery/ 199 200 .. _celery-documentation: 201 202 Documentation 203 ============= 204 205 The `latest documentation`_ with user guides, tutorials and API reference 206 is hosted at Read The Docs. 207 208 .. _`latest documentation`: http://docs.celeryproject.org/en/latest/ 209 210 .. _celery-installation: 211 212 Installation 213 ============ 214 215 You can install Celery either via the Python Package Index (PyPI) 216 or from source. 217 218 To install using `pip`,:: 219 220 $ pip install -U Celery 221 222 To install using `easy_install`,:: 223 224 $ easy_install -U Celery 225 226 .. _bundles: 227 228 Bundles 229 ------- 230 231 Celery also defines a group of bundles that can be used 232 to install Celery and the dependencies for a given feature. 233 234 You can specify these in your requirements or on the ``pip`` comand-line 235 by using brackets. Multiple bundles can be specified by separating them by 236 commas. 237 :: 238 239 $ pip install "celery[librabbitmq]" 240 241 $ pip install "celery[librabbitmq,redis,auth,msgpack]" 242 243 The following bundles are available: 244 245 Serializers 246 ~~~~~~~~~~~ 247 248 :celery[auth]: 249 for using the auth serializer. 250 251 :celery[msgpack]: 252 for using the msgpack serializer. 253 254 :celery[yaml]: 255 for using the yaml serializer. 256 257 Concurrency 258 ~~~~~~~~~~~ 259 260 :celery[eventlet]: 261 for using the eventlet pool. 262 263 :celery[gevent]: 264 for using the gevent pool. 265 266 :celery[threads]: 267 for using the thread pool. 268 269 Transports and Backends 270 ~~~~~~~~~~~~~~~~~~~~~~~ 271 272 :celery[librabbitmq]: 273 for using the librabbitmq C library. 274 275 :celery[redis]: 276 for using Redis as a message transport or as a result backend. 277 278 :celery[mongodb]: 279 for using MongoDB as a message transport (*experimental*), 280 or as a result backend (*supported*). 281 282 :celery[sqs]: 283 for using Amazon SQS as a message transport (*experimental*). 284 285 :celery[memcache]: 286 for using memcached as a result backend. 287 288 :celery[cassandra]: 289 for using Apache Cassandra as a result backend. 290 291 :celery[couchdb]: 292 for using CouchDB as a message transport (*experimental*). 293 294 :celery[couchbase]: 295 for using CouchBase as a result backend. 296 297 :celery[beanstalk]: 298 for using Beanstalk as a message transport (*experimental*). 299 300 :celery[zookeeper]: 301 for using Zookeeper as a message transport. 302 303 :celery[zeromq]: 304 for using ZeroMQ as a message transport (*experimental*). 305 306 :celery[sqlalchemy]: 307 for using SQLAlchemy as a message transport (*experimental*), 308 or as a result backend (*supported*). 309 310 :celery[pyro]: 311 for using the Pyro4 message transport (*experimental*). 312 313 :celery[slmq]: 314 for using the SoftLayer Message Queue transport (*experimental*). 315 316 .. _celery-installing-from-source: 317 318 Downloading and installing from source 319 -------------------------------------- 320 321 Download the latest version of Celery from 322 http://pypi.python.org/pypi/celery/ 323 324 You can install it by doing the following,:: 325 326 $ tar xvfz celery-0.0.0.tar.gz 327 $ cd celery-0.0.0 328 $ python setup.py build 329 # python setup.py install 330 331 The last command must be executed as a privileged user if 332 you are not currently using a virtualenv. 333 334 .. _celery-installing-from-git: 335 336 Using the development version 337 ----------------------------- 338 339 With pip 340 ~~~~~~~~ 341 342 The Celery development version also requires the development 343 versions of ``kombu``, ``amqp`` and ``billiard``. 344 345 You can install the latest snapshot of these using the following 346 pip commands:: 347 348 $ pip install https://github.com/celery/celery/zipball/master#egg=celery 349 $ pip install https://github.com/celery/billiard/zipball/master#egg=billiard 350 $ pip install https://github.com/celery/py-amqp/zipball/master#egg=amqp 351 $ pip install https://github.com/celery/kombu/zipball/master#egg=kombu 352 353 With git 354 ~~~~~~~~ 355 356 Please the Contributing section. 357 358 .. _getting-help: 359 360 Getting Help 361 ============ 362 363 .. _mailing-list: 364 365 Mailing list 366 ------------ 367 368 For discussions about the usage, development, and future of celery, 369 please join the `celery-users`_ mailing list. 370 371 .. _`celery-users`: http://groups.google.com/group/celery-users/ 372 373 .. _irc-channel: 374 375 IRC 376 --- 377 378 Come chat with us on IRC. The **#celery** channel is located at the `Freenode`_ 379 network. 380 381 .. _`Freenode`: http://freenode.net 382 383 .. _bug-tracker: 384 385 Bug tracker 386 =========== 387 388 If you have any suggestions, bug reports or annoyances please report them 389 to our issue tracker at http://github.com/celery/celery/issues/ 390 391 .. _wiki: 392 393 Wiki 394 ==== 395 396 http://wiki.github.com/celery/celery/ 397 398 .. _contributing-short: 399 400 Contributing 401 ============ 402 403 Development of `celery` happens at Github: http://github.com/celery/celery 404 405 You are highly encouraged to participate in the development 406 of `celery`. If you don't like Github (for some reason) you're welcome 407 to send regular patches. 408 409 Be sure to also read the `Contributing to Celery`_ section in the 410 documentation. 411 412 .. _`Contributing to Celery`: 413 http://docs.celeryproject.org/en/master/contributing.html 414 415 .. _license: 416 417 License 418 ======= 419 420 This software is licensed under the `New BSD License`. See the ``LICENSE`` 421 file in the top distribution directory for the full license text. 422 423 .. # vim: syntax=rst expandtab tabstop=4 shiftwidth=4 shiftround 424 425 426 .. image:: https://d2weczhvl823v0.cloudfront.net/celery/celery/trend.png 427 :alt: Bitdeli badge 428 :target: https://bitdeli.com/free 429 430 .. |build-status| image:: https://travis-ci.org/celery/celery.svg?branch=master 431 :target: https://travis-ci.org/celery/celery 432 .. |coverage-status| image:: https://coveralls.io/repos/celery/celery/badge.svg 433 :target: https://coveralls.io/r/celery/celery 434 [end of README.rst] [start of celery/backends/amqp.py] 1 # -*- coding: utf-8 -*- 2 """ 3 celery.backends.amqp 4 ~~~~~~~~~~~~~~~~~~~~ 5 6 The AMQP result backend. 7 8 This backend publishes results as messages. 9 10 """ 11 from __future__ import absolute_import 12 13 import socket 14 15 from collections import deque 16 from operator import itemgetter 17 18 from kombu import Exchange, Queue, Producer, Consumer 19 20 from celery import states 21 from celery.exceptions import TimeoutError 22 from celery.five import range, monotonic 23 from celery.utils.functional import dictfilter 24 from celery.utils.log import get_logger 25 from celery.utils.timeutils import maybe_s_to_ms 26 27 from .base import BaseBackend 28 29 __all__ = ['BacklogLimitExceeded', 'AMQPBackend'] 30 31 logger = get_logger(__name__) 32 33 34 class BacklogLimitExceeded(Exception): 35 """Too much state history to fast-forward.""" 36 37 38 def repair_uuid(s): 39 # Historically the dashes in UUIDS are removed from AMQ entity names, 40 # but there is no known reason to. Hopefully we'll be able to fix 41 # this in v4.0. 42 return '%s-%s-%s-%s-%s' % (s[:8], s[8:12], s[12:16], s[16:20], s[20:]) 43 44 45 class NoCacheQueue(Queue): 46 can_cache_declaration = False 47 48 49 class AMQPBackend(BaseBackend): 50 """Publishes results by sending messages.""" 51 Exchange = Exchange 52 Queue = NoCacheQueue 53 Consumer = Consumer 54 Producer = Producer 55 56 BacklogLimitExceeded = BacklogLimitExceeded 57 58 persistent = True 59 supports_autoexpire = True 60 supports_native_join = True 61 62 retry_policy = { 63 'max_retries': 20, 64 'interval_start': 0, 65 'interval_step': 1, 66 'interval_max': 1, 67 } 68 69 def __init__(self, app, connection=None, exchange=None, exchange_type=None, 70 persistent=None, serializer=None, auto_delete=True, **kwargs): 71 super(AMQPBackend, self).__init__(app, **kwargs) 72 conf = self.app.conf 73 self._connection = connection 74 self.persistent = self.prepare_persistent(persistent) 75 self.delivery_mode = 2 if self.persistent else 1 76 exchange = exchange or conf.CELERY_RESULT_EXCHANGE 77 exchange_type = exchange_type or conf.CELERY_RESULT_EXCHANGE_TYPE 78 self.exchange = self._create_exchange( 79 exchange, exchange_type, self.delivery_mode, 80 ) 81 self.serializer = serializer or conf.CELERY_RESULT_SERIALIZER 82 self.auto_delete = auto_delete 83 self.queue_arguments = dictfilter({ 84 'x-expires': maybe_s_to_ms(self.expires), 85 }) 86 87 def _create_exchange(self, name, type='direct', delivery_mode=2): 88 return self.Exchange(name=name, 89 type=type, 90 delivery_mode=delivery_mode, 91 durable=self.persistent, 92 auto_delete=False) 93 94 def _create_binding(self, task_id): 95 name = self.rkey(task_id) 96 return self.Queue(name=name, 97 exchange=self.exchange, 98 routing_key=name, 99 durable=self.persistent, 100 auto_delete=self.auto_delete, 101 queue_arguments=self.queue_arguments) 102 103 def revive(self, channel): 104 pass 105 106 def rkey(self, task_id): 107 return task_id.replace('-', '') 108 109 def destination_for(self, task_id, request): 110 if request: 111 return self.rkey(task_id), request.correlation_id or task_id 112 return self.rkey(task_id), task_id 113 114 def store_result(self, task_id, result, status, 115 traceback=None, request=None, **kwargs): 116 """Send task return value and status.""" 117 routing_key, correlation_id = self.destination_for(task_id, request) 118 if not routing_key: 119 return 120 with self.app.amqp.producer_pool.acquire(block=True) as producer: 121 producer.publish( 122 {'task_id': task_id, 'status': status, 123 'result': self.encode_result(result, status), 124 'traceback': traceback, 125 'children': self.current_task_children(request)}, 126 exchange=self.exchange, 127 routing_key=routing_key, 128 correlation_id=correlation_id, 129 serializer=self.serializer, 130 retry=True, retry_policy=self.retry_policy, 131 declare=self.on_reply_declare(task_id), 132 delivery_mode=self.delivery_mode, 133 ) 134 return result 135 136 def on_reply_declare(self, task_id): 137 return [self._create_binding(task_id)] 138 139 def wait_for(self, task_id, timeout=None, cache=True, 140 no_ack=True, on_interval=None, 141 READY_STATES=states.READY_STATES, 142 PROPAGATE_STATES=states.PROPAGATE_STATES, 143 **kwargs): 144 cached_meta = self._cache.get(task_id) 145 if cache and cached_meta and \ 146 cached_meta['status'] in READY_STATES: 147 return cached_meta 148 else: 149 try: 150 return self.consume(task_id, timeout=timeout, no_ack=no_ack, 151 on_interval=on_interval) 152 except socket.timeout: 153 raise TimeoutError('The operation timed out.') 154 155 def get_task_meta(self, task_id, backlog_limit=1000): 156 # Polling and using basic_get 157 with self.app.pool.acquire_channel(block=True) as (_, channel): 158 binding = self._create_binding(task_id)(channel) 159 binding.declare() 160 161 prev = latest = acc = None 162 for i in range(backlog_limit): # spool ffwd 163 acc = binding.get( 164 accept=self.accept, no_ack=False, 165 ) 166 if not acc: # no more messages 167 break 168 if acc.payload['task_id'] == task_id: 169 prev, latest = latest, acc 170 if prev: 171 # backends are not expected to keep history, 172 # so we delete everything except the most recent state. 173 prev.ack() 174 prev = None 175 else: 176 raise self.BacklogLimitExceeded(task_id) 177 178 if latest: 179 payload = self._cache[task_id] = latest.payload 180 latest.requeue() 181 return payload 182 else: 183 # no new state, use previous 184 try: 185 return self._cache[task_id] 186 except KeyError: 187 # result probably pending. 188 return {'status': states.PENDING, 'result': None} 189 poll = get_task_meta # XXX compat 190 191 def drain_events(self, connection, consumer, 192 timeout=None, on_interval=None, now=monotonic, wait=None): 193 wait = wait or connection.drain_events 194 results = {} 195 196 def callback(meta, message): 197 if meta['status'] in states.READY_STATES: 198 results[meta['task_id']] = meta 199 200 consumer.callbacks[:] = [callback] 201 time_start = now() 202 203 while 1: 204 # Total time spent may exceed a single call to wait() 205 if timeout and now() - time_start >= timeout: 206 raise socket.timeout() 207 try: 208 wait(timeout=1) 209 except socket.timeout: 210 pass 211 if on_interval: 212 on_interval() 213 if results: # got event on the wanted channel. 214 break 215 self._cache.update(results) 216 return results 217 218 def consume(self, task_id, timeout=None, no_ack=True, on_interval=None): 219 wait = self.drain_events 220 with self.app.pool.acquire_channel(block=True) as (conn, channel): 221 binding = self._create_binding(task_id) 222 with self.Consumer(channel, binding, 223 no_ack=no_ack, accept=self.accept) as consumer: 224 while 1: 225 try: 226 return wait( 227 conn, consumer, timeout, on_interval)[task_id] 228 except KeyError: 229 continue 230 231 def _many_bindings(self, ids): 232 return [self._create_binding(task_id) for task_id in ids] 233 234 def get_many(self, task_ids, timeout=None, no_ack=True, on_message=None, 235 now=monotonic, getfields=itemgetter('status', 'task_id'), 236 READY_STATES=states.READY_STATES, 237 PROPAGATE_STATES=states.PROPAGATE_STATES, **kwargs): 238 with self.app.pool.acquire_channel(block=True) as (conn, channel): 239 ids = set(task_ids) 240 cached_ids = set() 241 mark_cached = cached_ids.add 242 for task_id in ids: 243 try: 244 cached = self._cache[task_id] 245 except KeyError: 246 pass 247 else: 248 if cached['status'] in READY_STATES: 249 yield task_id, cached 250 mark_cached(task_id) 251 ids.difference_update(cached_ids) 252 results = deque() 253 push_result = results.append 254 push_cache = self._cache.__setitem__ 255 decode_result = self.meta_from_decoded 256 257 def _on_message(message): 258 body = decode_result(message.decode()) 259 if on_message is not None: 260 on_message(body) 261 state, uid = getfields(body) 262 if state in READY_STATES: 263 push_result(body) \ 264 if uid in task_ids else push_cache(uid, body) 265 266 bindings = self._many_bindings(task_ids) 267 with self.Consumer(channel, bindings, on_message=_on_message, 268 accept=self.accept, no_ack=no_ack): 269 wait = conn.drain_events 270 popleft = results.popleft 271 while ids: 272 wait(timeout=timeout) 273 while results: 274 state = popleft() 275 task_id = state['task_id'] 276 ids.discard(task_id) 277 push_cache(task_id, state) 278 yield task_id, state 279 280 def reload_task_result(self, task_id): 281 raise NotImplementedError( 282 'reload_task_result is not supported by this backend.') 283 284 def reload_group_result(self, task_id): 285 """Reload group result, even if it has been previously fetched.""" 286 raise NotImplementedError( 287 'reload_group_result is not supported by this backend.') 288 289 def save_group(self, group_id, result): 290 raise NotImplementedError( 291 'save_group is not supported by this backend.') 292 293 def restore_group(self, group_id, cache=True): 294 raise NotImplementedError( 295 'restore_group is not supported by this backend.') 296 297 def delete_group(self, group_id): 298 raise NotImplementedError( 299 'delete_group is not supported by this backend.') 300 301 def __reduce__(self, args=(), kwargs={}): 302 kwargs.update( 303 connection=self._connection, 304 exchange=self.exchange.name, 305 exchange_type=self.exchange.type, 306 persistent=self.persistent, 307 serializer=self.serializer, 308 auto_delete=self.auto_delete, 309 expires=self.expires, 310 ) 311 return super(AMQPBackend, self).__reduce__(args, kwargs) 312 [end of celery/backends/amqp.py] </code> Here is an example of a patch file. It consists of changes to the code base. It specifies the file names, the line numbers of each change, and the removed and added lines. A single patch file can contain changes to multiple files. <patch> --- a/file.py +++ b/file.py @@ -1,27 +1,35 @@ def euclidean(a, b): - while b: - a, b = b, a % b - return a + if b == 0: + return a + return euclidean(b, a % b) def bresenham(x0, y0, x1, y1): points = [] dx = abs(x1 - x0) dy = abs(y1 - y0) - sx = 1 if x0 < x1 else -1 - sy = 1 if y0 < y1 else -1 - err = dx - dy + x, y = x0, y0 + sx = -1 if x0 > x1 else 1 + sy = -1 if y0 > y1 else 1 - while True: - points.append((x0, y0)) - if x0 == x1 and y0 == y1: - break - e2 = 2 * err - if e2 > -dy: + if dx > dy: + err = dx / 2.0 + while x != x1: + points.append((x, y)) err -= dy - x0 += sx - if e2 < dx: - err += dx - y0 += sy + if err < 0: + y += sy + err += dx + x += sx + else: + err = dy / 2.0 + while y != y1: + points.append((x, y)) + err -= dx + if err < 0: + x += sx + err += dy + y += sy + points.append((x, y)) return points </patch> I need you to solve the provided issue by generating a single patch file that I can apply directly to this repository using git apply. Please respond with a single patch file in the format shown above. Respond below:
celery/celery
6592ff64b6b024a4b68abcc53b151888fdf0dee3
CELERY_RESULT_SERIALIZER = 'json' breaks Exception marshaling Setting `CELERY_RESULT_SERIALIZER = json` and raising an exception in the worker leads to this: ``` /path/to/lib/python2.7/site-packages/celery/result.py in get(self, timeout, propagate, interval, no_ack, follow_parents, EXCEPTION_STATES, PROPAGATE_STATES) 173 status = meta['status'] 174 if status in PROPAGATE_STATES and propagate: --> 175 raise meta['result'] 176 return meta['result'] 177 wait = get # deprecated alias to :meth:`get`. TypeError: exceptions must be old-style classes or derived from BaseException, not dict ``` where the contents of `meta['result']` are (in my case): ``` {u'exc_message': u'unknown keys: nam', u'exc_type': u'ValueError'} ``` so it _looks_ like celery could convert the dict to a real exception before raising, but it does not currently. Changing back to `pickle` works as expected. bug can be reproduced with the following: ``` python # jsonresults.py from celery.app.base import Celery CELERY_RESULT_SERIALIZER = 'json' CELERY_RESULT_BACKEND = 'amqp' app = Celery(config_source=__name__) @app.task def hello(): raise ValueError('go away') ``` worker: ``` # python -m celery --app=jsonresults:app worker ``` caller: ``` python import jsonresults jsonresults.hello.delay().get() ```
This is biting me as well. Any news?
2015-04-29T14:52:17Z
<patch> <patch> diff --git a/celery/backends/amqp.py b/celery/backends/amqp.py --- a/celery/backends/amqp.py +++ b/celery/backends/amqp.py @@ -195,7 +195,7 @@ def drain_events(self, connection, consumer, def callback(meta, message): if meta['status'] in states.READY_STATES: - results[meta['task_id']] = meta + results[meta['task_id']] = self.meta_from_decoded(meta) consumer.callbacks[:] = [callback] time_start = now() </patch> </s> </patch>
diff --git a/celery/tests/backends/test_amqp.py b/celery/tests/backends/test_amqp.py --- a/celery/tests/backends/test_amqp.py +++ b/celery/tests/backends/test_amqp.py @@ -13,6 +13,7 @@ from celery.backends.amqp import AMQPBackend from celery.exceptions import TimeoutError from celery.five import Empty, Queue, range +from celery.result import AsyncResult from celery.utils import uuid from celery.tests.case import ( @@ -246,10 +247,20 @@ def test_wait_for(self): with self.assertRaises(TimeoutError): b.wait_for(tid, timeout=0.01, cache=False) - def test_drain_events_remaining_timeouts(self): + def test_drain_events_decodes_exceptions_in_meta(self): + tid = uuid() + b = self.create_backend(serializer="json") + b.store_result(tid, RuntimeError("aap"), states.FAILURE) + result = AsyncResult(tid, backend=b) - class Connection(object): + with self.assertRaises(Exception) as cm: + result.get() + self.assertEqual(cm.exception.__class__.__name__, "RuntimeError") + self.assertEqual(str(cm.exception), "aap") + + def test_drain_events_remaining_timeouts(self): + class Connection(object): def drain_events(self, timeout=None): pass
1.0